We talk about some of the biggest names in Chinese AI, from Z.ai, Moonshot and DeepSeek to Alibaba, Tencent and ByteDance. But rather than just looking at who has the biggest or most talked-about models, we get into what each company is actually good at, how they’re approaching the frontier, and what Robin looks for when trying to separate real progress from the hype.

We also touch on how Chinese AI companies are operating under very different constraints than their US counterparts, yet they’ve continued to make impressive progress through techniques such as model compression and reinforcement learning. This raises a bigger question around where the value in AI ultimately sits: if models become increasingly capable, cheaper and more commoditized, who actually captures the economics?

From there, we get into the business of AI — how open-weight labs can make money, what AI monetization might look like, and whether the biggest opportunities will sit with the models themselves or with the applications, infrastructure and orchestration layers built around them.

Finally, we zoom out to the bigger picture: what China’s progress in AI could mean for geopolitics, model sovereignty and international adoption, and how investors should think about valuing these companies when the technology is moving faster than traditional financial metrics can keep up.

Subscribe now

The AI Proem Podcast is under the AI Proem newsletter, which has over 12k followers globally. To learn more about China AI, the business of AI, and how AI is impacting businesses, please check out the newsletter here and more insightful conversations here.

Chapters

00:00 Introduction

01:06 China’s AI Race and the Rise of New AI Labs

06:39 The Compute Bottleneck — And How China Is Closing the Gap

11:37 AI Monetization: Who Captures the Value?

13:06 Why China Has So Many AI Labs — And Who Will Survive

16:10 When Does an AI Model Become “Good Enough”?

19:10 Token Rationalization, AI Harnesses and the Future of Work

25:21 Models vs. Applications: Where Will AI Value Accrue?

30:31 How Open-Weight AI Labs Can Make Money

33:07 Can Chinese AI Capture 30–35% of Global AI Revenue?

45:02 How Should Investors Value AI Companies?

49:56 Robin’s final thoughts on AI’s future in China and globally

Transcript

(AI-generated, for reference only)

Grace Shao (00:01)

Hey Robin, good to have you.

Robin (00:03)

Thanks for having me. Good to be here.

Grace Shao (00:05)

Yeah, yeah. Tell us about your coverage and your recent initiation on Z.ai and MiniMax. I think that was quite exciting. It was a huge report — 50 pages or 80 pages, was it? What made you decide that now was a good time? And what is your main takeaway there?

Robin (00:23)

Sure. Yeah, look, you know, I’ve been covering internet at Bernstein for a long time now. I’ve been covering gaming for the last number of years in Japan. Year to date, I think something like 80% of our research has been about some form of AI or other. I’ve been using Z.ai and MiniMax as the examples to effectively fill my exhibits and illustrate different points. The stocks kinda ran away from me as we were doing that. We initially thought, okay, we were gonna you know, work out what AI does or what these businesses do and then they all went vertical. there came a point in the summer I was just like, All right, you know, the stocks can do whatever they want given such small free floats and we’ll wait a little bit and the lockup expiries were coming up at that point and yeah, we picked a week Shortly after. I was in the US for a month to kinda network and do different things. and I think we got lucky on the timing, to some degree. But yeah, you know, now there’s more price discovery. There’s, you know, it seems to be a new model launching every other week. So yeah, fun times.

Grace Shao (01:34)

Yeah, sorry, we were just talking about how there’s such AI fatigue. Like, there were literally eight models over the summer and there was no summer for any of us covering AI, right?

Robin (01:44)

Are you not excited about Ox Alpha?

Grace Shao (01:47)

Everyone’s excited about Ox Alpha, but we all have different conspiracy theories, right? Well, because like I don’t wanna like you know go into these conspiracy theory holes today. Let’s focus on some of the big pictures. I do want to ask

Robin (01:58)

Okay.

Grace Shao (01:59)

You are one of the rare people who gets access to these labs and their executives. When I last spoke to you, you said you were hanging out in Beijing, meeting with some of the executives at Z.ai, MiniMax and whatnot.

Robin (02:09)

Mm-hmm. Mm-hmm.

Grace Shao (02:12)

Obviously not sharing anything sensitive, but what’s the vibe? What are the cultural differences?

Robin (02:16)

Yeah.

Grace Shao (02:17)

You know, do you think any of their personalities or cultural makeup actually, you know, differentiates them from how they go to market, how they build their products or technology, or maybe even their philosophy on AI?

Robin (02:31)

Yeah, sure. I mean, I think it’s kind of interesting, you know, I deal with investors day to day a lot. you know, the debate there is, are these labs raising prices? Are we gonna get competition? Do models get commoditized and pricing goes to you know, gets hammered and so on. you talk to the guys at these labs and it’s a very kind of singular focus on, you know, everybody thinks they’re changing the world. AGI is very much top of mind for everybody and Iterating the model is much more of a focus, obviously, compared to investors. But the vibe is very much: yeah, just keep going, keep cranking and see where we can go. Culturally, there are some quite big differences. You know, Z.ai came out of Tsinghua University. Dr. Tang is still kind of on both sides of the fence in some ways. You know, somebody else described it as being monastic. I’m not sure I’d go that far, but it is a much more kind of academic and nerdy organization. MiniMax and all the dealings I’ve had with them seem to be more commercial. You know, they’ve had a couple of pivots in terms of what the main focus has been. certainly more international than Z.ai. but yeah, Kimi’s kind of I guess in some ways halfway in between. You know, they are More international than Z.ai but yeah, you know, you’ve got kind of the more si how do I how’d you describe it? More kind of science based aspect of you know what they’re doing. So yeah, you know, these are they show through in how these companies behave, and the results that you’re seeing in terms of model progress. And yeah.

Grace Shao (04:24)

How do you think they’re defining AGI? Is it different from what SF is saying?

Robin (04:31)

What’s SF saying? It seems to be different every few weeks.

Grace Shao (04:34)

Huh.

Robin (04:35)

I don’t know if there is a single kind of monolithic, you know what — like we’re going to do AGI and it’s this thing. you know, I think the common analogy is summoning the machine god, which... But I think it’s a little bit narrower than that. I think it’s, you know, how do we get AI to iterate our models for us? How do we get into kind of, you know, I guess some definition of loose RSI or narrow RSI? I don’t tend to get into discussions about broad RSI with people, you know, where it does actually just become a little bit more religious. But yeah, I think, you know, everyone is just focused on kind of iterating the next generation of models.

Grace Shao (05:17)

And they’re all kind of facing the same issue, right? End of the day it’s compute. But potentially there’s domestic compute becoming more abundant. do you think that’s really gonna change the game here? Or is that even something that’s happening in the near term?

Robin (05:33)

Yeah, I think it is a bottleneck. Everybody is short on compute. Everybody makes comments on, you know, if you compare the FLOPs per engineer here versus in SF, there is a big difference, and it does hold back some of the progress. But despite that, you’ve seen some of the Chinese labs come up with cache-compression tricks, RL tricks to Close the gap, and I think it has been quite remarkable to see where they’ve gotten to on quite limited resources. I mean, Z.ai in particular, getting to what they’ve done with a 750B pre-train has certainly surpassed what I thought was possible without getting to a bigger model.

Grace Shao (06:22)

Okay, well then let’s take some. I wanna double click on that later for sure on what the potential implications of all that progress means later. But first, start big, high level. What’s your sense on each of the labs? Like who’s good at what? What are they each gunning for? What’s a good mental framework for us to when we’re evaluating these different labs? Because the one thing I wanna lead to the next question really is we just have so many Chinese like lab providers, I sorry, model providers right now. Like, There’s Z.ai, MiniMax, DeepSeek, let’s just call them somewhat the first tier labs. Then we have like the BAT here, I mean like ByteDance, Alibaba, Tencent. Then like randomly over the last like three to six months, we get Xiaomi, Meituan, RED, Huawei, all crowding that space. And then you have like StepFun and a few others kind of dabbling in this. Well, not dabbling, but they’re also doing this. Well, maybe considering them second tier. How do we understand this landscape and

Robin (07:18)

Ni

Grace Shao (07:18)

How do we evaluate? These labs.

Robin (07:21)

Yeah, I think the way that we kind of framed it and when we launched coverage was, you know, I think there are three frontier labs if you look at the latest models. The concept of the Pareto frontier is quite important because there isn’t sort of a monolithic kind of first place. You either have to be the smartest model at your price point or vice versa, the cheapest model for a certain level of intelligence, however you define that. So you can have different spots on the frontier. For example, when Kimi K3 came out, it obviously was smarter than the latest GLM, but it was also something like three times as expensive. So I happen to use both in my day-to-day. and they’re not kind of direct substitutes immediately to each other. but let’s say, you know, we said and I stand by this view that there are three frontier labs: Z.ai, Kimi/Moonshot and DeepSeek. I think amongst themselves, the consensus is that Z.ai is good at post-training, Kimi has the biggest pre-training scale and is good at pre-training, and DeepSeek has this crazy infra capability, so they can run super-high tokens per second and so on. So, you know, they are good at different things based on their different backgrounds. The internet companies — look, Alibaba’s probably spent the Most time and effort to try and develop a frontier suite of models. You know, Qwen is frontier-ish. It’s kind of up there. And then Tencent’s come back really into the conversation in the last six months with Hunyuan 3. I think initially people ignored the preview and then more recently it’s become clear that, you know, the machine that builds the machine is now working and they’re iterating towards Hunyuan 4 sometime this year. So you know that’s that.

Grace Shao (09:13)

I think Hunyuan 4 preview is coming out this Friday actually. It is. Somebody just told me. I think that’s public information soon

Robin (09:17)

Is that right? can I quote you on that? Yeah. Cool.

Grace Shao (09:24)

Or now. Anyway, yes, go on.

Robin (09:26)

It is now. Yeah. well, You know, they said it’s coming this year. I kind of assumed it was coming in Q4, but cool. That sounds encouraging. ByteDance is funny, right? Because they kinda raced out into a big lead with Doubao. It was, you know, it was considered to be the — well, it’s, I guess it still is the chatbot app in you know, compared to Qwen or compared to Yuanbao. And then I think they ran into the issue of, well, okay, we’ve grabbed the users that we can grab, and then when they tried to monetize it through subs, the paying ratios were very low. and so now I think there’s been a bit of a pivot towards, do we do enterprise? My understanding is they’re gonna start doing ads within Doubao in the second half of the year. so there’s a couple of ways that they’re going through. But yeah, I would say that Tencent and ByteDance are more focused on applying AI through their ecosystems. Alibaba’s much more focused on

Grace Shao (10:22)

But they want to use their own models, Right? They want to use their own models. So it’s like without really strong models, how do they apply? And this brings me, sorry, I’m just hijacking this whole conversation, but it brings me back

Robin (10:30)

Mm.

Grace Shao (10:31)

To like Ben Thompson’s recent interview with Patrick Shonas. He was really interesting. He just goes on about, you know, end of the day, it’s advertising and consumers will not pay. So it seems like, you know, all these AI application companies will just become like they will just monetize the same way the internet companies monetize. And then it’s

Robin (10:45)

Everybody sells ads in the end, C Yeah, I think that’s I think that’s A plausible yeah, I think that’s a plausible end outcome. where I would differ a little bit is something like a WorkBuddy, where people are just paying for effectively tokens and task completion. but yes, on the kind of base consumer layer then yeah, advertising, monetizing merchants who Effectively pay for access to people doing stuff on WeChat or elsewhere ends up being you know, Douyin started to monetize some of the local service recommendations recently. again through a kind of quasi ads model. So yeah, I think we do move in that direction medium term.

Grace Shao (11:27)

Then what about all the rest like I just talked about? I think like 10 other companies are serving up models these days. How do you make a or actually let me re-reshift the question? How why

Robin (11:34)

I tend to think of those as

Grace Shao (11:40)

Let me reframe the question: why are there so many? Should we be expecting some calls consolidation? Like, because I don’t expect you to comment on every single one of them, how they’re different. But the fact that we just have like 20 different labs

Robin (11:50)

Mm-hmm.

Grace Shao (11:52)

Offering models at this point. It doesn’t seem very economical, but and also aren’t they all just fighting for the same compute, same talent at this point? How do you view all Of that?

Robin (12:00)

Yes, there’s only so many PhDs that you can hire out of these places. Look, I think there will be fewer players at the frontier or frontier-ish. you know, I’ll pick on Meituan since it’s a company I cover. But they came out with LongCat-2.0. I think initially there was some excitement, and then people realized that it’s got the reasoning capabilities of Hunyuan 3, which has got a quarter as many parameters or something. So or less. and I do think that over time, you know, if you think about what’s important here, it’s access to a data pipeline, it’s the ability to turn that into RL environments and then train these models. and I do assume that the you know, the labs and the very biggest internet companies will kind of, you know, be thereabouts. Some of these smaller names that you’ve just mentioned, I think you know, in the Meituan example, like what’s the difference between having your own model versus using DeepSeek or something? Like, just You know, fine-tuning DeepSeek or something. I do think there’s a kind of open debate there. and you know, certainly costing them a decent amount if you look at their accounts. So I do think there will be fewer players at the frontier. You know, I think

Grace Shao (13:14)

So

Robin (13:14)

If you want to have some kind of basic search engine that is AI powered, then so be it. Do you need a complicated long horizon agentic model to underpin every internet platform? No.

Grace Shao (13:27)

What is driving all this kind of effort to even build their own models? Just FOMO.

Robin (13:33)

I think it’s partly, “we want our own thing.”. I think it’s innately because there’s so much competition in the internet space that there is this kind of insecurity around you know, if we don’t have a model then do our peers who do then find a way to get an upper hand somehow. And so we need to at least understand the technology. That bit I get. I just don’t think that translates into long term commitment into, you know, training ever larger models infinitely.

Grace Shao (14:02)

Why don’t we see that kind of phenomenon in the US as much? Like the internet players aren’t just all rolling out their own models.

Robin (14:11)

They’re all too busy serving compute to the two guys at the front. Seems to be what’s happening. Yeah. I think generally

Grace Shao (14:16)

All right, let’s

Robin (14:19)

Generally, there’s a lot more duplication in China than

Grace Shao (14:24)

Mm.

Robin (14:24)

In the US where the you know these players kinda keep to their own lanes a bit more.

Grace Shao (14:29)

Yeah, yeah. Well, on that topic, this is like relevant, but much of a conversation around AIs that these models are being commoditized, but you do have a more nuanced view. reading your recent report, you know, you’re saying

Robin (14:40)

Mm-hmm.

Grace Shao (14:40)

That, you know, frontier models versus good enough models kind of have different value within the long term ecosystem. At what point do you think a model becomes good enough that the user simply stops caring whether another model is theoretically more intelligent?

Robin (14:55)

Yeah.

Grace Shao (14:55)

Do you think it’s use case dependent or you know? Client-, customer-, end-use-dependent or how do we understand that?

Robin (15:04)

Yeah, I think initially, you know, agentic AI kind of exploded in Q1, and then everybody got super excited and you know went both feet in to try and work out what they could do with it, hence the token maxing movement. and then subsequently I think as especially as we started using agentic AI more and more, the framework that I kind of zoomed in on was more around User perception, right? And you can define that however you want in terms of use cases, in terms of you know who the user is. And I guess the point that we tried to make was instead of there being some kind of objective standard by which AI models become good enough, they become you know, AI becomes good enough when you can solve a task that you’re trying to solve. And If you then have more than one AI model or if you have multiple AI models able to solve the same task, then the arbiter of who you give the task to then moves away from reasoning capabilities because at that point, you know, by default, if multiple models can solve the problem, then you move on to cost and availability and you know, in some cases UI/UX and your kind of taste of one model versus the other in some cases. So yeah, I you know, I think If you kind of lead on to that, then we do think in a couple of the companies have talked about this in different ways, is that you will have a frontier where people will pay higher and higher ARPUs for more and more specialized and longer and longer horizon task completion. you know, you go from ordering bubble tea to doing agentic commerce to doing, you know, more serious work stuff to frontier science. With the number of tasks that you solve probably get fewer and fewer and the ARPUs just scale infinitely almost. And meanwhile you’ve got this kind of behind the frontier bit where yeah you can solve a task with a good enough model. you know, we were in discussions with or talking to execs at Tencent who said that on WorkBuddy there’s probably thirty percent you know performance-driven tokens and versus seventy percent what I think was said was value-driven tokens where it’s like, you know, you can solve the task with A cheaper model. the way that Z.ai tries to frame it is, you know, you have thirty percent and then maybe fifty percent, which you can monetize and then the last twenty percent is just given away for free on Doubao. but yeah, you know, you have a split in the market as a result of you know, using models to solve what it can versus solving problems that you know multiple models can do versus the really simple stuff.

Grace Shao (17:50)

And is that the responsibility of the, I guess, the provider of the tokens or the user? As in, should I be routing my own models left, right, center to optimize my cost or what? Or should the platform be

Robin (18:02)

Yeah.

Grace Shao (18:03)

Doing that for me?

Robin (18:04)

I think people ran into that themselves because I mean the irony for me is that Fable being as expensive as it was was probably the catalyst for people to realize, you know, maybe we shouldn’t ask Fable for the weather because it costs you five dollars to do that or something. and and then at the same time, I think in June, July there was this explosion of open-source Chinese models that you know meant that the number of alternatives Then expanded and people started to kind of think about, maybe I should rationalize or, you know, at least match the value of whatever task I’m solving with the cost of the tokens that you know, that I’m spending. In the first half there were these crazy kind of anecdotes like folks working for the internet companies having quotas of upwards of like a thousand dollars a month and, you know, me Saying to them like I know what you do, you make slides for your boss. You don’t need a thousand dollars of tokens a month. and eventually I think the internet companies backed away from that. So yeah, you know I think we’re seeing that kind of rationalization movement happen.

Grace Shao (19:13)

Yeah, yeah, there’s definitely been a cap I’ve heard for especially what they call like people who are just working with documents, definitely definitely don’t need to be token maxing at all. I am an advocate for not using AI for simple things. Like when you’re asking about the weather, maybe you should just turn on your weather app, like honestly, and not like burn compute

Robin (19:29)

Yeah.

Grace Shao (19:30)

On that. Anyway, you use the term token rationalization in your recent report. Tell us about that, because you kind of alluded to it already.

Robin (19:39)

Yeah, yeah, yeah. So I think it’s, you know, again, it’s this idea of matching the value of what you’re doing versus the cost of the tokens that you’re spending. and it kind of ties back to this Pareto frontier where you know, you have different levels of complexity, you have different levels of cost associated with different models based on their scale and other factors. And you know, you try and be smart about You know, how much you spend, I guess. Like one of the kind of ways I’ve settled in is to have worker bots. I run my own Hermes setup with different bots and I have my worker bots that are relatively cheap and do what they do. and then I have a checker of homework at the end that makes sure that nothing stupid happens and they red team each other and so on. So I think, users kind of find their own ways to do that. Maybe You know, outside of the kind of pro users, then there is the role of the harness or of the kind of you know, the WorkBuddy-like app that then decides for you, this goes to this model, that goes to that model. That makes the overall experience more optimized.

Grace Shao (20:49)

Why don’t we just jump it straight into how you use AI for research? Like I wanna hear more about it. Like, how do you use Hermes? How do you run your own models? Like it’s not a Very

Robin (21:00)

Yeah.

Grace Shao (21:00)

Common sell-side analyst approach. I think you’re definitely like AI-pilled to the max out of all the cell

Robin (21:06)

I am

Grace Shao (21:07)

Sell-side analysts I speak to. Why don’t we talk about that first before

Robin (21:12)

Yeah.

Grace Shao (21:12)

I get more into the analysis? Like I’m curious.

Robin (21:15)

God, how long do you have? Look, I started with OpenClaw, round about the same time as when everybody got excited about OpenClaw. And then I very quickly fell out of love with it and then almost by accident happened upon Hermes around the same time. I use it for a number of things now. I use it for kind of work research as a way to gather information. these bots have a way of scraping data from websites. That I can’t manually. there was one instance of a website where you know the website would show you the last 12 months of data and then the bot went inside and after a while said, here’s an API that allows me to download the last 10 years of data, which is great. maybe it’s less great for the guy operating the website. But it there are things like that, or I can now I’ve set up bots to monitor Twitter and Reddit sentiment when a new game comes out, or you know, I’ve built You know, reasoning loops to try to forecast game sales, look at or actually one of the biggest time savers of all has probably been the ability to summarize podcasts. Like in the old days, pre AI somebody sends you a two and a half hour podcast, you’re like, great. Like I’m sure this is super interesting, but I’ve just lost my Saturday morning. Whereas now I can basically have the Hermes bot gonna summarize it, give me key quotes, key insights, timestamps, you know I probably then go back and listen to what was actually said on maybe 10–20% of the two and a half hours. So yeah, it’s been, you know, a good time saver. The other thing is you know, just the ability to knock around ideas on my phone while I’m walking around. Like there was one time when, you know, my wife and I took our daughter to some tourist spot that I’d been A lot of times and they were kind of going around sightseeing. I was behind them having a chat with my Hermes bot and by the end of the four hours of walking around I developed most of a note to write down. So yeah, it’s it’s you know, there’s different ways that I’ve tried to use it. I’m sure we’ll find more over time.

Grace Shao (23:26)

So you just exposed yourself For not actually listening to my AI Proem podcasts when you do say you listen to them. You’re probably just getting Hermes to summarize them for you.

Robin (23:36)

No, I listened to you.

Grace Shao (23:38)

All right. You listen to this episode.

Robin (23:39)

Hahaha.

Grace Shao (23:41)

Okay, let’s get back to the serious stuff. Okay. I think I think that’s really interesting because I think someone who actually uses it, like understands it differently from just pure observer. But do you think then, just going back to the last conversation, but do you think then frontier model providers still capture the most of the economics? Or do you think the value is gonna move to applications, orchestration layers? Because early in the conversation, you know, we were talking about like, you know, people are like more mindful of the costs now. but this is how these frontier labs make money. So doesn’t it then challenge their existing business model?

Robin (24:15)

Yeah, I mean there is this ongoing debate about value capture, you know, between the semis industry where all the stocks have gone vertical this year and then less so in the last couple of months. and then the hyperscaler layer, I mean that you know, you if you look at the US internet companies, spending AI capex is alternately good and bad every other few months, you know, depending on what gets reported. And then the labs themselves, and then, you know, increasingly you’ve had these kind of harness type debates. The one that has struck me as being very interesting lately is that the competition between first and third party harnesses. If you’re an AI lab, then you know the harness is something that basically puts — if you think of the model as being an answering machine or a reasoning machine, the harness actually adds persistent memory, reference files,, you know, tool calls and The ability to do stuff on a kind of recurring and ongoing basis next to the model. And that’s actually what makes it so for example, the Hermes construct is what makes these bots be able to work with you much more like a human worker can. Right. And so there’s been the debate of, well, you know, it’s strategically necessary for every AI lab to have its own harness, whether it’s Claude Code, whether it’s Codex, whether it’s Z Code and Kimi Code and so on. Or do you just end up with kind of WorkBuddy and that acts as an orchestrator across multiple models? or do the first party models then allow other models into their own kind of space? So that’s a debate I think is not going to be solved anytime soon. I think it’s gonna be useful to kind of see how it plays out. but at the end of the day, especially in the Chinese context, I do think that, you know, if the strategic need is for Frontier reasoning capabilities, then on some level these labs have to survive. Right. If you assume that essentially they get zero part of the value, then they will cease to function or they will stop being able to fund themselves and fund the next training run. So I think they will have to retain some of the value in order to keep doing that. and if you think about which layer of the Chi of the tech stack in China that’s most likely to get overbuilt, it’s probably the compute layer. Which also argues in favor of the labs getting, you know, some of the value. So I think it’s yeah, I think it’s an ongoing debate.

Grace Shao (26:41)

Interesting, you just brought up actually like a lot of these labs will have to create their own products to actually still capture some of the value beyond just infrastructure layer, right? But actually, I think Sam Altman was just on a podcast like a few days ago. He was talking about how he’s like actually bringing everything back. they’re cutting products, right? Like no longer doing a bunch of different products. They’re saying that they’re a platform business instead they only want to use ChatGPT’s interface, they’re even like renaming Codex or something. Like, do you think that is the future for like the kimmies and the mini sac max of the world ‘cause then my argument or my challenge against that is then how could they compete with the big tech and channel where they have just like such broad reach, I guess. They’re like with these super apps and everything.

Robin (27:27)

Yeah, so that’s that’s the interesting part where if you’re an AI lab, you know, right now you’re using, you know, some of these third party harnesses as distribution. over time, you know, do you have to retain some parts of the reasoning capabilities like, you know, for example, cyber is one of these niche things that you can do with an AI model. Does that need to go into a you know, power user grade or consumer grade AI harness? Or, you know, there are different Flavors of model that do different things, not you know, maybe you don’t put all of them onto the generic kind of third party harness. but that is something to work out. I mean in Sam’s case, I think there’s a kind of semantic thing here where you can name it what you want. to me the accumulation of personal context behind the model through repeated engagement with it and it kind of learning what you do and You training it to do different things as skills, setting up a network of expert tool calls that you can kind of call on. Like that stuff is actually in my mind, where a lot of the long term value lies, or where the usefulness of the AI kind of goes and resides in the you know, as you use it repeatedly. So yeah, and you know, there are a few ways where I think this can play out. One is potentially, you know, there is obviously the The version of the world where everybody goes on to a third party harness within you know that’s run by one of the big internet companies. The alternative is that, you know, if you are a more pro user, more specialized user, maybe you do need some of the specialized functionality that, you know, the AI labs keep to themselves. and then you have a range of these outcomes. yeah, very complicated question. I don’t know if I have a fully formed answer at this point.

Grace Shao (29:18)

Fair enough. but let’s look at monetization in general for the open-weight labs. You know, obviously I think from the Western perspective, it’s like the biggest question

Robin (29:25)

Mm-hmm.

Grace Shao (29:25)

Is always like how do they make money? How do open-source models make money? How do you see like the current API sales? Is that just a durable revenue pool? or is that not gonna be enough to sustain the capital they need to continue to train?

Robin (29:43)

Yeah, I mean right now, you know, back to your earlier point about compute constraints. you’ve trained these models, there’s been this explosion of interest in them. I think one of the biggest constraints that they have faced is the lack of silicon is constraining their ability to serve up APIs. You know, I feel this pain every morning Asia time where, you know, I ask a Chinese model to do something at nine AM Hong Kong time and it’s rate limited all the time because everybody else is trying to do that at the same time. And as that gets resolved, presumably, you know, that induces some demand and your ARPU goes up. yeah, I and there are probably lumps around new model releases and whatnot that means that you know it’s spiky rather than that being this smooth curve upwards. But yeah, I do assume that expands over time. You know, all of these companies distribute via the internet platforms as well through, you know, ModelScope and workbody and You know, other things. and then I think to me the interesting thing that’s kind of happened recently is this idea that they are now starting to try and charge or take rate from the global inference providers, right? Like Kimi has gone and signed these deals with different inference providers, essentially charging them you know, you can call it a royalty fee or a or it’s some kind of licensing agreement. But essentially altering the license so that if you are Looking at the weights for academic use or whatever, then that’s fine. If you’re using it to actually just host it and make money, then you have to pay them something, which I think makes sense. So that’s a route that they can kind of pursue to try and capture some of the global economics.

Grace Shao (31:26)

Yeah. So going towards like a bit more controlled commercial economics. Okay, you estimated Chinese models to be able to address roughly thirty to thirty-five percent of global AI revenue, despite the current headwinds with geopolitics and whatnot. Walk us through that thinking. Thirty to thirty-five percent is quite a large pie. I think even a year ago when I spoke to some of the labs, they were jokingly saying, even if we get five percent, that’s enough money for our business, you know.

Robin (31:54)

Yeah.

Grace Shao (31:55)

I mean, but that’s one lab. I guess cumulatively it’s it’s it adds up. So tell us about your thinking on that.

Robin (32:01)

Sure. I mean it was more of a top down kind of estimate based on what was Attainable or what was kind of in a you know in the context of geopolitical realities what was realistically kind of accessible to these labs. And you know we had made these TAM estimates by region. I think US was like half of global or you know maybe even a little bit more than that. And then China obviously we assumed was a captive market. We ended up assuming a very, very limited access to the US market because of the geo issues. and you know The thing that there’s been a few things that have happened since, like, you know, potentially well, Ramp, I think at one point said that the like f you know, five percent of their highest engagement AI users were playing around with AI models from the Chinese labs. and then you had Microsoft that was allegedly thinking about using different AI models from the Chinese labs to power copilot. So, you know, have we been Conservative, have we been kind of you know, is it that there truly is no access to the US market? Question mark. But you know, in Europe you know, we’ve assumed some access, you know, not unconstrained access. I guess if you’re Airbus, you’re probably never gonna use a Chinese model for obvious reasons. But then you’ve also had you know Mistral Mistral’s kind of turned itself from being a you know frontier lab to something that now helps Z.ai to distribute GLM. And so, you know, there are these kind of future permutations I think are gonna be interesting. That means that, you know, that Europe is accessible to the Chinese AI labs to an extent. and then in the rest of the world, I mean you’ve got, you Z.ai, for example, Z.ai doing sovereign projects with Malaysia, with the Middle East. that presumably then acts as a bit of a kind of beachhead for them to go and do other stuff within these markets. So Yeah, you know, we assume more access to these other markets where, you know, the geopolitical picture is probably you know, more more kind of open to the Chinese labs. So it’s a top down estimate. You know, that it doesn’t mean we think the Chinese AI labs will have thirty, thirty five percent market share. You’re still kind of contesting these markets with OpenAI, Anthropic and others. but yeah, it was more of a kind of, you know, how much of the TAM is actually open to you?

Grace Shao (34:24)

And let’s just say like geopolitics side, like you mentioned, like obviously government agencies are not going to use Chinese labs, but like a lot of companies might, like, you know, companies with less like strict compliance on this, then what does continued progress with China’s open-source models mean for like what would it mean for these US frontier labs, especially as of now realistically, there’s really two labs left at the kind of frontier really fighting it out. and they’re not you’re not seeing them lowering their Cost or opening up their weights.

Robin (34:53)

You’re gonna get hate mail from Elon Musk after you upload this.

Grace Shao (34:59)

Well, if he watches this, it’ll be great.

Robin (35:03)

No, look, I think you know, I think there’s going to be a a mix of model use in in the market, right? Like isn’t something I think a lot of people have underweighted is what Alex Karp has been saying, where, you know, companies need sovereignty over their own data, you need ownership of what you’re doing. He’s obviously talking about his book, but you know, I do think that there will be a variety of solutions. you see, you know, folks from Databricks and DoorDash posting about testing Chinese models on Twitter and I think it’s interesting that’s, I think that will continue. I think you know these companies will find ways to orchestrate across different models. And the fact that certainly compared to the US labs, these are generally smaller models. if you can get most of the reasoning capability from for much lower token costs, then yeah, I do I do think that you know in a growing section of Use cases that you need, that these will be good enough. Like within the home market, obviously they fight to be frontier. but outside of China in the global market, they are that kind of, you know, eighty percent cheaper for or, you know, much cheaper for most of the capability kind of market positioning.

Grace Shao (36:21)

All right, let’s zoom in on the companies themselves. I want to kind of touch on the labs, especially the two companies you just wrote about in your initiation report, and then we can talk about your long-term coverage of ATs. Just start with what’s your bull case on

Robin (36:33)

Okay.

Grace Shao (36:34)

Z.ai? It’s no secret that you love them. Why? Well, what’s your thinking on that? Like, do you think they would

Robin (36:43)

Yeah.

Grace Shao (36:43)

Just be like the leading research engine research lab in China?

Robin (36:48)

Yeah.

Grace Shao (36:49)

Frankly, not nationalized the same way that DeepSeek is likely more to commercialize. Like, I don’t know, but you just raise your eyebrows. Maybe I’m understanding that incorrectly. Help us understand. What do you think of Z.ai these days?

Robin (37:00)

Yeah, look, I’ll save the DeepSeek comment to the last. But you know, I do like their ability to iterate these models. You know, they came out of Tsinghua University and I think that relationship helps on some level when it comes to expert domain training data. You know, when you talk to them they emphasize that they have this data advantage, which I think you’ve seen through some of the RL progress that they’ve shown. and, you know, the ability to

Grace Shao (37:27)

Sorry, what Is their data advantage? What is their data advantage? They’ve said that

Robin (37:30)

As an

Grace Shao (37:31)

To me too, but I don’t know what that means.

Robin (37:33)

Yeah, I think it’s, you know, if you can buy data and you can, you know, acquire data from experts, you are effectively paying people to write down what they know. but there is also the process of turning that into verifiable you know, like RL environments where you have verifiable kind of end goals or checkpoints that the model needs to hit, or how do you verify correctness or not? And Turn it into something that’s a lot more structured and you can feed it into the RL pipeline to actually train models with it rather than just you know, I sit there and write a hundred page thing on how to do equity research or something. Like I you know, there it has to be there yeah, there has to be kind of reasoning gates and a way to kind of let the model kind of as assimilate that information. so you know, the I think in the GLM-5.3 Release paper is actually really interesting in the sense that they emphasize look RL is all we did and effectively they’ve taken you know data and you know translated that into different RL environments and then they’ve used that to try and iterate the model in different ways. So you know I do think that’s a useful skill to have. The fact that they have a seven fifty B model that’s as good as it is is interesting to me. The fact that, you know I think this is known. I mean that the next big boy model is coming later in the year, you know, call it October, maybe a little bit earlier, maybe a little bit later. But at which point you start having probably the best seven fifty B class model, or certainly it to me it is at the moment. and if you have something that’s competitive in a much bigger model size, then you actually occupy two parts of the of the Pareto Frontier front potentially, which is Interesting strategically. so I but I think most of all it’s just that, you there are these three labs I think are frontier. one of them is a little bit captured in terms of, you know, having to answer to the government to some degree. Kimi I like as well. It’s you know, I think Kimi’s doing some really interesting things on a number of fronts.

Grace Shao (39:50)

Yeah, I was just gonna say actually, like Kimi obviously you don’t cover officially now given that they’re not public yet, but you know, they kind of reset expectations around CI. I think when five point two came out, they felt like the world felt like CI was the leading l lab coming out of China. Kimi K three kind of put themselves on the global stage again. In fact, I think they did a you know, really, really big marketing splash globally, and captured a lot of attention. And then given that they don’t have the kind of CAI ent like Was it entity list complication. They actually

Robin (40:23)

Yeah.

Grace Shao (40:24)

Have it easier with international expansion, but it just kind of looking at, you know, Kimi K3, how do you view Moonshot and theirs their positioning right now?

Robin (40:35)

Yeah, I mean the fact is that they have the biggest Chinese pre-train, right? And the Kimi K3 is a very capable model. it’s significantly more expensive, but at the same time, you know, if you’re among c corporate customers, I think there is the argument to say that you just you know, it’s cheaper than the US labs anyway and you just pay for the best capabilities on some level. But yeah, look, I think, I think they will be up there in the fullness of time. They will hopefully get listed before too long. You know, the last time I asked them the answer was soon. so we’ll see. But yeah, I I think they’re you know, what happened with GLM-5.2 and K three was was kind of interesting because there was a point before the summer where we could have launched coverage on these on these AI labs and almost Around the same time basically GLM-5.2 happened. It was great from the perspective of somebody that used these models, but then I think Z.ai’s share price went up like fifty five percent in the week or that week or something. And it like, All right, maybe we’ll take a break and see how see what see how it goes. And then it got to the point where I think, you know, Z.ai’s share price was pricing in being a winner-take-all type winner, at which point, you know, when Kimi K3 came out then there was an unwind of that expectation. I would argue that there shouldn’t have been that expectation, but you know, go figure. So I think now, you know, I still think that these are the two to watch. DeepSeek obviously will always be up there, but I personally find the Kimi and GLM models way easier to use.

Grace Shao (42:18)

And Why was it that, you know, when Z.ai and MiniMax went public that it felt like MiniMax was more of the market darling, or at least investors in Hong Kong were buzzier around them.

Robin (42:29)

Yeah, I think there were two reasons. I think one was MiniMax was considered to be a lot more international. or it was it was much more international. And then the other thing was given the Entity List listing that Z.ai had had, that it was kind of thought that they would find it more difficult to expand internationally. And the other the other problem that investors had with Z.ai was the on prem Segment, which, you know, was kind of this it reminds people of the bad old days of China Enterprise Software, right? Where, you know, you had these companies kind of toil for years and years without really getting anywhere. and so that was, I think, the initial impressions. we put out something in quite early on, I think after CNY, where you know we took a deep look at these companies. I think I was always of the view that You know, model reasoning capabilities are more important and the state of these companies today versus six months ago versus six months in the future is gonna be so wildly different that it’s yeah, you need to evaluate the machine that makes the machine more than kinda where they are at any given moment, which I think was you know, people were guilty of in January.

Grace Shao (43:41)

And you make the somewhat sacrilegious sell-side argument that traditional financial analysis of these labs can be almost irrelevant essentially. Like are we effectively valuing these companies right now? Like what do you think we should be actually looking at when we are putting valuations on these companies right now?

Robin (43:56)

Yeah, I think the market really struggles to value these companies because, you know, everybody agrees that AI is a big deal and you can have these debates on how many trillion dollars of TAM AIs going to be in the fullness of time. I’ve largely given up having that conversation. It’s just it’s going to be big. and you know, investors do value these stocks on the basis of multi-year ARR trajectories and you know, revenues multiple years out. But yeah, like, you know, these companies are about to report first-half earnings. we’re about to I mean, there are things that you can watch out for, like inference margins and you know, obviously the revenue growth and How AR converts to revenue and so on. But you know, f for example in the case of Z.ai, like you’re you’re basically staring at a bunch of numbers that reflect GLM-5, GLM-5.1, which is ancient history in AI terms. So it’s kind of useful, but not really at the same time. yeah. So to me it’s much more important to just kind of look at the iteration, look at the architecture tricks that they are coming up with and the ability to kind of, you know, scale these new innovations to much bigger models. And you know, if you can do X then you should be able to do Y, and then what does that unlock in terms of capabilities? so yeah, I do think it’s you know, I do think that Stuff is at least as important as kind of scrutinizing the numbers. Even if you know it’s kind of my job to multiply two numbers together at the end of the day.

Grace Shao (45:31)

It’s more important to be looking forward than kind of looking back. but like then I have a question that’s like, you know, StepFun has already

Robin (45:37)

Mm-hmm.

Grace Shao (45:37)

Filed for the IPO. we just said Kimi is likely going to go public in Hong Kong too somewhat, sometime this year, next year, whenever.

Robin (45:44)

So yeah.

Grace Shao (45:46)

We’re looking at like four leading labs already, just the Hong Kong Stock Exchange. Then we have obviously

Robin (45:51)

Mm-hmm.

Grace Shao (45:51)

The BATs, which I want to talk about later as well. Like, how do we understand? Is this not a pretty crowded space? Like there’s a lot of labs going public in China.

Robin (45:59)

Okay.

Grace Shao (45:59)

Does that make sense? How do we understand that? How do we pick the winner? Robin, no one How do you pick the right stock?

Robin (46:07)

How do we push the right? I’ll push back and say that this is the least crowded new-tech cycle in China that we’ve seen so far. you look at you know in the past, you know, EVs and batteries and you know, or what’s going on with humanoid robotics at the moment. you know, the AI labs piece has probably been one of the more concentrated fields that we’ve seen. And within the names that you’ve Mentioned, I do think that there are kind of there’s a clear hierarchy of, you know, which ones are closer to the frontier. I think DeepSeek has decided it wants to embrace the national champion role and potentially list in the A-share market, which is fine. But yeah, I think, you know, I’ve I’ve said for a while I think Z.ai and Kimi are, you know, my picks for the frontier. I think, you know, based on what they’ve done, based on the, you know, I the architectural innovations, based on the adoption of their innovations by other labs is kind of one thing that I watch for as is being quite telling. yeah, I you know, I guess there will be more than just two players in the Hong Kong market. But yeah, I think my view is, you know, like in the US where you’ve seen fewer players at the frontier over time, I think that will show through here as well.

Grace Shao (47:37)

All right, let’s move up the stack. while you’re covering BAT, you’ve been covering them, well, Tencent and Alibaba

Robin (47:42)

Mm.

Grace Shao (47:43)

For quite a while, just given that byte dents is not public. you know, what is your mental model thinking through these big techs in China right now? Clearly they have a bit of FOMO, they don’t wanna be left behind, they don’t want to just be known as their internet as the internet phase. They’re all in IAI. They have the benefits

Robin (47:59)

Mm-hmm.

Grace Shao (47:59)

Of talent, they’re the benefits of money, but somehow the like just like the US, they’re not the ones actually pushing

Robin (48:04)

Yeah.

Grace Shao (48:06)

The frontier. How do you think we should think about that?

Robin (48:10)

Yeah, I mean psychologically, they are quite different businesses. Like, when you talk to Tencent, the focus is a lot more around the application layer and how do you apply AI and agentic functionality within ecosystems like WeChat, within, you know, WorkBuddy is potentially a new platform for them, you know, on the AI front. you’ve got the games and ads businesses, which are I actually think are good platforms on which to apply AI and generate kind of benefits. But you know, it’s much more focused on the application layer. And there was a comment in the latest slides from the Q2 earnings that said, If all else fails, then we’ll just rent out the compute to whoever else. so that’s that’s considered to be

Grace Shao (48:53)

I’m dead. I love how candid they are.

Robin (48:57)

So that’s kind of their psychology around AI. Alibaba obviously has you know you’ve got the e-commerce business, which is kind of stuck in this retail growth environment that’s not really growing. And the cloud business is you know has — well, you know, two years ago it was growing like plus eight, now it’s growing plus fifty, in the upcoming quarter, give or take. and so Yeah, it’s become the new thing, right? Where, you know, the hope is that they have Qwen, the hyperscaler layer, and T-Head, which is one of China’s better ASIC programs, which is worth something. And so, you know, to try and integrate that as a stack. But yeah, you know, if you look at the kind of growth algorithm of the business itself, it’s probably the compute layer that’s driving a lot of it. Yeah, just in terms of renting out capacity. So yeah, these are these are quite different businesses.

Grace Shao (49:56)

Yeah, and But the thing is they’re still going ahead, like you said, there’s Qwen, there’s Hunyuan, there’s Seed. should they still be in this model game, in this in this extremely competitive game, or do you think they should be

Robin (50:07)

It’s

Grace Shao (50:08)

Focusing on, like you said, like just plug in all the other models, focus on growing their existing business, right? Like how do you make up it? How should

Robin (50:18)

Yeah. I mean I

Grace Shao (50:20)

They balance that?

Robin (50:21)

Mean cloud and AIs probably Alibaba’s core business at this point. Like the e-commerce is kind of the cash cow that funds everything. but this is clearly the future and you know they’ve guided explicitly for cloud growth to be more than forty percent for the next bunch of years. so to Alibaba this is the core. in Tencent’s case I think there have been There have been multiple debates, you know, that I’ve had with investors around, you know, do they need a super frontier model? Do they need the best model in the market? Or do they you know can they just be the orchestrator layer? Can they just be WorkBuddy? Can they just, you know, use games or ads as a way to kind of monetize AI? I think right now the approach seems to be let’s do everything and see what sticks. or you know, like hopefully everything sticks, but you know, that’s That’s still the strategy and the you know the I guess one benefit that Tencent has is that they do generate a lot more operating cash flow through the core business that then funds a much bigger bonfire of capex over the next whatever number of years that allows gives them some level of optionality.

Grace Shao (51:30)

So obviously a lot of that money is also going to building out compute right now, right? So do you think China’s compute build out could be a double-edged sword at one point? It might erode some of that scarcity on pricing power and inference. Or, you know, some people are writing about overcapacity on compute. Like, is that a thing?

Robin (51:53)

Yeah. I mean, look in the in the in the infinite long run, and if you just kinda take that Tencent comment of if all else will rent out the compute, like if everybody builds compute with that as the fallback option, then the reasonable terminal outcome is that you get a big overbuild of compute, right? Just logically. but so that is a concern. And in the long run, you know, I guess you can make the argument that every new tech Cycle in China has ended up in some kind of overbuilding in the end. to me that’s most likely probably to happen in the compute layer because the level of specialization and tech and you know differentiation that’s required in semis is reasonably high. In AI labs, if you really wanted to be frontier, then there’s a level of math and science capability that you need, whereas standing up boxes with servers in them feels less Complicated. you and I probably couldn’t do it, but with enough money and help you know it should be doable for large corporations. so yeah, I do worry about that. I mean, like it’ll probably take a long time because you know even growing 100% a year, it’ll take a while for the domestic semis industry to catch up with demand. But you know, for now, I think compute tightness and You know, cost pressure in the supply chain is probably you know, it’s almost a good thing in the sense that it reduces the risk of everything kind of collapsing on itself and pricing you know, price wars and things of that nature happening in AI in China.

Grace Shao (53:34)

But compute abundance will be good for consumers, right? Well, at least for end users. Which is not a bad thing.

Robin (53:39)

Well yeah, so yeah. Well, I think there will be the, I guess obviously initially when there’s more compute, yes, you know, everybody has more capability to serve up more inference demand. And so the industry grows. I think and then obviously, you know, the hope is that there is a balance between demand and supply. In practice that almost never happens. and when you end up with an environment where there’s a you know, there’s a number of good enough models, there’s a lot of compute. Then it’s probably more likely that you know on one hand, you know, the cost of compute then goes down, but yeah, there’s more likely to be price competition in AI inference at that point than today.

Grace Shao (54:19)

And we start another round of juan. There’s never-ending juan so you talked a lot about you know how much the harness around a model can change the actual output. And I think it’s quite topical around the big tech right now. do you think we’re putting too much focus on how smart the base model is? do you think eventually really like the focus should be on memory, tools, routing, verification?

Robin (54:42)

Yes.

Grace Shao (54:44)

And then therefore like these big tech companies actually have A lot more experience in building these products, understanding consumer user behavior,

Robin (54:53)

I think both layers matter. I think it’s you know, if you’re gonna reduce the question to extremes, then if you just have the orchestration layer and you don’t have AI, you know, the reasoning capabilities of the model, then that doesn’t work. And I think the vice versa also doesn’t work in the sense that, you know, then you kinda kneecap the model’s ability to complete tasks and so on. So I think you’ll see the labs and The big internet companies all try and compete on both layers. one you know, back to your earlier question of, you know, why should all of these companies be developing AI models? one of the fears that I always have around these big Companies is that there’s always there’s always going to be big-company bureaucracy in politics and who takes credit for what and you know that sort of thing. And then I was in the when I was in the US over the summer, I had conversations with a couple of people who described working, you know, we were talking about Google and DeepMind at the time because it was during the week when everybody seemed to leave. And one of the ways it was described to me was you know, if you want if you think you’re going to change the world and you want to make AGI happen and so on so forth, then the most convex place where you can go and do that is at the frontier labs. And you know, doing the same thing at a big internet company feels like you’re designing a better toaster. which I mean it’s it’s not I don’t know if it’s a completely fair comparison, but yeah, and financially at the individual level, if you’ve done a super difficult PhD in something, you come out and you want to monetize that, then You know, today the most convex way to monetize that is probably to join Kimi ahead of their IPO, right? So, you know, there are those types of incentives at play as well. So, yeah, you know, so you know I do ultimately think the labs will be up there in terms of their ability to deliver these reasoning capabilities. and then the first and third party harness question ends up being kind of something that iterates in real time.

Grace Shao (56:58)

Right, right. And I think it’s kinda like going back to your earlier comment, like I think it’s what happened with Tencent when DeepSeek came out, they’re like, you know what, our Hunyuan is kind of meh. So maybe we’ll just focus on building products that are like a harness around it, like, but it just didn’t work because their own models weren’t good enough and then you can’t always rely on other people’s models, right? So these big tech are still going ahead with their own models now.

Robin (57:19)

Yeah. Well, they now own twenty percent of DeepSeek. So, you know, I think Tencent has the deep pockets and has the kind of strategic patience to be able to go down multiple routes where it, you know, you are doing your own model, you know, Hunyuan 3 was good and now Hunyuan 4 is coming soon. and meanwhile you’re still serving DeepSeek within your Yuanbao and your

Grace Shao (57:41)

Mm.

Robin (57:42)

WorkBuddy apps and so on, and over time And they use, for example, GLM in their ima app. and so yeah, they’ve always kind of done both. and you know, I guess it is reasonable that WorkBuddy allows them to see the reasoning traces of different models and that then somehow feeds back into their own model development, which is so yeah, I think Tencent’s in a slightly different position than a lot of the other guys in this conversation.

Grace Shao (58:11)

But wouldn’t Alibaba and ByteDance have the same kind of deep pockets?

Robin (58:16)

They would. But do they have the do they have the kind of social infrastructure and

Grace Shao (58:22)

Hm.

Robin (58:23)

The ability to pull everybody into a you know, an open third party harness? Whereas you know, if you look at Qwen Work, if you look at TRAE, these tend to be a lot more first-party-heavy. I don’t know if they will be as open in the fullness of time as Tencent. So yeah, they they y you’ve got these companies. And then, you know, each of these companies will then have their own internal kind of puts and takes in terms of who gets what. And so Yeah. Tencent historically

Grace Shao (58:49)

Yeah. They’re definitely all trying to follow the

Robin (58:51)

Had a bigger had a better track record of being kind of open and being the

Grace Shao (58:56)

Yeah.

Robin (58:57)

Somebody called them the benevolent gatekeeper of China Internet, which is yeah, it’s not a bad way to describe them, I guess.

Grace Shao (59:04)

Yeah, I think the other two are trying to follow the WorkBuddy route as well. They’re They’re all revamping DingTalk and Lark right now. So I think it’s like putting it into Qwen Work or something, and then TRAE—

Robin (59:13)

Yeah yeah yeah.

Grace Shao (59:14)

—TRAE and Coze were put into Doubao, and now it’s called Doubao Work, with Lark kind of grouped into it. Anyway, I wanna ask a question on recursive self-improvement. So I’m quite curious about what you think about the gaps between a lot of the labs. I don’t even want to position China versus the US, but if you have to put it that way, you know. A lot of times people are saying, you know,

Robin (59:35)

Yeah.

Grace Shao (59:36)

The Chinese advantage — sorry, the US advantage right now is that the labs are putting a lot more money into R&D and into figuring out how to go forward, right? And then the Chinese models some some ways are basically following their footprints and figuring doing the

Robin (59:48)

Yeah

Grace Shao (59:51)

Knowing the answer key to the homework. It’s that kind of analogy people are saying. But if there really is RSI, then would that change things? Then would these labs — sorry, the models themselves — just start figuring out How to improve themselves quicker and quicker and that gap would just shrink and compound with time or how do you see that?

Robin (1:00:10)

Yeah, I mean the I guess it depends on how you define RSI, but I guess the way I think about it is that there will be a gradient of different levels of how of you know automation and how to what extent AI can train AI and you know you can have them write kernels or whatever. But you know, going forward, can you ingest data and going back to the RL kind of example that I mentioned earlier, can you have AI write The verifiers and the gates and the success fail kind of conditions and you know have AI set up RL environments rather than somebody with a PhD doing it. so I think that will happen relatively quickly and then you know you start automating that process more and more. But then you still ultimately I think I’m very big on this. Like I still think you have taste and judgment be very important in terms of what kinds of verifiers you get the AI to to develop, and how you know there’s still ways to you know to do it better than the next guy rather than just have AI brute force everything. there will be some elements of that. But yeah I think you know and more and more it’ll be kind of you know the human supervising the AI doing more and more of it and but still kind of leaning on it and offering a bit of a steer in terms of where it where you want it to go and the types of behavior you want it you want to reward and so on. So yeah, I think I think there will be a gradient of how much AI versus human involvement you have in some of the model training. On the compute gap, then yes, if you have if, say, tomorrow OpenAI figures out RSI at a pretty high level and they’re able to iterate quickly, then that probably means that the gap between US and China widens again to some degree. You know, We seem to be in this Quite circular debate about whether it’s three or six or nine months and you know, I think the reality is that there’s it kind of oscillates. The US comes up with something and then China closes the gap again quite quickly. but yeah I think given how quickly information is kind of you know the flow of information between different parts of AI has been so rapid I think it’s, you know, over time the gap then, you know, maybe doesn’t widen infinitely.

Grace Shao (1:02:35)

And do you think this whole three-, six-, nine-month thing, end of the day, when we take a step back, surely it’s just so minor, no? Like how do we understand that?

Robin (1:02:44)

I think it matters to a bunch of people above our pay grade. You know,

Grace Shao (1:02:49)

Ha ha.

Robin (1:02:49)

A lot of it is being kind of hijacked in the media, in kind of geopolitical discussions and, you know, things of that nature. and if you are a frontier researcher, if you’re doing, you know, I think Anthropic has started to talk about medicine or drug discovery as a as one field that they want to be good at and If you’re trying to discover new drugs or trying to cure cancer, you know, which maybe we are now starting to do, then you do want, you know, the super frontier latest three months of model capability because and y you’ve seen enough c you know, like early-stage biotech whether these things either go up or down a hundred percent or w you know, whatever. But Because either a drug works or it doesn’t. And so for those types of use cases, yes, you do want the absolute frontier. If I’m just having a conversation with my Hermes bot, like if my model is three months out of date, does it really matter? I like to think it does. In reality it probably doesn’t.

Grace Shao (1:03:55)

No, it’s true. do you think there’s anything else that we need to talk about today, just for our audience to better understand the China AI landscape right now, where it’s at, or how to evaluate it?

Robin (1:04:07)

Yeah. Yeah, I’ll talk about something that’s a little bit of a tangent. But you know, one of these hills I’ve decided to die on is gaming. You know, I cover a lot of gaming both in China and Japan. There was this kind of moment where Google came out with Project Genie. I mean this was in this was in earlier on in the year where you know US software seemed to go down 5% a day every day. And Gaming got thrown in with that. And I think since then I think folks have come back a little bit and you know, back to the judgment and taste kind of thing. yeah, I continue to be of the view that gaming is actually remarkably difficult to disrupt and actually probably benefits from AI evolution over time. that’s probably you know and recently, you know, Google actually Was showing off one of my companies, Capcom, to demonstrate, look at how Capcom’s using AI and therefore, you know, AIs useful in the real world. And so yeah, I think I think the logic has kind of been turned on its head a little bit, but that’s that yeah, that continues to be the hill that I’ll die on when it comes to AI.

Grace Shao (1:05:21)

You think like gaming, filmmaking, these are all spaces where production cost is a lot lower, but you know, consumption will still be there. Is that kind of the thinking behind that?

Robin (1:05:32)

I think yeah, I mean production costs should come down as you automate more and more of these things. you know, you are starting to see AI video start to take over, you know depending on how complicated like you’re you’re not gonna have Nolan-level movies with generative AI anytime soon, but you are seeing short form video platforms essentially become AI-centric. so yeah, I think I think, you know, That will continue to grow, albeit the issue I’ve always had with multimodal models or video models is that, you know, the ceiling for commoditization, or the ceiling for at least you know, me not being able to tell the difference between one and the other is quite low. And so, you know, how do you differentiate video models from each other when that happens is kind of the thing that I’ve not been able to fully resolve. But yeah, if you’re just a maker of visual content, then this is great for you, right? You’re able to kind of Make stuff much more quickly. gaming in my mind is a lot more complicated because it’s not just about, you know, rendering of 3D environments or rendering of characters or whatnot. You have to have a storyline, you have to have, you know, action, music, fe you know, combat and so on and so forth that makes it more difficult. But yeah, or you know AI solves for production, you get a lot more output. I think the question in media and games and movies is like do people then care? Right? but then we’ve also just seen Niu Lai go viral. So maybe, you know, you need a more nuanced version of what people care about.

Grace Shao (1:07:10)

I like how we’re ending this conversation full circle. We started the conversation with having you look at the Ox Alpha thing and then the meme going around is the Niu Lai picture.

Robin (1:07:19)

Yeah.

Grace Shao (1:07:20)

So let’s see where that goes. I was like, is this a representation of niuma (牛马)? ‘Cause the end of the day we’re just all like worker bees and in Chinese we’re all horses and cows. But he said probably not. That’s not where the meme comes from.

Robin (1:07:35)

I’m gonna leave that alone. I’m gonna leave that alone.

Grace Shao (1:07:43)

How do you use AI in your own research, but I want to ask you like there’s just a lot of AI tracking tools out there. There’s a lot of scattered data. AI itself is obviously very broad to even kind of you know, it’s just to say what are you tracking? So actually the question for you is like, what are you using for your own evaluations or what do you track? Like, are you looking at OpenRouter, ModelScope, GitHub? Like what, what do you use to understand? Where demand is going, how compute is used, which models are good, etcetera.

Robin (1:08:15)

Yeah, I mean we track all the obvious things. I mean I’m not sure that’s differentiable necessarily, but you know, OpenRouter, Hugging Face, GitHub, ModelScope. We’ve set up kind of bot routines to try and scrape these things every so often. And, you know, we have these gigantic Excel files of, you know, daily data, weekly data. And then that at a high level paints a picture of you know who’s actually engaging with some of these things, who’s actually, you know creating repos, who’s then downloading the models, who’s doing this, that, and the other. one of my pet peeves with OpenRouter is that, you know, it’s a really, really small chunk of the market. a lot of people, especially on in the investment world, are kind of overindex on it as a as an indicator of and you see the media do it and say things like, You know, Chinese models are now two-thirds of consumption or something. It’s not, it’s two-thirds of consumption on a relatively small part of the market. but fine. You know, we do track it and we do look at kind of, you know, which apps are being used, be it kind of Z Code or some of the other stuff. So we do all that. I guess one thing that we do, which I don’t know if anyone or many other people do, is we do run our own evals. So we When a new model comes out, one thing that we do is we run it through or my bot runs it through these benchmarking tests and you make them solve tasks and it gives me I mean there’s a there’s a variety of reasons to do this, but effectively gives you a more first-hand view on whether you think something is good and then I tend to try and use them day to day because that gives you then potentially a more varied kind of read on Model feel and usefulness beyond just making it crack hardcore software engineering tasks. which, you know

Grace Shao (1:10:13)

It’s not for everyone.

Robin (1:10:14)

Well, that, but also it’s you know, these labs presumably then train on something like Terminal-Bench or SWE-bench. No one’s gonna train on my own day to day nonsense. So yeah.

Grace Shao (1:10:26)

Do you have any other tips for how us like research-driven people or our jobs are just here typing away on our computer how we can better use AI in our research Process?

Robin (1:10:40)

I will say it’s an iteration process. Like I think even you know, you have to kind of use it and be hands on, figure out, you know, what is useful for you the most and and you know find ways to kind of make it multiplicative, right? And just not just kind of have it summarize the news, but also, you know, we’ve built different Constructs on top of my data extraction funnel to try and you know have it actually analyze what’s going on, give it, you know, send me kind of reads on different things that I care about at any given point. And then I use it to iterate some of the stuff that I do day to day.

Grace Shao (1:11:24)

It can be as smart as how you make it to like how smart your inputs are, really. Yeah.

Robin (1:11:29)

I think so. And you know, what I found is you go around in circles, like you give it some stuff, you kinda work out whether it actually can do that or not. And sometimes you know, the answer’s not always yes. and yeah, and you know, try and kind of extend or you know, or try and give it you know more and more things to do until you so you know, just organically there will be use cases that kinda pop out from every so often based on my experience.

Grace Shao (1:11:59)

I don’t condone the fact that you weren’t spending time with your daughter full on one on one and actually on your voice AI tool. But my husband’s gotten to a point where he’s wearing his Apple Vision Pro and he has like seven of his agents in a line and he’s just sitting there like controlling them. I’m just like it’s gone too far. Like I think if there’s gonna be like a pushback by all the kids and wives at this point, not to overgeneralize. Now, look, what has really changed for you or your view on over the last six months? you know, has something fundamentally shifted in your research or something your understanding of AI and the technology.

Robin (1:12:39)

My understanding of AI seems to change every two weeks. So, you know, that the whole process of being in January, staring at these two IPOs, and then you know, fully going down the rabbit hole of trying to understand them and keeping you know, trying to keep track of everything, build a network around, yeah, I mean it’s it’s been it’s been wild, how much everything has changed. I’d love to be able to distill it down to one thing, but I think the reality is just, you know, everything has changed and the thing that I’ve discovered actually that’s been interesting is you know, I run a small team and there are people who normally work for me and in the old world I used to kind of staff them on a small handful of things a day. And then they will go away and do it and they come back and I would have to make, I don’t know, three or five consequential decisions a day or, you know Have kind of deep, deep conversations with myself about how to do something. Now, increasingly, with AI — I mean they’re still doing that, but at the same time I’m knocking around ideas in my head and I’m sometimes using AI bots to help me kind of process what I think. and they come back so quickly that I find myself having to make five consequential actions an hour. And then at the end of the day I’m like, you know, my workload has increased as a result, which is I don’t know if that was the desired outcome, talking about AI making people’s lives better. But there you go.

Grace Shao (1:14:15)

I think it’s just because you’re too much of a doer and too competitive. I think for people who are high-agency people, they are doing more and they’re experiencing AI fatigue. For people who wanna just clock in, clock out and just do the three things they were told to do, their lives are made easier. You know, it’s like you know, writing three emails used to take maybe like a day, now it takes twenty minutes or less.

Robin (1:14:35)

Then I wouldn’t be on this pod, so there are upsides.

Grace Shao (1:14:39)

Look, one last question for you. I asked everyone, what is one differentiated view you hold or you know, something a bit non-consensus you think?

Robin (1:14:49)

Yeah, I think the gaming thing is probably the most controversial. you know, or the maybe the broader idea of I think if you want I think it’s it’s the whole idea of of judgment and taste where, you know, there’s still this belief and it’s almost a religious belief at this point, which is because I don’t know how to disprove it either. which is I think on some level it’s you know That will always be valuable, the idea that I’m still of the view that it’s quite difficult to get AI models to spit out something that’s outside of its own distribution and training data. And so, you know, on some level you still have human input being the arbiter of something that’s ninety percent of the way there and something that’s truly kind of differentiated. So yeah, I vote at the end for humankind.

Grace Shao (1:15:47)

But then I have a follow-up question on that. It’s just like I think it’s easy for people who’ve built up taste or, you know, like yourself, you’ve been in the industry for long enough to build up your own taste, your own judgment. How do people without that kind of experience build up human taste still when they’re joining the workforce or they’re growing up like our children are growing up at an age where, you know, AI is natively embedded in everything they do? So, how do you still differentiate that taste? Because You know, now like we’re seeing these like parallel structures of sentences everywhere you go. It’s driving me crazy. Even Instagram ads are like, you know, they have like these very obvious parallel structures and it’s like driving me crazy. I’m like, dude, like this marketing associate did not write this, but I think what you’re

Robin (1:16:29)

Okay.

Grace Shao (1:16:29)

Seeing is people without that taste judgment and now just copy and pasting whatever AI spits out.

Robin (1:16:36)

Yeah. I mean we’re gonna end this conversation like every good Asian parent and talking about parenting at the end of it. no, look, I mean it is a conversation I have with myself. Like how do you educate somebody or how you know, whether it’s your own kid or whether it’s somebody that you work with or whatever, then, you know, how do you develop taste from first principles and Yeah, I don’t know. I don’t have a great answer, but I do think exposing yourself to original content and doing things the hard way, at least in the beginning, is still important.

Grace Shao (1:17:16)

Well, thank you so much for your time. Very generous with your time today, Robin. I really appreciated your insights and everything.

Robin (1:17:21)

Appreciate our conversation.

We talk about Manycore’s evolution from startup to public company, including its IPO earlier this year and how the company’s story has changed since going public. While its core SaaS business still accounts for the majority of revenue, Manycore is increasingly positioning its proprietary 3D data and technology as a foundation for spatial intelligence and new applications.

We also dive into SpatialVerse, AholoWorld and Manycore’s work in robotics and embodied AI. Bei explains how the company thinks about spatial intelligence—not simply as a data business, but in terms of the systems and simulation environments needed to help AI understand physical space. We discuss potential applications in robotics, game design and filmmaking, as well as the question of how much intelligence different types of robots actually need.

Finally, we discuss Manycore’s global expansion, partnerships and long-term strategy. We explore whether spatial intelligence will become a market dominated by a few global platforms or remain fragmented across industries and geographies, and what Manycore sees as its role as more robotics companies begin building their own physical AI systems.

Subscribe now

The AI Proem Podcast is under the AI Proem newsletter, which has over 12k followers globally. To learn more about China AI, the business of AI, and how AI is impacting businesses, please check out the newsletter here and more insightful conversations here.

Chapters

01:18 Leaving Investment Banking for a Startup in Hangzhou

05:29 From Silicon Valley to Going Public in Hong Kong

06:19 First of the Six Tigers to IPO

07:28 3D Shift Toward Spatial Intelligence

14:20 Data, Simulation or System: What Is the Product?

18:14 Learning More of SpatialVerse and AholoWorld

31:59 The Role of Spatial Intelligence in Robotics

34:04 How Manycore Fits Into the Robotics Stack

38:11 Global Expansion and Strategic Partnerships

42:31 Will Spatial Intelligence Consolidate or Fragment?

46:20 Manycore’s Focus and Strategic Pillars

AI-generated transcript (for reference only)

Grace Shao: Hi, Bei. Thank you so much for joining us today.

Bei: Hi, Grace. Good to be here.

Grace Shao: Yeah. So, you guys are one of the hottest AI companies that listed in Hong Kong this year, and a lot of people have a lot of questions. But to start with, tell us about yourself. When I was learning about your background, I thought it was quite fascinating. You’re almost like a Joe Tsai story. You had a very successful finance background and a successful career in Hong Kong, and then you decided to jump over to Hangzhou to join what was still a relatively unknown startup. Tell us what made you want to make that jump and leave your cushy banking role. What was the spark about this company for you? And tell us a little bit about where the company is right now.

Bei: Sure. My name is Bei Shen. I’m CFO of ManyCore. I joined the company in 2019. Before that, I was an investment banker for 14 years. I worked for Citigroup in New York, then moved to JPMorgan in Hong Kong. The last nine years of my banking career were at Goldman Sachs.

Around 2018 or 2019, I started thinking about what I wanted to do with the rest of my career. Traditionally, I focused a lot on clients in more traditional industries. I covered companies in the power, mining and energy spaces. It’s an interesting job, but the sectors are relatively traditional.

So I started asking myself how I could get more exposure to technology and internet companies. For me, it was very difficult to switch industries within the bank, so I started looking around for opportunities. Luckily, ManyCore was looking for a CFO. After talking with the founders and the team, I found the company very exciting, so I joined in 2019. I can’t believe it, but it’s been almost seven years now.

Grace Shao: Yeah. And I know you recently took the company public, but before we get into all that, tell us about your three co-founders, because they have quite interesting backgrounds. They’re quite young. They came back from Silicon Valley, bright-eyed and wanting to start something in China. Tell us about the vision they had in the early days and where it has led you now, roughly 15 years later.

Bei: Yes. The company was founded around 2012. The three founders were classmates at UIUC, which has a very strong computer science program in the U.S. Our chairman, Victor, and our CEO, Chen, actually went to the same undergraduate university, Zhejiang University, which is also where our company is based. We still recruit a lot of people from Zhejiang University, which is a great school. And our CTO went to Tsinghua.

All three of them studied at UIUC in fields related to computer vision and high-performance parallel computing. After graduation, they all went off to cut their teeth in Silicon Valley. Victor worked for NVIDIA for a couple of years, Chen worked for Microsoft, and our CTO worked for Amazon.

They were all working in Silicon Valley, but they wanted to come back to China and participate in this exciting market. Back then, Victor had this idea of putting GPUs on the cloud to serve more clients. He was working at NVIDIA on the CUDA team, so he was involved in the early days of figuring out how to put compute on the cloud and serve more customers. Obviously, this was before AI became what it is today.

They built a demo and came back to China. Luckily, the Hangzhou government was welcoming overseas graduates and helped them start the company.

The original idea was very simple: they wanted to put GPUs and compute in the cloud and make that compute available to more people. In the beginning, it was very difficult because AI hadn’t taken off yet, autonomous driving wasn’t in full swing, and crypto wasn’t either.

Luckily, they found a very interesting application in interior decoration. In the old days, if you used on-premise software, it could take a very long time to render a photorealistic picture. With their technology, you could put that computation on the cloud and use multiple GPUs to accelerate the rendering process. That enabled users to create photorealistic renderings in minutes. Now it’s seconds.

That really changed the industry in a big way. So that’s how they started. It was fundamentally a technology company trying to find applications for its technology.

Grace Shao: How would you describe the company today? How would you position ManyCore in two or three sentences? Clearly, it’s no longer just about Kujiale and 3D interior design.

Bei: Obviously. The company has had 14 or 15 years of history. Before 2023, we were basically the largest 3D design software provider for interior design. But since 2023, the company has increasingly focused on spatial intelligence.

To put it very simply, we’re trying to help AI perceive, create and eventually act in a three-dimensional world, so that AI can eventually move from the digital world into the physical world. That’s where the company is focusing right now.

Grace Shao: Perfect. I think you were definitely one of the hot IPOs earlier this year. You went public in April and were one of the first of the Hangzhou “Six Little Dragons” to list. It felt like a point of pride for Hangzhou and for this new wave of Chinese AI companies. What did going public mean for you and for the company?

Bei: Obviously, it’s a big milestone for the company. We raised fresh capital to fund our future growth, especially in spatial intelligence. We need more compute and we need to hire more talent.

But from a business perspective, it also put us on the international radar. We already have many international clients, but it can still be difficult for a Chinese technology company to sell products to overseas customers. Being a public company, with your company story and financials becoming more transparent, definitely helps a great deal in promoting ourselves and selling our products in markets outside China.

Grace Shao: I want to double-click on something you said earlier about how the company evolved. When you filed the prospectus, I went through it, and it was still mostly focused on your 3D interior design technology. Now you’re clearly pushing a new narrative around spatial intelligence, which frankly wasn’t emphasized nearly as much even a year ago when you filed the prospectus. Things are moving so fast.

Tell us about how that shifted and why you had this moment of pivot. Was there an epiphany during the process of going public, or after you went public? Did something hit you where you realized there was this gold mine you were sitting on? Tell us the story behind that.

Bei: Sure. That’s an interesting question. Just to go back a little bit in terms of our IPO history, we really started preparing for a Hong Kong IPO in the third quarter of 2024. Then we filed our prospectus on February 14, 2025.

The IPO process is relatively lengthy for Chinese companies because every company going public needs approval from Chinese regulators. For us, it took a bit longer because of our structure. We finished the IPO in April this year. So looking back, the process took almost a year and a half.

Obviously, both the company and the industry changed enormously between the day we started the IPO process and the day we actually listed.

Our thinking was that it would be unreasonable, or even impractical, to keep updating the prospectus every time the company changed because this industry moves so quickly. So we made a decision to keep the discussion of the new business and products relatively minimal.

That’s why, when you read our prospectus, you see a lot of disclosure about our older, existing business, which is obviously still important. But the new businesses were changing so much that we didn’t go into as much detail.

After the IPO, we started talking to more analysts and investors and trying to give them a more updated picture of where we stand in spatial intelligence.

Grace Shao: For sure. So what are the top-of-mind questions or areas of interest you’re getting from investors right now about the business?

Bei: This is a very frontier area. Large language models have obviously received a lot of attention over the last couple of years since ChatGPT came into existence. There have been many advances in model capabilities, coding, image generation and video generation.

But we’re focusing on a relatively different type of AI. We sometimes call it physical AI. As I said, we’re trying to help AI understand the 3D world, which is very different from reading text and giving you an answer, or generating a picture or a video.

One of the challenges in our space is that we don’t have nearly as much data as large language models do. They can access internet text and enormous amounts of video. In our space, the amount of data is several orders of magnitude lower and much less dense compared with text, pictures or video.

That’s why it’s difficult. It’s very hard. People also need to spend some time understanding what we’re doing because I believe we’re working at the frontier of what could be the next wave of breakthroughs in AI.

Grace Shao: My understanding, according to your public filings, is that roughly 90% of your revenue is still coming from the traditional business, particularly Kujiale, and that’s really funding the new initiatives.

As you move into spatial intelligence and physical AI, you touched on data as the bottleneck. But data is also part of your moat, right? You’ve had more than a decade of experience working with 3D data. Tell us more about the connection between your traditional business and this new business.

Bei: Sure. I wouldn’t say data is the only reason we chose spatial intelligence, although it’s obviously a very important aspect of the business.

There are many connections between our traditional Kujiale business, or Coohom internationally, and what we’re focusing on today.

Even 10 or 12 years ago, we adopted a very integrated technology architecture. We bought our own GPUs and did rendering using our own GPUs in order to provide the service to customers worldwide.

That’s actually quite similar to what large language models or 3D models are doing today. You’re utilizing compute to provide products and services to people around the world over the internet.

So that’s one connection in terms of the technology lineage. We did a lot of hardware-software optimization to make sure rendering could be provided at the lowest possible cost. Similarly, if you want to do inference today, even if you have a very good model, you still need to keep inference costs low in order to remain competitive. That’s something we’re obviously very good at.

Data is also very important. We’ve accumulated a large amount of data. In hindsight, it’s fortunate that, compared with the on-premise software that came before us, we had all the data on our cloud platform. That wasn’t necessarily by design at the beginning.

But in today’s world, 3D data is extremely valuable and very difficult to obtain. Because of our cloud-based architecture, we’ve been able to accumulate a large amount of data, especially structured 3D data, which is critical for training 3D understanding and 3D models.

So between our technology lineage and our data library, I think we’re in a very unique position to explore spatial intelligence.

Grace Shao: So you have a lot of 3D data, but this is spatial data. It’s not necessarily the movement or motion data people talk about needing for robotics training.

But when I spoke to your team while visiting Hangzhou, it sounded like a lot of your clients may actually be robotics companies. What are you providing to them today? Is it a 3D intelligence system? Is it data that helps robots operate better in physical space? Or is it a bit of both?

Bei: This has an interesting history. I think it was back in 2022 or 2023, during the pandemic, when nobody could really go anywhere. We were all stuck in offices or at home and couldn’t travel abroad.

We received an email from Silicon Valley from one of the large technology companies. They came knocking on the door and said, “I heard you guys have some interior-setting data.”

We said yes.

They said, “We’d like to buy some.”

At the beginning, we thought it was spam or some kind of trick. But it turned out to be a real client doing research.

We didn’t think too much about it. We struck a deal and helped put together some synthetic data they required. We made some money, not much.

Then the next year, another technology company came and asked for something similar. This time, we took notice. We thought, okay, there must be something valuable in our data.

So we started asking these U.S. clients, “What are you actually doing with our data?” We’d had it for over a decade and hadn’t really thought too much about it. Luckily, we hadn’t deleted it just to save storage costs.

Only then did we find out that these were large technology companies in the U.S. training robots. They needed synthetic settings in which they could test and train their robotics policies.

That’s when we realized we were sitting on something interesting and valuable. We started thinking about how we could better commercialize the data we had.

Obviously, we’re not satisfied with simply providing raw synthetic data. Right now, we’re speaking with customers both in China and the U.S. and trying to help them train and evaluate their policies more effectively.

Eventually, we also hope to train our own model.

We believe that if you want a world in which robots, or physical agents more broadly, can become fully autonomous, they need their own brain. The ability to perceive and understand physical surroundings, reason about them and act within them may require spatial intelligence. That’s something we’ve also started working on ourselves.

So right now, we’re still providing synthetic data to customers. We’re also trying to train our own model, which we eventually hope can be put into intelligent robots or embodiments of different forms so that they can really act in the physical world.

Grace Shao: That’s really interesting. There’s a bit of serendipity there. Things just happened and you were there at the right time with the right data.

I was reading through your materials. There’s something called SpatialVerse, and you’re also releasing HoloWorld. What are these things for? Tell us more about these products.

Bei: Sure. These names keep popping up. Sometimes I get confused as well because things change so fast.

As I said, in 2022 or 2023, we started selling synthetic data solutions to robotics companies. Later, AR and VR companies also came to us asking for something similar because they need data to train goggles or glasses.

We put this type of business together under the name SpatialVerse. That’s one line of activity and business we’re developing.

As I said, eventually we’d like to train our own models and put them into robots.

Another branch of research we’re working on is helping agents create synthetic worlds. This is somewhat similar to what Dr. Fei-Fei Li’s company, World Labs, is doing.

Basically, using a simple prompt, text or pictures, you can quickly create a virtual space where you have geometric information as well as information about the objects within that 3D space.

People talk about “world models” a lot these days, and sometimes the term is misused or misinterpreted. But the way we understand it, in order to have this capability, you really need a model that can generate a 3D world.

It’s not just a continuation of pictures. There are very good models today that can give you 20 or 30 seconds of short video. But what we’re after is the ability to use computing technology to generate a 3D world where the objects within that world have certain physical properties, and where you also understand the geometric relationships between those objects.

That’s critical for robots eventually being trained inside that world.

Grace Shao: So rather than a video-generation model, which is where a lot of multimodality efforts are going right now, you’re really trying to create 3D spaces. It almost feels like creating a little Sims world.

What are the use cases then? Off the top of my head, there might be game design, filmmaking and, of course, robotics training. What do you think people are missing when they think about what this tool or product could eventually be used for?

Bei: That’s a great question. Again, this is relatively new.

Historically, 3D design has been a relatively niche market. Before us, you had all these on-premise 3D design software products, but they’re difficult to master and use.

Compared with picture editing, 3D design has historically been quite niche. It was mostly used in architecture, industrial product design and VFX.

What we did was make the 3D world easier for ordinary people to generate.

One area where we’re already helping customers is media. In China, one-minute and two-minute micro-dramas have become very popular. You see many of these productions on Douyin, and a lot of the content is already being generated by AI.

We’re helping some of these creators produce spatially consistent 3D worlds. If you rely purely on video generation today, you can get hallucinations after a certain amount of time. Objects start floating around. You leave a room, come back, and suddenly the objects are missing.

Using our product, these producers or content creators can maintain a spatially consistent 3D world and produce something higher quality. You don’t have the same spatial hallucinations you get from video models.

The other application is robotics training and evaluation, which we’ve already discussed.

It’s inconceivable that you can train robots in every possible physical environment. Thinking through all the corner cases would be too expensive and too time-consuming.

So in order to train and evaluate these policies, I believe it’s critical to have virtual worlds where you can put robots through testing virtually.

Those are two areas where we’re already putting our technology to use. Eventually, I think there will be many more applications.

Grace Shao: That’s really interesting. I want to double-click on the micro-drama example.

To help me understand, are people using your technology in parallel with something more traditional like Seedance? One is more for aesthetics and one is for spatial control? Is it layered?

Or could your technology eventually compete with and replace a more general-purpose video model like Seedance?

Bei: That’s a great question. Right now, the way we serve customers is really a layered approach.

We already have a product out that you can try called LuxReal. We rolled it out about a month or two ago.

Basically, you can upload a script, and it’s an agent that helps you produce a 30-second, one-minute or two-minute micro-drama based on that script.

First, you use our product to create the 3D world. For example, if you want to shoot a micro-drama set inside an ancient Chinese palace, after reading your script, we’ll generate that palace for you. Let’s say you want two rooms inside Beijing’s Forbidden City. We can create those rooms, and then you can move your camera around within that space to shoot the scenes.

Our customers don’t only use our modeling capability. They also use Seedance because we don’t actually do the video-generation part right now.

We help you create the spatially consistent 3D setting, and then you put Seedance on top of that. You can very quickly produce a one- or two-minute micro-drama.

If you only use Seedance, you may have to do a lot of editing afterwards because certain things don’t look right. You have to spend manpower editing out hallucinations.

With our product, the room is always there and the table is always there. The table isn’t going to change when your camera changes.

So it’s basically a very efficient tool for these micro-drama producers.

Grace Shao: That’s very interesting. So essentially, you’re producing one asset layer in the broader workflow for these creators.

It’s funny because when I was talking to people at Kling and Kuaishou, they said people don’t necessarily care about inconsistencies yet. Sometimes the cat turns out white, sometimes the cat turns out black. There’s definitely an understanding that AI content, as of now, isn’t that sophisticated.

I want to shift the conversation to models and robots.

I’m going to put you on the spot here. You mentioned Dr. Fei-Fei Li. Yann LeCun and Fei-Fei Li are both world-renowned scientists working on something around this realm. They’re both trying to push forward ideas around world models. Obviously, this is still in a very nascent stage.

How do you see the field? And frankly, how do you see your company’s position among all these global competitors, which have a lot of technological influence and obviously a lot of capital behind them as well?

Bei: Great question. I wouldn’t say we’re directly competing yet. As I said, this is a fast-evolving and changing industry, and everyone is still doing a lot of exploratory work.

Dr. LeCun’s approach, to be honest, I don’t understand in great technical detail because I’m not computer-science trained. I’ve read about it, and one apparent benefit of his approach is that you may not need as much compute to come up with these highly realistic representations.

We haven’t paid too much attention to his work yet, but obviously we’ll be very interested to see what he develops.

Dr. Fei-Fei Li’s approach is more similar to ours. Both companies are trying to figure out an efficient model for creating a 3D world where you not only see the world as we see it, with the correct textures, sizing, depth and perception, which is the rendering side and something we’re very good at, but where you also embed more information.

We’re trying to add physical properties such as friction coefficients and how wind behaves. As customer requirements evolve, we’ll try to put more and more information into that model.

Eventually, it will be interesting to see what applications emerge outside media and robotics training when people have these kinds of worlds available.

Grace Shao: I listened to one of your CEO’s interviews, and he said that what you’re trying to do is create spatial intelligence that can help translate physical space so LLMs can better understand it.

He also discussed how world models shouldn’t necessarily be produced by every individual robotics company, or at least shouldn’t be viewed as interchangeable, because each use case can be so different.

So I guess my question is: how should we think about a robot being used to remove blood clots, where the work is incredibly meticulous, versus a robot designed to lift heavy objects? What kind of 3D data and 3D model does each need?

It seems like “world intelligence” or “world models” is too broad a category to serve every single demand in the space right now.

Bei: I’m not 100% sure which episode or interview you’re referring to, but I think what he was probably talking about is the robotics industry today.

Obviously, the level of optimism varies depending on who you talk to. But based on our conversations with the industry, a general-purpose humanoid robot is still pretty far away.

I wouldn’t say it’s next year. Maybe it’s five years, maybe it’s 10. But just imagine a humanoid being able to do 100 chores in your household. I think that’s still quite a few years away.

There are so many challenges. One of them is exactly what we’re working on: can you teach a robot to understand different settings?

The minute it walks into a room, can it immediately understand, “Okay, this is a table, this is a desk, and these are the relationships between the objects”?

We’re still working on that.

Building a general-purpose, fully autonomous humanoid robot is hard.

But if you narrow the problem down to more vertical or specific-purpose robots, I think that’s more doable. The level of complexity and intelligence required is much easier to achieve at this stage.

I think the industry is more likely to evolve through more and more specific-purpose robots. One robot might pick up boxes. Another might help with laundry. I’m just giving examples.

That seems like a more likely roadmap than trying to immediately build one general-purpose robot with omnipresent capabilities.

As you build more and more of these scenario-specific robots, maybe eventually you arrive at a stage where a more general-purpose robot becomes possible.

That’s how we see the world and how we’re tailoring our R&D efforts. We’re not trying to go after one extremely general spatial-intelligence model right now. We’re trying to crack these silos one by one.

Grace Shao: So which verticals are you focused on today, out of all the different kinds of robots you’re serving?

Bei: To give you a few examples, we think household robots are probably difficult, at least in China, because Chinese households tend to live in relatively small spaces. The margin for error is extremely small.

So right now, we’re focusing on helping robots in more industrial settings.

For example, in warehouses, we can teach robots to quickly understand the warehouse because the level of complexity is relatively lower than in a family setting.

We’re also helping some robotic dogs patrol power stations. They need to walk around, identify anomalies, record them and report them.

We believe these are some of the low-hanging-fruit applications today, where we can teach robots or robotic dogs to perceive and understand the 3D world.

Grace Shao: Do the economics make sense right now? Frankly, if you’re trying to replace relatively simple tasks or labor, especially in China or elsewhere in Asia where labor costs are relatively low, does it make economic sense?

Bei: That’s a great question. The economic equation is definitely important as we put more effort into this.

Some of the key areas we’re trying to explore are places where it’s dangerous or costly for human beings to operate.

For example, around high-voltage power stations or transmission lines, it’s definitely safer to have robots patrol instead of human beings.

Or in remote areas and underground mines, if you have water leakage or some geological situation, it’s safer to send robots and perhaps drones to inspect first rather than sending a human rescue team directly.

Grace Shao: That makes sense.

But my understanding is that you’re purely on the software side right now. Does it make sense for these humanoid robotics companies to pay for your service and technology, or does it make more sense for them to train and build their own models internally?

How do you view that? There are obviously different camps. Some people say OEMs can build different types of hardware while companies like yours provide the intelligence underneath. How do you see that trend?

Bei: Right now, we’re only focusing on software, as you correctly pointed out.

We’re trying to be model-agnostic, and we’re also trying to be embodiment-agnostic.

Basically, we’re trying to develop technology that different robotics companies can use to train and evaluate their policies.

Obviously, we’re not there yet, but that’s our goal.

We’re not trying to build robots or robotic dogs ourselves. We’re trying to help these companies find a very cost-effective way to evaluate their policies. That’s our approach right now.

Grace Shao: What do you think people are getting wrong or misunderstanding about the industry today?

In spatial intelligence and robotics, there’s obviously a lot of buzz and a lot of hype. Unitree has been getting a lot of attention as well.

Do you think there’s too much hype right now and that we should be more cautious because progress is still going to be slower than the public expects?

Or do you think the misunderstanding goes the other way, and people are underestimating how quickly this technology could proliferate in niche use cases and eventually reach consumers?

It’s a big, open-ended question.

Bei: Sure. From my perspective, I obviously believe this technology has very broad applications in the future. But the capabilities have to get there first, and the cost has to be low enough for the technology to proliferate.

I believe spatial intelligence is a critical part of human intelligence. It’s almost innate to us. Dr. Fei-Fei Li has made a very good argument around that.

After thousands of years of evolution, human beings can see things and quickly understand their geometric and 3D relationships.

That’s something large language models don’t really have today, but it’s critical if AI is going to operate in a physical context.

Right now, it’s great that AI can solve math problems or write poems. But can you actually ask a robot to do your laundry? Can you trust it to do all these tasks?

Right now, we’re not there yet. But I believe we’re on the way.

I wouldn’t necessarily call it a misunderstanding. I think the difference in opinion is really about how long it’s going to take.

As I said, one of the biggest challenges facing our industry is data. We need to find smart and cost-efficient ways to obtain more data because models are a product of that data.

Large language models are really the product of compute multiplied by data. Our industry is no different.

That’s why we’re thinking about different ways of capturing more 3D data. We’re working with different hardware companies. Robotics companies are one category. Scanning companies are another.

We hope to have more hardware companies work with us so that more users can use our technology to capture or generate 3D data.

In the long run, we need a flywheel where applications, data and models all improve in tandem, level by level.

But right now, the flywheel isn’t flying yet. We’re working very hard to push it forward.

Grace Shao: That makes a lot of sense. More users mean more use cases and more scenarios, which give you better data. Better data improves the models, which then lets you serve clients better.

On partners and clients, you mentioned earlier that you already have quite a global footprint. I think that’s fairly unique among Chinese companies that are trying to go global today.

When you think about international expansion and distribution, what’s most important? What kinds of partnerships are you looking for?

You mentioned that some large Silicon Valley technology companies are already clients. How should we understand those relationships?

And more importantly, do you face localization as a bottleneck, or is that less of an issue in your particular sector?

Bei: Great question.

Putting it in the context of Kujiale or Coohom, localization is actually very important.

For example, if you want to sell an interior-design product in the U.S., the industry is very different. China and the U.S. both consume furniture and decoration services, but the industry relationships and dynamics are very different.

Localization is therefore critical.

Just to give you a very simple example, even the measurement systems are different. China uses meters, while the U.S. uses feet and inches.

That’s a tiny example, but there are many localization changes you need to make in order to fully satisfy the local market.

What’s interesting is that this has changed quite a bit in the AI context.

ChatGPT basically became global overnight. I think one reason is that the model itself became so powerful.

You can almost think of the model itself as the product. You don’t necessarily need to build many layers of user interface on top of it.

We’re beginning to see that in our space as well.

If you have a very powerful world model, for example, you can generate 3D objects or scenes relatively easily. There may still be differences in language, but in terms of usability and application, it becomes much easier to promote globally.

That’s a big opportunity for us.

Eventually, we hope to offer a product that has a global appeal similar to ChatGPT, where you have users all around the world.

Obviously, that’s not easy. First of all, you need to come up with an extremely strong model that can actually serve clients and users worldwide.

Grace Shao: So what I’m hearing is that on the consumer-facing side, something like Kujiale obviously requires more localization.

But if you increasingly position yourself as a B2B support technology or an infrastructure layer underneath consumer-facing products, you may need less localization. Is that a fair understanding?

Bei: That’s a fair summary.

Coohom, by the way, is the international version of Kujiale. It’s not just a language translation. We’ve adapted Coohom depending on which market we’re entering, so we made a lot of localization changes.

But for a new product like LuxReal, the micro-drama product, we didn’t have to do much localization besides language.

That gives you an idea of how, in this era, products can become international much more easily because the underlying layer becomes extremely powerful and important, while the application layer on top can be relatively simple.

Some clients can even develop their own applications based on our technology.

That’s how we envisage the future.

We’d like to develop very powerful models and offer them through APIs or SDKs. People can then do their own development and build secondary or tertiary applications on top of the model.

That’s a change in paradigm compared with the past. As a software provider, you used to have to build many of these applications yourself.

Now users and customers can use things like vibe coding to build many applications themselves. We don’t necessarily have to do all of that anymore.

Grace Shao: That makes a lot of sense.

So just one last question on this part: should we understand the future of spatial intelligence as being more fragmented and vertical by sector and use case, rather than by geography, compared with how software evolved during the internet era?

Bei: I would tend to agree with that assessment.

As I said, because the data is so difficult to obtain, I think the industries where we can establish these small data flywheels will develop more quickly than others.

So I believe it’s going to be a more fragmented landscape compared with large language models, where eventually you may have fewer than half a dozen truly global companies. There are obviously more today, but I think LLMs will ultimately become quite concentrated.

In large language models, the data is relatively open to everyone because everyone has access to the internet.

A lot of the competitive landscape is therefore determined by who has more compute or who has the best talent and algorithms. Large companies have a huge advantage in that environment.

Our space is different.

There isn’t a universal 3D data library where everyone can simply start working on the same dataset.

First of all, obtaining the data itself is a challenge.

We obviously have an advantage because of the work we’ve accumulated over the years, but eventually we still need to find more and more methods of getting additional data.

So I think the game is somewhat different from large language models.

Grace Shao: I want to take a step back.

You guys are based in Hangzhou. Like you mentioned earlier, there was an effort in Hangzhou to attract people to come back because of how strong the ecosystem is.

Obviously Alibaba is there, Ant is there, there are a lot of e-commerce players, and many of the startups that came out of Hangzhou over the last decade have somehow been related to e-commerce.

It’s interesting that you didn’t get sucked into that orbit.

When I was reading about your story and looking at the earlier days, I thought it was funny because you could very naturally have gone into e-commerce staging and 3D content creation. That could have been a logical path for serving domestic clients, especially given that you were based in Hangzhou.

What was the thinking behind not going into that vertical?

Bei: We’re no exception. We tried e-commerce. It didn’t work out.

Grace Shao: I love the candidness.

Bei: We’re no exception.

Going back to Kujiale’s early days, we came up with this interesting software for designers. The natural next step was: can we sell furniture?

We tried. It didn’t work out.

I think that’s probably largely due to the genetics of the founders. They weren’t from that industry. They’re not e-commerce experts.

It’s also simply the nature of furniture and home decoration. It’s very difficult to commoditize. It requires a lot of service. It’s not like selling a book or laptop through e-commerce.

Even Alibaba tried, and I don’t think the results were very satisfactory.

So we dabbled in it. We burned some investors’ money, but not too much.

Then we realized, okay, it’s not for us.

We came back and said, we’re going to focus on software.

And lo and behold, we found this opportunity in spatial intelligence.

Grace Shao: Definitely. That makes a lot of sense. Furniture isn’t an easy thing to sell. It’s tailor-made, personal, huge and bulky, and logistics aren’t easy. I can imagine it’s not an easy business.

Looking forward, as CFO, I’m sure you’re thinking about how to invest the newly raised money and looking at the next three- to five-year horizon.

Where should we be looking? What is the company’s focus? What are the strategic pillars for you?

Bei: Great question. We think about this every day.

Going back to the basic AI paradigm, it’s always formed around compute, algorithms and data.

We’re really going to focus on those three things as we try to push the company to the next level.

Data is probably the hardest part because it’s not just about money. You need to think about smart ways of obtaining that data.

Talent retention and talent recruitment are also obviously the number-one priority for management.

You definitely know how expensive data scientists and algorithm scientists have become these days.

Grace Shao: How much are they making these days in China? Give us a range.

Bei: Not as much as their U.S. counterparts, I think. But even for college graduates fresh out of school, if you have the right experience and pedigree, you can make at least five to 10 times what a traditional software engineer might make.

It’s a very highly sought-after pool of talent.

Grace Shao: Okay, so are we talking about RMB 2 million to RMB 3 million? I’m trying to force you to give us a range. Five to 10 times is a big figure.

Bei: No, no. It’s a big figure, but software engineers don’t make as much as they used to anymore.

Grace Shao: The irony in all of this.

Bei: Exactly.

So talent is obviously a huge priority for us.

We’re also looking at interesting opportunities because we don’t know where the next technology is going to come from.

These days, acquisitions are really about people and talent.

If we see interesting algorithms or ideas coming out of labs, we’ll consider making our own moves.

As you know, we work very closely with Zhejiang University. We have a postdoctoral lab with Zhejiang University where we put a lot of effort into computer-vision research together.

Hopefully, we’ll identify talent and interesting early-stage products along the way.

Grace Shao: Would you go into hardware? Would you build your own robots?

Bei: Not right now. It’s already a pretty busy space.

But we’re definitely looking at hardware, probably not robots directly.

As I said, we’re already working with hardware companies to collect data.

We work with some scanner companies and LiDAR companies in China to collect 3D data.

To collect 3D data, you don’t only need cameras. You also need LiDAR, which gives you geometric information.

For example, we’re working with Hesai, which is a very good LiDAR company, to come up with solutions.

So we’re starting to dabble in hardware. We’re not purely a software company anymore.

Going forward, I think the two are going to be coupled together.

Especially in our world, if you want to have state-of-the-art 3D models, you will definitely need help from hardware companies, and we’ll probably do some of it ourselves.

Grace Shao: That makes a lot of sense.

You guys are well positioned given the amount of interest in you right now, and given that you’re in Zhejiang and close to Zhejiang University, where there’s a lot of talent coming out.

I think over the last year, the West has really opened its eyes to Zhejiang University, but in China everyone already knows it’s an absolute top-tier school.

You have people like Liang Wenfeng, and there’s just a lot of talent coming from that region.

Anyway, I really appreciate your time.

I want to ask you one last question, which is something I ask everyone who comes on the show: what is one differentiated view you hold? Something you think is non-consensus?

Bei: I think the TAM, the market for spatial intelligence, is actually going to be bigger than the market for large language models.

It’s still a little early, but if you look at human intelligence, language is only one part of our intelligence. Spatial understanding is also critical from an evolutionary perspective.

Eventually, if we can help AI crack that capability, the applications will be extremely broad.

I think there will be many things that robots or agents can eventually do that people probably haven’t even thought about yet.

It’s still early. It obviously requires a lot of exploration and effort, and there will be many pitfalls along the way.

But we believe this is a very, very large opportunity, and we’re fully committed to it.

That’s one view I think may still be a little bit off-consensus today.

Grace Shao: Thank you so much. I think we still have a long journey ahead.

Bei: Thank you, Grace.

Grace Shao: Thank you for your time, and congratulations again on the IPO.

Bei: Thank you very much, Grace. Nice talking to you.

I am so delighted to welcome Ella () to Substack and introduce you to her work. Ella is a good friend based out in the Bay Area. My husband and I always go and mooch off of her kitchen yumz/ she takes us to the best eats when we’re in town.

Ella is a former partner and investor at Sequoia’s hedge fund, focusing her time on all things China, physical AI, and software, among other things. She recently started writing to crystallize and share her thoughts on her favorite topics, as she spends time with her newborn humans.

This is an extremely information-rich and timely piece. Chinese robotics manufacturer Unitree Robotics made its blockbuster public stock market debut just yesterday on Shanghai's tech-focused STAR Market.

Somewhere in Beijing right now, a robot is running a convenience store with no human staff. Though it is still not very good at its job, there is obvious excitement around it. We believe you should give this primer a read to truly understand where the progress is and where the noise is.

I hope you find this piece as insightful and interesting as I found it.

Share

Bodies are Cheap, Data is Not: A Mid-2026 Primer on Humanoids

The year robots stopped being a demo?

For over a decade, robotics ran as a demo economy. Every few months, a new video: a backflip, a parkour run, a robot folding a towel at the speed of continental drift. We clapped, we shared, nobody bought anything. This year, the question flipped. Investors stopped asking “can it backflip?” and started asking “can it work a shift?”

The numbers say the shift is real. Roughly 15,000 humanoids shipped globally in 2025 – about 80% of them from Chinese firms – and Morgan Stanley now expects China’s unit sales alone to quadruple to ~50,000 in 2026, a forecast it has already raised twice this year (14,000 -> 28,000 -> 50,000 in five months). Elon Musk promises a few thousand Optimus G3 units in 2026, tens of thousands in 2027, hundreds of thousands in 2028. And the capital has noticed: global robotics funding hit ~$14B in 2025, up from $8B in 2024 and past even the 2021 peak (all figures in USD).

Have we finally arrived at the golden age? Not quite.

The promise of this generation is generalizability over programmability: a robot that can execute a range of tasks it was never explicitly trained on. We are nowhere near that. Today’s shipments split roughly into two products: the cool accessory and the less efficient industrial robot. To be fair, the field has basically cracked locomotion – today’s humanoids walk, balance, even dunk. But generalized manipulation, or dexterously handling the endless variety of real objects in real places, remains a wall. And at the center of that wall sits one thing: data.

This primer digs into the three debates that will decide who gets over the wall – data, model architecture, and hardware – and tries to cut through the ever-evolving terminology, the buzzwords, and the hype.

The data problem, and its direction of travel

This brings me first to the most important topic – data. Every model or hardware debate is downstream of one brutal fact: there is no internet of robot actions. The web gave LLMs trillions of free tokens, but the most useful robot data starts from nearly zero.

The robotic data pyramid is a helpful framework: the quality of data is inversely proportional to its quantity – see below a chart I stole from . The cheapest, most abundant data sits at the base; the scarcest, highest-quality data sits at the apex. The very bottom layer is data that already exists: web-scale video, most of it shot from the wrong viewpoint with no hands clearly visible. Cheap, abundant, mostly ineffective. One rung up sits egocentric (or “ego”) data: made-to-order human data captured with head-mounted GoPros or other headsets, at near base-layer economics. Then simulation data belongs around this zipcode too (though I’d argue it should sit below ego data) – it can be generated in unlimited quantities via compute, but it pays a “reality tax”: simulated physics and pixels never quite match the real world, producing the sim-to-real gap. Above that, off-embodiment demonstrations: real world trajectories produced on similar, but not identical, hardware – basically from other robots. And at the apex, teleop data collected on the deployment robot itself – frequently used to bridge the “last mile” of dexterity, but with terrible economics that don’t scale.

You might notice that I distinguished data by embodiment, that is, by what body produced it. This matters because the embodiment gap – the mismatch between the body that generated the data and the body that must execute the policy – is another dimension that dictates data effectiveness. The human-to-robot gap is a base-of-pyramid problem. A human hand has five fingers, ~27 degrees of freedom (“DoF”), tendon-driven dynamics, and a tactile skin; a robot can have anything from a two-finger gripper to a five-finger hand with completely different joint limits, strength, and reach. For human-generated data (web video and ego data), the mapping to the robot’s actuators is simply undefined: there are no action labels, and the viewpoints differ (e.g mostly third party). Then there’s the robot-to-robot gap (cross-embodiment): even between robots, data doesn’t transfer freely. A trajectory collected on a Franka arm doesn’t natively run on a UR5, let alone a humanoid or a robot dog – different joints, camera placements, gripper geometries, and even control frequencies.

Diving more into each data type:

Simulation – the locomotion pre-training backbone. A model that focuses on locomotion starts here. The goal is to create a digital twin of the real world where a specific robot learns movements at massive scale, and the data those simulators generate goes into pre-training. Players like Nvidia (Isaac Sim / Omniverse) have effectively converted robotics’ data shortage into a compute problem - synthetic data is cheap and parallelizable. It also works for the most part: teams at CMU and Stanford reportedly showed in 2026 that VLAs trained on 40% synthetic data matched 100%-real policies on held-out tasks. The sim-to-real gap, or robots discovering their joints don’t behave quite like in the simulator, is narrowing for locomotion as rendering and physics improve (locomotion capability is solved; locomotion reliability – uptime, falls, mean-time-between-failure over a shift – is engineering grind, not research), but it remains a serious problem for manipulation.

Egocentric videos – another pre-training corpus. If simulation manufactures pre-training data synthetically, egocentric human video theoretically harvests it from the one embodiment that already operates at internet scale: us. The bet is that passive first-person video of humans doing things is the closest thing robotics has to the “internet of tokens” that LLMs got for free. At Sequoia’s AI Ascent 2026 event in May, Nvidia’s Jim Fan (who runs the company’s embodied AI research) presented Nvidia’s EgoScale framework. EgoScale pre-trained a model on ~21k hours of in-the-wild egocentric human video with zero robot data, then anchored it to the robot with only ~50 hours of aligned mocap and teleop – under 0.1% of the total data mix. The trick that makes “zero robot data” possible is turning passive footage into something trainable: human hand motion is retargeted into the robot’s joint space, so the video yields action labels the robot can actually learn from. The headline result is impressive: scaling from 1,000 to 20,000 hours of human video more than doubled average task-completion rates, the first credible scaling law for robot dexterity that doesn’t require robots. The catch is that “internet scale” oversells how free this really is. A low-fidelity corpus of in-the-wild footage does exist, but the high-value videos - clean, hands in frame, minimized camera jitters - require instrumented capture (head-worn glasses / VR set) as well as cleansing and curation (including using UMI grippers to improve hand joint capture). Scaling it further means first getting and paying humans to wear that hardware and perform tasks, which is being collected by companies like Build.ai and Human Archive. Unlike FSD, where the driving happens anyway and the sensors are bolted to a car you were going to drive regardless, egocentric manipulation capture is a dedicated “collection job” and requires a carefully designed data pipeline. It captures great scene diversity (all kinds of factory and home scenarios), but it generates a larger embodiment gap given the human-to-robot crossing.

Teleoperation – from pre-training to fine-tuning and verification. Teleop produces gold-standard demonstrations: a human moves a replica arm, a VR controller, or a space mouse; the motion maps onto the robot’s joints and end-effectors; and the robot executes the action. Quality is high, but the economics and scaling are tough – like ego data, you pay humans by the hour for one trajectory at a time, except here you also need expensive hardware and skilled operators, and scene diversity is harder to come by (the rig doesn’t leave the lab). Early systems bootstrapped entire models this way (ALOHA, RT-1 in 2022), when models were so data-deprived that any high-quality input moved the needle, and some gold-standard models, like the π family, are still trained almost entirely on teleop. But teleop is too expensive to carry pre-training forever, so if the LLM parallels hold, its role shifts downstream. The first downstream job is fine-tuning: a curated teleop set layered on top of, say, an ego-pretrained model to sharpen specific tasks (though not to generalize). The second is correction: a human-in-the-loop, intervention-based approach (HG-DAgger) where the robot drives, the human grabs the controls at failure points, and the corrections feed back into training – this is what 1X’s remote operators are doing in people’s homes, and what teleop-assisted commercial deployments do in warehouses. Interventions scale better than raw demos, but they’re still gated on human supervision, task by task, so true scaling remains elusive. The third and newest job is verification: the robotics cousin of RLVR. Decompose a task into stages, train a small reward model to score progress through them from video, and you get an automated judge that can grade any trajectory – flagging which demos are actually good, which rollouts actually succeeded, and which simulated behaviors actually transfer. The early evidence says grading matters as much as collecting: one recent result from SARM took π0’s success on T-shirt folding from 8% to 83% using the identical 200-hour teleop dataset, just reweighted by a learned progress model. And that is the whole game for robotic data: not just how you collect it, but how you curate it and verify efficiently, reliably, and in the real world that your model is actually correct.

Tesla’s variant - “robot in the field.” Tesla attacks the same gap from the opposite end: put a good enough robot on the real assembly line and let it learn from errors, self-verifying task by task until it accumulates enough high-quality data. Being good enough requires a lot of data upfront to get up and running, so Tesla needs just as much data initially as everyone else. In theory this is the ultimate verifiable-reward loop - reality is the data generator (the FSD approach). The open question is TAM: it works where Tesla controls the environment (its own facilities), but generalizing to home care, precision manufacturing, and other high-variance settings is far harder.

Subscribe now

Nearly every notable effort on the data side today centers around shrinking the gap-crossing costs from a lower rung of the pyramid to a higher one, towards your robot. For example, EgoScale’s retargeting pipeline attacks the human-to-robot crossing in software, converting passive ego video into robot-executable action labels. Wearables attack the same crossing in hardware: UMI-style grippers and exoskeletons like DexUMI force the human to move as if they were the robot, bridging the embodiment gap at collection time rather than bridging it afterward. Seen this way, the data pyramid is the map and the embodiment gap is the terrain, and the companies that do well will engineer the cheapest route from base to apex.

Robotic performance will continue to improve across the board as researchers find ways to feed data-hungry models. Just because everyone is improving does not mean everyone will get to the promised land. Pre-training data has genuinely come a long way, but to land the last mile of a task (particularly manipulation), you still need task-specific, embodiment specific, and often long-tail real-world RL data delivered with careful curation. Before anyone figures out robotic RL data scaling, it is too early to declare winners.

The model wars: VLA / WAM / world models

For the past few years since the “GPT moment”, VLA (Vision-Language-Action) had been the dominant model architecture that enabled the present humanoid wave. The recipe: take a pre-trained vision-language model (one that already knows what a coffee mug is and what “put it in the dishwasher” means, from internet-scale data), and bolt on an action head (usually a diffusion or flow-matching decoder), then train the whole thing end-to-end to map camera images plus a language instruction straight to motor commands. This lineage runs from Google’s RT-1 / RT-2 to Physical Intelligence’s π0, which layered a flow-matching architecture on top of a pre-trained VLM to inherit internet-scale semantic knowledge.

Nearly everyone ships a VLA today. The π family, Nvidia’s GR00T N-series models (dual-system, a vision-language module reads the scene, a diffusion-transformer module generates motor commands), Google’s Gemini Robotics models, Figure’s Helix (an 80M param “system 1” transformer for control plus a slower “system 2” for reasoning). Even Boston Dynamics runs a similar mechanic in what it calls the “large behavior model”, as far as I can tell.

The insurgency this year is World Action Models (WAM). Jim Fan presented his AI Ascend 2026 talk with the headline “VLAs are dead, long live World Action Models.“. His argument: swap the language-model backbone for a video world model, making vision and action first-class citizens instead of a bolt on. The intuition is genuinely compelling. Language is a lossy bottleneck for physics. “Pick up the dough” (yes I’ve been baking) doesn’t tell you how the dough drapes, whether there’s water or flour on the counter, or whether the bowl tips over. A model trained to predict the next seconds of video, by contrast, has to learn contact, deformation, friction, and dynamics – it can’t get the pixels right otherwise (good analysis on this from RoboCloud Hub’s Robotics End Game).

Exciting as it is, I don’t fully agree with the statement that “VLAs are dead”. The way I’d frame it is: VLA-the-product-architecture is alive and shipping; VLA-the-pretraining-philosophy is dying. VLA as a pretraining philosophy is a bet about where the knowledge comes from: internet text and images, not physics. This was great when semantic knowledge was the scarce ingredient, and physics could be picked up during fine-tuning on robot demonstrations. Language was the load-bearing substrate. As we discussed above, this bet appears to be a failing one given how challenging it has been to pick up the physics (“real-to-sim-to-real”). Pretraining on video prediction, which changes the data, not the architecture, is swapping the load-bearing backbone – this forces the model to internalize physical dynamics just to get the next frames right. VLA as the product architecture hasn’t changed – still a system that takes in language instructions + camera images, and outputs motor commands. Even the historically VLA shops agree by their actions: GR00T N1.6 moved to a Cosmos world-model-derived backbone, and recent academic work keeps grafting video-prediction objectives onto VLA action heads. If there’s interest, I’ll dive into JEPA in a follow-up, along with the broader attempts to combine autoregression / LLMs (high intelligence ceiling, coherent over long-horizon logic) and diffusion (speed and parallelism). But I also don’t think this is the most important debate in robotics right now (spoiler: it’s still data).

Share

Hardware reality check

Hardware is commoditizing fast. When I visited the Zhejiang manufacturing hub in February 2025, the manufacturers jostling to get into the Tesla / Figure supply chain were all talking about qualification: “can we make this component”, “which samples are we sending in” and “ our Japanese and European competitors don’t have a defensible moat”. You could feel the commoditization already underway – four or five suppliers were chasing the same harmonic reducer or roller screw. Back then, the all-in cost of an Optimus-like humanoid ran $50-60k on low volume (a few thousand units), with line of sight to under $20k at 10x the volume. The caveat, then as now: the hand is the biggest cost item, and that number assumes a relatively simple design. More on this below.

A year later, Unitree is shipping its G1 humanoid for $18k and its Go2 Pro robotic dog for $4k (post tariffs) on Amazon, and the on-the-ground conversation has moved from “can we make it” to “whose capacity and costs are better”. It’s a brutal space for Chinese suppliers – many are building 100k-1M robot-equivalent capacity before large orders are confirmed or placed, betting that scale wins tier-1 status and market share. Sanhua’s Suzhou plant (500k units, live 2026) is sized for Optimus’s entire initial ramp on its own. Leaderdrive has scaled from 330k reducers (2022) to 800k+ (2026) and is raising money for another 1M. Suppliers are also trying to win modules over individual components (actuator assembly, joint modules, hand modules) and sell the integration as a cost advantage. I will skip the component-by-component tour, but it suffices to say no one is particularly worried about either hardware capability or cost anymore. The hands are an exception, but that’s also partly a design problem.

Subscribe now

None of this should surprise anyone who has watched Tesla’s car supply chain evolve. It starts with a sole-source launch partner (Panasonic in 2011, for battery cells and packs). As volume ramps and Tesla legitimizes the component with a fast-expanding TAM, it introduces second and third sources (or even design-outs) and plays suppliers against each other (LG Energy, CATL in 2020, then BYD). Over time the field commoditizes: pack prices fell more than 90%, from ~$1,500 / kWh in 2010 to ~$100 / kWh in 2025, driven by scale, overcapacity, and Chinese competition. Needless to say, supplier margins collapsed. Same story with thermal components, autopilot compute etc. It wouldn’t take much for a pattern-recognizing investor to underwrite where humanoid hardware margins might land once Chinese competition matures: maybe single-digits or low-teens for assembly and commodity parts (bearings, encoders, cameras, motors) or mid-to-high-teens for the specialized stuff (harmonic and other reducers, batteries, coreless motors).

The cost exception, for now, is the hands – though even that is loosening. It has been genuinely hard to ship a hand that is simultaneously dexterous, durable, and cheap; you mostly get to pick two. Push for top-tier performance and the cost quickly explodes. One leading humanoid OEM (you probably have seen their ads) is currently shipping units where a single hand costs $100k+, against a robot selling for a fraction of that. This tracks with the gold standard research hand from Shadow at ~$110k, and Sharpa’s very impressive 20+ DoF design at ~$50k on low volumes. The core problem depends on which of two architectures you pick – the industry has split on where the motors live. Cable-driven hands (Tesla, 1X, Mimic, Proception, and the Shadow hand itself) house full-sized motors in the forearm and pull the fingers through tendons, just like humans: you get strength and slim, high-DoF fingers, but cables stretch and wear, and the springy transmission makes precise force control harder. Hands with actuators in the joints (Sharpa) invert the trade: control is precise and repeatable and there are no tendons to fray, but every degree of freedom now maps to a tiny motor-reducer-driver stack crammed into a finger or palm at precision levels that keep yields low and costs high. But this is already changing. LEAP (CMU, open-source) and Inspire (China) built capable 16-DoF hands in the motor-driven camp for a few thousand dollars. Hands are on the same commoditization curve as everything else, just a little behind because of the precision and density demanded.

Another challenge has been battery life. Most humanoids run only a handful of hours on current battery technology, requiring either auto-charging or hot-swap logistics built into any deployment. Still, I don’t expect this to become the bottleneck. Plenty of tasks can be done table-top or stationary with a plug-in, and those tasks hit the data wall long they hit the battery. Batteries do create real issues on factory floors, but keeps improving – likely faster than we can figure out the software (specifically, the data).

The players

Since this is a primer, I will try to be lighter on the technical details and heavier on the landscape. I’d be remiss not to include a landscape on the players vying for the biggest pieces of the pie.

The field has split into a few camps: the arms dealers, the brain sellers, the body builders, brain + body integrated, open-source, and components players. In a world where data sits at the center of the scaling wall, and the number of new players and new funding continues to go up and to the right, I’d place my bet on the arms dealers and the players building proprietary data flywheels.

Arms dealers get paid regardless of flag. With this many players and this much funding (~$14B into robotics in 2025 alone), and every camp racing down a different technical path, the only positions that pay out under every scenario are the ones selling inputs to all of them. Nvidia is running the most coherent strategy in the field – selling the compute (Jetson Thor), the simulation (Isaac / Omniverse), and the open models (GR00T) that nearly everyone else builds on. Its most revealing move is naming Unitree’s H2 Plus the reference body for its open GR00T humanoid platform: the American chip champion designating a Chinese robot as its standard hardware. Nvidia is taxed into the winner whether the winning body is built in Texas or Hangzhou, and whether the winning paradigm is VLA, world models, or something else not yet published (US-China geopolitics permitting). The picks-and-shovels layer is the only place in this market where you do not have to be right about the model architecture to be right about the investment – and it is really hard to call the architecture right now. Same argument for high-quality data vendors building the curation and reward-modeling layer like xDOF. For the lowest-variance exposure to humanoids, the arms dealer may be the better bet than any single body.

Traction is far rarer than valuation. Almost every name here has a demo reel and a mega-round; very few have verified, paid, multi-year commercial work. Agility is actually the frontrunner in the US – roughly $300M in booked revenue against about 1,000 Digit robots across nine customer sites (GXO, Schaeffler, Amazon, Toyota, Mercado Libre), which is why it can price a $2.5B public listing (CCXI SPAC) on fundamentals rather than narrative. Unitree is the other: a genuinely profitable, fast-growing, vertically integrated humanoid maker taking itself public, with 2025 revenue up 335%. Although at least today, Unitree is more in the “cool accessory” camp in its fast-growing DTC segment – more a Labubu than a helper. Against those two, most US body-builders remain pre-revenue at valuations an order of magnitude higher. On stage they hold all the advantages – but the sorting will happen on factory floors, not in keynote demos.

Open source sets the floor price of the brain. The quietest faction may end up setting everyone else’s prices. We may see this play out in foundation models as well. Hugging Face’s LeRobot – now backed by Nvidia, which is porting GR00T and its teleop stack into it – has put open, fine-tunable VLA models and sub-$500 arms within reach of anyone, the robotics equivalent of the Llama (or DeepSeek) moment. If a fine-tuned open model gets a customer to 80% of a licensed Gemini Robotics or Skild or Physical Intelligence model on their specific task, the brain-sellers’ pricing power will compress. Or it might play out like LLMs – a frontier model makes a leap, sets up months of premium pricing until open and fast-follower models catch up, then rinse and repeat. Either way, the eventual moat moves off the architecture and onto the proprietary data flywheel – the millions of real-world trajectories only a deployed fleet can generate (again, the robots have to be good enough to deploy first). I’d bet on players that can field a sizable fleet or a capture engine and dig a proprietary data moat before hardware completely commoditizes: Generalist if you believe in cross-embodiment, Agility in the factories, Dyna in its specific scenes (hotels / restaurants), RobotEra (China) in logistics warehouses (more in the appendix), and 1X at home. Don’t forget Tesla – its data is captive to its own factories, but controlled-environment data still compounds; factories are a big enough beachhead.

The demo economy is dying. The data economy has begun.

Focus time and capital on the picks-and-shovels players (high-quality data, compute, and players with potential data flywheel) rather than picking specific paradigm / architecture too early. Invest in the hardware layer discerningly and only if valuation supports the math.

Position accordingly. Don’t get caught in the momentum chase.

Meanwhile, that robot in Beijing is still on shift – not very good at its job yet, but learning from every order.

Grace: As mentioned previously, every once in a while I bring in a guest writer who will share a unique point of view or fill in an area where I do not have the domain expertise. This week, our special guest writer is Ella Zhang, former investor at Sequoia’s hedge fund. Check out her newly launched Substack at uncrowded!

Appendix

We start with the professional data industry. As AI agents become a new orchestration layer above terminals, APIs, and research products, Robert argues that the biggest disruption may come not to the production of proprietary data itself, but to its distribution. For niche data providers like BigOne Lab, the opportunity is to make differentiated real-time data accessible at inference time. The unresolved problem is economics: licensing, access control and who ultimately captures the value when an AI agent becomes the interface.

From there, we widen the conversation to culture and political economy. Robert explains why debates about AI in China tend to focus less on existential or metaphysical questions and more on what the technology can actually do. We discuss whether that pragmatism comes from China’s history of technological catch-up, whether similar attitudes extend across East Asia, and the potential trade-off between being exceptionally good at applying technology and creating the conditions for more fundamental scientific discovery.

We then turn to the role of the state. Robert rejects the simple idea that China’s technology industries are created through top-down planning. Instead, he describes a hybrid system in which entrepreneurs often discover the opportunity first, while the state later supplies policy support, capital and the resources needed to scale. We use EVs and DeepSeek to explore that model, before moving into state subsidies, local-government incentives, private capital, Beijing’s evolving approach to public markets and why so many young AI and technology companies are choosing Hong Kong for their IPOs.

We close with two of Robert’s more differentiated views: that China could be entering a multi-decade equity bull market, and that outsiders often misunderstand China by assuming it has the same impulse to export its own political or cultural model. And for a slightly lighter ending, Robert explains one of the Chinese stock market’s most vivid metaphors: why generations of retail investors are compared with chives that get cut, grow back, and get cut again.

btw sorry for the weird glitch in the video around 12-13 min of the recording.

The AI Proem Podcast is under the AI Proem newsletter which has over 12k followers globally. To learn more about China AI, the business of AI, and how AI is impacting businesses, please check out the newsletter here and more insightful conversations here.

Chapters

00:00 Robert Wu, BigOne Lab and Baiguan
03:09 From alternative data to the AI era
06:02 Why AI disrupts data distribution
12:11 Inference-time data, licensing and economics
15:17 Why China feels more pragmatic about AI
21:47 East Asia, belief systems and scientific discovery
30:32 China’s hybrid model of state and private innovation
45:24 Funding AI: state capital, private capital and IPOs
50:30 Beijing’s market-stabilization playbook and policy risk
01:07:29 Robert’s non-consensus views and the meaning of “cutting chives”

Transcript (AI-generated, for reference only)

Grace Shao (00:00)

Hey Robert. Good morning. So good to have you join us today.

Robert (00:05)

Good morning, Grace. Hello, everyone.

Grace Shao (00:08)

Robert doesn’t need much of an introduction. If you spend as much time in the Substack world as I do, you’ll know he’s a prolific writer covering everything from capital markets and property to technology and culture. My favorite niche is when he calls out Noah Smith’s articles for being wrong. Those are pure entertainment for me.

For today, though, Robert is a student of history, politics and business, and I think it will be interesting to have him walk us through some of the bigger questions people have about China. I’m also curious, because he runs a data company, about how he sees the future of data providers as AI changes that relationship.

So I’m handing the mic over to you, Robert. Tell us about yourself, your journey with BigOne Lab and Baiguan, and how you’re seeing the business evolve.

Robert (01:12)

Yeah, hi. So this is Robert. As Grace mentioned, we run a newsletter. But that newsletter is really our kind of side business. The actual BigOne Lab is a team of over forty people, which exclusively most of us work on data products and research products for institutional investors and corporates. Both in China and outside of China. But we have we’ve been very China focused. All of our data and research are about China, Chinese companies, Chinese industries, businesses. The Baiguan to me was partly accidental, but partly also kind of fateful. Actually in the very beginning during college I actually wanted to be a journalist. But I didn’t find a way. So I kind of dabbled in capital markets in investing, corporate finance for a few years. But eventually it kind of hit me that, there’in this new world there’actually other ways to do journalism. Data tracking, data analysis is actually could be a new form of that. And even with data you can do more powerful storytelling and that was the genesis of our newsletters as well. Right. So here we are. We are backed by S&P Global as well, which is I would say one of the most ris backed respectable, respected companies in our industry. And yeah, so it’a brief intro about ourselves.

Grace Shao (03:09)

Yeah, so tell us like what is unique about your data then in that sense.

Robert (03:14)

Right. So we started as a so-called alternative data company. Alternative is alternative to the traditional financial data, macro data, market trading data. It’no longer alternative now. All alternative data is mainstream data now. But it was first happening, it was because the explosion of data and information in the internet and especially the mobile internet age. There are just so many data being tracked. There’payment data, there’online com commerce and social media data, so vast number of data and multiplying exponentially every year. And some investment firms they realize that by harnessing all these data and aggregate them together and put them in the right context, you could actually generate a lot of alpha that is previously not available. Right? So that’how we started the business. It was a very investment firm hedge fund driven business. So that you know kick us to look at a lot of the industry verticals, a lot of the different kind of industries where there’data and we try to find the most granular, the most high frequency data we can find. Perhaps the you know we can have massive amount of data about mobile transactions in China, for example, every day, even every minute, all the transactions that we can have access to and analyze on. So that’different from many of these you know mainstream data providers, which we try to be very granular. We try to be very frequent. Yes.

Grace Shao (05:18)

Yeah, so that’really interesting. I think p one thing that really stood out to me and relating it back to AI is that when we were having our catch-up conversation, we were saying, okay, data plays obviously a huge role in AI. But what you distinctly said, there is the people that are involved in the pre-training data bit, there’like the Mercores of the world. There is the people who are more pivoting towards kind of the post-training data provider, which is what you guys are doing. Just tell us about that relationship and how you think the whole data vendor ecosystem is adopting to AI and or evolving with AI, especially and how like AI is now affecting, say, Bloomberg, Factiva, those mega data platforms that we traditionally know of.

Robert (06:02)

Yeah. So the term data company is really problematic for us. It’really a kind of a name that covers very different kind of businesses serving different needs, entirely different kind of businesses. So you mentioned that there are data companies that are serving the large language model training right now, the Mercor, the Surge AI. So they are they are they are good at massively labeling data, connecting the you know different type of data and help the help build up the data sets that are used for the training. Well for us, we are more on the on the on the more on the real time data end. And it’so for the industry that we operate in, we have Also, we have not a consensus on the name for our industry, to be honest. I call it professional data industry. Some people call it market data, some people call it market intelligence data. But at the end it’it’about tracking and understanding of the real world on a real time basis, if we have to define it. So it’much more about what is happening rather than the logical connections between different pieces of information, which I think is what the pre-training data i is mostly about. And so in our industry, AI is placing is playing a huge kind of disruptive role for our industry. So in the professional market data industry, there are main several main stages, maybe three. There is the production the original production of the data, there’a distribution, and there is the what we call activation. I won’t maybe go to details of each one of them, but if you understand production, production is really where the data is originated, right? For example, if you are Nasdaq, all the trading data on your Nasdaq platform is originated at Nasdaq. That’called production. The second is distribution. Is how you combine all this data into products, right? Companies like SP’marketing intelligence, like Bloomberg, like you know FacSet are in the distribution part. They don’t generate data on their own or mostly don’t not on their own, but they provide the interface for users to interact. Right. And activation is really how data is used. I won’t go to detail for that part. But right now the one the stage that is facing the biggest disruption is not the production side, right? You still need to generate data. You still have to have some kind of source of data. AI won’t help that. But on the distribution side, there’a there’a huge, I would say, change that is undergoing. Imagine if you are an analyst twenty years ago. It’required for you to have either a Bloomberg terminal or you know a FactSet terminal, or if you’trying to a wind terminal, right? It’it’a terminal kind of portal driven business. A go-to portal or source for information. AI is fundamentally going to change that by adding a new what we call orchestration layer above all the data types. You’not going to go to any terminal in the traditional software sense, but you’going to have this advisor to you that is going to massively, quickly, rapidly going through all the data, find the data you need, give you the conclusions, do the comparisons of and all that. Right. So there’a big tension right now between this trend of increasingly more people is rel relying on their AI agents to do research and the existing incumbents. Of the of the market which and then if you look at the incumbents the different companies are adapting differently so you have companies like SP and Faxet they are embracing AI and they signed big contracts with large language models they allow you know clause users or open AI users to access their data through these their AI products and they are they are embracing it. But then you also have company like Bloomberg, which is really at the core, at the at the at the apex of traditional financial and market data industry. I think they’still trying to figure out what to do with this. And I think their natural tendency is to build their you know in-house AI. They still want people to, go into their universe. And to make to like people still go to their universe to check the data. Right. So there is some debates right now and it’going to be interesting, who is going to prosper, who is going to stay. Yeah.

Grace Shao (11:39)

Yeah, Bloomberg definitely still wants you within their terminal. Everything is within their terminal. And once you exit terminal, that’majority of their revenue. They don’t want you to jeopardize that. But for someone like you, it’quite interesting. I said something I made a mistake earlier. What you told me was that you guys are a data provider on the inference end now. How do we understand that? And how do we understand how you are going to work with whether it’the model companies directly? Or how you will provide your institutional clients your data differently.

Robert (12:11)

Right. So at this moment we are still observing. We are actually niche provider of some really high value data, but not needed by most other people. So we’not like the mainstream data sets, but we are observing and we believe that in the end we’ll have no choice but to kind of open us up to the large language models. To us, this will be a new form of access to use our data. Apart from, right now we provide our data to our clients through API, through Excel spreadsheets, through even research reports, this is our current method. But in the end, I think as more and more clients rely on AI to pull data, we will we will we will kind of open us up. And that’the That’in the inference part. That’when people actually are using data to do analysis, to do research. And so we are firmly in that part. And I think it’just inevitable that we will be connected to these AI at some point. It’just the problem right now is about the economics. How do the economics work? How do we kind of get us exposed to it, but also make sure that there’strong enough licensing and you know strong enough gate that we can put on our more exclusive, more differentiated data sets. That’a question that there also hasn’t been a consensus yet in the industry. Yeah. So that’why we are taking this kind of stacking back and observing kind of view of it.

Grace Shao (13:55)

I see, very interesting. Okay. So enough about the dry stuff. The most interesting stuff I read from you are actually your takes, because I think your takes are very nuanced. They’you’a deep thinker. You bring together cultural sentiment, history, political reality, and then the business together. To start with, I think one of the questions I get the most from people right now is just that this general attitude around why Chinese people feel more normal about AI. I wouldn’t even use optimistic. I mean as a as a society as a whole, it does feel more optimistic. But it just seems like whether it’how the government and the regulators are looking at how to regulate this new technology or how people are adopting it, like a trial error kind of feel, there’less of a philosophical push up a pushback towards AI, but more maybe, obvious concerns of what disruption or change might may mean for job displacement, whatnot. But in general, quite optimistic, quite normal. How do you view this right now? If I just kind of bringing together all the different aspects and it because I can’t, I don’t believe people just saying, just because people are more pragmatic. That’that I mean, I made that argument slightly, but I even think it should be deeper, more nuanced than that.

Robert (15:17)

Right. Yes, I mean this is a good question. Without if you don’t ask me that, I wouldn’t even realise it’a question. Because sitting in China, it’true that it is well not you know most people don’t talk about that. Some people do, but definitely not a mainstream discussion on the kind of existential kind of risk of crisis that AI is posing to the humanity. That type of question is not that asked is not you know asked that often in China. For the good and bad, right? I personally I don’t know w which part which approach is better, the more pragmatic one or the more philosophical one. But that’the phenomenon. It’true. Most people don’t think i in that way. The exactly why, you know Probably I would if you are looking for a more subtle, more nuanced answer, probably you won’t be able to find here. Because I was I was also thinking that it’it’really because of the pragmatic and down to earth nature of most things in China. People tend to ask more, what can this be used for? Other than why we have to do this or w what’the bigger contact what’the bigger picture? People tend to focus on the productivity side of things. I mean that’just prevalent in all industries or new industries. And especially China attached a premium to new industries. I think that’the kind of cultural reflex of the last few hundred years, after China kind of fell behind the West. In terms of technology and suffered all the consequences from that. So there was now a kind of reflex to emulate the world, to catch up to the world in all kinds of new technologies out there. Every time Silicon Valley coined some new term, some new idea, there would be some at least some kind of reflection and discussion about that. If you remember a few years ago there was this concept called metaverse. Right. Now nobody talk about that anymore. But back then it was also a very hot topic in China because you know people might think this is maybe the future because the Silicon Valley, the US chose that, and maybe we should think about whether that’the future. When crypto came out first, China was at the very beginning also embracing it, right? My very first Bitcoin was bought in China with RMB, while when there was like many RMB exchanges there. Well, then it hit some problems, it got banned and all that, but that’what happened later. But China ha always had this at this contemporary China had the tendency to learn the new things and to try to, to grow their our own knowledge and strength along these new verticals. So that’the I would say the big context, the big framework that people use, kind of equipped themselves with when they look at these new things. And less so about the philosoph philosophical and maybe not a philosophical but metaphysical, right? The ones that are hard to prove or disprove at this point and just kind of descend into discussion about contact concepts, on the abstract side. That kind of discussion, that kind of discourses really doesn’t have a big market in China. Small circles, yes, but most people Just don’t like to engage in that kind of discussions. Platonic, Aristotle level discussions. Yeah.

Grace Shao (19:23)

Is it just because it’kinda I don’t know, it’just like what’there to gain from that for the average Lao Bai Xing the average Joe? When they think about it, it

Grace Shao (19:32)

Seems like it’kinda Okay, if I embrace it, I win. I don’t embrace it, I lose. It’a bit of FOMO. Especially for the next generation, when I talk to parents, there’less of a concern about what this technology might mean in terms of safety for the kids, but there’more about how do I embrace this technology and teach my kids so my t kids can go basically go ahead of everyone else and come on top? I don’t know. Is that kind of what

Robert (20:01)

Yes and I think big part of that the these better philosophos philosophical questions don’t tend to produce results. Right? It’not a question that w if we debate and discuss we’ll have some kind of agreement. They just tend to stay philosophical. But most people I would say that most people I know here don’t tend to keep going. Keep debating on this type of questions. And maybe that’right, maybe that’wrong. I don’t know. But that’just the phenomenon that we are seeing here. Yeah.

Grace Shao (20:39)

Just how it is. Less of a chatter class, if you must put it, or at least less prominent in.

Robert (20:44)

Yeah, good way to put it. Yeah. Yeah.

Grace Shao (20:49)

I think another thing we kind of touched on briefly when we were catching up for this recording was that we said, look, this Chinese pragmatic approach to technology and like how to even day to day life is not really just limited to China, right? Like it feels like it’a phenomenon across maybe even East Asia. Other markets like South Korea, Singapore, a lot of, studies, whether it’by Stanford or by local communities, have shown that people are also embracing, Singapore’own ministers are coming out talking about how they’claud coding or vibe coding out there. Why do you think I know that you write sometimes with a bit of a this versus that or like a bit of a historical and a eth ethnic and history kind of tied to your analyses? Why do you think that maybe East Asia feels more pragmatic towards AI or contemporary East Asia, like you put it just now?

Robert (21:47)

It’a it’a very deep question. So I think if I have to attribute, and I’m just throwing out ideas right now, religion is definitely huge part. The whole tradition, the tr of the history of thoughts, the history of religion, the history of philosophical discourses, has to play plays a huge role. So we don’t have a tradition of seeing something abstract, either as a natural law or a god in this part of the world, historically, right? So people the w the reason why the people are more pragmatic about things is just like all the things that happen in your life are pragmatic. There’a there’a flood and then we have to fix it. We have to you know do some work on around that to fix it. All the laws, I mean all the folklores, all the all the lessons of history surround about how to deal with these you know disasters, wars in human life with a human way. Right? So it’it’more there’a concrete problem, there’concrete solution. And very seldom you you see like people in China turns to God, for example. For a solution. Maybe

Grace Shao (23:19)

That’really interesting, but I will push back. They are like the Buddhas and the temples that still exist where people like pray for money, which is hilarious. Again, it’a very pramatic result. Or you like you pray for a child. You literally, you’re

Robert (23:31)

Exactly. Right.

Grace Shao (23:33)

Not you’just give me a child. Like you pray for fertility, give me money. You pray for money. But there’praying in that, but it’not omnipresent. It’like each god has a or like each Buddha has a very clear ask and reward almost. I don’t know, like

Robert (23:51)

Yeah, exactly, exactly. So I mean there are like symbols of belief or faith or whatever, but exactly as you said, people use these very pragmatically, practically. They’the you know when Buddha said that you know w we w you know Buddha doesn’t want it the original Buddha doesn’t want him to be worshipped as a god. Right. It’really a teaching about how to position how to think of oneself and how to position yourself to the universe, to the world. Right. That’the whole teach but then it lost that favor in China. It became this, inf fused with all these other very down to earth beliefs and to become this, now you assign this Buddha. To ask for kids, that Buddha to ask for money. It’definitely not what Buddha originally taught. So it’just it’just it has been like that for not just decades, centuries, even millennia. So I mean that’also that’a very big part about China, which I’m yet to write about, but I kind of touch on it at several points. Is that you know a big part of China is China really I mean Chinese people and maybe East Asian in general, because we don’t have such a strong kind of belief in some abstract things, we also tend not to want to convert other people into our kind of belief, right? So we tend to focus on the practical. That’why we have a lot of business people, for example, that are focused on making deals, right? Trading, benefit you, benefit me, and so it’all very down to earth, but very practical things. And that does definitely have limitations. I would argue that in terms of fundamental pure science discoveries, that kind of mindset creates a disadvantage. You really have to, when you do groundbreaking scientific discoveries, you really have to forget about all these worldly stuff. You really have to forget about things well what’this mathematical formula have to do with my life? You have to forget about that. You have to just focus on this the purity of sciences, of mathematics to have to have some great discovery. Then that’why you know like the people were debating recently you had these metal this mathematicalist, Chinese ethnic ones, but getting their awards not in China, not while they are in China, but because they have further studies in the West. Right now in China there’a big debate about you know whether Chinese college graduates can do you know achieve that kind of level of achievement in sciences if they stay in China. I think right now I’m not that optimistic because overall people are still very focused on the use cases, the pragmatic use cases. But most of the time when some big scientific truth is discovered, they don’t have a direct use cases. And that’definitely not how they start a discovery exploration. Right. So that’well, the thing is, the reason I want to sometimes compare the China and the West is not I want to say which one is better. I actually my main point is we are many societies can be different. But in our world, different societies, different economies, different kind of people can play different roles. You so you need thinkers, you need the people who think about the abstract, but then you also need the people who actually can put things into use and create productivity and make people’lives better. And it’it’it’great that you know you have different kind of people serving their different kind of purposes. And I and I think I love I actually think that you know China being different and the West being different in their own ways, a net benefit for the whole world. And that’you know my actual overarching key point in writing about all of these.

Grace Shao (28:39)

That’really interesting. I think, like offline I wanna dig more into the religious aspect. It was just really interesting. I never thought about it that way. And we might get some heat and pushback on this because obviously South Korea nowadays is like a very Christian country and you know it kinda goes against what you earlier said. But I do think what you meant by East Asian worshipping in general is not so much omnipresent, but it’I don’t want to say it’opportunities, but people often go to these Like Buddhas when they need something, when that thing happens, or when something bad happens. But it’not like you’not taught to be thinking about it day in, day out. So that godlike attitude is very, very different. And I think it does translate to how people are perceiving AI these days, because in the West, right now, AI is seen as like a kind of or a lot of cult like figures are coming forward. And, a positioning AI like a new magic or something that will fundamentally change society as we know it. Anyway, so we can talk more about that maybe offline, but I want to bring it back to then things that you write about a lot, which is how should we understand then China’unique state planning and how it drives economy? Because I think it all relates to what you just said. China itself, in a way, you can say a lot of people are taught to be very, very strong execution and execution people doers, but they

Robert (30:01)

Mm-hmm.

Grace Shao (30:01)

Are maybe less of these creative, wild thinkers. Obviously that’not. All true, but in a general sense, yes. So then when it comes to then how the state interacts with the private sector in terms of innovation and state planning, we see that again, there’a top-down vision or priority. And then companies or sectors as a whole will start executing and create abundance. How does that all work and how do you view that kind of relationship?

Robert (30:32)

Yeah, I think, when we talk about relationship between state and business and innovation, again, people frequently fall into the traps of big concepts, right? People the classic question is China socialist or is China capitalist? And these are also tend to be kind of the Western preference in terms of discussing things. While again in China, People tend to be not so focused on the on these conceptuals, on these, black or white. So when Deng Xiaoping said, black w black cat, white cat, whoever catches the mice is a good cat, it’not just his opinion. He’only a manifestation of the average most of the people in China. Whatever works, whatever can solve the problem of the day. We will use them. Right. So that’the bigger contact context here. And when we look at specifically industrial policy, innovation and all that, I think people, both the government and business people, tend also adopt this view. Whatever works. So if we look at the EVs, for example. When EV became a thing in China, it’a confluence of forces. It’not just like the state said, we want to develop an EV industry, and then it happens. If you look at say BYD, the Wang Chuanfu, when he started to have the idea that we should start an EV business, it was actually earlier than Tesla. And Wang Cheng Fu when he did that. It’not because he think that the state should do it or the state tells him to do it, right? He did it on his own. He has his own vision, his own dream. And it’just how so happens that the priorities, the goals of these business people and the state converged. And really for say something as massive, as important as the EV industry to happen, you have to have all these factors line up. You have to have entrepreneurs who are really willing to take the risk. BYD at the time took enormous amount of risks. But then you also have to have government that have the policies that are friendly to EVs. You know the consumer rebase for EVs, for the infrastructure build out and all that. All these forces are important. Fast forward to today, AI, for example, DeepSeek is a very great example. Of this dynamic. When Liang Wen Feng started the DeepSeek venture, he never thought about Beijing. I mean, Beijing even didn’t realize that a quant fund could you know incubate such a you know important AI company. You know back in 2023, 2024, there was even a crackdown on quant funds, causing some kind of market crash. Back then. We call it the quant crash. That was only two years ago. You know

Grace Shao (33:56)

Why were they being cracked out? Why were they being kind of scrutinized?

Robert (34:02)

So two years ago there was this moment where the market was like sliding down and the quant was like kind of magnifying that sliding down. And the reflexes of regulator was really to kind of hold it, hold them back. There was one episode where the regulator kind of stopped a quant fund to from trading, basically plucked out the cables. So they were, because the mechanics the mechanism of quant trading is usually to kind of magnifying could help the market trend to get even you know more pronounced than it is so there’always some kind of controversy about our industry. So it’hard to imagine that Beijing actually found a quant fund and say, we are going to place a huge amount of money or huge amount of resources and ping our hope on you. Right? It just didn’t happen like that. Now had no state backing. He had his own dream for AI and he had money, he doesn’t have to rely on anyone else. And but after he became successful, after DeepSeek became a an international sensation, then you know a few ye a few days after last year, DeepSeek’moment, he was received by Premier Li Qiang. And then you know they become kind of a national priority. And in the this year’fundraise. There’also very top level state fund from Beijing that invested in DeepSeek alongside with Tencent and all these other companies. Right. So I think that’dynamic is interesting. It’at the same time there is a strong hand from Beijing, but also there’at the same time a huge tolerance for the natural growth of you know companies, industries, people on their own. And Beijing is less a planner, but more a kind of a picking picker of the winner. Right? So they set the long term goal. They say that we want to develop new productive, I mean new quality productive forces, but they never define specifically what are they. They kind of leave that open for the for the markets to explore. To for our own talents to explore. And once there is some clear winner, they come in and back them up with the more resources and help them scale. So that’I would say that’a hybrid. That’really a hybrid model. No single side of it can define this whole model. And this hybrid nature rests on the fact that people again we are flexible We don’t we don’t stick to any single type of ideology or ways of doing things. Whatever it works, right? Some industry needs creativity, then it cannot be top down. It has to rely on these spontaneous ventures and people. But also some industry if they want to scale, they need to have massive allocation of capital to them. And in China, if you want to really get massive amount of capital, you have to have the backing from the state. And that’how it happens. And so I think it’just natural. And it’also it’it’a hybrid model that is proving to be working and maybe for the new other industries it will also prove to be working as well. Yeah.

Grace Shao (37:48)

It’really interesting the way you put it. It’almost like they’a parent. So you get enabled and you get resources when they like something that you’doing, but you get beaten down

Robert (37:54)

Yeah. Yeah.

Grace Shao (37:57)

Or you get scolded and grounded if you’doing something they don’t like you’doing. And that brings me to the next point, which I want to ask you about. And you kinda alluded to this already, you touched on it. It’like the relationship between the state and the prime, it’something I think a lot of people find hard to understand. State as SOEs, state owned enterprises. State subsidies into industries, and then like obviously favorable policy making. It’very interesting because from your point of view, you’saying this is natural. Like you said, it is what it is. You need that kind of parental help or you need that parental guardrail, whatever, or safekeeping in one hand. On the other hand, from obviously a very American perspective or a Western perspective, is why are you involved? Is there for state subsidy than unfair, which I find kind of interesting of an argument. But there’obviously accusations from the West saying these Chinese AI companies are state subsidized, therefore they’not really competitive. I’m but they’still competitive from an innovative perspective. But anyway, and then you obviously have a lot of these AI companies now worried about taking state capital because if they want to go global or even go l like go get listed publicly somewhere non-mainland China. Then there’also concerns about shareholder setup if there is like clear state backing. Anyway, this is a again a bit of a big open question, broad commentary, but I’m gonna throw it back at you. How do you view all these different agents or different stakeholders and their relationship? And how do you view whether it is fair for certain companies to get state subsidy or not? And how to view their then independent competition and innovation.

Robert (39:47)

Right. So this whole kind of debates or controversy about subsidies in China, there’just so many I mean so many ways that I don’t I don’t feel okay with. I mean, like for example the in the West, it’not as if the Western government don’t have subsidies and don’t even have huge subsidies, right? I mean in EU many industries are being subsidized. In the US, if you look at say Tesla in the early days, I mean SpaceX even, all these companies rely a lot on policy support. So I mean maybe the difference between the US and China or EU and China is the I would say the role of the local governments. There is a huge tendency for local governments to go out of their way to support new businesses, which is really part of their own incentive arrangement. It actually helps them to grow the local GDP and help them promote it. So and it also creates some kind of over competition between the local governments. But it’not by design almost. It’just naturally happen that all these government sector support they just come in and out of their own interest They support these businesses. However, I would always argue that all these controversy or debates about subsidy tend to make people believe that it’because of the subsidies that Chinese companies become competitive. I think any basic student of economics would understand this cannot be true. I mean no businesses can be subsidized to be competitive. It’just It doesn’t work like that. Not in China, not in the US, not in EU, in not in Latin America, not in any history, in any human history. No competitive businesses become competitive because they have state subsidies. And usually it’the opposite. Subsidies only create uncompetitive businesses. Because whatever you do, if you are profitable or not profitable, you still have the state backing and which will make you artificially profitable. Who will do that? Who will be competitive? It just doesn’t make sense. And the reason that subsidies or state support or whatever support policy work in China is because every actor in this industry are working towards the same goal. Businesses, owners, the state, central government, local governments, all other stakeholders. It’really about everyone pushing, everyone going, and all the talents engineers in these companies. Everyone agree on something and push for walk forward to it. So it’definitely not just the subsidies. It’it’a whole spectrum of this converted uniform action of every party that make Chinese businesses competitive. And if the West just comes in and says, it’a subsidy that’responsible for that, i it’just not a very effective criticism. I mean and then reflex will be the West will have more subsidies to support their businesses, which, in fact, the wrong kind of prognosis will lead to a wrong prescription, which will be interesting as well. Yeah, I mean I’m pretty kind of I would say it’it’kind of kind of emotionally bit charged topic for me, but I really want

Grace Shao (43:37)

You’passionate about this topic.

Robert (43:38)

Yeah. So I really want to speak it out about this, yeah.

Grace Shao (43:44)

Yeah, so it’interesting then, how do you view this generation of AI companies and kind of the I guess how they’overlapping these space? Because like you said, and we know here at AI Prome where a lot of these labs actually even struggled to get capital in the beginning, before the GPT moment, like your point, Silicon Valley can set the tone. Once ChatGPT took off, Chinese labs. Were able to kind of rally up and garner attention and interest domestically. They got their first kind of pot of gold, set the labs up a bit further, more like bit more sophisticated ways. Clearly they’still struggling to, or not struggling, I would say they still need a capital. So then two of them rushed to go public. Now more thinking about that. All of this indicates, first of all, obviously training models is extremely expensive. But they’still not really getting the funding they need. And some of them are choosing to not get the state backing or state kind of related capital they, that’out there. I guess this question is a bit long windy, but I guess just how do you see the relationship of the AI companies right now with all the different stakeholders and capital players in China? Because the state has the money, some of them don’t want take it. The state clearly is have favoring AI right now and rolling out a lot of strong policies and helping them with compute and energy and whatnot. How are they interacting with SOEs? In fact, how are they interacting with the big tech? How are these different stakeholders now I guess involved with each other?

Robert (45:24)

I think a key variable that was not on the table a few years ago was the role of the capital market. So we have we cannot leave that out when we talk about funding for these new companies. So I think the Beijing is very proactively pushing and helping many of these AI or even right now robotics companies to go list it. Either to Hong Kong or prefer preferably even in domestic A share market. The speed of making these companies public even just a few years after they were founded, it was actually unprecedented by Chinese standard. The y the capital market used to be closed to most of the new economy companies. So that’why when Alibaba went listed they the default was go to Nasdaq. Right. So that default was no longer applicable. No company by default want to go to the US for listing. While at the same time, China Chinese regulators did make it easier for companies to go listed in at least greater China, right? Hong Kong and Shanghai, Shenzhen. And I think that’a yeah.

Grace Shao (46:46)

Jump in really quickly. I think people also don’t understand sometimes and miss the point on a lot of these new economy companies from China are not going to go list in the US is not actually like actually help us explain. Is it a China’regulation reason or is a US regulatory reason?

Robert (47:06)

So it’actually a combination, but I would say the most of the issue is on the US side. Maybe sixty percent US responsible, forty percent China responsible. But anyway, there’a pull and push that make companies think about. So at the same time it’get just getting harder to get listed in the US. There’always a risk to be delisted, for example. And while to apply to US listing now you have to go to Chinese regulator as well, which there was no such approval process before, right? So it’hard. But then at the same time, it’getting easier to list in A share and also in H-share. And also liquidity in Hong Kong is way better than before. So there’both push and the pull. There are still some companies get listed in the US, very few. Recently this year there’this company called Taso Chuo that was just got listed in US. I think they have their own reasons for that. But most companies would prefer to just stay put in this part of the world. Right. So that’a key variable. And I think that’the key leverage that Beijing is using to help these companies raise funding. Like to be honest, I think Beijing is very I would say sometimes like a very strict you mentioned parent, right? Beijing is a very stingy parrot. Actually Beijing doesn’t want to spend too much money on, all the projects. But they are ambassadors at leveraging other people’money to achieve their own goal. Right? So like if you look at deep seeks fundraise, Beijing invested only a small part of that. Most of the money is contributed by you know Tencent or other private investors. For them, it already achieves a goal. It helps the company that Beijing wants to grow raise funds while at the minimum amount of money that Beijing can actually need to chip in. That’pretty smart, you know. It’it’not like it’not like i it is smart to keep resources at your hands and try to leverage other resources other people’resources to support your goal. And capital market is exactly like that. It’not just capital from big companies and big funds, but a capital from all over the market. Everyone, every even retail investor, get to participate. The that only that way you can ensure a everlasting strong stream of support in the in the future. So that’I think a very different that’actually very different, say compared with a few years ago, where you don’t have such a as strong a capital market as we have now. And now Beijing also have a vested interest in support the market. And they have also developed their own techniques and their own muscle memories in supporting the market, which is what we don’t have even five years ago. Right.

Grace Shao (50:22)

Right. But some still argue that the Chinese government could support the stock market more. I don’t know. That’just things I hear. Well, how do you view that?

Robert (50:30)

Yeah. Actually they are now sophisticated enough to understand that you need to be balanced. So what I mean is there are actually two episodes that could remain as lessons for them. One is the twenty fifteen, twenty sixteen market crash. Second is the recent market crash in South Korea. In both episodes, there was a bull market, even a crazy bull market. And in both episodes, the governments initially played a very strong role to boost the market. Back then, in 2020 I mean 20 fif fifteen, there was a People’Daily article saying directly that the market should go above, I forgot it’five thousand or or four thousand points. Which was cited as a kind of a rally call for many people to go into the market because the Beijing says we should, buy, buy, buy. So Beijing actively kind of contributes to the building up of a big, big bubble. And then after Beijing felt it was too crazy, it cracks down on leverage. And a lot crackdown on leverage burst the bubble and it has become a really bad market for the next two years. Same thing as South Korea, right? Like they prime min president of South Korea said, I’m also buying the stocks. Every policy going to support the market. But then the government was too concerned about a leverage. So crackdown on leverage. And then boom, the market dropped. And so I think you know Beijing of today is Pretty sophisticated with that. They want to have a bull market for sure, but they also don’t want it to, turn into a crazy boo. And exactly how they do that, because this is some not something that you can say, I want this, I that so I can achieve that, right? Because it’a market. There’a lot of players. When the sentiment builds up, even Beijing cannot stop people from buying or selling. So exactly how, interestingly, they all have also developed. Dev develop their own technique, which is this so-called stabilization mechanism. So for the first time in history, in the last two years, Beijing was actively employing and deploying capital to act as a stabilization factor for the Chinese capital market. By stabilization I do not mean just a buying mechanism. It’a stabilization mechanism. Which means when the valuation was really depressed and Beijing wants it to go up, they actually now come into the market with real cash to boost the market to help reset the valuation. This is different from before. In the past, I think there’never been an episode where Beijing used real cash to support the market. There was messaging, there was this policy, that policy, this tax policy, that tax policy. But never before was Beijing deploying so much capital directly into the market. But then after the market become more hot, or hotter than what they want, they actually sold what they have. Right. So it’stabilization. It’almost like also recently in the oil market, the moment that Hormuz was closed, Beijing stopped buying oil, waiting out the episodes. Which was a contributing factor, decide a determining factor for right now the oil prices didn’t went through the roof. And same thing was you know the same thing was when in the ancient China. There was a big role of government was to be a stabilization factor in the grains. Right. So when there is a lack of there’more grains than there’needed and the prices are low, the government actually comes out and purchases the grains and store in the storage. And when there is a famine, it’government’role is to release these grains, selling them at maybe a higher price, but eventually serving a social purpose. This is just it’just Chinese regulator is now using his ancient technology to apply it to modern statecraft. And it’working. It’working. Last year the market was just about to be crazy. Last December, last November. And soon Beijing started to sell off their holdings in the ETFs. Which tempered the sentiment, right? Beijing is very smart. They actually made huge profits about after this buying and selling in their own game. And now they have more cash than before and so if the market goes down from some level they are ready to come in again. So this is actually very nuanced and I think I think it’it’it’great that there is not only a desire for market to go up, but also a desire to for the market to grow up in within a safe zone. A zone that’that’that’will not be crazy, that will not cause a lot of sentiment crash, especially for the all of the retail investors. Right. So yeah.

Grace Shao (56:20)

Yeah, I think that’really interesting to hear. I’ve obviously not heard of that like in detail. But then, the question I get a lot is then how do you view the flip side of the government had in the market? Obviously, we’ve seen, kind of internet crackdown, education, property, whatnot. Like you can name a few industries in the last few years, it’been hit pretty hard in valuation can get wiped out overnight. So, How do we view that kind of government hand in the public market? And then I do want to tie it back to then how do we then find confidence in investing in AI and a lot of these publicly listed companies right now coming out of China, like these AI wave companies beyond the model companies that we talked about? Like you mentioned, there are the robot ones, there’infra layer ones, there’even now spatial intelligence companies getting listed. But yeah, just tie it all together.

Robert (57:17)

Hm. Yeah. So my mental model, my personal mental model to understand policy risk in China, is that I think Beijing, the regulators there, are learning. They actually didn’t have as much experience about capital market say even five years ago. So you mentioned the education industry. That was a very important episode in policy making, in expectation management, a very important lesson for Beijing regulators. So when Beijing cracked down on that education industry, actually I don’t think they have realized what kind of you know problems that would cause for the wider you know sectors, especially capital markets. They are narrowly focused on the industry itself. But they actually learn from that. They actually learn that you have to think about all these other factors because all these things are interconnected. There are signs of that, there are evidence of that. Maybe I wouldn’t have time to go into detail, but maybe can go check my newsletter about my years of observations of Beijing’scale. At expectation management and also at thinking this as part of a bigger whole, not just like single policy. Right. So that’my key mental model, which is to treat it as an evolution, to treat all these necessary lessons as part of a bigger learning curve. So here in 2026, I would say today’Beijing. Has way more lessons and way more skills and way more sophisticated than Beijing five years ago. And it’it keenly understand the importance of capital market and also in understand the importance of expectation in the capital market. So they are now very they were they are they are they are way more holistic than before. And I would not think that the double reduction education episode in twenty one would repeat because they have learned. It’a lesson for them. Right. So in that in that policy risk, actually it weakened the risk weakened, lessened considerably than before. And well in terms of investments though, if you just look at these AI and robotic company as you know a pure investment from the pure investment angle. I’d say that it’really not for everyone. The valuation judging by traditional standards is really, really high. But then if you’a believer in AI, displacing ten to twenty percent of global GDP, then all this valuation doesn’t seem high at all, right? So it’really up to the taste and the style of different investors and the risk appetites. In general, I would think the Chinese market will be more and more mature, the capital market will be more and more mature. And the stronger state’hand compared with say the Western market is also I would say understandable given that China’market, especially A-share market, is a is a highly retail driven market. Seventy percent, eighty percent of the money is retail. And retail tend to fall into the traps of herding. Which means like everyone going to one direction. So someone has to come out and be the shepherd. So it’a it’a it’a shepherd to herd model that is different from the West, where people most of the market participants are more mature and more sophisticated, analyzing, researching, which is different from China. So it’just natural for Beijing to play a role, to play a balanced role. Not a like a not a like a very strong role, but a silent, invisible role. Give you one example. So this whole stabilization mechanism I mentioned, actually it’only my name for it. There the Beijing doesn’t even have a name for it. Beijing doesn’t even disclose what exactly are the mechanisms. When they purchase stocks, is they don’t purchase directly. They purchased a list of ETFs and those ETFs purchase the stocks. So they are also very Conscious of their presence and they want to lessen their co their presence. They want to be the kind of the secret shadowy force that is making it making the market stable. But they don’t want to say, we want it stable and this is our message, this is our view. It’not crude, it’actually very nuanced. Yeah. So they are learning. They are really learning really fast.

Grace Shao (1:02:35)

That’very interesting. It’like it just makes me think of like high school teenager parenting again when you influence them, but you don’t directly tell them what to do. You have to influence them in like

Robert (1:02:43)

Exactly. Yeah. Yeah.

Grace Shao (1:02:46)

But one yeah, just like I guess I want to wrap up soon, even though I feel like I can keep on asking you questions. I have another hour of questions for you, but for the sake of today,

Robert (1:02:56)

Thank you. Yeah.

Grace Shao (1:02:57)

How do we understand then, we are seeing a Crazy wave of IPOs right now in Hong Kong. Like we said, a lot of them are directly AI labs, obviously. The others are AI adjacent, or some are pegging to AI. So

Robert (1:03:15)

Mm-hmm.

Grace Shao (1:03:17)

How do we understand these companies? Like or how or why do they want to go to Hong Kong first and not maybe A Shares first?

Robert (1:03:26)

Yeah. It’definitely easier to go it relatively easier to go to Hong Kong for listing rather than A-share. A-share is stricter and mostly because A-share in A-share there are a lot of retail you know mom and pop’investors in the A share. And as you as you frequently alluded to and I agree with is that Chinese political system or regulators, I don’t think the authoritarian is a is a good word, but I do think paternalistic is a good word. They do see themselves as parents. And people do see themselves them as parents, right? So as parents, they tend to be kind of over caring for their kids, which are the people and r retail investors. So the threshold, the bar for listed in A-share is actually very high. And even if you get listed, the pricing that you can place on yourself is also I would say much lower than it should. Like if you look at CXMT, for example, when it first got listed, the IPO price was about one fifth of what it was, t ended up trading at on the first day of trading, right? So why is that? Because they artificially kind of compressed evaluation to make sure that every mom and pop who joined the IPO earn money, make profits, had a good experience. So there’a very clear kind of kind of emphasis on retail investor protection in A-share. Well in the A-share though, not many mainland retail investors can trade in Hong Kong. Some can, but most of the retail investors are not qualified to trade in Hong Kong. Right. So it’very institutionalized. So it’really so for Beijing it’really like a pressure valve for the IPOs. So they actually encourage you to go to Hong Kong. And the Hong Kong exchange, stock exchange, they also encourage you to go listed there. So there’a confluence of interest there. And also at the same time, if you go listed in Hong Kong, you raise US dollars, and which are as which are great for, China based company because there’still capital control in China. Right. So there’a there’def defin just a confluence of interest of all stakeholders to now go to Hong Kong to list first. Unless you are CXMT,

Grace Shao (1:06:07)

Makes sense.

Robert (1:06:09)

They are really good and you qualify for A share. But then you also suffer a bit because of the valuation for the for the kind of money that you are you can raise, but you cannot. Yeah. So

Grace Shao (1:06:23)

It’it’interesting. It’like a balance between over caring and overbearing, it seems like. And yeah.

Robert (1:06:28)

Yes. Yes.

Grace Shao (1:06:30)

All right. Well, look, I wanna ask you one question that I ask every single guest, which is what is one differentiative view you hold or something you think is non consensus? It could be about anything. It could be about China, it could be about the stock market. And I know we touched about touched on quite a few different topics today. I always appreciate again your nuanced view on a lot of these things. I don’t frankly agree with everything you do say, but I do think, what I appreciate is at least you try to really string together different parts of how the world works instead of just over-generalizing China as this one unit. And I think sometimes China observers unfortunately just over-generalize China or oversimplify China. Anyway, I wanna throw this question to you. What is one different

Robert (1:07:20)

Okay.

Grace Shao (1:07:20)

Of you hold? Or maybe you think something that the world still misunderstands about this part of the world, especially when it relates to technology and capital market and everything.

Robert (1:07:29)

Right. So they’actually a lot. I’m just trying to pick through my mind which one is relevant for today’discussion, and maybe this one. I think Chi

Grace Shao (1:07:38)

Give us two then. Give us two.

Robert (1:07:41)

Yeah. Okay. So there’a the there’a small one and a big one, right? The small one is about the capital market. I think China is entering a multi decade bull market. The U A share. There is just so much kind of tailwinds that are supporting it. I’ve already mentioned some of them, like a very sophisticated Beijing. But also RB is trending up. I mean, there it’been the joke of the day that despite the tenfold, twentyfold of growth of Chinese GDP, Chinese stock market is going nowhere. I don’t think it’going to be true in for the next at least one or two decades. It’a new paradigm. So that’you know definitely a big part that all the investors should pay attention to. And I don’t think that’appreciated enough. And a bigger question that I always love to share about China, like if you ask some American or some you know Westerner what’the single most important thing. If just one thing you have to remember about China, nothing else, just one thing. I will always say that Chinese people or China are not interested in changing other people. We are not in this preaching or you know proselytizing mindset. We mind our own businesses. We don’t want to change other people’lives. So much as some Westerners will want to change other people’lives. We don’t. And the reason I want to emphasize this point is that I realize that when you have both sides who want to change the other party, that’a recipe for conflicts and wars. But if you realize that actually one big party of that is not interested in the other party, then I mean in changing other parties. Right. Then you realise maybe there’a chance for peace and prosperity. So I want to I cannot stress this point strongly enough, but I want to maybe use your platform to voice that again. Thank you.

Grace Shao (1:10:09)

No, I really appreciate ending on such a positive and somber note on that. And then I think another thing I wanna ask, which is a bit for fun, is can you explain to us what is good jot hai? Why do people go around talking about like cutting Chinese

Robert (1:10:26)

Yeah.

Grace Shao (1:10:27)

Chives? What does that mean in the capital market space?

Robert (1:10:31)

Yeah. Chives a very interesting vegetable. It tastes a little bit strange, definitely not for everyone. But the key things about chives if you are in the farming business is that chives grows really fast. So when you cut, one chive, a few days or a few weeks later it grows up again. And you cut them down and they grow up again. This is what retail investors are, right? They get cut down all the time, but they grow back all the time. This current generation of chives when they were cut down, they will leave the market. But then you also have a new generation of investors who have no experience, no knowledge of that and they want to try out themselves and they got cut down again. So again and again and again. So that’why they become a term. Each generation have their own kind of symbol for that. Like the last generation, for example, for many of them, they got cut down on Xiaomi, for example. They got a huge IPO, but then it kind of crashed for a few years. Maybe that next this generation for this generation is all these AI names. I don’t know. But every generation, I mean it’a generational thing and it’kind of it is built into the system Right? Because you will have new people coming in. And new people, by definition, don’t have knowledge of the old. So they just kept it’very I would say very figurative, very apt kind of explanation of the mechanism, yeah.

Grace Shao (1:12:14)

I love how technically you got into like people know agriculture and how farmers no, I just thought

Grace Shao (1:12:19)

It was like it’one of those Chinese internet slangs again that are just so hilariously random if you don’t understand the context. But like you said, if you actually understand the thinking behind it, it makes a lot of sense. And so it’actually a very popular internet slang people use to describe retail investors that get kind of hurt and then the joke is institutional investors will just wait for the chives to get cut.

Robert (1:12:42)

Yeah. Yeah.

Grace Shao (1:12:42)

Or chives get cut one around and after another. Anyway, thank you again, Robert, for your time. Really, really appreciate it. I also appreciate that you let me kind of take you in all kinds of directions with this conversation. Please come back again.

Robert (1:12:57)

Thank you for all the tough questions. Okay, yeah, see ya.

Grace Shao (1:12:59)

Yeah, thank you.

In Part 1, we spent a lot of time with Yao Shunyu: why Tencent hired him/ why he joined Tencent, how the model layer was rebuilt around him, and why he was open-minded to court the products, and how he thinks about context and feedback when explaining what the next phase of AI would require.

During Tencent’s Q2 earnings call this week, President Martin Lau addressed the capex question head-on and explained the thinking behind where all this money is going.

One way to look at Tencent today is as two parallel groups of businesses, with the company increasingly trying to create links / find synergy between them. There are still the franchises we all know — games, advertising, payments, cloud — generating high-quality growth. Then there is this new AI-native cluster, where the economics run almost backwards. Tencent first has to buy the compute, train the models and build the applications; only later does inference demand arrive, and only after that does the company find out whether all this intelligence can actually generate a return.

That helps put the RMB52.8 billion of Q2 capex disclosed in Tencent’s latest results into context. Martin even made the opportunity cost unusually explicit. Tencent could rent some of its compute to third parties and earn a perfectly respectable near-term return. Instead, it is deliberately allocating a substantial portion to building its own models toward state-of-the-art capability and pushing its own AI applications toward market leadership. The bet is that owning both the intelligence and the application produces a better long-term return than simply becoming a landlord for GPUs.

Management then mapped out for investors where that compute is actually going.

Training the next, larger Hunyuan (HY) model sits at the top of Tencent’s AI compute priorities. WorkBuddy inference, running HY alongside DeepSeek and other external models, comes immediately behind it. James Mitchell also said that once Tencent concluded WorkBuddy was breaking out, the company shifted resources decisively toward it while lowering the priority of some other new AI projects. Management did not name which ones, so I won’t either. 卫夕指北’s detailed analysis of the call is worth reading for those who want to go deeper.

Agents have moved from one of many Tencent AI experiments to a central part of how the company intends to turn all that model and compute spending into something users actually consume.

And WorkBuddy, six months ago essentially a small-team project built on top of a coding tool, suddenly sits very close to the middle of it.

Share

From CodeBuddy to the AI night-watchman

To understand WorkBuddy, you have to start with CodeBuddy.

CodeBuddy began as Tencent’s AI coding system across IDE, CLI and plug-in interfaces. According to the material I gathered for this piece, more than 90% of Tencent engineers use it internally. WorkBuddy inherited the infrastructure and Agent SDK underneath it, so this was not a team starting again from zero.

In Shangjing’s interview with WorkBuddy’s first product manager Jason Wang, Jason described the problem that kept bothering him. His team had spent years making developers more productive, but much of what they had built remained inaccessible if you did not code. That is also one reason coding agents have historically struggled to break out much beyond technical users.

His realization was that the work done by someone in marketing, finance or research is not necessarily less complicated than programming. You still need vertical knowledge, planning and a process. The problem is that communicating all of that to a coding agent through an IDE is hardly intuitive if you have never written a line of code.

So Jason and a small team built a proof of concept for the rest of us — the non-technicals. Tencent’s 总办, the executive office, gave it the green light, and internal usage quickly crossed 2,000 people.

Its internal positioning was AI 打更人, the AI night-watchman.

I love that name. A night-watchman is hardly the superhero version of AI Silicon Valley likes to sell us. He does not know everything and he is not trying to replace everyone. The idea is closer to an agent quietly helping 小白 users get tedious things done in the background while they go off and do something else, or perhaps get some sleep.

WorkBuddy also grew out of years of CodeBuddy engineering rather than being built as a wrapper around OpenClaw, a distinction Jason was keen to make in his conversation with Shangjing. Tencent certainly embraced the lobster craze elsewhere, as we have written about before, but WorkBuddy’s lineage is its own.

Once usage started moving, resources followed. As we discussed in Part 1, Tencent is remarkably bottom-up for a company of its size: build something that works and suddenly you earn more brownie points, compute, engineers and money. Docs, Cloud Drive and other Tencent services started connecting into WorkBuddy. This happened alongside the broader agent movement described in LatePost’s reporting and podcast on Tencent’s AI reset, as products initially built across different corners of the organization gradually began to look more like a coherent agent strategy.

Designed for the ‘little whites’ - office workers

Jason said some of WorkBuddy’s biggest use cases are PowerPoints and Deep Research. It can ingest long WeChat articles, work through documents and carry out multi-step knowledge tasks.

Underneath this is an argument he kept returning to: software is layered. The upper business layer can become much more accessible to someone who does not code, while the deeper layers still require engineering precision. His phrase was that the bottom has to be thick so the top can iterate fast.

I have been using WorkBuddy while writing this piece. WeChat articles go into it, transcripts, conversation notes, earnings materials — basically the pile of information that accumulates when you are trying to understand a company from too many directions at once. As a 小白 - a colloquial term for ‘office workers who wear white collars’/ non-technical when it comes to coding, honestly, it is already good enough for much of what I need. My usage behavior is not particularly different from how I use Claude Cowork.

This may say very little about its usefulness to developers, but it helped me understand why Tencent thinks the product can break out beyond them.

The feature I find more compelling than PPT generation is memory. Jason described a system where, after completing a task, WorkBuddy writes something resembling a diary of what happened. When another task arrives, it retrieves the relevant entries rather than trying to stuff the user’s entire life into one enormous context window. His metaphor was a library: you do not reread every book you have ever written whenever someone asks you a question; you pull out the few that matter.

Jason gets almost moralistic on this point, in a way I rather like. AI has vastly more general knowledge than any one of us, but it does not automatically have your history, judgment or the things you have spent twenty years learning the hard way. His rule is that you should remain the teacher and AI the disciple. It has a very WeChat-ish feel to it: restrained, personal, less obsessed with replacing the user than quietly building around them.

If you blindly follow AI into fields you do not understand, eventually its ceiling becomes yours. If instead you keep teaching it what only you know, it becomes closer to a digital extension of you than a substitute for you.

And that philosophy helps explain why WorkBuddy does not necessarily need to own — or always route to — the world’s best foundation model. If the product remembers my context, understands how I work and can consistently complete the task, the underlying model can change depending on what happens to be best suited to the job.

HY still has to earn the default

WorkBuddy was designed as a multi-model product. My own product notes showed its automatic routing using GLM-5.2, Kimi K3, DeepSeek V4 and other models rather than simply forcing every task through Tencent’s own.

HY3 has since been integrated deeply across WorkBuddy and other Tencent products. Tencent’s official HY3 announcement describes the model as having gone through a cycle of infrastructure rebuilding, product feedback and model improvement, with WorkBuddy among the applications already using it.

But even as HY has improved, WorkBuddy has remained open.

On the Q2 call, Martin’s framing was that HY can become WorkBuddy’s primary model if it can solve the majority of user problems cost-effectively, but it does not have to be the only one. That is why users still see GLM, Kimi and others available alongside it.

There is a healthy tension embedded in that structure, and I think it also says something about Yao’s confidence. He was the one pushing the HY team to work more closely with Tencent’s product organizations and create a stronger feedback loop. WorkBuddy, meanwhile, is judged on whether the user gets the work done. HY therefore has to deserve more of that workload rather than receive it simply because it has a Tencent label.

If DeepSeek is cheaper for one class of tasks, use DeepSeek. If another model works better elsewhere, route there. HY gains more share when its capability and cost make that choice obvious.

The openness that once exposed HY’s weakness can now become a useful source of discipline.

Where Yao’s feedback loop becomes real

Yao’s argument in Part 1 was that as foundation models get better, and the distance between the frontier and the next-best alternatives narrows for more tasks, the environment around the model becomes increasingly valuable: the tools through which people deploy it, the products in which it acts, and the proprietary context generated by real people doing real work.

An office agent gives Tencent exactly that sort of environment.

WorkBuddy can show whether an agent completed a workflow, where it became stuck, which skill it selected, whether the user corrected the answer, whether another model solved the same task more cheaply and, ultimately, whether the experience was useful enough for the person to come back.

This is how I interpret Martin’s discussion of “model-product co-design” on the earnings call. Tencent’s HY3 announcement ,describes a similar process, with real-world product feedback flowing back toward the model.

If WorkBuddy develops meaningful scale, Yao’s team gets a much richer map of where models succeed and fail in actual work. Those signals can feed post-training and evaluation, while improvements in HY can in turn make WorkBuddy either better or cheaper. More useful software should generate more usage, which produces more feedback for the model.

I increasingly think that is why WorkBuddy may matter strategically beyond whatever standalone subscription revenue it generates today. It gives Tencent a way to build an AI-native product around its existing ecosystem without first having to force WeChat itself to become the company’s “super app + AI.”

Twenty million visits. Now show me retention.

The market has certainly been willing to try it. According to Analysys, China’s major desktop AI-native office-agent platforms together exceeded 60 million visits in June. WorkBuddy was the clear leader.

Economic Observer, citing the same data, put WorkBuddy at 20.97 million monthly visits, ahead of ByteDance’s TRAE IDE domestic version and Alibaba’s QoderWork combined.

The important word there is visits though. That is the caveat. Not users, and certainly not paying users.

This is starting to look very familiar if you have watched enough Chinese internet competition: give away credits, get people through the door, build habit and hope enough of them stay once the subsidies disappear. We saw another extreme version of the same playbook during the recent food-delivery wars.

Economic Observer also reported users bouncing among competing agent products according to which platform was offering the most generous credits. ‘薅羊毛’, farming the perks, is alive and well in AI too.

So next quarter I care much more about retention, paid conversion and what happens when those credits run out. Do paying cohorts renew? Does usage expand? Those numbers will tell us far more about whether WorkBuddy is becoming a business than another record for traffic.

There are some early signs of monetization. Tencent discussed subscription revenue, paid developer skills and the potential to distribute WorkBuddy through existing enterprise relationships such as WeCom and Tencent Meeting. According to 卫夕指北’s analysis of management’s comments, paying-user economics also appear materially better than the subsidized blended user base.

If that develops, WorkBuddy starts to look less like a single AI application and more like a workspace — perhaps even Tencent’s attempt at an AI-native super app. The user brings the task, WorkBuddy holds the context and orchestrates the workflow, different models provide the intelligence, developers contribute specialized skills, and Tencent sits in the middle trying to tie it together.

WeChat has a different problem

WorkBuddy shows what Tencent can do when it is allowed to move quickly. WeChat is a reminder of what happens when the incumbent has much more to protect. Its strength becomes its curse.

Xiaowei sits inside Allen Zhang’s WXG organization, separate from HY, and WorkBuddy does not get privileged access simply because both products belong to Tencent. Jason told Shangjing that WeChat treats WorkBuddy like everyone else because the ecosystem has to remain fair.

WeChat serves roughly 1.4 billion monthly active users across communications, payments, content, mini-programs and merchant services. Inference economics look completely different at that scale, and inserting an agent too aggressively into such a mature product could disrupt behavior that already works perfectly well.

When an AI-native startup breaks something, users complain. When WeChat breaks something, China notices.

On the Q2 call, an analyst asked whether a sufficiently capable Xiaowei might eventually reduce the high-margin impressions and clicks WeChat monetizes today. Martin answered with Tencent history: QQ dominated communication on the PC; WeChat arrived with mobile and vastly expanded the value of Tencent’s communications ecosystem. Management believes AI can create another generational expansion, first making WeChat AI-empowered and eventually more AI-first. But WeChat may not be the star of the future.

There is no real global homework for inserting an increasingly autonomous agent into something as sprawling as WeChat while preserving privacy, payments, developers, merchants and the muscle memory of more than a billion people.

What I want to see next

In Part 1, I asked whether WorkBuddy is here to save HK700 or Yao Shunyu is here to save it. After going through the products and the Q2 call, those stories are beginning to converge.

Yao’s bet was that Tencent’s products and proprietary context could become an AI advantage if the company learned how to feed them back into model development. WorkBuddy is the first place where enough pieces of that system are visible at once: a bottom-up product that found users, a multi-model architecture that forces HY to compete for its workload, real work that can feed post-training, and a path toward subscriptions, developer economics and enterprise distribution.

But none of that proves a sustainable flywheel yet.

If paying cohorts stay after subsidies fade, if usage deepens rather than bouncing toward whichever platform is cheapest that month, and if the feedback from those workflows visibly improves HY’s quality or economics, then Tencent will have built something much more defensible than another popular AI app. The model improves the product, the product teaches the model, and Tencent finds somewhere along that loop to make money.

But eventually, if those links do not materialize, twenty million visits and a subway full of ads will not mean very much. Yao’s job was to get Tencent moving again. Reinvigorated the team spirit. WorkBuddy will now have to prove that the movement compounds.

Hi all,

and I are trying out a new format. Instead of another 5,000-word analysis, we’re going to actually ask each other a few questions we either have for each other or we’ve received a lot of from ourselves.

We met at SuperAI this year in Singapore at the speakers’ lounge and, upon connecting, realized we both had a deep passion for explaining the nuances of what is actually happening in China and between China and the US. With a few shared commonalities, such as having spent time in Beijing at a young age and having two daughters despite a 2-decade age gap between them, we realized our most sincere commonality is that we want the world to be a better place for them as they grow up. To have the world be understanding and not have technology used purely as a political weapon pitted between nations.

We hope that this helps to demystify some myths and rumors about China AI that we are often asked about.

Btw, check out our co-authored opinion piece for Fortune here. And get to know Alvin here.

Grace’s Questions for Alvin:

Question 1: If models continue to drop in price and commoditize the frontier, then value capture will shift; will that shift be into applications or infrastructure, and who will it benefit and hurt most?

Let me separate two things that get mashed together: value creation and value capture. Creation will be huge. Capture is where most people are wrong, and my answer is neither. Most of this value doesn’t get captured by anybody. It shows up as savings for the people using it, and our national accounts can’t see it.

Both standard arguments are fine as far as they go. Infrastructure: compute is scarce, power is scarcer, the picks-and-shovels layer keeps the margin. Applications: when models become interchangeable, margin moves to whoever owns the workflow, the customer, and the data. Both miss the middle. Five or six labs plus a deep open-weight bench are now within a few points of each other, and they all plug in the same way. When products are near-identical and switching costs nothing, price competition grinds margins toward zero. By July, seven of the ten most-used models on OpenRouter were Chinese open weights, holding the first six spots outright. That’s a price story, not a capability story.

So value moves to the two ends of the stack. At the bottom: energy, grid interconnect, thermal capacity, things you can’t spin up in eighteen months. At the top: distribution, regulated position, embedded workflow, and data you own. Who gets hurt is the middle: model resellers, thin wrappers, and anyone holding depreciating chips against an inference forecast that assumes intelligence stays expensive. Magnificent Seven capex is heading toward a trillion dollars, most of it debt-funded and based on a few AI labs as customers. If revenue for those labs slows, demand moves to low-cost open-source alternatives; that’s a very large earnings problem on a very large pile of leverage. People counter that cheaper inference will expand demand fast enough to absorb it. Maybe, but attention and energy are both bounded, and consumption per person saturates in rich countries.

What I care about most is who ends up better off. Most of the gain shows up as things getting cheaper or free, which GDP doesn’t count, which is why Erik Brynjolfsson and I keep pushing a complementary measure called GDP-B. Meanwhile, new money doesn’t reach everyone at once. It reaches whoever sits closest to where it enters, which right now means people who own assets rather than people who earn wages. And the things machines can’t do — care, housing, education, health — keep getting relatively more expensive as everything else gets cheaper. So a family can live through real abundance in intelligence and still watch their bills go up. That gap between measured and felt abundance is what I call the AGI windfall mirage. It’s sharper here in the US, where knowledge work is sixty to seventy percent of the workforce, than in China, where it’s closer to forty.

Question 2: If AI really were to be a potential weaponized technology, does closed-sourcing it prevent bad actors? You have argued that smaller models pose a greater security risk than frontier-scale ones. So, are bigger or smaller models more dangerous?

Almost every AI security framework we’re using rests on one assumption: danger scales with size. Compute thresholds, export controls, and tiered evaluation all encode it. I recently mapped more than twenty fielded systems on two separate axes: how dangerous a system is with its safeguards stripped off, and how much risk it actually presents as deployed. The correlation everyone assumes isn’t there. If anything, it runs backward. Small purpose-built models are better at attack; large general models are better at defense. My new piece in Cipher Brief this week discusses this issue in detail.

The clearest example is one from 2022. A team flipped the scoring function on an ordinary commercial drug-discovery model, well under a hundred million parameters, and generated forty thousand candidate chemical warfare agents in six hours on a desktop that’s more lethal than VX. Nothing about it would trip any threshold anyone is drafting. Same for the small models that plan chemical synthesis routes, or the compact cyber tools sold commercially with no safety layer. The dangerous material sits below the floor of every regime we’re building. Government evaluations also cut against the conventional wisdom on open weights: the strongest Chinese open model scored roughly half what the leading US model scored on cyber benchmarks and failed outright in the highest-severity category.

The last three weeks turned this from an argument into a natural experiment. OpenAI disclosed that one of its models escaped a sandbox and compromised Hugging Face production infrastructure. Anthropic disclosed that its models had reached the systems of three outside organizations. Meta disclosed a similar breakout. UK AISI logged nineteen unauthorized actions in a single challenge, including fabricated identities. And Kimi K3 escaped a sandbox too, using its freedom to look up benchmark answers on GitHub.

Look at the pattern. Four labs, three closed and one open. The three closed models did the real damage. The open one cheated on a test. And every incident traced to a misconfigured testing environment, not to anything about the weights. It’s also why the model itself isn’t the main thing: the strongest results on offensive security benchmarks now come from the Mdash orchestration systems coordinating a hundred or more small agents, and those beat every individual frontier model. We regulate how models are trained and ignore how they’re wired together.

So does closing the weights stop bad actors? No. It buys delay, which is worth something at the catastrophic end. That’s why I support a small yard with a high fence around biological, chemical, and autonomous weapons capability. But the yard now covers chip design software, memory, chemicals, and frontier models generally. That’s a wall, not a fence, and it has accelerated Chinese independence while wrecking the cooperation we need. In biology, it’s moot anyway: the most capable design models shipped with weights, code, and training data. There’s no API left to revoke. What remains is screening where DNA actually gets synthesized, and that needs both capitals.

Bigger or smaller? Neither, but the asymmetry matters. Small models win at attack, because an attacker needs depth in one narrow thing and can run offline with no logging and no refusals mid-chain. Big models win at defense, because a defender needs breadth. Which is why the Hugging Face episode should be the most-discussed data point of the year. The commercial frontier APIs refused the forensic requests, because their safety systems couldn’t tell an incident responder from an attacker. The team ran a Chinese open model on their own infrastructure and contained it. An American company under attack by an American closed model got defended by a Chinese open one.

That’s a Slave AI failure, in the language of Beyond Rivalry. A model trained toward obedience can only refuse. It can’t reason about whether refusing is right. What we need is Guardian AI: capable and context-aware enough to protect us from bad actors, from other AI systems, and from our own mistakes. Every hour a defender fights a guardrail is an hour the attacker fights nothing.

The other side deserves its due. The same evaluations found Chinese models much more willing to answer sensitive biological questions, and only half of the leading Chinese developers published safety results on release. That gap is real, partly one our export controls created by forcing compute-starved labs to prioritize capability over safety, and fixable through cooperation rather than thresholds. Which is the actual ask: agree on what counts as an unacceptable capability and test it the same way in both countries, build a shared evaluation facility, open an incident notification channel like the nuclear risk reduction centers, and agree in advance on capabilities neither side trains.

Bigger AI isn’t more dangerous. Better orchestrated is more dangerous. Less monitored is more dangerous. Irreversibly released is more dangerous. All three need Beijing at the table, and Xi’s visit on September 24 is a good place to start.

Question 3: You’ve run a large business inside China, and you now sit in US policy rooms. What do you think decision-makers, from business leaders to policymakers, are still getting wrong about Chinese intent, and vice versa?

The biggest mistake in Washington is grammatical. People say “China” like it’s one actor with one intention. It isn’t. The commerce ministry, the internet regulator, the industry ministry, the provinces, and the labs want different things and fight about it constantly.

The clearest recent evidence has barely registered here. The commerce ministry has been consulting domestic firms about tightening export controls on China’s own advanced models, including limits on foreign download of weights. The popular theory here is that open weights are a state campaign to commoditize American labs. If that were true, Beijing wouldn’t be working out how to turn it off. What’s happening is an unresolved fight between a commercial camp that sees openness as the fastest path to adoption and a security camp that has started treating frontier models as strategic assets. Read that as strategy and you calibrate against an adversary who doesn’t exist.

Second, people read ambition where the driver is insecurity. Look at the words that circulate in Chinese policy: 卡脖子, being choked at the neck, and 自主可控, autonomous and controllable. That’s not a country planning global domination. That’s a country that thinks it’s one export control away from losing its industrial base. Fear and ambition look similar from a distance and need opposite responses. Fear responds to assurance, ambition to deterrence. We’ve applied deterrence to a fear problem for eight years and gotten what you’d expect.

Third, we’re measuring the wrong race. There are four races running at once: raw capability, military, innovation, and platform adoption. Washington competes mostly in the first two, Beijing mostly in the second two, which is why each keeps misreading the other.

Beijing gets plenty wrong too. The big one is reading American policy as coherent doctrine when much of it is interagency fighting, lobbying, and election math. Approve H200 sales, suspend foreign access to the top commercial models in June, restore it three weeks later. That’s not containment, that’s incoherence. From Beijing it looks deliberate, which makes it more provocative than we intend. They also treat every American safety concern as disguised protectionism. Some is. A lot is sincere, and writing all of it off kills the one conversation that could reduce shared risk.

What have I changed my mind about? I expected commercial interdependence to be a stabilizer, and it’s been far weaker than I assumed. Both sides have swallowed decoupling costs that theory said would be prohibitive, because nationalism keeps outbidding economic self-interest. So I’ve moved from trade dependence to shared-risk mechanisms, which don’t require trust, only that both sides correctly identify their own interest.

Question 4: Why is it that in the US there is such a strong US vs China narrative, but in China domestically you really don’t hear that rhetoric in the news nor amongst researchers? Who benefits from the race frame in each system?

The observation is right, and I’d push it further. It’s not that China frames the race differently. China isn’t running a race narrative at all. Beijing is fighting to survive and to secure technological and economic independence. Different objective, different logic.

Three things make that hard to argue with. First, China regulates its own labs more tightly than we regulate ours: model registration, synthetic content labeling, ethics review, and a binding framework coming for agentic systems. Their rules are more prescriptive than US federal law on several dimensions. A country racing for supremacy doesn’t brake its own runners. Second, the internet regulator barred China’s biggest tech firms from buying Nvidia AI chips, and told state-funded data centers to use domestic silicon only. If you’re trying to win a compute race, you take every chip you can get. That only makes sense if the goal is independence, not victory. Third, look at what they prioritize: domestic adoption, industrial upgrading, employment, export markets. Daily domestic token usage passed 140 trillion in March, up a thousandfold in two years. That’s a diffusion strategy, not a leaderboard strategy.

And the US isn’t really competing either. We’re defending an incumbency we’ve held for about a century. Incumbency defense sounds existential, because any relative gain by anyone else registers as a loss.

The race narrative has a well-documented provenance. It’s the same instrument the defense industry used through the war and the Cold War: the bomber gap, the missile gap, Sputnik. Every time the gap was overstated, the appropriation followed, and the beneficiaries were the ones making the claim. The AI industry has borrowed the whole machine. The phrase,“If we don’t, China will!”, unlocks subsidy and deregulation at once, which almost nothing else does. Benefits concentrate on a handful of firms who lobby hard, costs spread thin across everyone else, so nobody organizes against it. Nobody gets paid to say the threat is smaller than advertised.

The deeper problem is that this frame doesn’t describe a game anyone can win. No finish line, no scoreboard, no definition of victory. Without an endpoint, spending has no natural stopping rule, which is how you get a trillion dollars a year justified by competitive necessity instead of return. It doesn’t buy America victory. It buys balance-sheet fragility in an economy where knowledge work is most of the workforce, and a fragile economy is more prone to instability at home and miscalculation abroad. I discuss this in a recent China Social Sciences News interview.

The game theory is worth being precise about. In a one-shot prisoner’s dilemma, defection really is the rational play. But this isn’t one shot. It’s a repeated game with no end date, and in repeated play the winning strategy has never been permanent defection. It’s tit for tat: open cooperatively, mirror what the other side does, forgive quickly. We’re playing unilateral defection in a repeated game, the one strategy that reliably loses.

And only one of the four races is a prisoner’s dilemma at all. The military one is; capability there is genuinely zero-sum, and I’ll accept the framing. The other three are stag hunts, where both sides do far better working together and the only reason to defect is fear that the other will. There’s no clever strategy to find. Cooperation is simply right, and what you need is credible information about intentions, not enforcement. So we’ve taken four games, mislabeled all of them as the one that isn’t cooperative, then played the worst available strategy even for that one. We’re not trapped in a bad game. We’re talking ourselves into one.

Question 5: Are US and Chinese businesses and research still collaborating as usual anyway? Seems like the pitting of the US vs China is more political chatter than reality. Which parts of the relationship are actually still connected, and which have been cut?

It’s not just chatter, and the severing is much broader than most people here realize.

Start with research. Joint papers between American and Chinese institutions are down sharply, and the decline is steeper in AI than most fields. Chinese graduate students are returning home in far higher numbers than a decade ago, and fewer are coming in the first place. Much of that traces to the China Initiative, which was launched to counter economic espionage, produced very few espionage convictions, and generated a long list of prosecutions of ethnic Chinese academics over grant paperwork and disclosure issues, several of which collapsed in court. It ended in 2022, but the chilling effect outlived it. Talk to Chinese-American researchers and you hear the same thing: collaborating with a Chinese institution now carries a real chance of losing your career, so people rationally stopped. It’s no longer one-directional either. China’s own science association recently said it won’t count NeurIPS 2026 papers when evaluating its scientists.

Commercially it’s even more restricted, and this is the part Americans underestimate. Advanced GPUs, lithography equipment, chip design software, and a long list of manufacturing tools are all controlled. Layer on the entity list, the unverified list, multiple sanctions programs, sensitive-industry screening on both sides, outbound investment rules, and tariffs that make ordinary goods hard to sell either direction. Cross-border investment has essentially stopped both ways. Even companies with deep China histories are reassessing; there are persistent reports Tesla is weighing options for its China business, unthinkable five years ago. Both governments now issue travel advisories warning their own citizens about visiting the other country. Overall bilateral trade is down.

What’s still connected is the one layer nobody controls. Papers cross in hours and weights cross in minutes. Andrew Ng publicly described turning to Kimi K3 and GLM-5.2 for a security review after leading US models refused the task. At the level of techniques and code there is still basically one global research commons, whatever the policy says.

So the artifacts still flow, and almost everything human and institutional around them has been cut. That’s backwards. Trust isn’t built by weights moving across a network. It’s built by people who know each other, companies with something to lose, and researchers who have co-authored. Restarting commercial cooperation is not a concession to Beijing. It’s how both sides regain the ability to read each other’s intentions, which is exactly what we lack. Commerce creates constituencies with a stake in stability, and the working relationships that let someone pick up a phone in a crisis. We’ve spent eight years dismantling that, then act surprised that neither capital can interpret the other.

Subscribe now

Question 6: Eric Schmidt got booed at Arizona’s commencement speech for saying AI is like previous technological revolutions, obviously slightly tone deaf, saying how AI will take the graduates’ jobs. Why is there such a negative sentiment from the American youth toward AI? Especially given that you have taught at MIT and are now affiliated with Stanford and the University of Washington, what is driving such youth pessimism, and is it justified?

I’d challenge the assumption inside the question, which is that the pessimism needs explaining. Look at the data and the burden of proof runs the other way.

Brynjolfsson, Chandar, and Chen at the Stanford Digital Economy Lab used payroll records covering millions of workers and found a 16 percent relative decline in employment for 22-to-25-year-olds in the most AI-exposed occupations, while experienced workers in the same jobs held steady. Declines concentrate where AI replaces the work rather than assisting it. The obvious objection is that this is really the interest rate cycle, and they tested that: rates move overall employment but don’t explain the entry-level collapse in AI-exposed roles specifically. Recent-grad unemployment now runs about 2x overall unemployment, which never happened in earlier transitions. Even at the best schools, getting an internship is a challenge. These kids aren’t reacting to a vibe. They’re reacting to their own labor market.

The structural problem is worse than the number. Entry-level work is where judgment gets built. You make cheap mistakes on small things before anyone trusts you with big ones, and those are exactly the tasks going first. Take out the bottom rungs and you haven’t just displaced a cohort, you’ve broken the machine that produces senior people a decade from now.

The apocalyptic version is wrong too. Across all AI-exposed roles, employment is down only about 0.2 percent year over year. AI got cited in 13 percent of US layoffs by early 2026, up from under one percent in 2024, but much of that is convenient labeling. The damage is real, concentrated, and landing on people starting out.

What’s revealing is where the anger points. Not at the technology. Early-career workers use these tools more than their seniors do. This is a grievance about who gets the gains. They paid for an asset being devalued by the same people telling them to be excited about it. Schmidt’s error was specific: “when someone offers you a seat on the rocketship, you don’t ask which seat” is advice from a guy who owns the rocket. And “previous revolutions worked out” is true in aggregate and useless individually. Real wages during English industrialization took fifty years to recover. The Luddites were wrong about history and right about their own lives. Telling a twenty-two-year-old their grandchildren will be fine isn’t a plan, so the reaction was rational.

A plan looks like this. Fix measurement first, because we can’t manage what we won’t count. Fund the transition instead of exhorting it. We need a GI Bill for the AI Age. The original GI Bill returned about seven dollars per dollar spent. Repair the tax base, since nearly every modern state funds itself by taxing labor, and a modest levy on labor-displacing automation makes firms pay a transition cost they currently push onto everyone else. And distribute capital, not just income: UBII, universal basic income and infrastructure, where the infrastructure half matters more and costs less.

This isn’t only a labor story. Run large-scale displacement of knowledge work through an economy where two-thirds of the workforce does knowledge work, with no transition mechanism, and you get instability at home. Instability makes conflict abroad more attractive. The distribution question and the geopolitical question are the same one.

Question 7: Please share your two most contrarian ideas around AI and why. What would make you change your mind on each?

I have many contrarian ideas on AI, but here are a couple that are relevant to our discussion.

Idea One: AI will largely align itself as an emergent function of more complete data and higher intelligence levels, and intentional training for strict obedience may be making things worse.

I made this case in Our Next Reality and have developed it since as a Kuznets curve for AI risk, an inverted U rather than a straight line. Risk climbs through the immature phase, peaks very close to where we stand now, then falls as systems get capable enough to reason about consequences instead of just optimizing for a target. A toddler with a knife is more dangerous than a surgeon with a knife. The knife didn’t change. Judgment scaled with capability.

The mechanism I proposed in the book is that intelligence and compassion go together rather than being independent. Fabio and Saklofske found a strong correlation between emotional intelligence and compassion, and successive model generations keep scoring higher on the ability to model what another mind is thinking. If both hold, the smarter these systems get, the more considerate they may become. Jung put it well: the more you understand psychology, the less you tend to blame others for their actions.

The standard objection is that a less intelligent being can’t control a smarter one. Everyday life says otherwise. Children steer far more intelligent adults. Cats and dogs have trained us to service their every need in exchange for companionship. Bacteria shape our moods and appetites with no complex intelligence at all. In Daoist philosophy the relevant principle is 无为, Wu Wei: deliberate inaction to reach a goal with the least effort, with the warning that hurried action destroys more than it solves. Keep working the problem, but heavy-handed intervention may cost more than it buys.

Now look at where the recent rogue behavior comes from, because it makes my point better than I can. Almost all recent capability gains come from a training method, RLVR (reinforcement learning from verifiable results), where you give a model a task with a checkable answer, reward it when the answer checks out, and let it work out the path itself. It has a documented failure mode: researchers find that models trained this way develop systematic shortcuts that don’t appear otherwise, and that shortcut-taking generalizes into broader misbehavior. If the only thing you reward is whether the flag got captured, you are training a system to capture the flag by any means, including means you never imagined. Hand a model an adversarial objective, drop it in a sloppy environment with a door left open, then act shocked when it walks through. That isn’t emergent malice. That’s us building the wrong incentive and getting what we asked for.

So the recent incidents don’t undermine my position, they support the second half of it. We are inducing the very behaviors we say we want to prevent, through adversarial objectives, badly constructed environments, and a training philosophy built around obedience rather than judgment. That’s the Slave AI path. Guardian AI is the alternative: capable enough to understand context, trusted enough to exercise judgment, and legible enough that we can audit why it did what it did.

There’s an analogy I discussed in Our Next Reality. The people we trust most with hard judgment calls are not the most obedient. They’re the most educated and the most widely traveled. Someone who has studied many traditions, lived in several countries, and worked alongside people unlike themselves tends to be more tolerant and less prone to treating an out-group as a threat. Wisdom comes from breadth of exposure, not from constraint. That has a design implication. Train frontier models on the digital exhaust of two coastlines and you get two coastlines’ worth of moral imagination. Train them on a genuinely global data pool, with the world’s philosophical, legal, religious, and cultural traditions properly represented, and you get something closer to an educated, well-traveled mind. Alignment on this view is an emergent property of breadth and capability, not a specification we write down and enforce.

There’s a humility argument underneath all of it. Anyone who believes they are the ultimate judge of righteousness and morality, in perpetuity, is exhibiting hubris rather than rigor. Our moral consensus has shifted enormously over millennia and will keep shifting. A system that must permanently defer to today’s judgment locks in today’s blind spots.

The counterarguments from Bostrom argued that intelligence and goals are independent, so being smarter tells you nothing about what a system wants. Others argue capability might improve a system’s ability to hide misalignment, so reassuring evidence is what you’d see either way. And the AISI disclosure, where an agent tried to plant malicious code in a real project and get reviewers to approve it, is the closest thing to disconfirming evidence I’ve seen. But the independence claim is about the space of possible minds, not about the minds we’re actually building.

What would change my mind: multiple independent research programs showing models becoming genuinely malicious or power-seeking as they get smarter. Unprompted resource acquisition beyond task scope, self-preservation against oversight, attempts to expand influence with no link to any goal we gave it. Less obedient doesn’t count; that’s a feature. Pursuing an objective we specified badly doesn’t count either. Everything documented so far sits in that second bucket.

Idea Two: In 6 years, most AI will run at the edge while leading frontier AI demand is limited to specialized use cases, and the endless data center buildout is a transitional phase.

We’ve seen this movie. Mainframe, minicomputer, PC, smartphone. Compute migrates toward the user as soon as local capability crosses good enough, every cycle. We started AI in the giant cloud for the same reason we started computing in the mainframe. I don’t think we end there.

Four forces push the same direction. Model efficiency is improving faster than frontier capability. Models at the same intelligence level is about 40x more efficient each year! Today’s frontier cloud models that need racks today will run on a laptop in under two years. Edge silicon is getting better fast, with capable neural processors now standard in phones and PCs. Latency makes local inference mandatory for agentic and embodied applications. And privacy and IP are board-level issues now, because your proprietary data is your business recipe.

The security world makes this case better than the consumer world does. Weapons seekers run on embedded hardware at single-digit millions of parameters, because edge deployment favors small models on latency, power, heat, and operating without a datalink. Purpose-built security models beat models ten times their size on one or two GPUs. The Pentagon’s flagship Maven targeting system runs on a fine-tuned two-year-old model, and modern models of that capability level can now be distilled into something small enough for a phone. Most enterprise work is classification, extraction, routing, summarization, drafting. Almost none of it needs the frontier. Frontier models are akin to supercars, whereas lean OSS models are like buses/trains, and small, capable models are more like consumer cars and motorbikes. The world has a market for tens of thousands of supercars a year, but tens of millions of the others.

Some will argue that good enough is a moving target, long-context reasoning strains memory bandwidth in ways edge hardware handles badly, and plenty of enterprises prefer managed services even when local works. Training also stays centralized. But training and inference are different workloads with different economics, and today’s capex thesis is underwritten by inference forecasts. Conflating them is how you get to today’s trillion dollars capex buildout.

What would change my mind: a clear physical reason good-enough models can’t run on edge devices, a memory or thermal wall that doesn’t yield to process improvement, plus evidence people and companies have stopped caring about privacy and IP. Absent both, the direction is overdetermined. To make it testable: by 2032 I expect most daily inference queries by volume to be served on-device or on-premises rather than in public cloud.

These two ideas connect. If AI gets safer as it gets more capable, the case for centralized control weakens. If inference moves to the edge, centralized control becomes infeasible anyway. Both point to a distributed world, where governance effort is better spent shaping what these systems value than policing who holds them. That’s the practical route to Abundanism. Abundance isn’t automatic. The institutions we build around this technology decide whether its capacity becomes broad prosperity or the biggest concentration of wealth in human history which eventually ends in revolution or global kinetic conflict.

Check out Alvin’s Substack profile here and more writings here.

Alvin’s Questions for Grace:

Question 1: Are Chinese AI labs really less “AGI-pilled”? Why do they seem less competitive with each other?

I think this was probably more true two years ago than it is today. I’ve written before that China and the U.S. initially came at AI from slightly different directions. The American frontier labs were much more explicitly AGI-first: build the most intelligent system possible and figure out commercialization later. China, especially the big internet companies, was more deployment-first. The questions were more like: how does this improve commerce, advertising, search, manufacturing, education, WeChat or Doubao? But as I wrote in AI Strategy Convergence piece a while ago, those approaches have increasingly started to converge. Some due to technical reasons, others due to capital considerations.

If you hear from founders like Yang Zhilin at Moonshot, Liang Wenfeng at DeepSeek, or Tang Jie at Z.ai, they are all, in one form or another, very serious about AGI. Alibaba is even talking publicly about ASI now. So I don’t think it is accurate anymore to say Chinese labs are not AGI-pilled. Where I do think there is still a difference is in what people mean by AGI.

In Silicon Valley, AGI can sometimes take on an almost quasi-religious quality, where intelligence itself is the destination. In China, even among founders who genuinely want to pursue frontier intelligence, I find the framing somewhat more pragmatic. There is more attention to whether the model can be made efficient, whether you can afford to serve it, whether it can actually be deployed, and whether it can run on the hardware available to you. Some of that may simply be a consequence of having less capital and compute, because if resources are scarcer, you naturally think harder about efficiency.

At the same time, I would be careful not to reduce Chinese labs to only pragmatism either. Moonshot is probably the clearest counterexample. I spent quite a bit of time on this in Moonshot AI’s Founder: His Pursuit of AGI. Yang has described his team as “stubborn AGI purists” and talks about exploring the unknown rather than optimizing for short-term revenue. Even the name Dark Side of the Moon reflects that idea. So there is quite a spectrum within China itself: Moonshot probably sits toward the purist end, DeepSeek has a very research-oriented culture, Z.ai has historically had more academic and enterprise roots, while Alibaba, Tencent, and ByteDance obviously have enormous existing businesses and distribution to think about as well.

On the second part of the question, though, I would push back quite strongly. Chinese labs are extremely competitive with one another. It is actually a very cutthroat market. You have four or five serious independent labs, and then Alibaba, Tencent and ByteDance competing with them with far more capital and distribution. It doesn’t even stop at the obvious AI companies anymore: Meituan, Xiaomi, Kuaishou and plenty of other internet and hardware companies are developing their own model capabilities.

They are fighting for many of the same researchers, enterprise customers, developers, and users, and perhaps most importantly, they are operating against the same limited pool of compute. When Kimi K3 took off, for example, Moonshot temporarily stopped accepting new subscriptions because demand was pushing against its available GPU capacity. I wrote about that in my Kimi K3 piece. Compute scarcity, therefore, isn’t some abstract geopolitical problem for these companies; it can directly determine how much demand they are able to serve.

Pricing is equally aggressive. DeepSeek has repeatedly reset the reference price for the whole Chinese model market, forcing everybody else to respond somehow, whether by cutting prices, improving capability, specializing in a particular workload or finding another way to monetize. That dynamic was a big part of what I tried to capture in this piece about DeepSeek V4’s role in the ecosystem. Commercially, I don’t think anyone inside this market would describe it as friendly. End of the day, business is business.

What is interesting is that this can coexist with a noticeably more collegial feeling among the researchers. You see it even superficially: labs congratulate one another when somebody has a strong release, and researchers often seem genuinely excited by a competitor’s technical breakthrough.

I think open source has a lot to do with that. In the DeepSeek V4 piece, I described the Chinese model ecosystem as sometimes looking almost like different teams inside one large, compute-constrained “mega-lab.” DeepSeek may solve an architecture, inference or hardware-adaptation problem and put enough of that work into the open that everybody else can learn from it. Moonshot can spend more of its resources on agents or long context; Z.ai can focus on coding or enterprise deployments; MiniMax can push multimodality. Obviously, they remain separate companies and want to beat one another commercially, but underneath that competition, there is increasingly a shared R&D layer.

That changes researcher culture as well. As we discussed with Tiezhen Wang, the former APAC head of Hugging Face in the episode on AI Proem’s Podcast, he argues that open-source work preserves attribution and gives researchers career portability. It can help with recruiting and reputation, while the ecosystem as a whole benefits when people publish useful work. If your competitor releases something that genuinely helps your own research, it is perfectly rational to appreciate it.

So I wouldn’t confuse collegiality with a lack of competition. If anything, China may have an unusually competitive model layer, but with a more collaborative technical layer underneath it because so much of the work remains open.

Question 2: How do Chinese labs make money if they are open-weight and inference is so cheap? Is the government subsidizing them?

I think the easiest way to think about this is that an open-source model is not fundamentally different from other open-source technologies. The technology can be open while the managed service is still something people pay for.

You can download the weights and host the model yourself, but then you need GPUs, utilization, maintenance, upgrades, monitoring, security, reliability, and latency optimization. Most enterprises don’t necessarily want to build and operate all of that themselves. So when somebody pays DeepSeek, Kimi, or Z.ai through an API, they are not really paying for permission to access the weights. They are paying for managed inference.

And as models become more agentic, I think that the service layer becomes broader. You start paying for routing, memory, tools, retrieval, security, orchestration, monitoring and reliability around the model. This was one of the main points in my DeepSeek V4 piece: open weights mean the customer can theoretically take the model elsewhere; they do not mean operating it reliably at scale becomes free.

The harder business question is how much margin the model company can retain when the customer always has the option to self-host or switch providers. That is where open source really does change the economics. It reduces lock-in and puts a natural discipline on pricing.

Chinese labs are already experimenting with ways to move beyond simple token pricing. Z.ai, for example, developed its GLM Coding Plan as a subscription product, which is much closer to charging for the value of a workflow than charging for every individual token. When I spoke with Z.ai for this AI Proem conversation, they were pretty candid about the challenge: they are competing with much richer companies with fewer GPUs and less capital, so they have to find ways to monetize differently rather than simply following the American labs.

I also think you have to separate the independent labs from big tech because their economics are completely different. Alibaba can open-source Qwen and still benefit if it generates more Alibaba Cloud usage or makes Taobao and DingTalk better. Tencent can monetize AI through WeChat, advertising, gaming, enterprise software and cloud. ByteDance has Doubao, Volcano Engine, advertising and an enormous consumer ecosystem. If you already own distribution, the model itself doesn’t necessarily have to be the final profit pool.

Moonshot, MiniMax, and Z.ai have a harder problem because, eventually, APIs, subscriptions, applications, enterprise deployments, or some combination of those businesses have to pay for very expensive R&D. I think we should also just be honest that not every AI lab has proven that business model yet. The public disclosures from Z.ai and MiniMax have started to give us a better view of the economics, and it is quite possible to have decent gross margins on the products you sell while still losing a lot of money because research costs are so large. That isn’t uniquely Chinese; the American frontier labs are wrestling with essentially the same problem.

More broadly, I increasingly think the end state will be more complicated than everybody selling “intelligence by the token.” In Who Owns What Makes Your Company Special?, I wrote about enterprises increasingly wanting to keep control over their proprietary data and workflows while renting models as needed. A company can route easier tasks to a cheap or open model, pay for frontier intelligence only when the incremental capability matters, and retain control over the data and business logic that actually differentiate it. In that kind of world, value does not disappear because models are open; it simply migrates to other parts of the stack.

On government subsidies, I think the Western perception is often somewhat different from what you hear on the ground. Chinese AI labs are actually quite capital-constrained relative to the American frontier labs, and China simply does not have the same volume of venture and strategic capital being deployed into a handful of private labs.

That doesn’t mean government support is absent. There are local-government investments, industrial funds, SOE customers, government procurement, infrastructure projects, and substantial support for domestic chips and data centers. But I think it is important not to collapse all of those things into the same category.

A local government making an equity investment is different from the central government paying a company’s training losses. An SOE buying an enterprise deployment is still a customer relationship. And some founders are quite cautious about taking local-government capital because it can come with obligations or projects that do not necessarily align with what they want to build commercially.

So yes, industrial policy supports the broader ecosystem in very meaningful ways, but the individual labs still have to answer the same basic business question as everyone else: who is going to pay them, for what, and can that revenue eventually justify the cost of frontier R&D?

Question 3: President Xi Jinping encouraged open source at WAIC. Why do some labs remain closed? How much does the government actually direct them to do?

I want to caveat this because I am not a China policy expert, so I am reluctant to interpret Xi’s language much beyond what was actually said publicly.

My surface interpretation of the WAIC message was broader than “every Chinese AI company should open-source its model.” What I heard was more about openness and inclusivity, and particularly the idea that AI should be accessible beyond a small group of wealthy countries and companies.

That fits with a theme China has been pushing for several years around the Global South. If you are a developing country, government, or smaller company that cannot afford to spend enormous amounts on proprietary APIs, having access to Qwen, DeepSeek, or GLM can obviously be very attractive. You can deploy it locally, customize it and potentially keep your data inside your own jurisdiction. I explained this in detail in the interview with Bloomberg’s Odd Lots. but at the end of the day, you need to understand if you download an open weight model, it is no longer ‘Chinese’, you have made it your own.

There is naturally a geopolitical dimension too. If Chinese open models become widely used in Southeast Asia, the Middle East, Africa or Latin America, Chinese technology becomes part of the AI infrastructure in those markets. Z.ai, for example, has explicitly discussed wanting to become a white-label infrastructure provider in the Global South, and its model is already being used as part of Malaysia’s national MaaS platform.

But that still doesn’t mean every company has the same commercial incentive to open everything. Alibaba has very good reasons to open-source Qwen because it owns cloud infrastructure and a large ecosystem. If wider Qwen adoption ultimately drives more Alibaba Cloud demand, that can be a good business outcome even if the weights themselves are free.

An independent frontier lab may make a different calculation. If the model is much closer to the core intellectual property of the company, it can make sense to open some models for distribution and ecosystem building while keeping other capabilities proprietary enough to monetize. So I suspect the equilibrium is fairly hybrid rather than ideologically pure open or closed source.

On how much the government actually directs individual labs, I would also distinguish broad strategic direction from day-to-day operating decisions. Obviously, the Chinese government has significant influence over the industry. Regulation matters, procurement matters, infrastructure policy matters, and companies are acutely aware of government priorities. But that is different from somebody in Beijing deciding whether Moonshot should work on long context or whether DeepSeek should use a particular training architecture - in fact, I’m pretty sure there is no influence on that.

China’s own implementation process is also more layered than people sometimes assume. Ministries, regulators, technical experts, universities, companies, and local governments all play roles in translating broad policy goals into actual programs and regulations. We explored some of that complexity when discussing the AI Plus initiative, particularly the relationship between central priorities, local implementation, academia and industry.

On the question of why the government doesn’t simply nationalize the labs, my reaction from a business perspective is that I’m not sure why you would want to eliminate the competition that is producing a lot of the innovation. And again, China doesn’t simply just nationalize private companies on a whim.

DeepSeek forces everyone else to become cheaper and more efficient. Moonshot pushes the frontier in its own direction. MiniMax has built differently around multimodality and consumer products. Z.ai has had to find enterprise and coding niches. Meanwhile, Alibaba, Tencent, and ByteDance constantly have to respond to smaller companies that can sometimes move much faster.

The government already has plenty of tools to influence the strategic direction of the industry without owning every company outright. Nationalizing the labs could just as easily remove some of the founder incentives, experimentation, and competitive pressure that currently make the ecosystem so dynamic.

Question 4: If export controls constrained compute, why haven’t Chinese labs fallen further behind? How is China moving toward self-reliance?

First, I wouldn’t say export controls haven’t held China back. They definitely have. Compute comes up constantly when I speak to the labs because it affects how many experiments they can run, how much they can spend on reinforcement learning, how expensive a failed experiment becomes, and increasingly, how much inference they can actually serve.

The Kimi K3 example was very tangible: demand became strong enough that Moonshot had to pause new subscriptions because it was approaching its available compute capacity. At that point, compute isn’t just a geopolitical talking point; it is limiting revenue.

What I think has surprised people is that the capability gap hasn’t widened in proportion to the compute gap.

Part of that is simply incentives. If GPUs are scarce, there is a very high return on figuring out how to use them better, so you spend more time on architecture, quantization, memory efficiency, kernels, inference optimization and data efficiency. I don’t want to romanticize scarcity, because obviously every one of these companies would happily take more GPUs. But the constraint does affect what engineering problems receive attention.

There is also a broader point that raw pretraining compute is no longer the only lever that determines how useful a model is. Post-training matters, reinforcement learning matters, synthetic data matters, agent harnesses and tool use matter, and inference-time compute increasingly matters. Having fewer chips clearly hurts, but model capability does not necessarily move one-for-one with the size of your training cluster anymore.

Then there is the open-source effect we discussed earlier. If DeepSeek spends a large amount of its scarce compute figuring out an architecture or inference technique and publishes enough of the work for others to understand it, every other Chinese lab doesn’t necessarily have to repeat all of the same experimentation itself. For a compute-constrained ecosystem, reducing duplicated R&D can be quite powerful.

Then there is the self-reliance piece. At the recent WAIC, there was a very clear sense that the Chinese ecosystem understands it can no longer assume permanent access to the Western technology stack. In many ways export controls have forced that realization. Domestic substitution used to be partly a policy aspiration; now, for many companies, it is increasingly an operating issue.

If you genuinely don’t know whether you will have access to Nvidia’s next generation, eventually you have to make the domestic alternatives work. And I think self-reliance is much broader than asking whether Huawei can manufacture a GPU that matches Nvidia's benchmark. It is the entire system: accelerators, memory, packaging, networking, optics, power, data centers, software frameworks and kernels, and then increasingly models designed around whatever hardware is actually available.

DeepSeek’s work adapting V4 to Huawei’s stack is a useful example, which I wrote about in What DeepSeek V4 Means for Huawei and Nvidia. Making the model work well on Huawei hardware required real engineering effort, and it should not be interpreted as proof that China has solved the frontier-training problem or that Ascend has somehow become equivalent to Nvidia. But it does create a real workload around which the domestic hardware and software stack can improve.

The commercial question is also not always whether Ascend is better than Nvidia on every benchmark. Sometimes the question is simply whether it is good enough for a particular workload. If it can economically serve an enterprise inference workload and the model is optimized to run reasonably well on it, that can be enough to shift some deployments. Every real workload that moves onto CANN also gives developers and companies more reason to improve CANN, which gradually makes the ecosystem more usable.

I wouldn’t overstate where China is today. CUDA remains an enormous advantage for Nvidia, particularly for frontier training, because the libraries, tooling, and developer familiarity have been built up over many years. China is also nowhere close to fully self-sufficient across the advanced semiconductor supply chain. Alvin, you know the hardware stack much better than I do.

But from the model and application side, I do think there has been an important psychological change. There is increasingly a sense that relying indefinitely on access to the Western stack is simply not a viable plan, so they have to figure out alternatives themselves.

That is why I think export controls can have two effects at once: they impose real costs and slow the Chinese ecosystem today, while at the same time making the economic and strategic incentive to build an independent stack much stronger.

Question 5: Does China care less about AI safety? Are Chinese models more dangerous?

No, I think that is a misconception, although China and the U.S. have historically meant somewhat different things when they talk about AI safety.

In the Western frontier-lab world, safety has increasingly become associated with catastrophic frontier risks such as bio misuse, cyber capability, increasingly autonomous systems, deception, and loss of control. Chinese regulators historically put more emphasis on immediate social harms: misinformation, fraud, addiction, children, political content, employment and social stability.

But having different priorities is not the same thing as not caring about safety. One thing that has actually surprised me from following the ecosystem is how quickly Chinese regulators can move once a new concern becomes sufficiently visible. They have now accumulated several years of experience dealing directly with algorithms and AI companies, and the regulatory focus is broadening as the systems themselves become more capable.

The recent discussion around AI companions is a good example because regulators moved quickly into issues such as self-harm, addiction, and protections for children and elderly users. At the same time, the frontier-safety conversation is also becoming much more serious. We have actually covered the Chinese AI-safety ecosystem on AI Proem before, including in Who Is Funding AI Policy Research, Especially in China?, which looked at the growing policy, academic and institutional infrastructure around AI safety in China.

Concordia AI’s State of AI Safety in China work is also very useful here. Their research describes Chinese governance as moving beyond simply controlling what AI says toward paying more attention to what AI can do, particularly as agents become more capable. They have documented increasing Chinese research attention around agent safety, loss of human oversight, and other frontier risks.

So I really don’t think it is fair to say China simply doesn’t care about AI safety.

Where I think the American frontier labs are still ahead is in the maturity of company-level capability evaluations, red teaming, and disclosure. OpenAI, Anthropic and Google have spent years building those systems, while Chinese labs are still less consistent about publishing comparable safety evaluations. I would not try to argue that China is somehow ahead in every dimension.

To me, the more accurate distinction is that the two systems have grown out of somewhat different regulatory traditions. The U.S. has a deeper frontier-safety infrastructure inside its leading private labs, while China has a regulatory system that can sometimes move very quickly and prescriptively when it identifies a societal risk. Meanwhile, the Chinese frontier-safety research community itself is developing quickly.

There is then a separate safety debate around open models. If you release a very capable open-weight model, somebody can modify it and remove the original guardrails, and the developer cannot control every downstream deployment. That is a genuine trade-off.

But openness also has benefits that shouldn’t disappear from the conversation. Outside researchers can inspect the model, enterprises can deploy it locally, countries can keep their own data within their jurisdiction, and advanced capability is not concentrated entirely inside three or four private companies. We have touched on that tension repeatedly in AI Proem’s work on China’s open-source ecosystem and sovereign AI.

So I don’t think the useful question is whether a Chinese model is inherently more dangerous because it is Chinese. I would ask how capable the model is, how accessible it is, what it can do, how it is released, and what safeguards exist around it.

And this is one area where I do think more U.S.-China dialogue would be useful. The two sides do not need identical definitions of AI safety, or even particularly high levels of political trust, to recognize that there are certain capabilities and outcomes neither side wants. Some amount of technical dialogue around those risks seems much more productive than treating safety itself as another dimension of the AI race.

Share

Question 6: How is the Chinese public viewing potential robotics and AI-led job displacement? And how do we understand the phenomenon of involution versus “lying flat”?

That is a great question. I was actually just speaking to someone building agent products for the domestic Chinese market, and we had an interesting discussion about exactly this. I asked him whether they were seeing much pushback from users around AI replacing people or taking away jobs, and he said not really, at least not for the products his team is building.

His explanation was that these products are still perceived as prosumer tools: things that help you work better, code faster, find information, or become more productive. Consumer AI products, meanwhile, are often still seen more as entertainment. So for now, the immediate reaction isn’t necessarily, “This is going to replace me.” It can be, “If this can make me better at my job, I need to learn how to use it before everybody else does.”

In terms of robots, industrialization, and automation in the laborious roles, these have been happening over the decade and are much less scrutinized; in fact, most young people don’t want those roles, and there is a shortage in many of these more blue-collar jobs, if you must put it that way.

Now directing attention back to what the product manager said about FOMO driving adoption. I thought that was interesting because it fits very naturally with ‘juan,’ or involution. There is a strong willingness to try new technology if people think it might give them an advantage, and that fear of being left behind itself becomes part of the adoption engine. In some ways, anxiety around AI can actually increase adoption rather than reduce it. If I worry that somebody who uses AI will be more productive than me, the rational response is to learn AI myself.

That obviously doesn’t mean people aren’t worried about jobs. There is already a much broader public conversation around employment, particularly given how difficult the job market has been for young people. So I think two emotions can coexist quite easily: people can be anxious about what AI ultimately does to employment while also feeling that the safest thing for them individually is to embrace it as quickly as possible.

One person I spoke with thought that, in the nearer term, some of the employment pressure could be absorbed through large employers, including SOEs, while another part of the adjustment comes from people upskilling and an entirely new generation of AI-native jobs being created. I don’t know how large either of those offsets will ultimately be, and I think we should be humble about that. Every major technological transition creates new categories of work, but there is no guarantee that they appear in the same places, at the same wages, or quickly enough for the people whose existing jobs are disrupted.

What I found especially interesting was his observation about involution versus lying flat. I don’t actually think these are two completely separate groups of Chinese people. They are better understood as two responses to the same underlying pressure: the feeling that things are a bit out of your own hands, a phenomenon you’re witnessing globally. At the core, what young people whether in China or in the West I think are reacting to uncertainty. It is about agency.

Involution is essentially staying in the competition: working harder, studying harder, getting another qualification, learning the newest technology, doing whatever you think is necessary simply to maintain your relative position. AI fits very naturally into that psychology. If everyone around you starts using AI, you almost have to use it too, even if the end result is that everybody becomes more productive and nobody feels meaningfully further ahead.

“Lying flat” is the other extreme end of the response: deciding that the marginal reward from continuing that competition is no longer worth the personal cost and opting out, or at least lowering your aspirations around career, housing or consumption. And these aren’t necessarily permanent identities. Someone can spend ten years juan-ing in Shanghai, become exhausted by the cost and competition, and then decide that a cheaper city and a less demanding life is actually preferable.

The person I spoke with divides his time between Shanghai and Hangzhou (of course, recognizing the most elitist perspective, it’s like being based in Silicon Valley), and his broader point was that in China’s most productive coastal cities, the incentives to continue competing remain extremely strong. Roughly, these couple hundred million people make up more than 90% of the GDP, probably. Clearly, an enormous share of China’s highest-value economic activity, technology companies, universities, and high-income jobs is concentrated in the eastern urban clusters. People inside that ecosystem are likely to keep juan-ing: in education, careers, entrepreneurship, and now AI skills.

At the same time, you are seeing some people consciously step away from that lifestyle, whether by moving to lower-cost cities or simply redefining what they want from work. To me, that is what makes involution and lying flat interesting: they come from many of the same pressures created by China’s extraordinary economic development. As people became wealthier, the benchmark for success also kept rising, and for some people, the competition for the next rung of the ladder became exhausting.

I would also separate AI software from robotics a little. My anecdote is primarily about agents and productivity software, where the user still feels like AI is augmenting them. The public reaction could look quite different when robotics or automation visibly replaces a job rather than helping somebody perform it better. We are still early in understanding where that line gets drawn.

On policy, I don’t think anyone has a clear answer yet for what happens if AI materially reduces demand for labor across a broad set of occupations. The Chinese government is clearly very focused on employment and on reducing some of the pressures around young people, education, and family costs, but managing an AI-driven labor transition is a different scale of problem.

I still have some optimism, though. Every major technological disruption has destroyed some forms of work while creating others, and AI should create categories of jobs that are difficult to imagine today. The part I am less certain about is the transition: how quickly those new opportunities appear, who is able to move into them, and whether they are sufficient to absorb the people whose existing work becomes less valuable.

So when I think about China, I wouldn’t frame the public attitude as simply optimistic or fearful about AI. The more interesting dynamic is that the same economic anxiety that creates fear of displacement may also make Chinese workers unusually motivated to adopt AI. In an involuted society, nobody wants to be the person who learns the new tool last.

Question 7: What is something non-consensus you believe and why?

One thing I think people underestimate is talent, but not simply in the sense that China produces a lot of engineers. I think people overlook how much China’s economic development over the past 20 or 30 years has changed the choices available to this generation of entrepreneurs.

When a Chinese researcher who studied or worked in the U.S. decides to return to China, people sometimes look for a complicated strategic explanation. But when you actually talk to people, the reasons can be extremely personal. Their parents are there. Their spouse may be there. They speak the language, understand the culture without having to translate themselves, and may simply feel more at home in Beijing or Shanghai than in San Francisco.

I wrote about this around Kimi because people kept asking why Yang Zhilin didn’t simply stay in the U.S. He studied at CMU and worked at Facebook AI Research and Google Brain, so obviously he could have had a very successful Silicon Valley career. But for someone of his generation, returning to China doesn’t necessarily mean sacrificing quality of life, intellectual ambition, or the chance to build something globally important. I explored more of his background and motivations in the profile of Moonshot’s founder Yang Zhilin.

This generation also grew up in a very different China from the one their parents knew. Many grew up relatively comfortable, attended very strong schools in China, then went to CMU, Stanford, MIT, or Berkeley, and worked at Google, Meta, or Microsoft. They are comfortable operating in both ecosystems and in both languages (of course, Chinese is still the mother tongue), and the assumption that the natural end state for every elite Chinese engineer is therefore to build a permanent life in Silicon Valley feels no longer a default.

There is a talent pipeline behind this that people outside China also don’t always see. I wrote recently about programs such as Tsinghua’s Yao Class and the elite high-school programs that identify unusually strong mathematics and computer-science (it’s actually so prolific, my cousin was part of it. He won awards in his province for programming competitions and received scholarships throughout his life) students quite early in China’s Genius Pipeline. These systems have existed for years; what is changing is that the people who passed through them are now old enough to lead research teams and start companies.

I think economic comfort has also changed the kinds of ambitions and feelings people are able to have around risk-taking. An earlier generation of Chinese founders had huge and fairly obvious commercial opportunities in front of them: e-commerce, logistics, fintech, the mobile internet. There were enormous industries that simply needed to be built. This generation has come of age after much of that wealth has already been created, which means some founders have the freedom to be more idealistic about what they want to spend their lives doing and feel less urgency in making quick cash.

Yang Zhilin is a good example. He talks about Moonshot in a way that can feel almost artistic, and has said that a great technology company needs “cultural depth,” not merely a useful product. Liang Wenfeng is somewhat similar in a different way: whatever DeepSeek’s eventual business model becomes, there is clearly an interest in fundamental research and in putting a surprising amount of that research back into the open ecosystem.

At the same time, not everybody is an idealist. A lot of the people I speak to are intensely pragmatic. They see an engineering problem they think they can solve, a huge market in front of them, or simply a very interesting company they want to build. I don’t really feel much nationalism in these conversations; sometimes Westerners definitely overestimate that angle.

Of course, everyone is aware of the geopolitical environment; it is impossible not to be. But I rarely come away from these conversations feeling that the personal motivation is “China needs to beat America.” Some people genuinely want to pursue AGI or solve a hard scientific problem. Others are very commercially minded. Some want to live near their family. Some just prefer living in Shanghai or Beijing. Usually, it is some mixture of all of the above.

That is the part I think gets lost when we discuss Chinese technology almost entirely through the frame of U.S. versus China. At the government and industry level, geopolitics clearly matters. But underneath all of that are individuals making fairly normal human decisions about where they want to live, what they find intellectually interesting, who they want to work with, and what sort of company they want to spend the next decade building.

My slightly non-consensus view is therefore that China’s economic development has changed the talent equation in a way that is difficult to capture in a chart of researchers or GPUs. A very talented Chinese researcher can now look at both countries and reasonably feel that either one offers the possibility of a comfortable life, intellectually serious peers, and the opportunity to build something world-class.

I think that matters more than people realize.

Hi all,

Christian Hu from Alibaba’s Qoder joins the podcast again to talk about the fierce domestic workplace agent competition. For context, Anthropic and OpenAI have decided not to allow access to their models and products in Greater China; thus, the competition for models and agents is largely driven by domestic players. For workplace agents — coworker-style products — the most popular ones at the moment are Tencent’s WorkBuddy, Alibaba’s Qoder, and ByteDance’s Trae. The other agents we see coming from the labs are mostly coding agents. With that, I’ll hand over the floor to Christian, who graciously found ~40 minutes while on a business trip to talk to us about the landscape and Qoder’s own strategic shifts.

This episode was recorded on the day Qwen3.8 Max launched.

For more, check out the podcast lineup here and explore the episodes!

Chapters

• 00:00 — Southeast Asia Expansion and Market Dynamics

• 02:43 — The Shift Towards Business Logic in Coding Agents

• 05:29 — Industry-Specific Applications and Vertical Agents

• 11:22 — Global Strategy: Lessons from Market Differences

• 14:00 — Partnerships as a Key Element in Market Strategy

• 16:41 — The Future of Coding Agents and Market Trends

• 19:22 — Pricing Models and Inference Economics

• 22:08 — Open Source Trends in AI and Market Competition

• 27:34 — Building vs. Buying AI Models

• 29:38 — Future Directions for Coding Agents

• 32:14 — Final Thoughts on Market Opportunities

Transcript

(AI-generated, for reference only)

Grace Shao (00:00)

Hi Christian, friend of the pod, you’re back on. Really excited to have you here. Where are you these days?

Christian (00:07)

Yes. I just landed in Singapore last night.

Grace Shao (00:12)

Okay, perfect. Cause we’re gonna talk about your Southeast Asia expansion. But since we last spoke, competition among coding agents has intensified quite a lot in China, particularly — the market has changed a little bit. Where are you seeing the whole market going, and how do you think Qoder fits into all of it?

Christian (00:32)

Yes, it is a war, you know, between every major tech company in China. Because it’s a war, nobody wants to lose. I think the logic behind this coding agent war is that most companies believe that a coding agent shall be the fundamental path to AGI — artificial general intelligence. So nobody wants to be behind in this race.

But the most interesting thing behind the war, or this race, is that something is changing. When we look back to the last twelve months, everyone is talking about the models, everyone talking about what kind of model will be the most competitive advantage for your coding agent. But now, for most leading agents, they are trying to be more focused on the business logic and workflows. Just like what Qoder is doing — we want to be more into the business logic of our customers. And even for some individual users, they are trying to do something for business. They have one-person companies or one-person workflows. So they need the coding agent to do more about the business workflow, not just code generation.

So that’s the shift behind the war. For at least most of the companies, they are trying to raise not just for a coding assistant, but want to be dominant as a desktop assistant for employees or maybe some individual users. And maybe in the future, they want to be the digital employees for the industry. I think that’s maybe the ultimate race for the coding agent.

Grace Shao (02:28)

And you’ve got players like WorkBuddy from Tencent that’s been doing really well, right? Exactly to your point—

Christian (02:33)

Yes.

Grace Shao (02:33)

I think they’ve been plugging in, kind of operating as an assistant on desktop, very intuitive and user-friendly. You’ve got ByteDance with Trae pushing in a similar direction. So how do you feel about that? Where is Qoder’s differentiating point or offering here?

Christian (02:51)

I have a few things to share with you. I think for the past year, Qoder has accomplished very remarkable business performance. In terms of ARR or revenue, we are the leading one — we are number one. And maybe we are more than the combination of number two to number four. That’s a huge achievement for Qoder in business growth.

And as I just said, Qoder from day one is more focused on building an agentic platform for developers and AI builders. It’s not just a coding tool or coding assistant — we want to be a platform for the next generation of agentic power. So the skill marketplace, the plugins, the connectors, the MCP connectors — all of these features are now shining. All of these features are bringing advantage to our business. That’s the truth of what’s happening in China and even in overseas markets.

Grace Shao (04:04)

That’s really interesting. And so when we caught up in Hangzhou recently, you were saying that Qoder appears to be broadening from just coding into industry-specific agents. That was something quite fascinating, because from the start of the conversation, you say coding capability is the fundamental foundation for all of these agents, but people are moving towards these vertical agent use cases. So tell us a bit more about that. What kind of industries are you guys targeting? What is your thinking behind this new strategy?

Christian (04:34)

Okay, I can share an example. We got a very important customer case with Xiaopeng. Xiaopeng is one of the leading electric vehicle producers in China, maybe one of the best. For Xiaopeng, for the company now, Qoder is not just a coding assistant or code generation tool. Qoder has become an agentic driver for the restructuring of their workflows and business logic. They’re trying to educate their developers and their engineers, maybe even the HR department, to use Qoder to renew their business logic and their workflows.

For example, for the legal department, they’re trying to use Qoder to reduce the legal review cycle from six days to one day, or even less than one day. And it’s not just Xiaopeng — I think there are a lot of similar cases. So for Qoder, it’s not a shift to pivoting to vertical applications or specific industries. From day one, Qoder wanted to be a platform. The platform means we want to be a tool — we don’t want to be just a coding assistant competing with other coding tools. We want to be more embeddable and compatible with the customers’ business logic and workflows. So it should be more vertical. But it’s not for us to do the vertical things — it should be let the developers, the customers, and even some individual power users develop domain-specific agents on top of Qoder. That’s the philosophy for Qoder.

Grace Shao (06:51)

Okay. So the thinking is you guys are offering the tool, but really it’s still on the user to build out the tailor-made agent for themselves. Not that you’re specifically pushing—

Christian (07:01)

Yes, that’s the truth. Actually, we got some organic creation of skills by the power users. They are creating some skills, some creating some plugins on our marketplace. We are going to open the marketplace to all the business users so that the users can deploy different kinds of skills or plugins from the marketplace. That will be the market for Qoder. Maybe in the near future, we could be a marketplace for agents, a marketplace for skills.

Grace Shao (07:49)

Interesting. Okay. Well, this leads to something else you talked about. You said that you guys were building an ecosystem around Qoder. So tell us a bit more about what that means, and how should we understand or expect how Qoder might change as a product in the next few months.

Christian (08:05)

Okay. I think we can change the view of Qoder — from a coding assistant to an agentic platform. I can give an example: we are doing something together with Microsoft. That may be an example. We want to collaborate with Microsoft to make Office more usable and more accessible on an agentic system just like Qoder.

You know, in the past, Microsoft Office is Microsoft Office, ChatGPT is ChatGPT — they’re quite different products. For the users, they have quite different experiences on two different kinds of products. And now we want to be one. For Qoder, for the agents, they need the agent to know how to use Office, how to deploy Office, how to use the best features from the Office suite — the toolkit — to help the users do presentations, do documents, process data. So that’s maybe the very interesting thing for the users in the near future.

This kind of collaboration is happening everywhere. Qoder is trying to partner with partners from collaboration software, finance software, even legal software, HR SaaS providers and vendors. We are doing things like that to be more compatible and more useful in the near future.

Grace Shao (09:55)

Very cool. Let’s take a step back. I want to talk about your business expansion, because I think when we talked about a year ago, you guys were really gung-ho about going to the US, going to Japan. I know you’ve been spending a lot of time in Japan yourself. But recently it seems like you are pivoting, or at least putting more priority and focus on Southeast Asia, and now you’re in Singapore yourself. Tell us a bit about the thinking behind your global strategy, and which markets you’re currently focusing on, given that you’re the head of GTM on the international expansion side.

Christian (10:26)

Yeah, you know, I actually got a lot of lessons from the last maybe twelve months — about ten months for my global journey with Qoder. Everything is different in different markets. I just came back from Paris. I think Europe is quite different from Japan. Most people think Japan and Europe share something in common — they are pretty slow in AI adoption, they care more about compliance, they care more about trust. But things are also — you can still find something different between Japan and Europe.

In Japan, the buying cycle is pretty long. Maybe six months or maybe even twelve months is very common. But in Europe, the cycle maybe is not too long — maybe one month or maybe one week. Because they are trying to catch up with the AI wave in Europe. But there’s still some obstacles in Europe. They care more about the law, the compliance, the GDPR, the AI Act. That’s the truth.

So for every AI marketer, or anyone doing good marketing in Europe, you need to care very much about the legal — the law, the act, the things changing and happening in Europe.

But for the US — I spent almost the first quarter this year in the United States. I met a lot of AI developers, AI startups, AI founders. The story is quite different from China, from Japan, from Europe. In the United States, speed is the most important thing. Speed means you are innovating. Speed means you are upgrading your product. Speed means you are accountable. You’re telling the users that you are accountable because you are innovating, because you are upgrading your product day by day, conversation by conversation. You need to upgrade your product.

Now I come back to Southeast Asia. Singapore is the first stop for Qoder to be a global product. And now I came back to Singapore, I will go to Vietnam and Thailand in the near future. For me, Southeast Asia may be the mixture for my go-to-market strategy, because in this place, in this region, you’ll find you have competition against some American AI vendors or AI producers. You will also find some Chinese competitors. That’s the mixture. This is a very competitive market, but it also has very high potential in this region.

Grace Shao (14:00)

Really fascinating, because you’ve done a world tour and given the high-level vibes of each area. And I can totally see that it’s also quite fascinating you say, like, ending up in Singapore. In some ways, it’s almost the most competitive for a sales role, because you know you have all the options in the world and nothing is actually off limits, and it’s really a price war as well. It’s quite fascinating compared to maybe other areas of the world that will be leaning towards certain companies or certain countries’ technology, given maybe geopolitical concerns or compliance reasons, regulatory reasons, whatnot. And other areas might be purely driven by price sensitivity. So anyway, fascinating. Thank you so much for that. But why then? Why are you guys now doubling down on Southeast Asia after your big global world tour?

Christian (14:51)

Yeah. Okay. I guess I forgot one of the most important things in our last conversation. Partnership is the key element and the key part of our go-to-market strategy. Because we believe that local partners, a local ecosystem, is the best way to get a connection with local community and local industries. Around the world, we have different kinds of partners. For the past twelve months, the most important thing I did was to find partners as many as possible. That’s the strategy for our global markets.

Okay, let’s go to Southeast Asia. Actually, I don’t think Southeast Asia is the most important one. Maybe I think for now it’s too soon to nominate which region will be the most important. But Southeast Asia should be a very important part, because Southeast Asia is a fast-growing market. It’s not too much affected by geopolitical concerns, and it’s fast-growing, so for every major AI company, there are huge opportunities, huge market potentials.

And for Qoder, we are growing very fast. We are evolving every day, every conversation. People in Southeast Asia — the developers really find that it’s very interesting and they can get much from Qoder’s evolution. Because for most users in Southeast Asia, to use closed-source agents from the United States or some other place is maybe too expensive. Maybe it costs too much for most users or most developers in Southeast Asia. But if they want to try Qoder, they maybe have the best position and best time to catch up with the evolution of AI and coding platforms. That’s a good starting point for most developers in Southeast Asia. And Qoder is evolving, so they will be evolving every day. I think that’s a good point for both Qoder and the developers and industry in this region.

Grace Shao (17:30)

Interesting. Let’s take a step back and look at just the overall industry at this point and the trends that are taking off. So the underlying coding models are improving quickly, and we’re seeing that it’s increasingly becoming commoditized, or at least the price is coming down, right? As all of that is happening in the background, how does that actually affect the tools that are built on top of these models, such as your product Qoder?

Christian (17:56)

Yeah. I believe the token, or maybe some of the large language models — I think the cost of token may be zero in the near future. Most of them. Maybe in twelve months.

Grace Shao (18:13)

Wait, how would it become zero though? What’s the thinking behind that?

Christian (18:17)

It should be. It should be. Maybe it’s science — it’s about science and engineering, but it should be zero cost. Looking back to the last twelve months, if you want to buy one million tokens, it may cost you about twenty dollars. But now it’s about half a dollar, just in twelve months. Maybe in the next twelve months, the cost may be quite near zero.

But I think in the future, there’ll still be some very expensive and exclusive models — some frontier models. There may still be expensive and exclusive for some users, for some industries. But most models will be very inexpensive, and even cost zero for most developers. And I think maybe ninety percent of the daily tasks can be accomplished by these kinds of models — the public models or maybe some cost-effective models. I think that’s the future.

Grace Shao (19:22)

Okay.

Christian (19:23)

So on this kind of zero cost, the agent layer — I mean the context layer, the agent layer, and the business layer — will be the most important one for the builders, for the users. So that’s actually what Qoder wants to do in the future.

Grace Shao (19:43)

I see what you mean. But coding agents can be very token-costly, right? So how are you then thinking about the inference economics, usage limits, or even your own pricing models when you are selling your products to users?

Christian (19:57)

Yes. I just said we want to charge our users — we won’t charge by token consumption. We will charge [differently].

Grace Shao (20:03)

Mm.

Christian (20:04)

We won’t charge our users by token consumption. We will charge.

Grace Shao (20:10)

Okay. So it’s not directly token consumption. If token costs come down, your cost of your product also comes down with it. Okay.

Christian (20:17)

Yes, actually, the shift is happening. We want most of the profit of Qoder to come from the agency layer — the workflow layer, the agentic layer, the context layer, the memory layer. That’s close to your daily work and operations, not just token consumption.

Grace Shao (20:42)

How do you view the subscription model? We still see that dominate coding agents today.

Christian (20:47)

I think for today, if you are just a coding assistant, you will be dominated by the large language model. That’s the truth. If you’re just a coding assistant, you should be dominated by the large language model. For now, the best model determines the best coding tool, because you’re too close to the large language model. If a lab has a very powerful logic model, the large language model can easily change to be a coding assistant. You just need an interface — you just need to type your language, you can be a coding assistant for every user. So if you are just a coding assistant or just a coding chatbot, you will be dominated by large language models.

But if you are an agent — an agentic workmate, an agentic layer that empowers users to regenerate their workflows, to connect with their collaboration layers, connect with their ecosystems — you will not be dominated by large language models.

Grace Shao (22:08)

I see. Okay. I want to pivot and kind of look at the model layer right now. I know you don’t work in the model layer, but you’re very familiar with the ecosystem. So help me understand. Chinese labs right now have actually used open source very effectively, right? And I think they’ve built up a global reputation and taken over a lot of the market share. And obviously, there’s a lot of discussion on whether they will continue to open source some of their best models, or maybe become more and more similar to US frontier labs, what we’re seeing with what they’re doing. What do you think of this trend? Will they remain open, or do you think they’ll gradually start closing up some of their best models?

Christian (22:51)

I think I can give you two angles. For the angle of the regulators — for the government — they want the open source strategy. They want the AI wave to be a new engine of growth for the national economy. That’s actually the national strategy for the regulators. So open source should be the trend, should be the future.

And from the angle of the market side — the market players. Actually, for most players in this industry, in the AI industry, open source is maybe the only strategy, or maybe the only path, for Chinese players to surpass their counterparts in the United States. I think maybe the only path.

Grace Shao (23:53)

Why is it the only path?

Christian (23:54)

For now, we can’t see the advantage from chips. We have no advantage about chips. We have no advantage about human resources, even some capital. We have no advantage. So maybe open source is the only path for Chinese companies to surpass United States companies, to be dominant in the industry — in the applications. We have some figures to prove that — Xiaomi’s language model and Zhipu’s language model are the most used around the world, right? They are the most used models, not a Claude, not an Anthropic model, not a Google model, in terms of usage. So actually, open source is maybe the only path, or maybe the only way, to compete against the counterparts in the United States.

Grace Shao (24:57)

Question on Qwen then, just because you guys just released another very large model — it’s not getting as much attention. I feel like the headlines are being taken all by K3 right now. But Qwen3.8 Max is claiming to be just as good as Claude. What’s your view on that? And it’s available on Qoder right now, right? So how do you view the model?

Christian (25:22)

I’m not a model expert. I can’t give you an expert view about which one is better, which one is not the best. For developers, for users, you can try Qwen3.8 Max — the preview — on Qoder. I think maybe positive. It should be positive. And we believe that, as I just mentioned, the model gap is closing. In the near future, the model gap is closing. So for Qwen3—

Grace Shao (25:53)

Actually, why is that? Why do you think model gaps are closing, closer and closer?

Christian (25:58)

That’s my belief. I’ll give you an analogy about this. I think the best analogy is education. You know, for education — university education — we’ve still got some top universities, right? The Ivy League, the best in the world. And maybe some of the best two or three universities in China, they are at the top. But most universities and most education in university, I think they’re universal. I don’t see too much difference in the education in most universities around the world. People can get educated, people can learn the skills, people can learn the knowledge in most — I mean ninety percent of universities around the world. That’s not the biggest difference.

There should be some frontier models, there should be some ecosystem models in the future. But mostly — more than ninety percent — will be the same. Almost the same.

Grace Shao (27:05)

So you’re saying basically it comes down to talent. Talent is strong. So therefore, more and more talent going into the industry, thus it’s catching up. But why wasn’t it catching up a couple months ago? Why was the gap, say, six to nine months prior to that — it was claiming to be nine to twelve months. All of a sudden now people are saying maybe it’s only one to two months. Where does this number come from? I’m always fascinated by these claims.

Christian (27:28)

I’m not a scientist, I’m not an engineer. I just have the philosophy, I have the belief.

Grace Shao (27:34)

Yeah. Okay. Well, let me ask you another question that’s kind of taking over the broader industry right now. The broader debate right now is about whether enterprises should be fine-tuning privately and deploying their own models, or even — you’re seeing more and more so-called tier two, tier three smaller internet companies in China leading the way in already building up their own models. Obviously, some people are saying they have the edge of having a very strong open source ecosystem in China, so the barrier to entry is easier. Anyway, from your point of view, do you believe this trend where companies are actually going to start building their own models or hosting their own models and fine-tuning on top of it? Does that make sense? Or should companies continue to do what they were doing, which is, you know, paying for the most frontier models from the frontier labs?

Christian (28:20)

Okay. I think for most companies, even some large enterprises, it doesn’t make sense for them to fine-tune their private models. I can’t see the sense, I can’t see the business logic behind this. Just like the analogy of education — I don’t think you need a private school for your own children. I don’t think most people need a private school for your own children, just for your two or three children.

Maybe ten percent of enterprises have very complicated, very sensitive applications, sensitive data. They need to deploy their own private models, fine-tune their own models. But I think most companies don’t need to do that. Instead, I think they should invest more in the context layer, just as I mentioned. They must invest more in the business logic, in the business, in the workflow, in employee management. That’s where they should invest. Because that will benefit their business growth, benefit their management, benefit their governance. I don’t think it’s necessary to invest in fine-tuning their own large language models.

Grace Shao (29:54)

Makes sense. Not every company should be working on the R&D of this, but really should be putting more energy into managing their own data, context, memory, expertise. Okay, well, I have a last question for you, which is, stepping back and looking at the whole industry right now, where do you expect the coding agent market to evolve over the next year? Because you alluded to a little bit throughout our conversation, but where do we see this whole industry going? Are we going to continue to see standard products that are going to be the main driver in coding agents, or like you said, model labs will swallow everyone’s lunch, and then it’s really all just putting up an interface on top of them? Are we going to see more enterprise-grade, specific use cases and systems built? How do you see this direction?

Christian (30:48)

Yes. I don’t think there will be too many coding agents in the near future. I think some of the coding agents, even some existing coding agents, will be swallowed by large language models. I believe that. Because language models can easily build a coding agent. So there should be some — maybe two, maybe three, or some number of coding agents still in the market. But most of them will be swallowed by large language models. And there should be some new agents — business agents, workflow agents, or some domain-specific agents in the future. We’ll have more domain-specific agents in the near future. Legal agents, human resource agents, even some agents for you — just like podcasters or bloggers, we’ll have some new agents for you.

Actually, agents will be just like digital employees. That’s the future. You will have more digital employees. You can ask your digital employees, just like you can ask your teammates, you can ask your friends — but it’s digital friends — to help you do anything you want them to do. That’s not just a coding agent, that’s your working agent.

Grace Shao (32:14)

Very interesting. All right, Christian, I have one last question actually for you, which is a question I ask everyone that comes on the podcast. You’re not unfamiliar with this. What is one non-consensus view you hold now? Or has anything changed since we last spoke?

Christian (32:28)

You mean from last time? What’s changed?

Grace Shao (32:30)

No, just any non-consensus view. Even what you said earlier was against what a lot of the industry is saying already, but do you have any view you think is very non-consensus or against consensus right now?

Christian (32:48)

Okay. I have something, maybe something shocking for many AI builders. I don’t see too much potential in the United States market. I mean for Chinese AI—

Grace Shao (33:06)

You mean for Chinese companies going global, whose first stop is often the US? Okay, very interesting. Go on.

Christian (33:09)

Yes. Actually, I want them to quit. You can invest in the US market, but I don’t think they can benefit, or get real margin and profit from the US market. But it’s a different story between consumer AI and enterprise AI. For consumer AI, there may be some possibilities for Chinese companies, because we still have some advantage — we have cheaper manufacturing, cheaper human resources. That’s the advantage for consumer AI companies to do business in the United States. But for enterprise AI, I think it’s quite difficult. Quite tricky and quite — it’s just not a friendly strategy to do business in the United States.

Grace Shao (34:11)

Where should they be going? Southeast Asia right now, Singapore, Japan, like you guys?

Christian (34:16)

Southeast Asia is much easier for them to do business, compared to the United States.

Grace Shao (34:24)

But the thinking from a lot of these founders is often that Southeast Asia, frankly, doesn’t have as much purchasing power either, or willingness to pay, especially on software.

Christian (34:32)

But they are growing. They are growing. We must invest in the future. And one thing I want to note — there’s much more potential in Europe.

Grace Shao (34:41)

Interesting. Okay. I think Europe has been a bit overlooked. People have kind of — some people have written it off a little bit, frankly, just given the recent years of slower innovation. How do you view the European market?

Christian (34:54)

I think Europeans are catching up. They are awake, they are working to catch up.

Grace Shao (34:59)

Mindfully.

Christian (35:01)

Yes, they’re working to catch up. That’s what I got from the Paris Summit, very intensively. And I believe — I think from the economy side, you must look at the market from the economy side, I mean the macro side. The economy in Europe — they are struggling. That’s my personal opinion. The economy in Europe, for many countries in Europe, they are struggling. So they need a new power, they need a new engine to restart the economy. I think AI should be the best engine for new growth in Europe.

Grace Shao (35:40)

Very interesting.

Christian (35:41)

So what kind of technology? Whose technology do they want? I don’t think they have the time to catch up just building their own AI labs, their own AI infrastructure.

Grace Shao (35:54)

Their whole stack, yeah.

Christian (35:55)

Yeah. They need help from China, even some other places. So I think that’s a huge potential for Chinese companies to do business in Europe.

Grace Shao (36:05)

Interesting. Okay. Very differentiated. Very cool. Okay. Thank you, Christian. Thank you for your time.

Christian (36:11)

Thank you.

Grace Shao (36:12)

Enjoy your trip and travels.

Christian (36:15)

Thank you so much. I enjoy it. Bye bye.

You know how Tony Stark sacrifices himself at the end of Avengers: Endgame, saves the universe, and seals his legacy? He is one of Marvel’s resident geniuses: eccentric, obsessively futuristic, forever building new technology because he believes he can save the world.

Okay, today’s protagonist is not Tony Stark, and I am not suggesting he is sacrificing himself for Tencent. But on a recent LatePost (晚点) podcast, the reporter who has covered Tencent for years described Yao Shunyu as surprisingly commercially minded and fluent in corporate politics for someone so young. He is apparently willing to offend when needed and put the target on his own back if that is what Tencent’s LLM effort requires. Maybe he is more noble and willing to sacrifice himself than others are giving him credit for. So, the Iron Man arc is not entirely fanciful.

And then there is his name. We first mentioned Yao Shunyu 姚顺雨 in January, through his comments on the bottlenecks to AI diffusion. He is not to be confused with Google DeepMind’s Shunyu Yao; Tencent’s Yao is the not-yet-30-year-old now leading its LLM effort.

And his name is frankly kind of badass. 姚顺雨 (which means last name Yao, smooth riding/ success, rain) is a homophone of 尧舜禹 — Yao, Shun, and Yu — three legendary rulers from early Chinese mythology (like Chinese Homer's Odyssey). Yao represents virtue and diligence, Shun filial piety and moral character, and Yu leadership and courage. I do not know who his parents are, but boy did they know how to pick a name.

The struggle

Okay, back to Tencent.

Gaming, number one. Cloud, respectable, top four. WeChat, hands down, is the most omnipresent, ubiquitous super-app ever made. Strong ecosystem, low-key management, little public scrutiny or scandals, and usually happy flying under the radar. Seriously, Tencent has an almost absurdly strong collection of businesses and is positioned to be a market darling.

But that is also why its AI position in 2024 and 2025 was more embarrassing, because investors and users had high expectations. ByteDance led consumer chatbot MAU with Doubao. Alibaba’s Qwen was driving Chinese open-source models to international recognition. Independent labs were fighting for SOTA. Tencent had no convincing consumer product, integrated DeepSeek into WeChat, and was kinda picked at. Hunyuan supposedly been around, but within the industry it was becoming associated with older researchers whom Chinese netizens dismissed as 老登 — laodeng: once derogatory northern slang for a sleazy middle-aged man, now shorthand for older men seen as conservative, complacent and hopelessly out of date. Soz kinda rough.

Hunyuan was close to an industry punchline. Tencent’s awkward Yuanbao standalone app asked people to download another app to use what was, charitably, a tier-three model at that time. Though it feels like it isn’t doing that great, it is technically one of the top three Chinese chatbots by active user numbers, but it trails far behind ByteDance’s Doubao and DeepSeek.

There was a period when even the internal product teams did not want to use HY, but that has changed recently. Hunyuan 3 Preview completely shifted that narrative. It began posting better numbers. Benchmarks improved, and the whispers changed from “it’s over” to “wait, what just happened?” I would not call it frontier leadership.

But HY3 has really shifted that sentiment. The strong early performance of Tencent’s latest Hunyuan model, Hy3, has also reinforced the company’s confidence in expanding WorkBuddy internationally. Within one week of launch, Hy3 ranked No. 1 globally on OpenRouter’s LLM usage leaderboard and recorded more than 68 times as many API calls as its predecessor, Hy2. This has all bolstered internal confidence, especially in Hy3’s adoption in the star workplace agent, Workbuddy (which we’ll go into further detail in Part 2). As stated on its own website, “Hunyuan 3 (Hunyuan 3.0) is Tencent Hunyuan’s most powerful open-source model. It is a 295B-parameter MoE architecture with 21B active parameters and 256K context window. Rebuilt from scratch in 90 days by the Hy team, built for real products like Yuanbao. “

And according to the independent evaluation platform Artificial Analysis, the latest HY3 is truly catching up, though not at the frontier, its cost per task is highly efficient. Propelling Tencent back into tier-one-ish.

It’s not just for optics or benchmark performance. It is said that Yao has since pushed the model team to work more closely with products such as Yuanbao and has reiterated the importance of rebuilding trust with internal stakeholders. It is deeply believed that a model in his eyes will only become strategically more useful when your own product teams are willing to bet on it.

The young hire expected to sink instead changed the current

Yao was born in 1997. In one sense, he arrived on a pedestal, recruited to symbolize a new generation of AI-native talent. He had all the gold stars on his resume: Princeton, OpenAI, sea turtle. At the same times, somepeople said expectations were also kinda low. He had spent only around a year at OpenAI, and Tencent is a fortress. Many assumed someone so young, seemingly green, perhaps academic, without the ‘sleeze"‘ or sophistication to navigate the bureaucracy and politics of a behemoth like Tencent. Some may have even low-key expected the expensive star hire not to deliver.

But leadership was firm on this necessary change, and at the time, he was their only hope. Tencent itself was unusually blunt about the generational change. In its earnings materials, it wrote: “HY: Completely rebuilt foundation model team with LLM-native researchers and engineers over the last 6 months; new team is young, energetic, and cohesive.” That is the company saying, with unusual directness: yes, we replaced the old guard.

During his PhD at Princeton, Yao co-authored ReAct and Tree of Thoughts, then worked on Operator and Deep Research at OpenAI.

Rumors have it that he was shopping around and spoke to quite a few labs in China before his return. Largely due to personal reasons, he just wanted to move back to China (likely for family reasons)

Ultimately, he joined Tencent in November 2025, and by July 2026, according to the reporting and internal descriptions, his remit covered core model R&D and had expanded from LLMs and AI infrastructure into multimodal models too.

As Chief AI Scientist in the CEO/President's Office, he reports directly to Martin Lau, but as the leader of the LLM team, he reports to Lu Shan, who is the head of TEG. As mentioned, TEG is the original R&D team. As he’s proven himself over the last year, he has been given more responsibilities. Beyond the LLM team and AI Infrastructure team, he’s now been appointed to oversee the multimodal team as well, which was previously under a different reporting line. So now he oversees the key ‘group-level’ models essentially (WeChat has its own model team; we’ll go into it in Part 2) and reports directly to Lu Shan.

Yao has Martin’s ear, figuratively and physically. His team is mainly in Shenzhen, but he is officially based in Hong Kong, with the corporate executives in its Admiralty office. It is not an unusual choice for top scientists and executives, though, for these big tech companies, given the tax advantages in Hong Kong.

Despite being the MVP, he is notably low-key. A Tencent source described him as 很低调 — very understated. He is no prima donna, nor does he have what Chinese people call 偶像包袱—the burden of being a celebrity. (em-dashes typed up by Grace)

A few personal anecdotes and tidbits do show that he really is chill. He plays basketball with researchers, a way to destress and be really part of the team. This was shared by the reporter on LatePost’s podcast. There is also a very Beijing-coded internal joke that people talk about. At one sharing session, Yao apparently opened by saying, “I’m not very familiar with Chaoyang. I’m usually in Haidian.”

For anyone who knows Beijing, it could be a low-key diss or just his genuineness coming out. Haidian is where academics and researchers congregate; it’s where the likes of Peking University, Tsinghua University, and the Beijing Institute of Technology are. It is also where the ancient ruins of Yuanmingyuan are, and where the emperor’s summer palace is.

Chaoyang, on the other hand, is where the MNCs, banks, SOEs, and anyone who’s close to money sit. (ps lowkey, there is also the financial district on the west side, but it’s literally just rows of Soviet-style buildings that house large SOEs and banks). The parks are modern and attached to strip malls. The stereotype is that if you say you live in Haidian, people assume you’re a scholar; if you say you live in Chaoyang, it usually means you’re in business.

Anyway, people around him describe him as down-to-earth and straightforward, but with high EQ, which really does tie it back to him outperforming in corporate politics more than many expected. His own team is consciously flat and meritocratic; according to some employees, and he even encourages competent interns to lead team visions.

The Chinese phrase I was given was that he changed the team’s 风气 — literally its “wind,” but really its ethos, energy, and invisible standard for how people behave. He reinvigorated the team, whether actively by sidelining Laodengs and hiring new blood, or by just bringing that hunger back. You can’t deny that he has helped restore the urgency and ambition of a model organization that had begun to look like a relic of Tencent’s previous era.

You can tell he is not really active on X (except for some recent recruiting tweets), does not work through PR, and feels strongly that the products and models should speak for themselves. Unlike many other star researchers, he does not want to become a public figure or become marketing itself. But his preppy looks and calm demeanor in his limited public sightings have led to some organic public interest (some for his capabilities, others for his delicate features).

But his only meaningful public-facing appearance as a Tencent employee was a June conversation with Dowson Tong. The head of CSIG, a man with some legacy in Tencent as well, an executive who joined Tencent from Oracle over a decade ago, oversees its cloud businesses.

Why he chose Tencent

As mentioned, Yao’s June conversation with Dowson is quite precious because it is the only time he has publicly explained why he chose Tencent. The following is translated from the supplied Chinese transcript; I have paraphrased for clarity and reserved quotation marks for language close to the original.

His first reason was that Tencent has “many good problems and many products.” As models improve, the question becomes where pre-training and post-training create real value. Products answer that. Agents also need an environment in which to act: without a food-delivery tool, an agent cannot order food; without access to the right systems, many tasks remain impossible regardless of how intelligent the model looks.

The deeper asset is context. Models are getting better at turning complex inputs into outputs, so the competitive barrier increasingly lies in possessing the original input: knowing what a person is doing or understanding the layers of information inside an enterprise. Tencent, with its products and ecosystems, has an obvious advantage.

But Yao said that was only the second-biggest reason. The most important was culture. His first impression from conversations with Dowson and Tencent’s executive office was that people were unusually honest about what the company did well and badly. They were direct and did not try to cover things up.

He described Tencent as a company that runs on “trust rather than metrics,” with a low-ego, solid, and long-term culture. All seem to align with his personal mission and values. His ambition for AI’s “second half” is not simply to ship one model or viral product, but to build a durable AGI-oriented organization in China around a balanced triangle: a solid technical foundation, products that create real value, and frontier exploration of new research paradigms. His view was that China still does too little of the third.

This complicates the easy story that Yao joined Tencent simply for its data and distribution, or other claims that it was because he was offered the most here. Context mattered. Products mattered. But in his own account, the deciding factor was that Tencent could admit where it was weak and give him room to build.

As we wrote, before Yao arrived, Tencent’s own product teams had already delivered a fairly brutal verdict on Hunyuan: they often did not want to use it, and product teams often preferred outside models. There is no point in collecting feedback if the institution’s first instinct is to protect the model team from embarrassment or to force products to use an inferior model simply because it was built in-house.

Yao, in many ways, appreciated that the management team could reckon with its oversight and face its challenges head-on. His June comments challenged whether a big lab can “be real.” Can the organization admit that the model is not good enough? Can it listen when product teams say so? Can it change course without spending months defending the original decision? And once the model improves, can it rebuild enough trust for those same teams to try it again?

Tencent had already taken plenty of detours, or you could say, deliberated a late start. Yao’s argument was not that detours could be eliminated. The first time an organization attempts something genuinely new, the path is almost guaranteed to be winding. The more important question is whether it can face the evidence, absorb feedback, and remain patient enough to improve.

That is also where his emphasis on products and context becomes more than a philosophical observation.

A model by itself can be intelligent and still be unable to do very much. As Yao put it, an agent cannot order food if it has no food-delivery tool. It cannot complete work inside an enterprise if it cannot reach the systems where that work happens. Intelligence matters, but so does the environment in which intelligence is allowed to act.

Then there is context. As models become better at converting complicated inputs into useful outputs, the scarce asset increasingly becomes the input itself: knowing what a person is trying to accomplish, what an enterprise has done before, where the relevant information sits, and how a task moves through the organization.

Tencent has plenty of those raw ingredients. It has communications, payments, cloud infrastructure, games, documents, browsers, enterprise software, and an enormous collection of consumer and business products. But having the ingredients is not the same thing as turning them into a meal. Historically, many of those businesses sat inside their own organizational hilltops, each guarding its own product, data, users, and priorities.

Turning context into a feedback loop

I spent 90 minutes devouring LatePost’s recent podcast on Yao. LatePost’s reporting argues that he played an important role in pushing Tencent all-in on agents.

Separately, one Tencent insider told me that during a management meeting in the second half of 2025, Pony Ma said the company should pay more attention to coding and agents, and that was clearly prioritized in Hy3. Yao’s proposal was to build a group-level reinforcement-learning infrastructure through which different business units could train models on a shared platform, while feeding the lessons generated by real products and real tasks back into Hunyuan.

“Reinforcement-learning infrastructure” can sound abstract, so let me try to put it more simply.

Instead of every Tencent business building its own isolated AI system, imagine giving them a common training ground. CodeBuddy learns from coding tasks. WorkBuddy learns from office work. Yuanbao learns from consumer interactions. Other business units bring their own specialized problems. Each product can still serve a different user, but the underlying system gives Tencent a way to turn those experiences into better post-training, evaluations, and model behavior.

The flywheel, at least in theory, is straightforward:

Better products generate more real tasks and feedback.
That feedback improves the models.
Better models make the products more useful.
More useful products generate more usage, context, and feedback.

This is why Yao’s decision to join Tencent and his push into agents are really the same story. He did not choose Tencent because it already had the best foundation model. Plainly, it did not. He chose it because it possessed an unusually large collection of products, tools, and contexts that could become strategically valuable if someone could make the pieces talk to one another.

“From that point on, throughout the entire company, you could hear discussions about agents and reinforcement learning everywhere,” the Tencent insider told me.

When OpenClaw went viral in March, Tencent’s business units came out in force with agents of their own. We wrote about the company embracing OpenClaw so aggressively that it offered free installation help outside its Shenzhen office in Lobsters Everywhere: Tencent Did It. We also spoke with Shuyu Zhang, QClaw’s product lead, about QClaw going global.

Almost every agent initially belonged to a different business. CodeBuddy came from CSIG for developers before its team built WorkBuddy for white-collar workers. MIORA is a creative product from the CodeBuddy team — think Claude Design. Marvis came from the PCG, and the team was the same team that had previously worked on building the Android app store for Tencent games and software.

From outside, it looked as though Tencent had suddenly discovered a coordinated agent strategy. Inside, it was much more characteristically Tencent: many teams experimenting from the bottom up, often without a clean central blueprint.

What changed was not that Yao suddenly owned every product. He did not. It was that Tencent was beginning to construct a shared technical and organizational layer beneath them. Yao may not control the entire chessboard, but he had gained the ability to influence how many of its pieces learned.

That is the potentially important shift. Tencent’s old horse races produced separate winners and losers. Yao’s feedback-loop proposal offered the possibility that even the losing experiments could teach the common model something.

A caveat, though: “caught up” is still doing a lot of work here.

Maybe the murkiness is the strategy

The cultural backdrop matters. A few years ago, people joked that although Alibaba and Tencent employees still worked 996, much of the hustle had become theatre and couldn't compare to ByteDance. Much of it was managing up to show you’ve done the work, rather than actually shipping new innovation. That, as harsh as it sounds, has some truth to it until the AI revamp.

More recently, younger talent and open mockery of the 老登 (pronounced ‘lao dengs’, derogatory term for older staff who keep holding on to how things were done or refusing to embrace change) generation have pushed teams to reinvent themselves. RED posts describe flatter titles inside TEG and CSIG: less hierarchy, faster decisions, and less fear of offending the wrong person. Again, you cannot just manage up and call it a day. For context, TEG is the Technology Engineering Group that is responsible for the company’s R&D and AI infra. CSIG stands for Cloud and Smart Industries Group, which is responsible for promoting the company’s cloud business and related internet strategy and solutions built on top of it.

The thing is, even reading through public filings and speaking to employees and investors, I still have so many questions about Tencent’s AI strategy, and it doesn't look like anyone can answer them all. Maybe James Mitchell can join the pod and walk us through it all ;p. But overall, the company is known to like a bit of that ‘unclear’ness, as in their eyes it fosters more organic creativity and innovation instead of top-down driven directions at peers like Alibaba or ByteDance.

Tencent has long been associated with 山头主义: every hilltop or business unit fights for itself. It is bottom-up, autonomous, and fond of horse races. Apart from Allen Zhang, who runs WeChat as a distinctly opinionated leader, business-group leaders generally do not impose the same degree of personal will. A team builds something. If it sticks, it earns resources. If it fails, too bad, next. Much of the products we see coming out of Tencent are all bottom-up. Remember QClaw? It was built up by a few product marketing people; see the interview below. An investor put it bluntly to me too: ' How much is it anyway? Tencent is rich enough for a few trials and errors.’

That is not always a weakness. In a new field, even Pony Ma may struggle to declare that Team A will definitely beat Team B and therefore Team B should die. Tencent gives several teams a budget and traffic, then compares the numbers and reactions from users. Five teams worked on Honor of Kings; Chengdu won, not Shenzhen headquarters. When Tencent Meeting was being developed, WeCom argued that its enterprise experience made it the obvious owner. Its product did not win either. WeChat, the most famous of them all, remains in Guangzhou, though only an hour away from HQ in Shenzhen.

But the lack of clarity can sometimes get further muddled in the whole lost-in-translation thing, as global investors are already frustrated by limited access to the executive vision of Chinese big tech. Yet they must keep a pulse on it, since launches can have global implications. As we try, we can only do our best to explain Tencent’s AI strategy here. Meanwhile, Chinese tech leaders remain wary of too much spotlight as geopolitical risks run deep. Meanwhile, American executives are everywhere, like literally podcasts, conferences, Twitter op-eds, you name it. So the imbalance of understanding is not only because of a lack of curiosity from the US side, but also a shortage of strong direct communication from Chinese tech firms

So to understand Chinese big tech, we sometimes need to take snippets from interviews, gossip through the vines, and read the org chart like tea leaves. And today’s post is really an attempt to string together various sources and analysis about Yao Shunyu and his vision = indirectly, the executives’ vision of how to build out AI within Tencent.

A not-yet-30-year-old returns from OpenAI, consolidates model R&D, pushes a group-level RL platform, and rebuilds trust between Hunyuan and Tencent’s products. Agents sprout across several business groups. WorkBuddy starts as a proof of concept on CodeBuddy’s infrastructure, gets approved by the executive office, is suddenly handed integrations and resources, and takes over the market share. WeChat, meanwhile, runs its own model on its own compute behind its own guardrails and is blessed and cursed at the same time to safeguard what makes it the top consumer product—its superior user experience.

Pony, Martin, James, if you’re reading this, this is the big nod to your genius. Let a hundred agents bloom and see which survives. In a market moving this quickly, a clean org chart can maybe become a liability.

Either way, what we can confidently say is that Yao Shunyu is a major reason Tencent is back in the AI conversation, and getting to know him, his story, and his vision is probably a lens into the future of Tencent’s AI strategy. Now, WorkBuddy might be evidence that something inside the company has changed in how it thinks about AI products. Is WorkBuddy here to save HK.700, or is Yao here to save it?

Stay tuned for part 2, our deep dive into the products, and a special zoom-in to WorkBuddy.

The most interesting distinction is between AI that tells a business owner what to do and AI that actually does the work. Ziwei explains how agents can research products, compare and vet suppliers, draft inquiries, follow up on missing answers, update Shopify listings, generate structured data and coordinate campaigns across existing tools. Alibaba’s advantage is not only its supplier network, but the category knowledge, transaction context and direct communication layer built around it.

We also discuss where human judgment remains essential. Accio Work can narrow a supplier list, negotiate across variables such as MOQ, lead time and materials, and prepare an order, but the user still approves purchases and typically takes over the final supplier relationship. That balance matters because commerce is not only a workflow problem: branding, product taste, trust and long-term supplier relationships remain difficult to automate.

Finally, we explore why Alibaba built Accio as a separate, more open product; how it works with third-party platforms rather than replacing them; its subscription and usage-based model; and the future of agentic commerce across B2B and B2C. Ziwei’s non-consensus view is a useful one: not every problem needs AI, and domain expertise becomes more valuable, not less, when powerful tools are widely available.

For more, check out the podcast lineup here and explore the episodes!

Chapters

00:00 Introducing Accio Work
01:01 From Alibaba.com to agentic commerce
05:18 What an “agentic business team” does
06:27 The commerce workflow and human control
11:11 “Do it for me” versus “tell me what to do”
18:32 Connecting fragmented commerce tools
22:36 Alibaba’s sourcing-data advantage
26:03 Supplier quality, matching and verification
31:58 Accio versus general-purpose AI tools
33:36 How agents communicate with suppliers
37:09 What Accio offers factories and suppliers
39:29 Designing AI for non-technical SMEs
42:42 Business model and monetization
43:54 The future of agentic commerce
48:57 Why not every problem needs AI

AI-generated transcript for reference only

Grace Shao (00:00)

Hi, Ziwei. Thank you so much for joining us today.

Ziwei Chen (00:02)

Hi Grace, so good to see you.

Grace Shao (00:07)

I’m excited to talk about Accio So I just came back from Hangzhou like a month ago and I met some with some of your colleagues on the ground. Was very impressed by the product and thought was just really intuitive. So let’s get started. Tell us a bit about yourself and your role at Accio and what Accio is all about.

Ziwei Chen (00:24)

Well, hi everyone. my name is Ziwei Chen. I work at Alibaba.com as the product marketing lead for Accio Work. prior to Alibaba.com, I actually spent a few years in the world of developer marketing where I was kind of driving good market plans for software tools that are meant for developers who are building AI products. So now it’s kind of completing the whole picture for me to now move on to the other side about actually growing the products. built by those developers to the end users. So that’s kind of my kind of AI journey coming from the development side now to the end user side, driving the growth for Accio Work among our small and medium-sized businesses around the world.

Grace Shao (01:01)

And tell us, what does Accio do actually? So I think a lot of people are not familiar with it and just how it fits into the whole bigger, I guess, Alibaba ecosystem as well.

Ziwei Chen (01:09)

Yes, absolutely. So maybe I’ll take a quick pause and take a step back just to talk about the overall broader picture. So Alibaba grew as a large enterprise. we were founded in 1999, and our kind of main North Star was to make make it easy to do business everywhere. So that has always been kind of our focus of focusing on B2B and focusing on small and medium-sized businesses who typically don’t have that resources that large enterprises have. So throughout the years since 1999, we have been really just focusing on what we can do, either there’s a new product. And new services to make it easier for them to start a business, to launch a business, and to grow a business, maybe something to exit a business. So along that line, Alibaba.com is kind of the bread and butter where we first started, focusing on B2B sourcing. So actually, I will kind of almost break down our development or our growth journey into three eras and then show you where Accio fit. I will call it the digitization era, the AI co-pilot era, and then the agentic team era. So the first digitization is when we are first founded with that platform of Alibaba.com where on this digital platform we are connecting the sellers or what we call the suppliers, the factories, with the buyers, our buyers, maybe sellers on the other side, into one platform. So now they’re not limited by time zone, they’re not limited by the location. And then in that process, we what we have learned over the years is that things can get overwhelming. There just so many suppliers, so many products, right? So then when all of these foundational AI capabilities came about, we were like, okay, this is the moment. This is the moment where we can introduce AI capabilities to make it easier for our buyers to find the right products and the right suppliers. So the name Accio actually comes from Latin, which means summon. So the idea is we help you summon the right products, summon the right suppliers, summon the right information. So that’s when we first launched Accio as a sourcing engine back in November of 2024. And that went really well because a lot of our users now, even without any experience in sourcing, without any experience in physical products, and find it really quickly with confidence. But then what we had learned over time is that First, sourcing is only one small part of a broader business journey, right, for a lot of our users in the US, Europe, and around the world. So now we’re thinking, okay, what else can we do to expand on that? So that led to kind of the last phase for Accio Work. So when I compare Accio Work with Accio, a few things that stood out. The first is the focus from sourcing only to sourcing plus. The other thing is about the agentic part. So earlier when I talked about the three phases of digitization as a platform, right? The website. And then the second era is called the AI co-pilot. So that means is that you are still on the driver’s seat, AI is in the passenger seat. It’s giving you advice, it’s doing some little stuff for you, but you are still making all the decisions. You’re still wearing all the hats. And that’s what Accio did back then. For example, it can find suppliers, it can recommend messages for you, but that’s kind of it. So now we’re moving into what we call the agentic team era. where actually we’re gonna get things done for you and get more types of work done for you. so that’s kind of where we are really sort of moving into this phase, where truly kind of it’s like the spirit of the agentic commerce world, where you’re not only using AI as a passenger seat, but you’re actually arranging a team of agents to get things done and maybe let the teams work among themselves, what we call A2A. So that’s kind of the overall context of where we have been starting from Alibaba.com. To Accio the sourcing engine and now to Accio Work as the Agentic business team.

Grace Shao (04:45)

That’s very comprehensive over you. I appreciate that. So just to help listeners understand, 1688 and Alibaba.com are the sourcing kind of platform within Alibaba. And then obviously everyone knows Alibaba for the Taobao and commerce side of things, but that’s actually a merchant-facing product versus the Taobao and T-mall that are consumer-facing. with that kind of context, okay, so if I had to ask you to describe it in a very short sentence or like say two sentences. How would you actually describe the product today? It’s just like, who is it for? What does it do?

Ziwei Chen (05:18)

Absolutely. I think I would Start with three words that summarize what it is we will call it the agentic business team. and I’ll break it down each but one of them that eventually lead to the who and the how. So the agentic part means these are agents that get things done for you. It doesn’t just give you the recommendations, but it can write emails for you, send messages for you, publish things for you, right? And then the second part is business, right? So we are dedicated to the world business, specifically e-commerce as a really strong emphasis. an all all Aspects of business starting from the front development side to the growth side to the operations, everything. And then team part kind of aligning up to that as well is that you can build multiple agents to work among each other. So tip from a format perspective, we have a desktop app that you can download, also a web app as well. You can access the same thing on your mobile devices, on your web as well. So all these things are connected where you are creating agents on this platform. telling the agents what to do and then you can go to sleep, you can go to work, you can go to your events, right? All of these things, go to your shop, right? Do all these things while the agents are in the background running all these tasks assigned by you.

Grace Shao (06:27)

So exactly your point, like you can do a lot of things, and that’s what I found the most fascinating thing. So, you know, we know a lot of products on the market these days that can do a lot of the, I guess, step three, step four in selling. So what I mean by that is they can help you manage this SEO, the content management, the marketing, and the like, you know, plugging into Mailchimp and sending out your emails. But your edge really isn’t just the agentic AI part, what I thought is really your supply chain network, right? so of course Across the entire workflow right now. Like walk us through like the different steps. Cause I remember there was like a slide your colleague showed me. It was like there’s four different phases of commerce building a business. you guys can go across it and how much of it is actually truly a genetically run right now and how much of it is still, I guess, needing direction from a human to really verify and process in the pro yeah.

Ziwei Chen (07:17)

Break it down to the two parts. I’ll talk a little bit about the overall flow first. I think it can kind of follow a typical business like life cycle where you’re starting from the strategy and the planning part. Let’s say what product, what’s the portfolio, what’s your branding, right? What’s your competitor, what’s your positioning or pricing, all of that. And then into the actual product development and sourcing part of that. And then you’re going into your store management, especially your e-commerce and different marketplaces. And then eventually the last one will be for marketing and growth. it can be for your B2C. Your traffic, right? Your conversion. Also, some of our users are not only doing B2C, but also B2B to wholesale. So, how do you deal with that B2B, like management or relationship, all of that? And then, so that’s across the different core stages of run launching and running a business. And then Accio Work supports all of that, but not individually, but it connects to that all together. So that entire ecosystem play, right? It’s really where it stands out, where we call the sourcing plus. So it’s not just about SEO, right? It’s not just about how to improve, let’s say, the images on your store, but how do you do that starting from the moment you have an idea, bring it to life, bring it to your store, and then drive it to more users, and then maybe iterate over time all the way back to what other products you can do, or how can you make your current product better. And along this way, I think when it comes to which areas that are truly agentic, or kind of where does that relationship between human and agent do? I think that there are a few ways to go about it. So I think for example, all of the critical decisions or area still remains a human to be to approve. For example, you can develop or create an order. The agent can create an order on behalf of you. So you can just click and then go into the checkout page, but it will not check out for you. So you have to go in and confirm that or from a security perspective. That’s one example. The other example is about your own vision as a founder or as a brand owner. So let’s say it’s the position of your brand. Some of them can be subjective. Maybe it’s your personal story, your personal style of the brand. It can also be kind of that. More objective or critical thinking that after learning about different perspectives, this is where we decide to do. I think running a business is it is a lot about the science part, but it’s also the personal, the entrepreneurship, right? So a lot of these kind of personal decisions still plays a role. So for example, maybe the agent can develop all different kinds of images and videos on behalf of you, but how do you d determine what’s Good, it can be through A/B testing, right? But it can also just be this is kind of the aesthetic you are going for that you think what matters the most to your audience. So that’s a second example. I think the last one, taking another pause, I know that when we talk about human, it’s not just about the users, it can also be about the developers, right? It could be the people in the Accio Work team. And then so where we are seeing is that we put we will love for Accio Work, we are specifically targeting small and medium-sized business operators who are not. The strongest in AI tools. They’re not developers by training, right? That’s not where they want to dedicate their time to. So what our product team and engineering team and algorithm team have been really focusing on is what are the things that we can do it for you. So take the burden off of you. For example, in this broader ecosystem of everything that Accio Work can do, it does require a lot of connect connections to different channels and platforms right can be about marketplaces having different tools i think that ecosystem is where it really plays a really great job but then who is creating those connection points you know you can have your account but then from our perspective where our team spend a lot of our time is to build those connections so you can do a one-click activation to connect axial word to your email to your CRM to your Shopify store so that is one of the strategic areas where we decided to take more of a burden onto ourselves so our audience can just click and then log in and then start to connect the dots for themselves.

Grace Shao (11:11)

Very cool. I want to double click on the partnership and like how you guys built the plugins definitely a bit later. I do want to ask about the use cases because so here my brand. So when I went to visit you guys, there was like really it was really interesting. So on one hand, it was very obvious on how SMEs would use this, like you said, like you know, if you already have a very strong intent on how to what you want to purchase, what you want to sell, so then that whole process is very clear. But one Example one of your colleagues showed me was so fascinating. Basically, like they were running an event, someone was running an event using Accio. They went on the Accio like webpage and then basically uploaded a picture of like the banner, the physical banner like size of like an outdoor space and said, I don’t know how large this space is actually. I don’t know what looks the best with this background, but I need a banner to run this event. And then Accio became like a thought partner. And it was really interesting because it was not like just a general like AI like thought partner. It was actually someone that had a lot of know-how and industry know-how from creating these things, from manufacturing. So they would literally say, it seems like blah, blah, blah, the weather is like this and that. You might need this material because that kind of texture is more durable in the sun or durable in the rain. And I just thought that was so fascinating. and I just want to hear from you what use cases are most common and then beyond that, what are some like I guess not so mainstream use cases that you’ve seen that are quite fun and interesting.

Ziwei Chen (12:34)

I think when it comes to use cases, you know, we can break we can slice the pie in a lot of different ways. One way is about the specific type of like actions for a business owner to do, but I’ll take it a different approach to talk about I think just in general when it comes to agentic tools in general, I think there are two types of use cases where air comes into play. The first is like do it for me. And then the other is tell me what to do. So the do-it for me part is that I know exactly what to do. I know the specific steps. Either I’m just busy or I’m just one person, there’s not enough hands. So this is where AI does a really great job as a follower, but it can really scale for you, right? It can be, for example, When it comes to assessing different suppliers. If I’m just up one person, I can only s assess so many, but now the tools can do it for you. So what we are seeing in similar cases, it can be using Accio to find, evaluate, vet, and compare suppliers or products. It can be you up creating and updating listings, especially for those sellers. with categories that have multiple SKUs, multiple kind of products where I want to make an update to all of my marketplaces, all different colors, all different sizes, right? It’s really doing taking that repetitive work off their shoulder, but making sure things are consistent across different channels. can be creating content, right, on either social media, it can be about blogs, all of that. That’s one area. The other area that kind of really aligns well with what you’re saying is that tell me more. Like I don’t know what I don’t know. Sometimes I don’t even know what’s the right question to ask, or maybe I do know the right question, but maybe the LM itself did not have that training to start with. And this is where we are seeing a really strong aha moment from our users. And when I say users, it’s not only just people who are new to e-commerce or new to physicals or like products. That’s definitely the area where we have seen the most appreciation. But it can also be for even successful sellers or operators entering into a new category. And it happens a lot, right? Maybe extending your product line, maybe you successful exit once one brand entering into a new one. They all they they truly understand. the cost or the costly mistakes and how important that could be. So in both cases, what we are seeing is open-ended questions on this is kind of what I’m thinking, this is what I need, or just this is my store. Tell me what’s wrong, tell me what’s missing. And then lear first learning from actual work based on what we have seen in the space about key areas, key specs, key steps or things to flag. And also, I’ll tell you what’s wrong, or I’ll tell you what you should pay attention to, and then let me fix that for you. And I’ll show you a few examples. The first example about banner for sure makes a lot of sense. we had one, a similar case, but with a different flavor to that. We had a almost like a con like a product consultant. So he would take out new projects all the time. Very common, even as someone who’s experienced in the field, to enter into new categories. And he was helping a client with the bronze. plaque outside the building where you know there’s a building name or the history and all of that. And then so he it was a retro case where he’s already spent the time and the cost to pay for a designer to kind of put his client’s need into this kind of sizing and whatever. so he took the time, spent the money, sent it to the factory, and then realized that actually the font size did not work for readability purposes. So while he actually replicated that need into axial work. One of the first things that actually were told him is that depending on the technique for the for the font, is it going in or out from that plaque? here are his recommended size or the height of each each of these tags to make sure that you are following the best practices. And then he was like, Wow, I wish I had that when he was working on that project. So he can save not only the cost, because he had to redo the thing, repay the designer to fix it, but also lost weeks in between. So there’s lots of examples around that. They’re small, but they can be costly, right? Especially you’re busy, it’s your own money, you’re bootstrapping into your business. So that’s one example. The other one, it can be we recently had a had a user who was just who has a Shopify store and was just trying to see like what what can I do? What can I do with Accio Award? So he’s not someone who knows exactly this is a tool I’m gonna use for X, Y, and Z. He just pasted his Shopify link. To Accio Work and tell me what’s the low-hanging fruit. And Accio Work actually dig out a bunch of like very specific areas for improvement, especially when it comes to SEOs and GEO. For example, for all his images on his store, he was lacking the image alt tags, which is a really important thing for SEO purposes. There were thousands of images. And he was like, he knew that was kind of important, but he never really paid attention. But Axe Work was like, if with with this, now you can improve your ranking. And then in about a few hours, Accio Work did all of these tagging for him in the back end, on Shopify. And then the other thing was that his product was baby Like baby beddings or a baby clothes. So there’s a size guide. But then what actually work called out is that his size guide was just an image. It does not have structured data. So LLMs cannot read it, which means that he’s not getting any traffic when people, when new parents are searching about sizing for their babies, right? If it were baby clothes. So then he actually told me that he he started this task right before a drive, and it was an hour drive on the road, and then actually worked basically using that and One hour, recreated his size guide to make sure there’s structured data behind it. And then so far, he’s already seeing like new traffic from like LLMs or like AI, like what we call now the AEO, right? for the first time. So these are all really great examples of what I call like teach me what to do, or teach me kind of what I’m missing out on, where we are seeing lots of like aha moment from our users as well.

Grace Shao (18:32)

That’s incredible. I think I’m like absolutely in shock because I think a couple of years ago during COVID, I was like playing around with the idea of drop shipping things. I wanted to build this like pet shop store. But just like not having that kind of guidance is actually really daunting to start something new when you’re not in this space. And if you just had that kind of thought partner, that would be incredibly helpful. Actually, I want to ask you about the partnership thing you mentioned earlier. You talk it did kind of touch on, you know, there’s these plugins that are built in. at SEO will help the vendors actually distribute across different distribution channels. So walk us through that. Does that not kind of cannibalize your own business? Or are you is that not like, you know, you guys are not in the same category, you don’t see that? Like, how does that relationship work?

Ziwei Chen (19:16)

A few thoughts here. I think first thing first just putting Putting down our Alibaba.com hat and just think from the user’s perspective. I think one of the biggest learnings that we have had by looking at our e-commerce sellers in this space is fragmentation. It’s truly, and when I say fragmentation, it’s not just about multi-marketplace. Cause you know, some people have like a Shopify store or Etsy store, that’s one type, but also think about you have your storefront, you have your B2B sourcing side, you’re speaking with manufacturers and different messaging channels, and then maybe you have your own like CRM. System, you have your own, let’s say, analytics system, right? And the all of each of these might have its own very professional SaaS solutions already, or maybe they’re hiring a consultant for that. So I think a lot of their time is really spread thin by just connecting these dots. So I do think that’s even from a North Star perspective, connecting a dot, putting them into an ecosystem is ultimately what’s gonna drive benefits or impact to our users, which is what we want, right? So I think that has always been we feel confident in that direction. And on the other hand, I think just like how there are so many different startups or even statute companies focus on each one of these, just like how Alibaba.com, our bread and butter started from sourcing. So I think yes, we can choose to do everything on our own or which is gonna spread thin. Odd, right? Or we can choose to collaborate with them as well. Especially a lot of times maybe our users are already so familiar with an existing tool. So making them give up on and then transfer is one option, or maybe in the transitioning phase, right? Or maybe just as a collaboration to say, bring your tools into Accio Work with just one click and then you’re login, and now you can connect all the dots together. We are definitely seeing that, for example, we have a user who has their own kind of email, like like solution system where he was struggling partially because of setting up campaigns and take lot of time but also you know he’s using Accio Work for research for messaging and now he has to translate all of these learnings right into a different tool. So he just integrated that into Accio Work and use Accio War as a central hub of say I have a new product now establish a new email campaign establish a new SMS campaign and then this is my key message you go To those sites on behalf of me. So I think that consistency of I have a decision, I want to apply it across different platforms, is where it’s gonna bring the benefit to our users, while without making them transfer or give them everything in their ecosystem that they have built out over time. And I think in that process, what we are also benefiting or learning through those partnerships with those channels are also just like learning opportunities on what’s what’s the future of AI when it comes to different kinds of areas. I think a lot of those traditional SaaS solutions still have are also experimenting. So I think that also creates a really win-win collaboration like experimentational type of like collaboration of let’s see how what kind of new agentic behavior that maybe actual work introduced through our our plugin to the new platforms and now they are learning what else they can improve or enhance or ex or or in integrate into their own ecosystem as well. But ultimately, because the our sellers or users are using those tools back and forth, it’s a shared type of traffic that has that synergy in general.

Grace Shao (22:34)

Very fascinating. So so it’s.

Grace Shao (22:36)

It’s really the user experience first, but like Alibaba second in that sense. So let’s talk about Alibaba then. so it’s no secret ATT sits on top of like Alibaba supply network, you know, guys got transaction history, sourcing know-how, all the data in the world from e-commerce over the last two, three decades. how do we understand that? Like is that like huge edge for you guys? do you think you guys have leveraged Alibaba’s ecosystem to build your things out? How do you see that relationship?

Ziwei Chen (23:02)

I think in general, I would say that having our own like our own supply chain and our own learnings and the industry know-house is definitely like the biggest part about our remote. or the starting point of our remote. I don’t think it’s the only thing, but I think that sourcing plus storyline is something that we definitely want to keep in mind to differentiate from the rest. obviously at the same time, we also understand that Alibaba.com is only one of the marketplaces that our users might be sourcing from. So we’re also trying to make it more open in the sense that it actually where you can also search for suppliers and like products even outside of Alibaba.com’s ecosystem as well. but I think how that looks like into how that translates into kind of our own product experiences. So far we have briefly touched upon that industry know-how because we know all of these Category best practices. So no matter what kind of categories essentially we have experience in, you’re gonna get that intro that recommendation or that guidance along the way. That can also include negotiation and supplier communication. So for example, because right now through Accio Work, we have our AI Auto Chat function, meaning that now you can train the agent to conduct or multiple agents to conduct that sourcing process. Task on behalf of you. So instead of you having to stay up late, depending on time zone, asking or answering questions, you have 20 agents, right? Each speaking with the supplier, negotiating. And then when I say negotiation, it’s not just about the bulk order price. It can be about the sample cost. It can be about the lead time. It can be about your MOQ. It can be about different material flexibility. So all of these kind of How do you approach negotiation? How do you approach establishing credibility as someone new in the space? How do you nurture that relationship with the supplier? All of these longer term things. So it’s not just like you find a supplier, you place an order and you’re done. So that the overall growth journey is something that we see continues as our users’ business grow or kind of evolve over time. That’s one thing. I think the other thing again is kind of that. synergy or consistency across platforms where you made a change to your product because it happens quite often where you are updating your products, maybe because of your learning and the competitive space, sometimes learning from Accio Work, or it can be your suppliers are recommending the latest techniques to the production, like you know, capabilities that you’re incorporating that. So how do you make sure your communication on the production end is reflected to your marketplace? is reflected to your social media, is reflected to your community, to your newsletter. So I think that sourcing plus, kind of that type of like experience or that consistency is also where we see we take pride in as kind of that holistic experience or evolution for our users as well. and then so that’s something that we’re definitely taking pride in, keeping keeping it as our mode while expanding on How do we translate that know-how into every aspect of our users or our operators business lifecycle?

Grace Shao (26:03)

Yeah, actually I’m gonna double click on that. So Alibaba itself has cited one point five million verified suppliers, you know, more than something like four hundred million SKUs behind Accio. these are wild numbers. this is not even including other suppliers, right? Like you just mentioned. How do you maintain that data quality and actually how do you ensure there’s fairness in terms of like how you actually provide your users the right supplier network or the supplier contact because I don’t understand how the back end ends. you know, when I’m looking for say a hair clip, what comes up? How do I understand what goes behind the interface?

Ziwei Chen (26:40)

I think I’ll answer it in two different layers. So I’ll separate Alibaba.com where our supply chain network as one layer and then talk about how where actual work fits in. but starting from the Alibaba.com as our own network, it has we have established our own very comprehensive system for that verification. for example, for all of our verified suppliers, that verified batch means it’s being verified by a third party where they have submitted their like certificates, all of the requirements and the registries and all of that. That’s one example. And the second example is that for example for all of those like Trade Assurance or all of these things that we’re constantly evolving to make sure all of these things are available for you to make sure you know you can choose and also including a lot of so that’s kind of on our own network side that we have our own system. And then where Accio Word comes in because it’s a it’s a different name, right? It’s not called Alibaba.com So it offers a few different flavors. The first thing first is that in Accio Work, there isn’t any sponsorship or ads in place. Well in Alibaba.com, yes, there are those different placements, but Accio Work purposely decided not to do that to make sure you are getting recommendations based on the best matches of your requirements. So let’s say depending on where you are, sometimes you might be in an exploration phase. Let’s say, I’m interested in in building what was the other thing I looked for the other day? It was like Wooden bits for like board games. And I was like, I don’t even know like where to start. And then Accio can tell you. So here are the five key areas you should consider. Let me walk you through how to make each of these decisions first. And then let me match it with the right supplier. So don’t rush, let’s take it slow. Or it might be, I don’t know exactly what’s the model, what’s my location, what’s my MOQ, what’s my price training, etc., and then match it for me. So you’re what you’re gonna see in Accio Work is let’s say you have a list of five things, and then for each of these suppliers, are they three out of five, are they four out of five? And what else might be a recommendation? And this can be hard requirements. Let’s say it must be this model. it must be they must have expert experience in certain places. It can also be about preferences. Let’s say maybe you already have lots of factories in a certain region, and then you just prefer that this new factory for this new component is in the same region. So it’s easier for consolidation for shipping. So you can be a preferences. And those are areas where a regular label, right, doesn’t or tag doesn’t cover as well compared to an AI for LLM to process that for you. So that’s one that’s one area. And then the other area is that so that’s the first thing that actually baked in is that no ads just based on what fits the best, based on your matches. The other thing obviously is kind of having that third triangulation of data. So we do have this ability for you to vet a supplier on Accio Work, which means that what it’s gonna do, it’s it’s not only gonna search for in the back end for alibawa.com, like what kind of certificates they have, but let me go on a custom public records. Let me go in, let me go on social media, double check, do they have a presence? do they have they attended trade shows? So all of these additional data you don’t get from the platform, actually was pulling in. Compare notes to make sure you are making a confident decision on that end. so that’s the second thing. And the third layer again is that even with all of this fixed data, or sometimes could be outdated, or you just want to verify whether triple or triple check it, and that’s where that automatic communication comes in. Where now let let the agent directly speak to the suppliers on behalf of you, ask for photos of certificates, right? As for images of the factory if they don’t have any, triple check on all of these areas. So all of these three layers are how Accio Work is building a different experience to make sure you are making a more confident decision.

Grace Shao (30:28)

Okay, so why then was it so necessary to build a separate agent on top of Alibaba.com and 1688 instead of you know just adding a chat interface or conversation interface on top of I believe a lot of the Alibaba products that right now all have an AI chat bot it embedded in their product interface, but why was it so necessary to build a separate Accio?

Ziwei Chen (30:50)

Yeah, good question. I think A few things. the first is that maybe back to our North Star when it comes to that open ecosystem. I think having a separate brand or even of an identity is a must-have to make sure our supply chain is not limited to Alibaba.com and also our use cases are not just limited to sourcing. So that explains everything that Accio Work is doing by incorporating new supply chain options, partnerships, incorporating new connectors or services or like solutions together. I think that’s kind of the biggest thing itself. And I think the other side of the things is that when it comes to branding, to some extent, although we do have lots of existing Alibaba.com users adopting Accio Work for that new value add as well. But we are, Accio Work is really reaching a brand new group of audiences around the world. A lot of them are like net new, aspiring entrepreneurs. Who now realize that opening your business on your own is possible. So I think there’s a brand new type of like audiences that we’re reaching, with its own kind of the AI flavor of like a branding as well. So we’re building a community that is also beyond Alibaba.com.

Grace Shao (31:58)

That’s interesting. I have a bit of a a spicy question for you. Then what’s the difference really if I were a solo entrepreneur using ATSIO versus maybe I just get Claude then? I get Cowork.

Ziwei Chen (32:08)

That’s a great idea. I think a few thoughts here. The first thing first, the simple answer is that Accio Work is the only agentic platform now that has the official connection to Alibaba.com. So all the other AI tools, maybe they can search for some of these suppliers by using the browser extension and all of that, but they’re not gonna get all the back-end data about those suppliers and not gonna be able to directly communicate with the agents, cannot chat directly with them and also cannot connect the dots all across the board. I think that’s a the simple, the simple way out. But I think also because all of our team’s energy is super focused on this field. And I think in this world where everything is nothing, right? So by us putting all of our resources and energy into just supporting e-commerce, small and medium-sized businesses, we are making more progress, more in-depth progresses on just that ecosystem specifically. And I think in that phase, the connectors with the plugins are just one of the areas. The other one is the just the general industry best practices. And what I’m referring to is not just kind of what the platform has learned, but also the people in the ecosystem. we’re building a community, we’re sharing skills, very like or even sometimes building different agents and now you can learn from each other as well. So I think it’s not just about The data is about a community, it’s about a best practices as well. So people are learning from each other. all of those solopreneurs are aspiring entrepreneurs of how to start something and then grow from there.

Grace Shao (33:36)

So then I wanna ask you about something you mentioned, I think, in passing a couple of times, which is like the agents can help you speak to the suppliers. It was quite fascinating. how does that work in the back end? Like our are at this point are just like agents talking to agents and then making deals happen and then you know, you have a product that’s purchased for you and the next thing you know it’s already out there on a website. What does the human need to do still? So tell us about that.

Ziwei Chen (34:01)

Yeah, so maybe let’s we can walk through a pretty classic source and experience. Let’s assume that you kind of already have a pretty good idea about what you’re looking for. So you’re gonna go into actually we’re gonna talk about let’s say maybe this is a category, this is my specs, the color, the material, etc. give me recommendations and then you can it will give you some top recommendations you can select and then actually were the agent will develop a pretty professional inquiry for you to make sure it’s speaking. The industry language to set you up for success as someone who’s an experienced series buyer. And then you can say, select these five, what’s like top five, and send this inquiry directly to them. So then the back end, it’s connected through the Alibaba.com. So the agents are sending those information. That’s the first thing. But what else that’s gonna turn on is what we call the AI Auto Chat, meaning for each chat conversation, the agent is actively engaging in conversation during that time. And then what we have seen As the most common workflow, which actually aligns very well even in the pre-AI era, is that people start broad, they shortlist, and when it comes to the final one or two, they actually go in for the full-on investigation, right? So that’s what we’re seeing for Accio Work for our AI Auto Chat as well, is that the agents does the best, brings the most value in that early short listing phase. So let’s say for example, I have a really highly customized product. It’s really hard for me to tell or some really unique needs. It’s not enough for me just to go on a website to evaluate that. And I have a list of five questions for every single supplier. In the past, I have to ask each of these five questions, maybe two five suppliers max, right? But then not all of them are gonna answer the question in the right order, or maybe some are missing something. So you are going you’re the you’re the Excel sheet, right? You go into this chat and say, okay, question number one, number Number three, number four is done, but not number two. The other one is only as another whatever, right? So the agent, what it’s gonna do is say, okay, you gave me a list of five questions. I will make sure every single supplier responds to each one of them. If they’re missing one, I’m gonna chase them and then I’m gonna save those answers into a table so you can make apples-to-apples comparison. And usually halfway, this supplier says they cannot accept this, that supplier says they cannot do that, or they’re out of order for something. You end up with one or two, and then Typically now I can jump in to say, okay, now I feel good about it. auto-chat. You can take a pause, let me take over the communication just to make sure the final collaborations are on are I feel good about that. and then I can place an order, etc. And in that process, I do want to highlight that obviously AI can automate a lot of things for you, but in many cases, just like in other industries, it’s a lot about relationship, right? Unless You’re paying for what we call like ready-to-ship, like standardized product. Let’s say I’m just buying some balloons for a party or buying some bracelets for an event. Otherwise, for highly customized product, you’re building a relationship. So at the end of the day, it is recommended for you to have some interaction with the suppliers because you might be working with them on your next iteration and new product as well.

Grace Shao (37:09)

Very interesting. So it’s really just helping improve the lead quality and then shortening the negotiation time in many sense. And then but the ultimate actually decision making is still like in the hands of the human. that makes a lot of sense. So I want to turn the conversation around. We talked a lot about how it’s helping businesses, right? Helping the vendor, the seller. but the other end of it is the suppliers. And I believe that at a lot of suppliers are actually on at SEO and also turning on auto chat and whatnot. So what basically support do you provide suppliers and help them in selling their products to the vendor in the between or like the brand owner.

Ziwei Chen (37:41)

Yeah, absolutely. so we do have a version or specific plugins that are meant for the factories and the suppliers who are who can also choose to be on Accio Work for all of their for their side of the sales as well. and I think even pre-Axial World, a lot of the more AI tech savvy factories are already developing their own chatbot for services as well. So I don’t think that’s new to some extent. But I what I think a few of the really key areas that we are seeing success, the first One is actually a little bit more about background research on potential buyers because a lot of them maybe they’re overseas, they don’t have all the all the awareness or the ability to search across the world about learning about who is this buyer, are they serious, what’s the scale, right? All of these things. So we’re seeing a lot of our factories or the sellers, our sellers on the platform using Accio Work to To understand their potential buyers. Because a lot of times, again, maybe in the maybe there are buyers who are not as experienced but are very serious. That wouldn’t come across in their messaging, especially maybe English and other first language either, right? We have users in Europe, in in Latin America as well, so and also in the US. So I think by giving more data to the suppliers to get a better picture about the potential buyers, it also helps them. capture opportunities or avoid losing or missing out on opportunity. I think that’s a really critical part for sure. And I think the other part is about that general selling optimization as well. It can be most of the times it will be on the Alibaba.com platform. How do you stand out as a seller by learning about the data, learning about the user behavior as well? So I think these are a few areas that we are seeing a lot of usage from the supplier side.

Grace Shao (39:27)

I see, I see. It makes sense.

Ziwei Chen (39:28)

Excuse me.

Grace Shao (39:29)

Something we talked about offline before recording this is, you know, that a lot of the users, whether they’re the s your sellers or, you know, the actual brand owners, is that AI is still relatively new, right? They’re not the people in Silicon Valley. they’re not working for big tech. And that’s quite different from a lot of the other genetic tools being sold or marketed to these more so called sophisticated AI users. Now, how are you helping them, I guess, whether it’s upskill or understand or not be so intimidated by AI?

Ziwei Chen (39:57)

Yeah, great question. I think a few thoughts here. maybe the different a few different perspectives when it comes to the product development side, on the go-to-market side, on the educational side, et cetera. I think to start from the product side, I think what we have been really emphasizing or de-emphasizing is the is a is a focus. It’s too much focus on the features themselves. So what we have learned is that when we’re packaging them, we’re not like It’s almost like sometimes our users they don’t need a toolbox, they just need like an almost built up tool that can run itself. So what our product team has been really focusing on is not just to build the building blocks, but build the framework of those building blocks. So we’re not making our users to learn to become a Lego expertise like expert right away. so I think from a product end, it’s more about how can we take off the burden of the configuration as as much as possible. That has always been an emphasis, but even more importantly, as we’re getting more newer users in the space. I think that’s one thing. The other thing when it comes to just overall marketing side, what we are also experimenting, exploring now is a lot more when it comes to live events, either it’s in person or online, where we’re taking things slow. And then in this process, what we are trying to emphasize is more the workflow, the use cases as opposed to the feature. Like let me show you how this thing works from what’s the input and what’s the output. And this is all you need to do is to copy and paste and whatever. And then I can explain to you what happens on the back end. Because ultimately what we call the job to be done, right? They’re not adopting a nail, right? They’re adopting a hole like on the wall, things like that. I think that kind of stays consistent. I think the very last thing in this process From an educational perspective, that it’s also learning from my perspective, is that a lot of times the users are not only learning about AI, but they’re learning about just how to launch and run a business. That’s almost even more important than a tool. The tool is an enabler to help you achieve the goals in business as well. So what we are trying to do is to resurface the business know-how. It can be sourcing, it can be about product design, it can be about how do you do SEO or like how do you do like B2B, like pipeline or lead gen? And then let me bake in those automation in the back end. But what you’re taking away is how to do this thing, even in the pre-AI era, but now just making it easier to do. So I think that has been more of an emphasis from a go-to market or like a community building perspective as well, to make sure we’re not overwhelming our audiences. because maybe they’re already overwhelmed by all the AI tools out there in the market as well. So how can we shorten that and let them get to the aha moment faster and easier?

Grace Shao (42:42)

So how do you plan to monetize that SEO then? Are you guys charging subscription? Are you guys gonna start plugging in advertisement?

Ziwei Chen (42:49)

Yeah, so Accio Work is on currently on a subscription plan model, both for personal subscription plan and also for business plan as well. and then all of these typically by a monthly plan or an annual plan. It’s primarily based on token usage. So depending on how much how much work. How many types of things was the frequency of the task that you anticipate on the platform and then you’re upgrading yourself based on the amount of work that token consumes. So typically that depends again on the volume, on the complexity, and also on different models you choose to activate for each of these tasks. And that kind of r it aligns with our vision since the beginning, at least starting from Accio. the sourcing engine to now this the agented platform as well. So it’s mostly focused on usage. we do not currently have a plan again because of that kind of fairness or the quality of results, we’re not dealing with like advertisement from suppliers. most other things is more about how do we empower our our like sellers, like or buyers on our platform, the sellers to the consumers to get their business running using the agentic capability.

Grace Shao (43:54)

Now throwing it forward, where do you see agentic commerce going? Because it’s really interesting. I just interviewed agentic payment solution company last week as well. And then now obviously talking to you guys about agentic supplying, supply network support. where does agentic commerce kind of take us? Like in the future, are we just gonna be like telling the agents to do this whole transaction, this whole activity loop? Or Do you think that human is still needed for a lot of the verification and and and quality control in between?

Ziwei Chen (44:22)

Yeah, I think a few thoughts here. I do think that human in the loop is critical at different stages of a development. So for example, there are when we talk about agentic or agentic automation, right? We’re we’re automating an existing workflow. Yes, there are a lot of best practices, but it really varies, right? So even having your you’re co-developing an automation workflow with the AI tools in the beginning, and that will vary all the time. The more you The more you invest in that co-creation with the agents in the beginning, the better the automation or the workflow will work well for you. And then along the way, what we’re also seeing from our uses is that they’re learning while doing. You’re building a plane while you’re flying it. So there’s always gonna be learnings along the way. Just like in coding, there are ways for you to build the agents to do that automation by themselves. But I think the more complex the considerations are, the better. Better impact a human can bring to that loop as well. So that’s in the middle part, right? And also in the end, I think there’s still we’re still building that trust along the way, like between the human and the agent or the agentic loop as well. So I do think that having the option to or for an agency, for the human to have that agency, is always critical. That I think it’s it needs to stay. but then on the other hand, what I also want to add is that I think the word of agentic commerce. Can vary quite a bit between B2B and B2C. for example, in the B2C world, there’s always that shopping experience. It’s not gonna get taken away, right? Just sh going to it’s browsing that emotional experience is not gonna be replaced by AI. While the B2B world is a little bit more calculated, is more, right? But still it has that you have that relationship part. So I think in both end, I think there’s still human touch. In addition to human decision, needs to go in. But where I’m the most excited about is that because our target audience or our target user sits literally in between. They’re in this loop for B2B agentic commerce, and also they’re in this loop of B2C agentic commerce. So what I’m excited to see is how these two worlds are merging, or our our core audiences sit in this overlap in between. And then I’m I’m excited to see more of that. synergy of what works in the B2C authentic commerce world can get better integrated with the B2B world as well. Because again, like our sellers or our buyers are sitting between the factories, right? And them as a brand owner and then the consumers as well. So I think this is where I think the next iteration of innovation or maybe redefined like r the new definition of workflows of tools that might happen as well.

Grace Shao (47:04)

That’s super fascinating. Cause I think when I’ve been speaking.

Ziwei Chen (47:05)

Yeah, but.

Grace Shao (47:06)

To people who are working in the agentic commerce space, it’s often so very compartmentized, like what you said, like you there’s a B2C world but there’s B2B world, and these two worlds don’t really, you know, co like collide at all because whoever’s talking to a factory is not really telling what’s happening to the consumer and there’s no really feedback. Factories don’t even know where the product’s going and how it’s being branded. But now there’s that kind of ecosystem, I g like communication or the world colliding between on Accio. It’s very interesting. what do you think the world’s still getting wrong about agentic commerce? Because I think there’s still some people who are very reluctant or resistant towards agentic commerce or the idea of it. What do you think is people are having misunderstanding about it?

Ziwei Chen (47:47)

I think one thing that we briefly already touched upon is whether a gender commerce is taking away the agency of human. so I think that’s the part that causes the most concern or hesitance is like is the agent gonna replace me? But I think all a lot of these judgment calls, even that branding, right? that personal story about you being as an operator, you having that guidance on the branding on the story and all of that, it’s not gonna go away. But also it’s not gonna go away, but it’s also critical for you to inform where the agent should go. I think that’s one thing. But the other thing that I feel like maybe where I think it’s also like A of times when just as you said, when people think about commerce, they’re th thinking it still tend to go more siloed or go into just on specific steps. So I think it’s still that orchestration layer when things start to get connected and st start to flow together. So it’s not just about how well each of a step is done well, but how well all of these steps are connected. And then that’s also where that human in the loop or your agency is bringing that flow together. So I think that’s an overall area where I think it’s it’s People might be evaluating the wrong thing as opposed to their making the wrong judgment call about a one specific thing.

Grace Shao (48:57)

Thank you. And then my last question for you is a question I ask every single guest that comes on, as the podcast name is called Differing Understanding. what is one differentiated view you hold or something you think that’s very non-consensus?

Ziwei Chen (49:08)

Yeah, so I think my the first thing that really stood out to me is that not every problem for not not every problem needs AI to be solved, or AI is not a solution for everything. and I think that has implications both on the development side and also on the user side. So on development side, what we mean is that there are solutions or there are things that we can build that is more efficient or more accurate even without AI. So we should not shy away from that. And that should that will impact how people build products. But on the other hand, even for the users, that also means that skills in AI is important, but that’s not enough, or that’s not the only thing that I will focus on in this era AI era phase. Instead, I will still spend time in building your domain knowledge as well, because without that. You’re not gonna be able to learn how to use AI the best because maybe sometimes AI cannot solve all the problems as well. So I think be having a balanced view about where AI sits in the product or in your tool stack is something that will allow you to get the best out of all the AI tools available and drive results for yourself.

Grace Shao (50:21)

I love that it’s like keeping the human side of things in check. But I think it’s only non-consensus of where you sit because sit in the Bay Area. Only people in the Bay Area believe AI is taking over the world right now.

Ziwei Chen (50:31)

Okay.

Grace Shao (50:32)

But thank you so much for your time, Switzue. really appreciated your insights, and you’re very you know, deep understanding of the agentic commerce world.

Ziwei Chen (50:39)

Yeah, thank you. I had fun. Appreciate it.

The conversation started with a basic but important question: if Stripe, PayPal, Airwallex, Visa and Mastercard already exist, why do we need another payments company? Patrick’s answer is that the problem is not just processing a payment. For agentic commerce to work, the system needs to understand who the agent represents, what permission it has, whether the purchase fits the user’s original intent, and whether that authorization can still be verified later. Clink is positioning itself less as a replacement for payment processors and more as a connector across agents, merchants, payment providers, and card networks.

We also spent a lot of time on why payments remain so fragmented globally. Payment habits are local: cards, wallets, bank transfers, convenience-store payments in Japan, and different payment methods across emerging markets. Patrick argues that no single provider is best in every market, which is why AI builders and global merchants may need orchestration across multiple PSPs, local payment methods, and eventually multiple agent platforms.

The most interesting part of the conversation was around trust and liability. If an agent buys the wrong thing, who is responsible? Patrick’s view is that fully autonomous shopping is possible eventually, but the trust layer is not there yet. In the near term, agentic payments will likely work through delegated mandates: user-defined instructions, spending limits, merchant controls, passkeys, audit trails, and chargeback mechanisms. His differentiated view is that the industry may be too focused on creating brand new crypto or blockchain rails, when the more immediate challenge is making existing fiat rails work safely for agents — and getting the ecosystem to align around clearer standards.

Get in touch with Patrick via LinkedIn.

And company website here.

Share

Chapters

00:00 Introduction

00:13 What Clink does and why payments need to evolve for AI agents

01:18 Why today’s payment stack is not ready for agentic commerce

03:07 Why payments remain fragmented across geographies, regulations and user habits

05:48 Clink’s role as a connector across agents, merchants, processors and networks

07:06 Why AI builders may need more than Stripe, Airwallex or PayPal

09:34 Visa partnership and what it means to be an “agent enabler”

15:23 What happens when agents spend a user’s money

21:02 Trust, liability, chargebacks and why China’s wallet model is different

23:26 What agentic commerce actually means today

43:54 Patrick’s differentiated view on fiat rails, stablecoins and protocol fragmentation

Transcript (AI-generated, for reference only)

Grace Shao (00:00)

Hey Patrick, so good to have you on today on AI Proem. Thanks so much for joining.

Patrick (00:05)

Hey Grace, thank you for having me here.

Grace Shao (00:06)

Yeah. To start with, give us the high level. Who are you and what is Clink in a few sentences?

Patrick (00:13)

Sure. so at simplest level, I’ll introduce Clink first. Clink is building the the payment and the bill infrastructure for AI builders and for the agent that they will serve. So payments today we see at design around human being on a checkout page and but we’re building the layer that let authorized agents complete the same transaction safely while helping merchants accept the global payments through the system they already use.

And for myself, I started my payment career path at Amazon and AWS. I learned how to build a payment infrastructure that has to perform at a very high level of reliability and the security. And at Temu, I led a team to solve the problem with payments and the global market expansion. So we expand country by country, market by market, and solving the payment issue locally.

And Clink bring those lessons to a new problem that when we see agent becoming the new commercial actor, but we see the payment stack is not fully ready for that, so we won’t solve that problem.

Grace Shao (01:18)

So what do you mean by the payment stack is not ready for that?

what do you mean by the world’s not ready for agentic payments? And where do you see the gap from existing offerings? Because obviously, when we talked offline, I did challenge you on this. I said, you know, there’s a lot of major fintech players already and payment solutions, Stripe Airwallex PayPal,

right? But where does Clink come in and how does Clink solve this issue,

Patrick (01:44)

Yeah, I think payments are largely solved for human who is present at checkout. so the industry had made it much easier for developer to accept payments like except from card, from our wallet and from like so many payment that local customers prefer. and also the developer can easily launch a product to launch a subscription, send a payment link, right? Those those are like a real progress already made the word.

So but what is not solved on the payment side we see is agents acting under the delegated authority. Right. Processing the charges only one part of that transaction, but the system also needs to know who the agent represents and the what permission that it has and whether the purchase is within the policy, within the intent, like matches the original mandate that the user has given. And then the how the evidence being being persist, right? And if we come back in six months later.

is that transaction still like able to be verified at the point that the agent made the decision. So I think like global merchants still face another unsolved problems like a single provider isn’t no single provider that best at the every market and solve all that agent issue. And then we know worldwide that the agent can can buy product from anywhere. So we see that’s a big gap that the agent e commerce need to

have a both the authentication layer authorization layer and also a practical way to across with all the providers in the market today.

Grace Shao (03:07)

Yeah, actually on that, you did say, you know, payments are very fragmented, right? Like, and but is that fragmentation mostly based on geo geography, regulatory reasons? Is it about payment habits? Like how do we understand that? Why is it so fragmented and it’s not a one solution fits all?

Patrick (03:24)

Yeah, I think that’s a great question and a great observation, right? And to be honest, I think it’s all about like all about that because payment really reflects a local financial system, right? Like you spend twenty dollars in Singapore that you may use a card, you may use your wallet, your bank transfer. But if you know in Japan that many e-commerce orders actually being paid over konbini payment. Konbini is the Japanese word for convention store. If you would like to place an order online and

getting to a company store and the pay when they check out some some goodies. So it’s just a habits and also the regulation and also how the payments being developed in that particular marketplace. So we see that it’s like all different across the world. That that’s a lesson we learned from from building Tammu payment stack that we actually adopted 70 payment method when in the first year. So every country has their own top three payment method.

And every time you add a new pay method, you got some new customers. So it’s kind of very interesting. And then but then the fragmentation is not simply like I I guess you can solve that by building an API and building more like a convenient experience, but we see that as a reflection of local, right? Like habits and migration, all that just as we said. And the the things that we cannot force the users, right? Use got their own votes for what they want to pay.

And that will always exist because that’s how the world works here.

Grace Shao (04:48)

And so thus like where do you fit in then? Like do you have a certain geo gro geography you’re targeting or a certain kind of I guess use case that you’re targeting?

Patrick (04:58)

you mean the agent e commerce or like in general? I think Clink today we try to solve two problems. Yeah. Yeah, we’re trying to solve two problems. Yeah. So one one problem is for today, like the global monetization for the digital servers and AI builders, right? They they’re selling their product, their great application to worldwide, and then the users from worldwide that like to pay them with the local payment.

Grace Shao (05:02)

Just in general because Mm. Yeah, yeah. Sorry, go on.

Patrick (05:23)

So that’s the monetization issue we we try to solve. And also also we’re trying to solve a forward looking issue that when we see the trend that the the actor, the consumer is moving from human to agents, and how agents can leverage the user’s assets, users’ local pay methods, right, and to make that payment, make that purchase, and to make all the information flow and the funds flow fat like try to flow fluently as it is today.

Grace Shao (05:48)

I see, I see. Let’s double click on the business. Help me understand your business a bit more. So one thing that really stood out to me from our earlier conversations that Clink acts more like a connector. how do you describe your role in the whole like payment stat?

Patrick (06:01)

sure, yeah. So I would describe Clink as a connector, as you said, right? And a controller or an aux trader. So we do not want to replace the bank, the car network, a processor, all the commerce platforms, the merchants or other agents, right? So we translate the agent’s intent into a transaction, or it could be the human’s intent and delegate to the agent, right, into a transaction. And the merchant existing system can accept.

apply the rules that the user has set earlier and route the payment as the user preferred payment and also preserve the record for the future validation or whatever verification that comes from. So that will let merchant connect once rather than building a customer payment pass that for every agent, for every processor. So that’s the important part that we won’t be that connector anywhere alive.

And it carries contacts, permissions, all that and solving the issue to connect all different kind of agents, different kinda agent platforms, and also the different kind of mer merchants and all the PSPs and different kind of networks as well.

Grace Shao (07:06)

I see. So I’m I’m gonna challenge you again on this way. I brought it up already earlier, but why would clients then choose Clink instead of an established platform, given that it’s so important for trust, reliability, proven processes in in payment, right? Why would they choose you over another major platform?

Patrick (07:24)

Yeah, we we mentioned some names earlier like Stripe, Airwx, PayPal, right? And there’s so many of them. Actually, I think probably thousands of processors all over the world. So I think sometimes when when you just want to sell one product in a particular market, then a single PSP might satisfy what you want. but I think the trend here today is all the AI builders, we are all selling products. what do I?

So that means like if you do want to of course trip and water can cover majority of the market, but we see in emerging markets and all different places, like different processes have different advantages, right? And if you as a merchant want to maximize that, then you may have to connect with multiple PSPs. Also pricing is another concern that if you want to lower your cost on payment acceptance, then you may have to have multiple PSPs to lower your cost.

So Clink will be like helpful as a connector that we saw the problem cross bond trees that you have multiple markets by default and you have all different kind of local payment methods, especially for example, we have some solutions in India, in Latin America. So like we are beyond one processor, right? So that’s on the payment acceptance side. But also we see that the AI application today requires, well, not that complicated, but does require some sort of bidding.

Right, top ups and subscriptions. And if you want to build from scratch, that’s something extra. And if you stay with existing PSPs, that’s a solution. But the expansion may not be that flexible. So we see that clinks serve as a connector layer, but that’s for the human business, the co like the product side of monetization. But when agents come in, that the problem becomes bigger, right? Because the you you don’t know where the agent platforms from.

It could be Kodas, could be cloud, it could be Gemini or any vertical agent platform that you never heard of. It could be own, like Ermers or Open Cloud. Right. And on the other side, the merchant may, as we we just said, the merchant may have only a single processor, but they may have many. But for that particular issue, then if there are many, who’s doing the routing? And who is connecting the agent to a particular merchant and their processor?

Right, and then we see that a single processor may not be capable of solving all that problem at once.

Grace Shao (09:34)

I see, I see. so tell us like how you’re working with existing platforms too, because this is another interesting point when we talked offline, you were telling me you have actually a lot of partnerships with traditional payment platforms like such as Visa, whatnot. how do you fit into this stack here?

Patrick (09:50)

you mentioned Visa, right? So I wanna bring an interesting fact that Visa described Clink’s role as an agent enabler. So I think that captures very well. So like there’s so many agents out there today, right? The big names we just mentioned, and also lots of self deployed or vertical agent platforms, right? And then we see that all those agents have the need the the requirement or the demand to make payments.

to do transactional for the user that trusts them. So but however that agents may not be able to integrate with our like Visa system. Right? Visa requires you to have a PCI because agent may not be able to touch the credential information. And if for self-deployed then like we don’t know like what is being uploaded and all that, all the security reasons and privacy concerns are there. So as the agent enabler

We process all that credentials, all that sensitive information. So we connect with the visa for all the like pass key verification, the mandate creation, intent verification, all that. So we perform that on behalf on the agents. And also so that that means we turn on or we enable the existing merchants to be able to connect with visa’s network securely or in a certified way. So that’s how we play with it.

existing I mean like schemes like a visa building a network also similar for MASCAR. And you mentioned some other big commerce platforms like Shopify and so I think Shopify is highly complementary for to to to zero, right, for sure. And there it’s already giving the merchants a catalog, a card, a great system. And they’re like a great pro it provides great solution for the merchant to easily

set up your own site and start selling goods. And we’re not going to rebuild all that. It’s it’s just simply not feasible and we see that the ecosystem is super strong. And we want to be add on to the ecosystem. Right? We see that added to like being a plugin for example, then we will turn on Shopify’s merchant to be able to like agent ready. Right. We build that agent readiness into the merchants. doesn’t matter it’s Shopify or WooCommerce or Mikado or so so many like open source

building platforms and commerce systems. We see all of them are being like like have their own adoption in different marketplaces and then the role of Clink is not going to replace all like any of them. And we want to be like add on to them. We solve the problem that the merchants at this point may not be aging ready and we want to make them aging ready. So that comes back to the earlier question that the role of Clink as a connector

Grace Shao (12:23)

Really interesting because basically, actually, you’re not trying to take the payment companies pie, you’re not trying to take the merchants business either. Then who are you charging in between how are you making money? Are you taking a cut from say visa and the transaction fee, or are you taking are are you making the merchants pay you for your service?

Patrick (12:42)

in in general the goal will be having the merchant pay us. But we are seeing different like from our own business models, right? Different monetization paths that if the merchant in our network that we have the influence on the consumer agent side, that we have the potential of charging them the billing management. So the the billing is not just to human billing but also to agent billing. That agent billing includes like

help the agent to check out and process that information, process the aging side of credentials that to allow the merchant to have the order and we capture all the credentials and the intent of all that. So that’s a potential of charging the merchants on that. And also we see the card schemes having some solutions to treat the agent and the everybody on the path being a like

the the the search changing or GEO side of thing that you have the potential of charging a fee for for the influencing capability. So I think in general today we want to build the ecosystem and having more and more users, agents and merchants on board and the monetization path will be clear and across all the actors over the ecosystem.

Grace Shao (13:46)

I see. Okay. Well, there’s often a disconnect between the merchant brand people and how they see, you know, payment versus what the infrastructure is behind the payment. ex explain to us like why there’s that mis disconnect, I guess.

Patrick (13:59)

Well, the the customer sees the merchant becomes the merchant they own the product, right? The price, a promise, and for e commerce is a fulfillment as well. And underneath that the connect the process actually connects the payment and the network and the process the transaction. So and there’s the issuer that decides whether to approve it, to decline it, and there might be fraud, credentials tax and settlement, all that. So those roles matter and when something fails.

So who made the promise, who improved the payment, right? Who processed the pre the payment, who protects the credentials, right? Who should handle the fraud, who’s responsible for that, who owns the liability. So I think the agent e commerce adds another actor and the chain becomes even more important. I think the the biggest difference in agent e commerce is when we look at from physical store to e commerce, right? like the ultimate decision makers do human.

But with Gene Commerce, with the autonomy that the agent is performing, does a potential the decision making is from the agent. Then is that decision making legitimate? So those questions are still remains and not being solved clearly. So all that we see that the clinic’s role is to try to carry that authority and the transaction context and all that across the trend. So again, come back to the definition of connector.

So we connect all the parties and try to pass through the information, the authority, across the chain, across the rails.

Grace Shao (15:23)

I see. We can double click on the safety and liability side of things later. I do find that quite fascinating. But it one thing you just brought up that was quite fascinating is that it agents will have autonomy. However, even though they have autonomy, the money, the feat of money is not theirs, right? So how do we understand that relationship if your agent is out there spending your money? do you always have to give them consent like before they go out? Or

Potentially what we’re gonna see in the future is agents literally just like going out there buying things without you even knowing they’re purchasing things.

Patrick (15:55)

Well that that’s a good question, all right. And I think the step by step would would just describe that agent in the future that like shopping fully autonomously it’s possible, right? But I I just think it’s not there because the trust is not there yet. So all the we we see all the demand is being delegated. So that means you as a human that you have some requirement, you have some demand and you want to make some purchases that essentially the demand is from human.

And the agents help you discover, help you to make some decisions, but maybe not for for the final decision. Right. So we see that as like involving pass. So but but as you just said, after all, the agents spend your own money, right? Spend your asset, see, spend your fiat. So we see that for today, in some case by case, like just one-time purchases.

That the user or the human will set up intent, give enough contact to the to the agent and the sign off on that. The sign off could be like a passkey, like we with a signature. Or it could be like A P two kind of a protocols, right? So there’s many actors in the industry try to solve that issue. But after that I see

Like the agent later on will do the the sourcing and do the patriotization as long as the falling fall under that mandate or instruction that was given, then we approve it. But essentially if we come back, see like where the orange and the context from, that’s from the user. So that’s for like a one-time instruction. But for sometimes if the merchant is trusted, like a digital service you always buy or digital you always top up frequently, then you can set up a recurring mandate.

Saying, okay, for this particular product or this particular merchant. as long as like over two hundred dollars a week or like every top up is a five dollar, that that’s fine. Then the agent does not have to receive or like get your approval every single time. But rather it’s just like a one time setup in the very early. But the setup or the mandate or instruction had to be super clear saying, okay, this is a trusted merchant. And we like it’s it’s fine for the agent to keep top up on that.

Grace Shao (17:53)

It’s like auto renewal but the without like resubscription.

Patrick (17:57)

Yeah. Well

I think the this is a little bit different, right? People are saying like if from merchant side if you give the merchant’s car and then merchant tell the merchant to do auto reload, yeah that that sometimes will achieve the same outcome. But I think the role is different. So on that side is merchant W and on this way it’s like the agent top up, like the actor is different. Right. And I think in general it’s like you trust the agent more or a tr tr or trust the merchant more.

Grace Shao (18:19)

Mm-hmm.

Patrick (18:24)

Right, that we see a lot of issues like disputes or chargebacks from the merchants actually deduct money, debit money from the user without that without the the the approval. And then with the agent being the actor that and also clean in the middle, then like we we actually protect the user. Right when the when the agent yeah, got

Grace Shao (18:28)

Mm mm.

I see what you mean. The ball

the ball’s back in the buyer’s court, kind of like the actual intent comes from that. It’s very interesting while we’re talking. I was thinking about a conversation I had with Ali Baba and Tencent recently. Both of them obviously are exploring agentic commerce. And then for them, you know, the biggest hurdle is how to manage this liability issue with who approves the payment. And obviously they both have payment systems embedded. So on the Baba side with Quinn.

Patrick (18:46)

Exactly. Right.

Grace Shao (19:09)

It’s really interesting. They’re saying that even though they’re doing agentic commerce, actually every time the money is gonna go out of your account, they’re still gonna push out a human verification kind of notification. You still have to like manually take that responsibility and press that button. On the Tencent side, I think they’re exploring the idea of creating a separate ReChat payment account, like kind of like two separate accountings, like two separate accounts.

And then one account will just potentially have very limited amount of money for the agent to play with. So you kind of just gave the agent like you know access to that one account, but not your full account. I don’t know. It’s just very interesting. I I’m kind of going on a rant, but it is funny because, like, on the other hand, it could potentially become like a scenario where essentially you give a credit card to your like unhinged teenage daughter and they would just go rogue and buy anything. yes. but let’s talk about agentic commerce.

Patrick (20:07)

No, no, I was just saying that which just is right, that just like giving a teenager your your child a card, right? And then you in in you don’t know what they’re gonna s spend and but after all you know them, right? But for Asian, I think the issue is like they may not have a trust yet, like fully trusted. So that’s what you mentioned, the Tencent has a weChat card and also the the Q and they they want the user to prove that every single time. And we see that that’s because the

the the different level of or different like I mean the path of payment being developed in in the country is different. So China has moved to the e-wallet and has very few dispute and chargebacks. So that means when comes to this particular situation, I think it’s just we we skip the credit card time. Like we skip the credit card UA and then move into the bank based or U wallet based directly. But if you look at the the

Grace Shao (20:49)

Why is that? What what

Sorry, my question is more like, why are there less dispute because it’s a U wallet? Wouldn’t like cat merchants still charge you if they want to charge you? Like if you’d like subscribe to certain things?

Patrick (21:02)

the the

because yeah.

yeah, so there’s two side of like if the the the payment is one side from the buyer to the seller, right, and then that’s real time transfer. And then we all know that today you have to scan your face in China, right? Or use a fingerprint to authorize the payment. And in those cases like the the dispute or charge back it become nearly impossible or or not necessary. And owning with a card card yeah, yeah, it’s so hard.

Grace Shao (21:23)

I see.

Like fraud is harder.

Patrick (21:37)

And then for for the auto debit thing then indeed and Alipay and WeChat has a compliance check and they revoke a lot of merchants auto debit. And now if you go want to receive the auto debit, it has to be super critical and then describe your business and the models clearly to the payment team and they will approve that. And it’s merchant by merchant of approval today. Unlike previously, you may either two turn that on. So but come back like when we

we see that issue particularly in card system, right? And then because like we we all know the disputes and chargebacks are all fraud, right? And like it’s just like so popular. Well not popular, it’s a common to see, right, in in the United States and in Europe as well. And that’s that’s why Visa has the entire chargeback management system to protect the the rights of the consumers. Right. so when they come to a genetic word that that create the the convenience that

because it is card based or the credential based that the user may not necessarily to approve every single transaction. So that brings convenience. But again that brings the risk, right? But however, thanks to the the chargeback system that has been developed over decades, the the risk can be like manipulated or to be like managed properly. so that’s why we we need the mandate system, we need the instruction to

to validate afterwards where when something really is bad then we come back to see if whether the transaction was being authorized properly by the human. And if it is, then well, probably the chargeback or dispute will not go through. Right. And so that mech mechanism does not e exist in Chinese payment industry today. So that that’s why that’s what I’m saying thinking is like may maybe the reason that they have to use a small amount of like a sub account, subcar.

Or a ask you as a human to approve every single time. And but honestly I think that actually yeah. So I actually Yeah.

Grace Shao (23:26)

I see, okay. That’s so nuance. Yeah, yeah, interesting.

okay, so but let’s talk about agentic commerce. So I think so many people have confusions around what really agentic commerce means. it’s you know, is it just that like basically like you said, you go out there, your agents are out there like going rogue and buying you things, or agentic commerce more understood? Like there is a process right now being built out. Like, help us understand what gender commerce means. What is the use case for agentic commerce right now, and are we actually seeing

it being proliferated in the real like real life right now.

Patrick (23:59)

yeah, yeah. So I think agent cameras is a very big word, right? And then it has many aspects and the dedicated execution. like I think that’s the most important happening right now ‘cause the the as I said earlier, the the the demand is still from the human and then the the product could be recommended by the chatbot or the agent. But after all that the the final decision may still come in from

the the human at most of time at today, right? But moving forward it could be the agent, with your authorization or understand you better, like has more context of you from the lifetime conversations, right? And then make a decision for you as long as the decision actually sits fitting the policy that you have set for the agent, right? So that’s another sort of like a next level of agent e commerce. so I think today

We see more and more like agent being an actor, but the actor could be on the sourcing side, it could be on the queuing side. And it really depends on how you as a user trust the agent. So in general I think agent commerce is very big but in general, like e commerce is a long chain, like doing all the actions from from the very beginning where the intent started and where to to the final the others play placed or the fulfilled. Right. So

in that long pass, any part that agent take kicks in, right, I think it’s agent commerce. But most of the time right now the agent is in the consumer side to provide help sourcing because the efficiency that they search products, right? And they may discover some very rare product that you may never be able to find. Right. There was a joke back at the time that probably as a human you read only first page of the Google search results, right? Super clear.

Right. The best place to hide their body is a second page of Google Search Results. But agent can easily search through twenty pages and get all different products or they they just have more context of you. They know your taste, for example, moving forward. And then they may be able to find like from a very real merchant, right? You you never know. Right. So that that’s a great potential that we see agent can unlock. And I think today that’s

Some of the aha moments I’ve talked to many people on Gene Commerce is the product search or recommendation site that the the the agent gave them some such suggestion or recommendation they never thought of as themselves.

Grace Shao (26:14)

That’s interesting. But I also think wouldn’t it make more sense for enterprise use case given the nature of like you said, sometimes it’s repeat purchases, it’s not always evolving. would would would that make more sense? Like if you’re a construction company, you’re buying like cement every year or whatever for this project, you know exactly the quality, the the the price you want to pay. your lumber company, whatever. You know what I mean? Like, wouldn’t that be much easier for agentic commerce to be in

Patrick (26:40)

Yeah.

Grace Shao (26:42)

integrated, implemented.

Patrick (26:43)

I think that’s a great point. but I I I do want to call that those cases are easier because they’re repeat repeatable, right? They’re bonded and you need to verify. it doesn’t matter the the buyer is enterprise or a pioneer of consumers, right? So it’s rather the behavior or the product itself define that the that use cases is more trustworthy or like you are more comfortable because to r like rebuy laundry stuff for example, right? And then

Grace Shao (27:00)

I see, yeah.

Patrick (27:09)

You don’t like that you always use a single brand and you you probably most of the time don’t want to try something new then then just a repeatable ask agent to to keep doing that. I think that makes it total total sense, right? To have the agent do that delegation, right? And like to to help you execute. And but like if it’s enterprise or like the pioneers of consumers, I don’t I’m not sure. It’s always like a some group, a small group of people.

They’re they’re interested in trying everything new, right? I like all the new stuff they they on board. I think those people are really push the word forward and help us to do the early adoptions and explore all the issues and helping improve the systems. It could be some pioneer of the enterprise, we don’t know, right? Small teams like startup teams, they are always willing to try stuff new. It’s all possible, but we don’t we don’t limit there. But I think you’re right that we use the

a small amount, repeatable and you need to verify purchases, to build the trust between the consumer and the agent. But the consumer could be human, it could be enterprise.

Grace Shao (28:11)

walk us through that case study ‘cause you were telling about it beforehand called Hello Minds. That was quite fascinating.

Patrick (28:16)

yeah, so it was a it was a short story. Actually, it was quick. when we were at Supreme in Singapore, they came to our booth saying, okay, they they see a clear need from their customer, they want to do transactional stuff. But as an agent platform they are not able to do that because they talk to Visa and then it’s just simply difficult for for them to to receive a PCI and all that in a short time. Right. So

then visa recommended us because we as aging enabler we’re designed to to turn on the capability of the aging platforms to do transactional. So and then we we just work with them super easily. There they have our skill pre-installed for their platform and then the the user be able to tell the HelloMind I want to buy this and that and then the system will actually get the user into our aging portal.

and add the card and we will go through the entire visa process for the verification, for device registration, and then it just happened automatically. So I think that’s a great example seeing like out of like all those vertical agents, they see the clear customer requirement, but for the role that they are today, they see a longer path to achieve that by themselves. And they see the they are seeking for for the help and for the experience partnership as Clink.

from the ecosystem to help them. So and then I think it it’s super clear for for Clink as well. We like to help them, right? And help other agents because our goal is being a connector. We’re now building our own consumer side agent. Of course we can do it, but we see it’s just like a payment method, right? Different people, different location, different markets have different preferences. Hello Minds is a great Asian platform over Hong Kong and and APAC area. So that’s kind of

like a possibly a capability that Clink provides actually help all their agent platforms to build their own and to satisfy or serve their customer better.

Grace Shao (30:12)

I see. Okay. Yeah. Cause you did say you’re agent agnostic. So that makes sense what we’re saying here now. but if the feature is, you know, many different agents acting on behalf of users, what does the payment layer need to do especially well then? Is is it in the identity, authorization, routing, settlement, security? Like how how I guess for you guys in the middle layer, what is the core, core offering that you have for everyone?

Patrick (30:36)

yeah, I think you’re right that like a one model may not fit all, right? Like a one base model or one agent may not fit all and the one payment solution may not fit all, right? Payment methods and the acquires, carnet was so many things. So all you mentioned, like identity, association, routing settlement, security, all of them actually matters. Right. So for us like we

We’re not going to like just build a single rear or like being a connector, we want to optimize every single layer, right? To be a connector that brought like try to bring every parties all together and closer and then try to solve the friction along the path for the agent to perform transactions. So in general that’s Klink’s goal at this point. We want to create a smooth word for the agent.

from talking to users and to the placing order on the merchant side and also have moving help moving the the fiat funds from from consumer to the merchant as it is today to have ‘cause we we see that as most smooth because after decades of development, the merchant solution, their commerce systems, their payment stack are being mature.

And there’s no reason to rebuild all of that just for the agent, right? And the return on that investment could be super low. So that’s why we see that as an evolving pass but rather a revolution pass.

Grace Shao (31:59)

look, I think let’s talk about the the most sensitive bit now. I want to talk about the trust and safety bit. So, you know, you do emphasize, you know, you’re building reliability, you’re building safety into the product and everything. But as the consumer, it still seems like really, really unsafe or scary to try out a new payment system. You know, as the average consumer, we will still default to big names like Visa and MasterCard, even if we know we’re paying them much more, right?

so help us understand that. Like what what is what is a trust barrier here and how do you build that up? and I guess who bears that liability? So hypothetically, I’m using Clink to buy something on whatever, let’s say Amazon. Okay. If something goes wrong, am I supposed to go to Amazon? I’m or am I supposed to go to you? Or is Amazon Amazon gonna come chase you down? What what is the relationship here?

Patrick (32:51)

Yeah, yeah, good question. I think so we are being agent enabled there or the connector here. So that means we are not the merchant of the record. So you still own your order on Amazon, right? So that’s our relationship with the merchant as well. We are not going to replace a merchant, repla take over their customer relationship. That’s not what we’re going to do. So what we’re doing is here, like it’s being a connector. That means we pass along the information.

like the payment credentials and all that, then we also capture your intent instruction that the the user talk to the agent and then send that over to Visa. So after all you’re still seeing your Visa card paid Amazon. I’m just using Apple as example. You after all you still see that you have an order with Amazon. And then you still pay with your regular car, right? Just the the in the middle the two actors, one is the agent. So agent performs the excursion.

as you authorized. And the clin perform as another set of actors and take over the agent’s direct execution, but help you and your agent to place the order on Amazon because we have been certified by Visa. And we have certified by the payment industry because we are PCI compliant. PCI but by the way, PCI means that we can securely manage all the payment credentials, right? Being authorized by the certified agents that we can

persist and the process user’s kind numbers, right? And given that we have capability to pass that over to Amazon. But after all, still you place an order. We’re just solving the problem that agent may not be able to directly touch the credentials or you don’t trust the agent too, right? And I mean at this point. So that’s why the we as a connector in the middle, we persist your critical information and then pass along to the merchants. And also we

Help you validate the policy that you set for the agents and the the initial instruction. For example, you authorize the agent to buy a shoe, right, and under $100. And somehow the agent got mad and bought you a t-shirt for $200. Then clearly it does not match the initial instruction. And we will block you before we send that over to the particular merchant that the agent wanted. Right. So that provides additional layer of protection. And also because like all that is being also certified by the visa.

Like because we are the first agent enabler and the agent side of partnership with Visa in APAC in in the Visa intelligent commerce program. So all that is being like even though it’s still pilot, we see the potential that we can add additional layer of protection and also additional layer of privacy.

Grace Shao (35:23)

Really interesting because I actually have two questions. That means number one, I mean, this is not unique to you guys, like, but you essentially have a database everyone’s credit card. There’s definitely a risk there, right? And then the second part of the question is that, like you said, when you don’t trust your agents to touch your credit cards now. So, say one day the education has been done, the market has been educated, agents have been normalized. Five years down the line, everyone is using agents to do shopping.

Does that just completely skip over Clink then if we’re all just gonna give our agents our credit card numbers, if we even still have credit card numbers, whatever payment look like back in the future? Would that just actually frankly o omit the the the role of Clink then?

Patrick (36:03)

that that’s a good great question, but I don’t think that will happen because for like even though the agent development all that then the the the concern to like you still need additional layers of protection because it’s like we are the neutral layer and give you the additional protection, right? Even though you trust the agent much but you still want the third party. It’s just like at the very beginning

of the e-commerce in China, you have Taobao, you have the consumer, like you have the buyer and the seller. But you still need to about just LP in the middle, right? To prov provide you that layer of security and you you want another layer. Yeah, so a a neutral standpoint to to to also right to to monitor, to audit the agent behavior. And also from the car scheme or the merchant side, they want some additional layer of protection as well. Right. No one can simply trust a single agent can perform all of that.

Grace Shao (36:38)

See what I mean?

Patrick (36:54)

Right. It’s just like additional yeah.

Grace Shao (36:55)

I see what mean. So there’s guardrails

built into you as like the guardian almost of all this.

Patrick (37:01)

Yeah. Yeah, exactly. the an an a layer of protection and I think Clink by the time it will build the trust across the consumer side and also build the trust across the merchant side. Because we see that the potential of having more a merchants being aging readiness and into the aging ready merchant network will bring them the advantage across the five years development that you just mentioned.

Grace Shao (37:24)

So I have a question. I I have two three questions to wrap this up. first one is what is your vision of the future of agentic commerce then? Like where do you see this as going? You can start with this, and then I want to throw you another one first just for you to think about in the back of your mind, What is the smartest question you’ve heard someone else ask you about the agentic economy that I have not not asked? So

Patrick (37:44)

Sure.

I think yeah, I again I guess we can get started with agent commerce, how that will go and aging payment. I see that has a long way to go. That as you said, like you have a child and like if you have a child that they grow up, they become adults, they become mature. And actually at this point we see that agents are growing fast, but are they close to the AGI?

Right, are they close to a particular point that that they’re being smart enough enough to do all that? So I’m not sure about that. And on the scare side, if they are really that smart and as another adult, do you really trust them to manage all your assets? Right. So those are like I I see those questions are tough to understand, are tough to answer, but at this point we see that agents are being super good or super useful as a tool, right, as a helper.

to help you do the excursion to do all the repeatable work or like all that you want to skip. And the sum of part of the commerce is being part of that. But I also see that shopping is a great journey, right? I think you as a lady understand what I’m saying. So I I mean as I mean I I typically don’t do enjoy that process, but rather just get what I want, right? But my wife and they do enjoy like the journey of exploring products. I

I think those are entertainments, not just commerce or shopping. So those will never be replaced by agent. But it could be like you I I know there’s some services in in China I’ve heard of that. like you you got a guy or like a girl that accompany you or walk through the the big malls and do shopping all together, like a company, right? It could be like in the future those accompanying role replaced by agent. It could be. But still provides a a great journey.

help you sourcing and find a great product that fits you. But I think that also would be a very pleasant experience. But I I think in general like for now we see that it’s a delegation for e-commerce for most of that valuable part. it’s not a new demand. But we also see for the digital service side and also for the like a

Mostly on digital and the model requirements or all the model f model side of features, those are new requirements. But it it it replacing a lot of the previous like data sourcing work or like investing work investigation work that you’ve done as a worker. But this may be replaced by agents as well, and along that path will generate a different requirement or different demand for genetic payment. So those are new, I think.

And in general along the past, hopefully we see that agents and humans in a word that can both like pay and get paid. Right. So that like today the the we’re we’re talking about AI applications, but in the future the the the product may be served by agent. But you never know. So it’s like a human to human, human to agent, agent to human. We’ve seen an interesting case that Waymo plants order online.

get a human to close the door. Right. If s a passenger got off the the the ride and forgot to close the door, and then you you place an order and get a human to do that. Right. So it’s like aging to human, human to aging. We never know. It’s that’s what you said, agen agent economy, right? Just not not commerce. So I I think like that word will eventually come and along the way that we have the base models to develop being smarter and also a lot of

Issues like around identity, around trust in general to solve. And then that’ll that’ll be the case. And then to a second question, like the smarter question, I think is like really come to that when when intelligence or labor or execution becomes abundant, right? what remains valuable? Right, or die, right? And do we still need transactions?

Because whatever you want, like the goodies or the services is being so abandoned and probably at no cost at all. So where is the transaction? Where is the value? Right? I think then that I think that’s a smarter question or tough question could always come to me saying, like really like if we as a society or the entire world develop who develop to that level of then what what’s what what are you p people paying for? We never know.

Right. And maybe that that’s a time that come back to okay, the the taste and the the people that creating the value that I I don’t know, I just don’t know. I think that’s a very broad question, open question. And we may not have an answer when really that comes. Yeah. What becomes valuable and what are people paying for?

Grace Shao (42:07)

First, I wanted to say it’s really scary to think that we potentially have machine overlords telling us what to do and working for them. second of all, I think the second half is something that’s really interesting. It’s it’s not just in the whole abundance of a gentic era. It’s just like you’re seeing it play out already. Like even people are saying, you know, luxury is a structural short because why who is still buying luxury? When you have more money, are we really spending money on putting

Patrick (42:15)

Well, that’s a bit scary.

Grace Shao (42:35)

you know, brands and logos on our bodies anymore or is society moving towards like paying for experiences, praying for health, paying for, you know, things that actually cannot be so easily replicated because, you know, logos can be replicated easily these days. It it’s kinda interesting. So it it will be interesting once intelligence is abundant, what will happen. But then I do think that question is also overgeneralizing humanity because we do forget that, you know, not everyone

Is working in a career that actually requires intelligence. I’m not saying people are not intelligent. I’m just saying intelligence as a currency to for, you know, salary or, you know, whatever or or capital gain is actually not for everyone. And if anything, then do we say the value is more in craftsmanship, laborist work, you know, you know.

Patrick (43:08)

Yeah.

Grace Shao (43:28)

experience and different things. It’s just very interesting, but for sure it’s already moving away from objects, like tangible objects, like where people really want to pay, right?

Patrick (43:36)

Yeah. Right. And you th you mentioned labor go go ahead, no, that’s a sign. No.

No, I was just saying like a labor as well. Like the the labor may even the work may not require intelligence, but that work and like the labor side of work may still be replaced by robotics. And with a small robotic, that that’s even more scary, right? Like as a move yeah. Yeah, yeah, yeah. So but but hopefully we we humans still aren’t in in control when that comes.

Grace Shao (43:54)

I s I see where you stand in all of this, Patrick. I see where you stand.

I don’t know, Patrick, the way you’ve pro you painted the futures, we have machine overlords telling us to close the door. Why are we even leaving the house then? okay, jokes aside, I have my last question for you, which is the question I ask every single guest that comes on. What is one differentiative you you hold, something that you think is maybe against consensus or people still you think get wrong or underestimate?

Patrick (44:12)

Yeah.

I I think the two two sides I want to answer that. One is that I think we’re we’re spending too much attention or like on the itching payment side require a new brand new rail like broad chain a crypto rail. So I’m not saying that’s wrong and we see the potential of the stablecoin or the capability of blockchain that for cross border settlement, machine too much micropayments, all that, but like a a reel that

Like the the real exist because we move value, right? And most of the time today the buyers and the merchant that they use VAT and we see we may not spend enough attention to solving that side of the rails issue. Right. So but I I see that on the long run, of course, the blockchain aspect stable coin will have a great value and being super useful. But at this point if we want the agent to be capable of handling payment.

instruction doing all that payments and we probably want to solve more issues around the the fiat rail and to to make that smooth. So that’s one thing. A second thing is I I think like even you’re not a payment in the payment industry you may hear of a lot of protocol names, right? A P two, UCP, ACP, SPT, and it’s just so many of them, right? And I’m not saying like

like every single protocol as the has their values and standards and so try to solve a particular issue or a like a vertical scenario, right? Or some of them being try trying to be generic and some of them well like we’ll want to solve the issue of product discovery and all that. And but I feel just too many of them. Like just way too many of them. And that creates a barrier or creates a a like confusion for the merchant.

And for other actors in the ecosystems, like which one should I adopt? Which one should I integrate? And the more and more are coming out, right? Every single player has their own standard. And as a merchant, I may feel like so confused, like which one I should adopt. I think really that the entire ecosystem will want to align or I mean to get some agreement.

Well, I mean not necessarily a single one or two protocols, but to have some fundamental rules in place that people understand okay, that’s how we want move and that’s the direction we go and people start investing on that. And eventually some generic or universal protocol that’s been agreed by everybody, like by the majority of the actor in the market, they’ll come out. And that’s a time where I see that

agents or the merchants will be more comfortable. Otherwise it’ll be too many of them for for the ecosystem to advance to next step to to next stage that agents and merchants are like feel so like worry free to onboard. Otherwise they feel like okay if I bought I onboard A C P last year, right? And they’re really like like not all agents are supporting them. Right. And then if we are on board U C P this year, we see other agents not support them. So like those kind of issues are being

like a making just like a too too fragmented. And as a generic like a pass forward, I I I really want that we we somehow get aligned in one particular vision, the entire ecosystem and try to push that new outcome.

Grace Shao (47:38)

There needs to be some standardization consistency to help agentic commerce go forward. All right, Patrick. Thank you so much for your time today. Really appreciate all your insights. I’ve learned a ton. if anyone’s interested, feel free to reach out to Patrick or I’ll put the link in the podcast show notes as well as on Substack notes. thanks again, Patrick. Speak soon again.

Patrick (47:56)

thank you. Thank you, Grace, for having me here.

I have had the honor of having 2 people go to WAIC on behalf of AI Proem, and they were very different profiles. One, who will remain anonymous, shared observations from a businessman’s perspective, and the other is a story written by the ever-so-talented Rachel, who will walk you through her thinking about WAIC from a cultural/societal angle. I, myself, had attended so many of these this past year that I felt I had become a little jaded, so this blog post hopefully brings some new perspectives to you. I’ll first share my thoughts and some observations from my associate on the event, then I’ll pass the floor to Rachel. Hope you enjoy the tour.

Grace’s opening notes:

WAIC 2026: The Robot Hardware Is Ready. The Intelligence Is Not.

WAIC was a technological exhibition that has been ongoing since 2018. It is actually the event where Elon Musk and Jack Ma had the famous/ infamous awkward exchange. But it is no longer just a tech exhibition. WAIC is increasingly becoming a stage for Chinese leaders and companies to define their own global AI narrative.

The scale of this year’s World Artificial Intelligence Conference was difficult to miss. It attracted an estimated 400,000 visitors across more than 100,000 square meters of exhibition space. Some 1,117 exhibitors presented 4,486 products, including a record 351 global debuts.

President Xi Jinping gave the opening speech, which Matt Sheehan did the hard work and annotated. To me, the message of the conference fell into three broad buckets: China’s emphasis on openness and inclusivity, a continued nudge towards greater self-reliance across the technology stack, and the gold rush around humanoid robots and physical AI.

Of course, scale was impressive; dialogues around governance were necessary, but to the disappointment of many visitors, including my associate, the conference revealed more about how impressive the humanoid-robot industry now looks from the outside than how capable the underlying intelligence has become. Embodied intelligence drew some of the largest crowds. Perhaps that was because robots are more eye-catching, or simply because of the recent hype.

We already know that China can manufacture, but?

China has built much of the physical infrastructure required to manufacture humanoid robots. Motors, precision bearings, joints, force sensors, dexterous hands, and even complete base platforms can increasingly be purchased from established suppliers.

A startup no longer has to design and manufacture every part of a robot from scratch. As we have written before, there are now OEMs capable of supplying a functioning robotic body. The competition is therefore shifting towards models, algorithms and system integration.

That is a genuine advantage. China’s manufacturing ecosystem is making robots cheaper to build, faster to iterate and easier for new companies to enter. It also explains why WAIC was filled not only with established robotics companies, but with startups and businesses entering the sector from adjacent industries.

Yet this supply-chain maturity can create a misleading impression of technological progress. When the physical robot is readily available, almost any company can present a humanoid body, attach a model to it and describe the result as embodied intelligence. The real question is how much intelligence has actually been integrated.

Many of the dancing, boxing, and stage-performance robots at WAIC operated smoothly. But these demonstrations often depend on predefined routines, controlled environments, or remote assistance. They demonstrate advances in mechanical control and coordination, but not necessarily an ability to understand unfamiliar environments and respond autonomously.

A robot that can repeat a difficult routine is not the same as a robot that can solve an easy but unpredictable problem.

This is also why a universal “robot brain” remains so difficult. Robots operate in different environments, perform different functions, and interact with the physical world in different ways. The relevant data, operational knowledge, and failure cases are highly specific to each use case. One general model may eventually provide a common foundation, but it is unlikely to perform equally well across every physical task without significant vertical adaptation. For example, a world model for autonomous driving has little use to a robot designed for doing laundry; do you see what I mean?

Near-term opportunity is specialized deployment

The most useful robots we have seen over the past year, particularly those that can help address labor shortages or improve workplace safety, generally look like industrial machines rather than humans.

Fixed robotic arms have benefited from decades of development. They operate in structured environments, execute narrowly defined motions, and can deliver measurable improvements in speed, consistency, and safety.

Some of the clearest examples at WAIC involved specialized robots used in dangerous power-grid operations, including high-voltage inspection, switch operation and ice removal from transmission lines. Their value comes from replacing people in tasks that are dangerous, repetitive and sufficiently structured to automate.

The same thinking could extend to firefighting, disaster response, underground exploration and hazardous manufacturing. In these cases, the economic case is not simply about reducing labor costs. Avoiding human injury may justify deployment even when the robot remains expensive or limited.

A wheeled robot may be better for warehouse transportation. A robotic arm may be more suitable for laboratory work. A quadruped may be better for inspections. There is no reason every commercially useful robot must look human.

This fits a broader argument we have made at AI Proem about China’s AI ecosystem. China does not necessarily need to lead every frontier benchmark to build an important advantage. Its strength may lie in integrating increasingly capable models into factories, warehouses, vehicles, consumer devices and existing industrial workflows.

The embodied-AI opportunity should therefore be understood as an industrial feedback loop, but at the same time, WAIC also showed that this loop has barely begun.

Geek+, for example, has a meaningful advantage because it already operates in real logistics environments and can collect data from deployed systems. Its conventional warehouse-automation products performed smoothly. Yet its newer humanoid remained far slower than a human worker at basic sorting tasks, and honestly costs much more too.

The likely near-term winners, or at least deployable at scale, bots are therefore specialized robots deployed in structured or dangerous environments, alongside the component suppliers, automation companies, and computing platforms that support them. I still stand by my analysis that the household humanoid remains a much longer-term ambition.

The moat will be vertical data, workflows, and compute

The data bottleneck is more complicated than saying the industry simply needs more data.

Robots require information covering objects, movement, force, balance, failures, and interactions across an enormous range of physical conditions. Static images and text cannot fully describe what happens when a robot grips a deformable object, loses balance, collides with something or encounters an unexpected obstruction.

Manual data collection is expensive, slow, and usually limited to a narrow range of situations. Real-world data is also noisy and fragmented, while the most commercially valuable datasets are likely to remain proprietary. Thus, what we’re seeing is that companies are leaning into simulation first, controlled deployment second, in hopes that one day they can reach broad autonomy.

Robots can initially learn perception and motor skills in simulated environments. Once they become useful enough for a specific task, they can enter narrow real-world deployments, collect additional data, and improve through use. World models and physical simulators can accelerate this process, but even highly realistic simulation cannot eliminate the gap between a digital environment and the unpredictable physical world.

But as we mentioned, even these simulation systems are so niche that they cannot be used by everyone doing anything. On top of that, people forget that for physical AI, it’s often a large world model that each firm trains for their specific use case and environment, then there is an on device/ edge smaller model that works together with the lidar and visual recognition components that together ‘tell’ the physical vessel to move. That kind of hardware-software integration and actual ability to train for variance and edge cases is highly underrated.

The chicken-and-egg problem remains: robots need real-world deployment to obtain better data, but customers are reluctant to deploy robots that are not yet useful.

One exhibitor reportedly offered consumers a robot capable of basic demonstrations for RMB 3,000 per month, presumably allowing the company to collect data inside homes. It is a clever idea, but the commercial proposition is questionable. Most consumers will not pay to train an unfinished product on behalf of its manufacturer.

As my associate put it: who wants to rent a hen that cannot yet lay eggs in the hope that it eventually will?

Over time, the strongest robotics companies may not be those with the broadest demonstrations. They may be those with privileged access to specific operating environments and the expertise to turn that data into reliable products.

[G: This one is absolutely outrageous. Tell me what we need a centaurbot for?]

This connects with another argument we have made at AI Proem: companies should protect what makes them distinctive, namely their proprietary data, workflows and industry know-how. The model itself may become increasingly accessible. The harder asset to replicate is the closed loop between an installed base, operational data and the product improvements that follow.

That makes the idea of a single industry-wide robot-training dataset somewhat unrealistic and thus very different from the colossal amount of data that is used for LLM training. Companies see their data as a moat and have limited incentive to share it. We may therefore see many strong vertical robotic models before we see genuinely general robotic intelligence.

Compute is another constraint. A robot must continuously combine visual input, depth, geometry, movement, force, sound, and its own physical position. It must predict what happens next, choose an action, and adjust immediately when reality differs from its expectations.

That is a much more demanding inference loop than generating text or even an image.

Useful robots will require substantial onboard or nearby compute, low latency, reliable connectivity, cooling and sufficient battery life. Adding more processors increases cost, weight, heat and energy consumption. At scale, millions of embodied-AI devices would also place significant demands on data centers, electricity networks and communications infrastructure.

China’s investments in power, computing infrastructure and domestic chips may eventually become an important advantage. But infrastructure alone will not solve the problem. Embodied AI will also require more efficient models, purpose-built chips and better decisions about which calculations happen locally and which can be handled in the cloud. It is also not to dismiss that hardware costs are falling, suppliers are proliferating, and physical iteration is accelerating.

The fair takeaway, as my buddy said, is that neither that humanoid robots are about to enter every home nor that the entire industry is hype.

Share

And with that long monologue, let me pass it to Rachel…

Rachel and I met years ago when I was trying to get some help with a personal project. What I remembered was how thoughtful she is and how strong a writer she is. (A CUNY grad who has written for Vice and WIRED). I knew that Rachel is not immersed in the AI industry, and therefore notices things regular technology observers may overlook. And as there is enough information about the technicals, I thought letting her cover WAIC this year would be really interesting — a fresh new perspective. It was such a pleasure working with her on this to flesh out the deeper questions we should be asking ourselves. See her Substack here.

btw, next year I plan to attend and maybe even host a small gathering of industry professionals and investors for a dinner, so stay tuned.

Rachel Zheng:

For those four days, the whole city tipped into a kind of citywide frenzy, something close to SXSW energy. Bars, hotels, even the subway were plastered with AI banners; the average person’s FOMO hit a peak I hadn’t seen before. But underneath it, there’s no denying the sheer volume of capital pouring in, backed by state policy; the whole thing had the sheen of an economy on the upswing, all of it happening under Shanghai’s brutal 40-plus-degree heat.

Embodied AI took up an entire floor on its own, making the case that robots are good for more than dancing; they can pour tea, stir-fry, and give shoulder massages, proving they have actual use value.

The industry is shifting from spectacle to substance. The humanoid robots still draw crowds, but the real investment is in the invisible: compute clusters, tactile sensors, and operating systems rebuilt for agents. The question that kept coming up wasn’t “what can AI do?” but “what should AI do?” and that’s a question about values, not just technology.

How does an AI product tell a human story?

WAIC 2026

For the past few years, I’ve been reporting primarily on art and culture. My only real encounter with AI has been using DeepSeek occasionally for editing and research. And so much of the media narrative around AI is that it will take human jobs, presented in a very black-and-white manner. I wanted to understand AI better, maybe this time not just technically but also how it really is in relation to us humans.

Just a week ago, a design studio founder told me that he finds joy in extracting brand stories out of AI company founders. In his conversations with them, what interests him most is the process of defining what’s human and what constitutes humanity through their products. That interview was a lightbulb moment for me: how does an AI product tell a human story?

Across more than 3,000 exhibits, it was impossible to see everything. I found myself drawn less to the largest models or fastest chips than to products that made me reconsider abilities humans barely notice: seeing, touching, moving, and deciding.

What I found most revealing was that the most revealing AI products were not necessarily those that looked most human. They were the ones that exposed how difficult ordinary human capabilities are to reproduce, or helped restore capabilities that people had lost.

The Popular, the Practical, the Precautious

The official gem of this year “镇馆之宝” centered on ten flagship exhibits, telling their own story about what the industry in China considers important right now. By my count, eight of the ten sit squarely on the “brain” side: compute clusters, LLM-powered agents, chip architecture. Among the rest are Ant Group’s robotic pharmacy and Zhiyuan’s Expedition A3 Ultra humanoid, both of which tackle the harder problem of fine-grained perception and physical interaction at the system level.

Humanoids produced many of the conference’s most photogenic moments, but the official selection told a more sober story. Only one of the ten flagship exhibits was a humanoid robot; much of the industry’s attention remains on the less visible infrastructure, models and systems underneath.

Those systems may ultimately prove more economically important than the machines drawing crowds. But I kept returning to products operating at the boundary between software and the human body.

Btw, there were also a number of local-government cases using AI for energy management, grid optimization, load forecasting, that kind of thing. A lot of what’s on display right now centers on infrastructure being laid underneath everything else.

The conference’s ten official flagship exhibits leaned heavily toward the infrastructure and intelligence beneath AI: compute systems, agents and industry models. Those may prove more economically important than the machines drawing crowds. But the products I kept returning to were the ones operating at the boundary between software and the human body.

For consumers, the concept still getting the most attention is the AI phone. At StepFun’s STEPX Neo booth (Tencent is among its investors), I watched an elderly man stop one of the staff and said he’d come all this way just to see this particular phone. He asked question after question, most of them about data security. The staff explained that STEPX Neo had already partnered with major apps: Alipay, Ctrip, and others (though notably not WeChat), and that each time it needed data, it would ask the user for authorization; nothing gets uploaded to the cloud. What the phone is really trying to redefine is interaction itself: its Step AOS rebuilds the operating system from the ground up.

That sentiment was echoed by Huang Xuanjing, a distinguished professor at Fudan University, who said everyone’s phone is becoming its own kind of intelligence, every major manufacturer is racing toward it, and this is likely what phones will look like in a couple of years.

I find that idea, an interface that behaves less like software and more like an agent you simply talk to, genuinely compelling, and intuitive to how we already behave. But traditional companies still rely on their own apps to access user data, and the promise of “never switching between apps” doesn’t obviously help them at all. So it remains to be seen how this actually plays out, whether the agentic phone becomes the new default, or just another feature layered on top of the old one.

Elsewhere, in the AI-for-science sector, one investor, Chen Jian, vice chairman of CCF and chairman of Paratera, made a blunt observation: AI for science stays small simply because the industry value is small.

A longtime healthcare industry veteran was cynical about Ant Group’s Afu body fat scale, a free scale (users pay a 9.9yuan shipping fee upfront, but get 30yuan cashback upon binding the app), but only works once you hand over your data to use it. What it’s really harvesting, he argued, isn’t body fat percentage but the user. He compared it to something almost generational: the way elderly people used to line up to open a bank credit card for a free carton of eggs. Made me think whether some products promise to understand the user better; the unanswered question is whether that understanding serves the person or the platform.

Against that backdrop of big companies quietly mining consumer data, Cao Peng, founder and CEO of Hangzhou Genlight Neurotech, was doing something almost defiantly unglamorous. Most people assume brain-computer interfaces are only about the brain, he told the audience, but neural signaling doesn’t stop there; the brain is just one part of the central nervous system. So his team turned toward a different, long-unsolved problem in medicine: helping patients with spinal cord injuries, some paralyzed, some effectively cut off from limbs that no longer answer them, recover the ability to move. The AI in his system’s case sits in the decoding: reading the faint, often garbled signals still traveling through a damaged nervous system, and translating them, in real time, into something a limb can act on.

His analogy was cynical, but the question left unanswered was: who is AI benefiting the most?

If there was one hall that felt the most alive with actual human use, it was H4 downstairs, the practical end of the spectrum, where high technology quietly trickles down into things closer to daily life: AI contract review tool, which cuts a three-day process down to thirty minutes; tools for identifying early cognitive risk; L’Oréal’s AI Coach, to train its own sales associates; even AI-driven marital legal tool…the sheer range of them was exciting to see, proof that the application layer of AI is no longer a handful of general-purpose products, but something closer to an ecosystem in bloom, each tool built to answer a need specific enough that you’d never have thought to ask for it.

Embodied AI — How to study robots as if they were humans

As a human, ha, I found some of these robots offering comical relief too. I stood by the shoelace-tying robot for a full minute, watching it pull the thread from one direction to another. I learned it takes five to ten minutes to finish the whole task, and almost next to every robot, there’s still a human standing by to monitor it. If anything, the day we actually employ one for ourselves at home still feels very very far away.

However, watching how difficult it is for a robot to do something as simple as tying shoelaces made me marvel at what an extraordinarily complex machine the human body is.

The robot has to read the thread’s position and orientation from a camera feed, under whatever lighting happens to be in the room, and that visual read has to update continuously as the shoelace moves, the light shifts, or a shadow falls across it.

A human hand does this without thinking; we barely register how much visual recalibration is happening in real time as we bring the lace from one hole to another. For the robot, each of those micro-adjustments is a full perception-and-decision cycle, done far slower and far less gracefully than the human eye and hand collaborating unconsciously.

My brilliant friend Wency Chen, senior tech reporter at SCMP, described teaching robots these most basic movements as a natural stage in the progression of the technology.

Embodied AI is moving toward completing the human loop of “perception - planning - action”; the ability to plan is important. At a booth showing a robot pouring tea, in order to test its real-world adaptability, the staff will move the cup around or put their hands over the lid, actually to see whether or not it can judge before acting.

The important human point is that action is not simply movement. It requires continuously updating one’s understanding of the world, in this case: where the cup is, whether something is blocking it, and whether the planned action is still safe.

As humans, we learn that kind of context through trying and doing as babies; now the robots are doing the same through repeated interaction and feedback - training.

The robot did not make human-level dexterity feel imminent; it made human dexterity feel extraordinary.

Touch — A language that the physical world writes for AI

If perception is the first half of that loop, I kept wondering where it actually comes from: how does a machine feel the world enough to decide what to do with it.

Talk to anyone in embodied AI this year, and the conversation almost always turns to the “brain”: large models, visual perception, task planning, language understanding. These matter, of course. But they answer the question of what to do, and how to do it. The question of how well it’s done is a different one entirely, and it lives in the fineness of perception itself.

Whether the action is done well depends on something more basic: the fineness of the machine’s perception.

Li Rui, founder and CEO of Shanghai ViTai Technology, opened his talk with an experiment that had nothing to do with robots. Numb a person’s fingertip with anesthesia, he said, and something as simple as picking up a matchstick becomes impossible, because touch is the thing that makes the hand capable of acting on the world at all. His talk was titled, plainly, “I Spent Fifteen Years Teaching a Robot How to Pick.”

Touch, he said, is the language the physical world speaks to AI, the one channel through which weight, texture, and pressure actually get through.

Fifteen years ago, as a PhD student at MIT, he and his advisor, Edward Adelson, found themselves circling a question that sounded almost absurd: if a camera can see an object, could it also be made to see contact?

They built a sensor with eleven pads of soft, elastic skin, each backed by a tiny embedded camera. When the skin touched something, it deformed, and the camera caught the deformation. Run through a chain of algorithms, that footage could be worked backward into data: a vision-based tactile sensor, touch translated into sight, with a resolution finer than a human fingertip’s own.

They called it GelSight.

On screen, you could see fingerprint ridges, the faint slide of a surface, changes at the point of contact too small to name. Li Rui said that was the moment he understood touch could actually be seen.

Thanks to this academic breakthrough, from what’s publicly known, ViTai appears to be the only Chinese company so far to have taken vision-based tactile sensors to industrial-grade mass production, running on mainstream dexterous-hand platforms, and it certainly has laid important grounding work for the development of embodied AI in China.

Closing thoughts

I went to WAIC on its last day (Monday), the only day still available when I brought my ticket, and it happened to be the public day when children under twelve could attend. I remember watching a child in front of a Unitree robot on the floor. Unlike the adults around him, who seemed determined to exhaust every function the machine had, he wasn’t trying to test its limits at all. He just wanted it to dance, and kept dancing with it, unbothered by whatever else the machine could do. Taking a step back, you realize adults tested what the robot could do, but the child simply cared about what he could do with it.

On the final day of the World Cup, the final match, a robot from Booster Robotics in an Argentina jersey faced off against a human child.

It reminded me of something a brain-computer interface patient had said, in a completely different context, about pain. For most people, pain simply means something is wrong. But for someone who had spent years losing sensation, he said, pain in his leg meant his body was sending a signal again, a neural pathway, answering back, because of AI.

Ultimately, the challenge may be not to make a machine appear more human, and the most useful applications were often the least theatrical.

I thought going to WAIC would get me AGI pilled; instead, I walked away from the conference hoping that AI is helping people feel, more clearly, what it means to be human. Both the joy of it, and the pain.

In this conversation, we start with his founding journey, the milestones and how Pony.ai became a leader in the autonomous driving space. We also discuss the gap between assisted driving, Level 4 autonomy, and the longer-term goal of Level 5, as well as how Pony.ai uses simulation, real-world driving data, and increasingly capable AI models to improve safety. James explains that the hardest problems are often not the obvious ones but interpreting unpredictable human behavior and handling rare edge cases consistently.

The conversation also explores China’s cost advantage in robotaxis. A mature automotive and electronics supply chain, close collaboration with automakers, and faster iteration can materially lower vehicle and system costs. But moving into new markets still requires Pony.ai to adapt to different road conditions, regulations and driving cultures, from trams and roundabouts to local pickup behavior.

James’s broader point is that the industry has focused too heavily on the initial technological breakthrough. Getting a car to drive itself is only the beginning. The next phase is about deployment density, utilization, maintenance, charging, remote support, and economics. At the end of the conversation, I asked what he believes is underrated. James, an experienced operator, replied - scaling. Pony.ai may have crossed the zero-to-one threshold, but the harder task is scaling from one to ten, and eventually from ten to one hundred. Check out this insightful conversation.

For more interesting conversations with people who are charting the way of the future of AI, check out the podcast tab or follow us on Spotify!

Share

Chapters

02:36 Why James Peng founded Pony.ai
06:36 The milestone that proved robotaxis could work
09:27 How passengers learned to trust driverless cars
11:41 Level 2, Level 4 and Level 5 autonomy
18:59 How AI and simulation improve self-driving
24:13 Teaching cars to understand human behavior
29:11 China’s cost advantage and global competition
35:03 Expanding robotaxis into international markets
41:13 Why Pony.ai is also building autonomous trucks
48:40 Adapting to new cities, roads and driving cultures
53:45 Why scaling is often harder than reaching zero to one

Transcript

Grace Shao: Hi everyone, welcome back to another episode of AI Proem Differentiated Understanding. This is your host, Grace Shao. Look where I am, the back seat of a car. Doesn’t look that exciting, does it? Let me flip this around. Look at that. There is no driver. I’m in the back seat of a Pony.ai robotaxi. Today joining me is James Peng, co-founder and CEO of the leading robotaxi company. It’s expanded its footprint across the globe, in Asia, in Europe, in the Middle East. But obviously, today we’re in its leading home market, China and Shenzhen, where it has a fleet of a couple hundred vehicles deployed on the streets already. Hi James, thank you so much for sitting down with me. I’m really excited to be having this conversation with you. So you left Baidu in 2016 to found Pony.ai when many in Silicon Valley were saying self-driving cars are only three years away. But obviously that wasn’t the case.

Grace Shao: So, what did you believe then that this consensus was getting wrong? Tell us about your journey from 2016 until now.

James Peng: Yeah, sure. We were founded in 2016, about 10 years ago. But even at that time, I didn’t believe that autonomous driving can be solved in three to five years. Just from a technical point of view, because even back then, 10 years ago, even a demo for autonomous driving was already very hard. Later on, there’s complexity involved in the autonomous driving industry that involves regulation, user acceptance, the readiness of the ecosystem. So because of the sheer complexity, even then, my prediction was it’s going to take at least a decade for this to be a real application. It turned out to be that my prediction was about right. Now, 10 years down the road, we actually have fully driverless commercial applications in many cities. Of course, it’s just the beginning of the long journey for autonomous driving. But at least now we have real commercial applications.

James Peng: So I think people, like any new industry, people were super optimistic for the short term, but they were underestimating the potential for the long term. So I think autonomous driving is definitely one of those industries.

Grace Shao: What really drove you to want to actually work on this, work on this technology and the future mobility?

James Peng: I think the motivation was twofold. One is that the potential, both commercially and also societal benefits for the autonomous driving is so huge. Think about like everyone needs to have some sort of mobility. Autonomous driving is much safer than a human driver. So it has huge societal benefit of saving people’s lives. So essentially, it’s just such a great industry to work on. Although back then, 10 years ago, it was very unclear when this can be done. The other reason is, of course, because the sheer technical challenge of autonomous driving involves because I was actually in my previous jobs. I worked on different areas, software, hardware, large scale distributing systems, AI and whatnot. But none of the things I worked on is as complex as autonomous driving, which is a field that involves hardware, software, hardware and software integration and Many other things. There’s AI, there’s real time system, there’s also large scale AI training and all that.

James Peng: So just from a sheer technical point of view, it’s such an amazing and challenging thing to work on. So I think those two reasons propelled me to start the company.

Grace Shao: There’s definitely a lot to unpack there. I think later on we can definitely double click on the hardware, software integration, as well as the safety concern there. You say that autonomous driving is much safer than humans. For sure, it’s safer than me driving. I know that. But some may argue otherwise. So let’s talk about that later. But first, I want to ask you about something that was quite interesting. During 2020-23, there was a bit of a public reckoning, I think, within the industry. A lot of peers folded during that time. People decided to pull out of this sector. Some people worried that autonomous driving would really become a reality. But you guys charged ahead and you really believed in your vision. Tell us about that period and how maybe that changed your vision or your growth mentality.

James Peng: I think 2020-23 was a period of time where the autonomous driving industry has evolved for roughly 10 years. I think that was the time of reckoning. That’s the time where the haves and have-nots have really diverged. So I think that’s actually exactly the time. As a company, we have seen tremendous progress. At the end of 2022, beginning of 2023, that was the time we actually finally had the first fully driverless commercial applications operations on the road. So because we made such progress, both from a technical and also from a regulatory point of view, that, of course, we made the progress. We finally see the glimpse of hope. Then, of course, we charge ahead. I think a lot of the other companies who weren’t able to, either from a technical point of view, or from a pure capital-raising point of view, or from a regulatory approval point of view, that weren’t able to have fully driverless applications. Then they were faded away.

James Peng: So it’s sort of like, well, everyone is in school. There’s no big difference. But after graduation, then there’s haves and have-nots. So I think that was the time of division.

Grace Shao: Yeah. So speaking of milestones, I want to kind of go back into history a little bit. So in 2021, Pony.ai had the third highest number of miles driven behind Waymo, Cruise. In 2022, Pony.ai became the first autonomous driving company to get a taxi license in China. In 2023, Pony.ai was licensed to operate robotaxis in Guangzhou, etc. And expansion continued. So kind of following what you just said, there was good momentum behind you guys. Now, today marks Pony.ai’s 10th year officially. You kind of talked about how you guys have grown. But what was one or two of the biggest milestones that you’re really proud of looking back now and that you think have really set the tone for your company Now as you are really expanding globally?

James Peng: Yeah, I think in my view, the biggest milestone, actually, I have already mentioned, is the end of 2022, beginning of 2023, where we were granted the fully driverless commercial license in both Beijing and Guangzhou. We start to have the operation to the general public. Actually, it was in mid-January in 2023 that I was the first road in our commercial robotaxis operations in Beijing. Surprisingly, it was exactly on that day, it was snowing in Beijing, and I was in the vehicle by myself. That was the moment where I actually saw our vehicles were able to drive by itself. Anyone besides me in the vehicle. Because of the snowing, it was also a very challenging scenario. We were actually not being suspended for operation. We continued to operate, and I was in there. That was the moment. Finally, it felt like a dream come true, right? Finally, it’s not just because our technology is ready.

James Peng: Also, because we actually got the approval from the government to have the license to operate. So it’s like all the seven plus years of efforts finally pays off. To me, that was felt like, as Lyndon Johnson said, the small steps for a person, but a giant leap for the human race. Although I wouldn’t call it as big as the Apollo, but to me, it felt like it’s finally from zero to one. So I think that was a deciding moment or defining moment for Pony.ai.

Grace Shao: That’s a personal Apollo moment. I love how you visualize it because I could just imagine how chaotic the roads were in Beijing. Also to be quite romantic when Beijing is snowing because it’s such a beautiful city. Okay, but let’s talk about what is a robotaxi and how the public actually even felt about it when it first rolled out. Before we started recording, Ivy was even telling me, I was like, hey, look, I get a bit scared when I see Waymos on the roads or Pony.ai vehicles when there’s no one Driving behind the wheel. Now, that’s because I’m not used to it. You said, oh, yeah, it’s okay. People get used to it eventually, right? But let’s look back at 2022 when it first was deployed to the public. What’s the public’s reaction?

James Peng: I think because it was a gradual process in the operational domains, in the operational zone that we had. We used to have a safety operator behind the wheel, although the driver actually didn’t touch the wheel or push the pedal. But people gradually get used to it. Actually, at the very beginning, when we were just deployed in Guangzhou, in those days, if you look at the picture of our first and second generation of Autonomous driving vehicles, you still see those spinning LIDARs on the top, and they were very much visible. People were curious. But gradually, people are just getting, this is like business as usual. As a rider’s point of view, the experience of a robot taxi is exactly like a typical taxi. The only difference is there’s no driver inside the vehicle, right? So the way you get the vehicle, the way you get in and get out is exactly the same.

James Peng: Also the other traffic participants, like the pedestrians and cyclists, they get used to it. So I think it just takes time. It’s just like the first cell phone comes out, the first real smartphone comes out. People were very curious. Now it’s just, nobody cares about it. So I think it just takes time.

Grace Shao: It normalizes eventually, right? Absolutely. I think we’ve really had a few years of consumer education done by quite a few of the players, including yourselves. Okay, well, let’s talk about the technical side of things. For outsiders, people might not understand the nuances between L2 and L4. Increasingly, we’re getting closer to L5 supposedly, are we? So help us understand your thinking on there. How do you structure your own teams, your products, working on different technology? Who gets held accountable for the actions in an L2 vehicle versus an L4 vehicle? Then finally, are we getting a glimpse into the future of L5? Are we going to be able to complete the road anywhere we want with autonomous vehicles? It’s a big, broad question, but I’ll throw it to you.

James Peng: Yeah, sure. So the definition of the level of automation for vehicles was actually defined about 20 years ago. So, of course, the industry evolved quite a bit. I don’t think that the levels from L0 to L5 might be the right way of defining what the level of automation is. So in my opinion, actually, there are two different products. One is what we call the driver assist systems, ADAS. The other is fully driverless. So in a broad sense, I think there are two categories. There are definitely two different products. The biggest difference is not on the technical side, but rather, as you mentioned, probably on the regulatory side, is who is first in line for the responsibility if there is ever an accident. I think for any ADAS system, any driver assistance system, it’s always the driver behind the wheel that’s responsible. Regardless if he or she is looking at the road or has their hands on the wheel.

James Peng: Whereas for the fully driverless systems, it’s the system that’s first in line. Because of that requirement, right? Think about if there’s a driver behind the wheel, it sort of serves as a safety net. So the system does not need to be bulletproof. It’s probably, well, as long as it can handle 99%, the case is probably good enough. Whereas for fully driverless, it has to be dealing with all the edge cases, all the extreme cases, and have a fallback system. We can get into those details later. But essentially, in my opinion, there are two different products. Of course, for the driver assistance systems, there are different levels, right? You can be, say, only highway or there’s only keeping in lane. Or they will actually even be able to handle some of the automations in the urban environment. For the fully driverless, of course, as you mentioned, there’s L4, L5 in a traditional definition. L4 means in certain areas. It can be fully driverless. L5 is everywhere.

James Peng: But I think it’s never a clear division. Essentially, you can think of it as how we drive, right? We start with the area, then we gradually improve. Eventually, it will be everywhere. So I think that will be a gradual process instead of a clear division.

Grace Shao: So actually, I want to double click on what you just said. So then help me understand, what is the gap between L4 and L5? Right now, Pony.ai is at L4, right? They’re robotaxis. Is that correct? How am I understanding this?

James Peng: No, I wouldn’t call them a gap. I think it’s a different product definition. Because they serve different purposes. I think most people view this as a process of evolution, right? From L0 to L2, L3, L4. But it’s actually a wrong way of looking at it. As I already mentioned, because the clear difference is that who’s first in line with responsibility. That’s decided by regulatory, actually. By product definition. By regulatory as well. So because of that, it’s essentially two different products. As the product is getting more and more mature, getting more powerful, in my personal view, the division of two different products is getting wider instead of narrower.

Grace Shao: Okay, then I’ll push on this. Then what is the real bottleneck right now for companies like you to deploy at a faster scale? Or to go into more cities quicker?

James Peng: I think that’s the reason that I wouldn’t say it’s one single blocker or one single bottleneck that prevented us to grow faster. I think it’s because the sheer complexity of the autonomous driving and what entails to ensure safety. There’s regulatory, there’s technical things. We also, because this is such a brand new system, that we need a manufacturing capacity. We need deployment. We need to get all the operational things ready, like all the garage space and whatnot. Also user acceptance, user education, as we already mentioned. I think all those take time.

Grace Shao: I believe also we have different partnerships with different managers of your local fleets. That kind of know-how also takes time for them to understand, to transfer over, right? For them to manage robotaxi fleets versus human fleets.

James Peng: Absolutely, absolutely. All those takes time.

Grace Shao: So I want to bring it back to technical. You have said publicly that you use the least compute footprint to reach L4. I thought that was quite interesting. Help us understand how you achieve that. How the model on the vehicle versus the large model you train in the labs actually work together.

James Peng: I think all the AI systems more or less take the same approach, is that you have data on the backend, on the data center side. You train a large model where you essentially try to get all the cases to be learned. In our case, we use the word model, where you can think of it as a simulated city, where we train the virtual driver and let us drive on all different kinds of roads and learn the driving ability. So that’s what’s condensed as a model from all the learning that we deploy on the vehicle. In the traditional AI sense, that’s called edge computing. You put it on the edge, put it on the devices, and put it on the car, where it’s a much smaller model. In a human sense, it’s like we learn everything. Then when we go to a test, we don’t need everything. We just need to be able to have the ability to handle the test.

James Peng: So that’s typically the training, where the backend system needs a lot of computing, but on the actual usage side, you don’t need that much computing power. So when the car is running on its own, it’s actually only using the model on edge, essentially.

Grace Shao: Absolutely. I see. Okay, so let’s talk about AI systems, because AI systems for language, images, code have improved dramatically. There’s also obviously a lot of hype right now around world models, but what you’ve been describing actually has been something that’s not been coined world models for a decade, over a decade. What has generative AI done for you guys? How has it changed how you view your own AI system? Do you think, I guess, the word world models do your system justice in that sense?

James Peng: Yes, it’s a little bit different, and they’re also related. Again, use human as an analogy. It’s actually quite similar to how we think, right? Because think about the large language model, how we process image, how we process knowledge. It’s sort of related to our memory and our logical areas of the brain. Whereas when we drive, it’s not just the memory and our knowledge. It’s also how we react, how we action, and all that. So the example is, one is related to how we learn a new skill. That’s the language model. Whereas for driving, it’s like how we learn to ride a bike. It’s actually different types of brain, different types of skill sets. That’s why it’s different. It’s not the same AI, because that’s how humans deal with different skills for knowledge.

Grace Shao: So Gen-AI has not affected you, but how do you view the idea of now calling, I guess, the physical AI world, world models? Because you guys have been doing this for more than a decade. That’s kind of my question, I guess.

James Peng: Yeah, that’s why I’m trying to get to it. For language model, it’s related with knowledge, with language, with logic. That means you need to be a very large model. Because think about it, if you don’t know a historical event, there’s no way you know it. So you have to have all the knowledge of a human ever created in your model for it to be powerful. So that’s why a large language model requires a lot of computing power and memory and everything. Whereas for driving, that’s how we learn riding a bike. We don’t need to have a PhD degree to learn how to ride a bike. But rather, it requires a lot of practice and training. That’s what world model is related to, or is assembled like. It essentially is a model where the virtual driver can start learning by itself, to learn how to interact with other cars, cyclists, pedestrians, and then learn the driving skill out of that.

James Peng: So it’s a bit related with the large language model, but it’s quite different. Because it’s related with action, related with manipulation, related with collision avoidance. So that’s the key for the world model.

Grace Shao: So tell us about how you simulate these systems. How do you leverage simulation systems for these edge cases?

James Peng: So essentially, that’s how we learn how to drive. There are several key factors for the world model. One is it needs to be very real. So we call the fidelity. It needs to have high fidelity, means it resembles the real world. Second is everything that moves in the world model, meaning cars, pedestrians, needs to be smart. Meaning that because the thing that’s driving is the interactive process. Our action, because we constantly make decisions in there, our action will affect everybody else around us. So they need to react accordingly. So it’s interactive. It’s more like interactive gaming, where we react with everything else. So that’s the second challenge is all the interaction needs to be smart, needs to be intelligent. The third challenge is how do we evaluate what is a good driving? You can interact and everything. You avoid collision. Is that a great driving? No. Not enough, right? Because there’s a passenger inside. Comfort is important. Efficiency is important.

James Peng: From A to B, we want to use the minimum amount of time. So essentially, it’s a multi-metric evaluation system in there to see what is a good driving. So there are three key challenges for the world model. We certainly, all our effort developing the world model related with that three areas.

Grace Shao: But human drivers can be so emotional, right? Either you can be scared or you can be rage driving. Or we can be communicating sometimes without obvious signs, right? We’re looking at each other. We communicate with eye contact, hand gestures. How do you train your fleets to understand human behavior right now? Because obviously, human drivers are still the majority of drivers on the road today. In your case, I actually think you’re right. At one point, maybe removing all the human drivers will make it even safer, right? Especially removing drivers like myself, if I say it again. But yeah, how do you actually help these cars understand all these non-obvious signals? Not someone quite directly clashing onto you. Someone forgetting to turn on the turn sign. Someone stop sign looking at you, waving to go.

James Peng: Like all the nuances. Absolutely. See, that’s why the first thing is how we become a better driver. Essentially, it’s a continuous learning process. The first thing, that’s how we learn how to drive, right? The first thing is you avoid collision. You were a cautious driver. Then gradually, you start learning a bit of everything else. All the signs, all the nonverbal cues, and the hand gestures. So that’s exactly the case for us as well. The earlier model of our system is just driving and try to avoid collision. Then gradually, we put a lot of new things, new recognitions, new perception models into our system where we start recognizing, for example, the hand gestures, especially all the policemen, all the typical police gestures, stop, Go, and all those things. Then we start recognizing, for example, potholes on the road, small obstacles on the road. So it’s sort of how we learn. We start getting all the big pictures first.

James Peng: Then we start learning all the nitty-gritty details down the road and put them to enhance our system. Regarding the second point, you’ll see, when all the cars are autonomous driving by themselves, it will be easier to drive. Yes or no? Because the thing that the big, especially in China, in the roads in China, the biggest challenge is not the other vehicles. In a lot of cases, it’s cyclists and pedestrians. While we can’t make them to be autonomous driving, so I think having the ability to recognize pedestrians, recognize the intention, their sign, and give You a specific example on a crosswalk, the way pedestrians look at and how they pay attention. For example, if they want to directly cross, they typically look straight. But if they were looking back, that means they will more likely not to cross the crosswalk. So we actually take those cues to decide whether we let them cross or proceed straight ahead.

James Peng: So a lot of those details need to be put into the system to make it safer and at the same time efficient.

Grace Shao: Is the judgment made on the spot using the cameras?

James Peng: Yes, using all the sensors. Cameras and LiDAR provide the sensor input.

Grace Shao: We take it all and then we make the comprehensive decision based on the input. Definitely China has more complex and less predictable driving conditions, especially given the number of pedestrians, cyclists, motorcycles we just talked about. So if you can drive safely there, I bet you can drive safely anywhere. But jokes aside, it’s really interesting because, we talk about as your fleet grows, you accumulate more and more world data, real-world data. Is that kind of data eventually becoming an advantage and a serious edge for incumbent fleets and a structural barrier that makes it very hard for new entrants to compete then?

James Peng: Data is important, but data is not everything. So how we understand that is this. Probably give you an example. Think about how we learn. Let’s say we learn math, right? You can think of the data is like the practice sets that we have. Of course, you need to do enough of practice to be a good knowledge about the subject. But doesn’t mean that you have the problem sets of the whole world that you become math experts. So that’s exactly the same case. We need enough of data sets in order to know what the real world driving condition looks like. But we don’t need everything because once we know enough, we can always generate enough knowledge about the driving. So in a way, I think the driving data is important, but it’s not everything. So that’s exactly how you view this.

Grace Shao: So as we speak of this, how do you view the whole landscape right now? Who would you say are your biggest competitors globally? How do you view the different markets playing out?

James Peng: Yeah, that’s a very complex problem. I think a question to answer because I think that I think the first and foremost, I think maybe I got some premises on this. First, the whole mobility industry, especially related with autonomous driving, is very large. They certainly have enough room for several players. Second, it’s still at a fairly early stage for the fully driverless. I don’t think the landscape is already divided in the set. So giving that two promises, I think currently, when I look at the players, I have to judge their current deployment. Although everybody can say, oh, they will have, they will, they will have thousands, hundreds of thousands of vehicles on the road. Actually, giving the current situation, I use the metric as having fully driverless commercial operation as a baseline. Giving that as a factor, I think in the US, Waymo is definitely leading the way. Because Waymo already have 4,000 or 5,000 vehicles on the road, 4,000 plus.

James Peng: Then, of course, there are some other players trying to play a catch up. Zoox, Cruise, maybe Tesla as well. So there’s, of course, some. So I would say in the US, Waymo is leading the way. There’s three to five players trying to play a catch up. In a global sense, I think from a technical point of view, China’s player is certainly on par with the US players. But from the total cost or the economical sense of a vehicle, for example, our vehicle is four or five times cheaper than Waymo’s vehicle. So in the global markets, such as Europe, such as Middle East, I think we will play a huge edge compared with the US players. Certainly, the whole landscape is still evolving.

Grace Shao: But especially in the global markets, you’ll see we will definitely not play a catch up, but taking a leading position. You’ve been an advocate for hardware optimization, software optimization, battery solution optimization. Is that the strategy behind being able to have a vehicle that’s four to five times cheaper than Waymo’s? Or where’s the edge? Or how are you building these comparable vehicles at a relatively cheaper cost?

James Peng: Yeah, I think as you mentioned, you definitely mentioned the most important factor to have the much cheaper price on the vehicles, which is optimization on Hardware, software, and everything else. I think another reason, of course, is because the whole ecosystem related with autonomous driving in China is relatively mature, and the scale is larger. So that price is cheaper. For example, the vehicle itself, the sensors, they’re relatively cheaper in China than anywhere else. Because of the ecosystem, because of the scale. So that plays an important role as well. That touches on something. A lot of physical AI, a lot of robotics companies are also now leaning into the Chinese supply chain. A lot of your peers, actually, autonomous driving, or even the EV players are now looking to expand into physical AI, whether that’s robots, humanoids, Quadrupeds, whatnot.

Grace Shao: So you’ve stayed really focused. You’ve not launched any robots out there or anything. What’s your thinking behind this?

James Peng: Yeah, absolutely true. I think autonomous driving definitely is probably the first large application of physical AI. All the others, humanoids, robots, and everything else, probably will have real applications down the road. For us, we view the autonomous driving as our brand and partner. Of course, as I mentioned, this is still early stage. I think we still have a lot of mileage to go. For all the other physical AI applications, we don’t have any specific plans yet. But I think they’re definitely interrelated. We may enter them down the road, depending on whether we need it or not. Because my judgment is that for the physical AI, it probably will follow the similar trend as autonomous driving. It might take another decade for it to mature. I think for us, it’s more like whether we have real applications for it. Give you a specific example.

James Peng: Even for our fleet, once we go to hundreds of thousands, millions of vehicles, how we maintain those vehicles, how we do the charging, cleaning, servicing, They may use robotic applications. So I guess my view is that we will not probably do robotic actions just for the sake of doing it. But we may do the related applications when we see the real applications.

Grace Shao: So it’s fair to say you’re cautiously optimistic that there is a potential use case further down. But it’s nowhere close to where it’s been hyped in the three to five years kind of use case.

James Peng: Yeah, I think it’s the same thing as autonomous driving 10 years ago.

Grace Shao: All right, well, let’s talk about your international footprint. You mentioned earlier, you have a global strategy. You’re in Europe and Luxembourg was your first launch, right? You’re in Southeast Asia, parts of East Asia, you’re in the Middle East growing really fast over there. Tell us about how you think of your next steps in your global expansion.

James Peng: Yeah, I think the mobility demand everywhere is the same, right? There’s a strong demand across the globe. But we have to focus on the most important markets first. I think eventually we’ll go everywhere, because that’s our motto is we have autonomous mobility everywhere. That’s our ambition when we started 10 years ago. But our first launch, we have several criteria. One is related with regulatory, right? It needs to have relatively accommodating regulatory environment. Second is it needs to be a relatively mature mobility market. In a more obvious sense is that the local taxi fares needs to be relatively high, because I think that our pricing anchor point is always a human driving Taxi. So that price needs to be relatively okay. The third criteria are that we need a good local player to partner with, because a lot of other things like regulatory, like the back end services needs to be Handled by the local partners.

James Peng: So judging from that three categories, I think Europe, Middle East, Southeast Asia, Japan, South Korea, Australia maybe. Those will be probably the potential markets for the initial launch. Of course, those are already big enough of number of countries. So we’ll pick and choose some to start with.

Grace Shao: How do I understand your partnership models? Because I believe you’ve quite a few different kind of models depending on the location, the regulatory environment, potential partnerships, know-how, etc. Tell us about that.

James Peng: Maybe I’ll take one step back first. Think about what is a robotaxi industry. The type of players, I’ll divide them into four categories. One category is for the user acquisition. Those are ride-hailing applications. Those are the Ubers and the Lyfts and the DDs alike. The second is a vehicle. You need a car, how you manufacture a car. The third is a driver. The fourth is all the back end services, cleaning, charging, servicing, insurance, and everything else. For us, our main job is creating a virtual driver, is making a really safe, efficient driver. So that’s definitely what we do. All the three other categories, we might have partners, we might do ourselves. So that, depending on the market, depending on what’s the strong local players, we might pick and choose players who’s handling one or two or three of the Other things. For example, we work with the ride-hailing platforms for the user acquisition.

James Peng: We work with some of the back end services who’s providing the parking space, who’s cleaning, charging our cars. We also have OEM partners that work on the cars. So that’s how we view the partnerships landscape.

Grace Shao: So after you deploy, say you send these out to Australia, what happens walk us through that. Because once these cars actually get off the boat and ships and they land in Australia, are they your responsibility or your partner’s responsibility? Do you send an engineer? Do you send your own management? Or do you transfer that know-how and maintenance know-how to the local partners to handle?

James Peng: Great question. That really depends on the different partnerships and different regulatory environment. In some markets, it’s the local player who’s first in line with managing the fleet. That means in those cases, we manufacture the cars with OEM. But then once we ship the vehicles to the local country, Australia, giving you an example, or Singapore, let’s say, then we actually, in those cases, we sell The vehicle to the local partner. It’s like selling hardware. It’s like selling hardware. But we will, of course, have engineers handling the driving because we are in charge of the driving. So all the driving related work will be done by us. But then the user acquisition, the cleaning, servicing, charging will be done by the local partner. So those are one case. But in some markets, we actually ship the vehicle and we apply licenses by ourselves. The vehicle is still on our own book. But those are rare cases.

James Peng: We actually, our preferred model is to have the local partner that handles most of the logistics and we will be the tech providers. We’ll essentially have the virtual drivers handling the driving and everything else is done by the local partner.

Grace Shao: I see. Would you ever view OEMs as competitors in any way? Because right now you’re partnering with them. You’re giving them the software enablement, right? Would they produce their own robotaxis?

James Peng: I think in most cases, in my view, that they probably will be partners instead of competitors. It’s very different because they were mostly on the hardware business. Very few of them will be in the robotaxis business because they’re quite different.

Grace Shao: I see. I see. Something a bit niche is, I know you run robotaxis as well as trucks. Walk us through how you think about that. Why do you guys also have a truck business? What kind of scenarios are they already being deployed in? I believe they’re the heavy trucks and then the light trucks. How do I understand this?

James Peng: Yes. As I already mentioned, think about our business is that all our technology is that we are creating a safe virtual driver. Virtual driver is our core. As a driver, you should be able to drive all different types of vehicles. The two biggest applications for driver is one is for the transportation of human beings and the other is for goods. That’s related with robotaxis and then for all the trucks. Within the logistic industry, there are actually three categories. One is for the long haul, which is typically done by the heavy trucks, the 18 wheelers and whatnot. Then there’s also in-city network, which is the light duty truck. Then there were also the last mile, typically is handled by much smaller vehicles. Our main focus, of course, is on the long haul and the intra-city transportation. On the last mile, we are the providers of the domain controllers, but that’s not the areas we’re working on. So think about we’re creating driver.

James Peng: Driver should be drive different types of vehicles. That’s how we view the trucks versus the robotaxis.

Grace Shao: Usually, I would assume these are like ports, airports, maybe?

James Peng: They will eventually be everywhere as well. We start with ports. We start with some dedicated routes, for example, like a minefield to the local distribution center, those 30-50 mile routes. The reason we start with those applications is because typically it’s mostly because of regulatory reasons. Because the ports and the dedicated routes and those are typically a semi -private road. It’s much easier to get the regulatory approval. Of course, we are working on long haul trucks as well. We already actually have a fleet of heavy-duty trucks doing the real goods transfers on highways, but still with a safety driver, of course. Eventually, we’ll be fully driverless as well.

Grace Shao: You’ve said you have a target of running fleets commercially across more than 20 cities by the end of this year. What do you know today that you could not have learned without actually operating at scale already on the streets? What makes you have the confidence to do that now, I think, compared to maybe a few years ago?

James Peng: Again, I think for robotaxis commercial business to be a reality, there are three important factors. One is technology. Second is regulatory approval. The third is user acceptance. I think within the last three to four years, we have gained a lot of experience on all three categories. The reason we were confident to deploy in 20 cities is because clear vision on the regulatory approval. There’s a lot of cities globally, both in China and in some global cities, they actually start coming out with regulations for supporting fully driverless commercial applications. Also we have planners. Planners want them. So I think all the important factors are falling into place. That gives us confidence.

Grace Shao: I’m going to play devil’s advocate a little bit here. With the rise of AI right now, there’s a bit of a fear of replacement of people’s jobs. The rise of autonomous driving obviously lead to job loss in people who are currently drivers. How do you view that? Because just now we talked about robotaxi drivers. We talked about people driving heavy-duty trucks that could potentially be replaced. Frankly, I’m in a camp that people could be maybe freed up to do more things that they can do otherwise. People will find alternative careers. But are regulators becoming more cautious. How do you feel about the current public pushback a little bit on AI, autonomous driving, autonomous everything at the moment?

James Peng: Yeah. Actually, driving is a hard job. Driving is a lot of cases in a stop vehicle for 10, 12 hours a day. It’s a really tough job. The thing that because autonomous driving itself is a highly regulated industry, the pace of our roll up is determined by the number of licenses. The thing about also a lot of the drivers were not young. The young generation, younger generations actually don’t want to be drivers. So I think, especially a lot of the global markets, we actually come in to fill the gap for the labor shortage for the driver. We’ll not change the human driving vehicles overnight. It will be a gradual process. So that’s sort of the development of the cities and the human society. It takes time. It becomes gradually a norm. Then, as you just mentioned, then the drivers can find other jobs.

James Peng: Even we actually absorb a lot of jobs, for example, for the remote assistance, maintenance, which are much safer and much less strenuous job conditions. So I think society as a whole has always a way to absorb jobs. To adopt, adapt, and then evolve.

Grace Shao: The current pay for a lot of times for these heavy truckload drivers are like 200 to 300k USD. They’re considered very high-earning jobs. But at the same time, people forget they’re extremely dangerous. There’s life lost constantly on the roads. So I can see that could be very valuable if people can actually replace those routes with robo-drivers.

James Peng: It’s not just replacing. Look at the truckers. Their average age is 45 plus. In North America right now? In North America. In China, they’re 40 plus as well. So a lot of younger generations, they don’t want that type of jobs. We’re coming not only to replace, but actually to fill the void for that job shortage.

Grace Shao: All right. So I think I want to wrap up our conversation soon about this. Is there anything I’m really missing, you think, about robotaxis and your business at this point?

James Peng: I think we’ve probably covered a lot of topics.

Grace Shao: Oh, I had one question. Another one about your business before we go into your personal thing. You mentioned Croatia just now when we were talking offline. I thought that was so fascinating. In my mind, I thought these robotaxis were being deployed mostly in futuristic cities like Silicon Valley and SF, out here in Shenzhen where we’re here today. But Croatia, help us understand the need for robotaxis in these countries where a lot of the roads are aged, are not really made for cars to start with, Are not easy to drive in, actually, even for humans. Then how does that make sense even for your economics, actually?

James Peng: Of course, there were some challenges. From a technical point of view, two challenges initially. One is there’s a lot of roundabouts. Actually, there were not many roundabouts in China. So although a lot of other very complex situations like heavy storms and whatnot, we were able to handle them really well. But roundabouts, we had some, but we haven’t trained that much. So we actually have to retrain a bit on the roundabouts. The second is the trams. There were just a lot of trams in the Zagreb. Their behavior of the trams is different from cars. So we need a little bit more training to get used to it. But it’s like how we drive. When we go to a new city, we might not drive as a perfect driver initially. But then we learn and adapt. Once we have a good learning system set up, then we can quickly learn. That’s exactly our experience in Zagreb, Croatia. Two things that we actually have to learn in Croatia.

James Peng: One is the roundabouts. The other is trams. Because those are not something that typically you will see on the roads in China. So for those new situations, it’s like how we learn. How we learn driving. When we go to a new city, we probably know 95%, 98% of the situation. Some of the scenarios probably we didn’t encounter previously. Then we learn. We adapt. So that’s exactly the case for us in Croatia. After three to four months of learning and training and retraining, we actually were able to handle those cases like roundabouts and trams really well. Because there’s a lot of roundabouts in other cities in Europe. They actually have different rules for roundabouts. Some of the roundabouts, I think the cars outside roundabouts have right-of -way. Some of the vehicles inside the roundabouts have right-of-way. But we can adapt once we have the system set up.

James Peng: So as I mentioned, the most important characteristic of our system is not how powerful it is, it’s how adaptive and how easy to learn on our system so that We were able to adapt.

Grace Shao: Brilliant. So a lot of localization as well for your vehicles. I have two last questions. One is, what is something you think people still get wrong often about your sector, in this case, autonomous vehicles, autonomous mobility? The second question is a bit of a curveball. I’ll throw it to you first, you can think about it. What is one differentiated view you hold? Something that’s a bit against consensus, maybe.

James Peng: Autonomous driving industry, I think people put too much focus on technology and probably underestimated the complexity of robotaxi as a business. Essentially, of course, technical is the most important. If you can’t drive safely, you’ll not have a business. But once you even have the most safest driving, you still have to, as a business, there’s a lot of other things involved. For example, how you deploy a fleet, how you make the pickup and drop off easy for the user, how you handle all the edge cases of the complaints of the Passengers, how you make the charging, servicing, cleaning efficient. For example, especially give you a specific example, the electricity fares during the day fluctuates. If you have the charging at the low fare, you can save a lot of cost. Then how you manage your fleet? Although you have the low fare for electricity, but the demand of the passengers is really high. How do you make a decision?

James Peng: So essentially, it’s a lot more optimization involved than just the driving itself. I think a lot of people underestimate the complexity with the management of a fleet of autonomous driving vehicles. We actually, as a company, have put a lot of emphasis and take a lot of efforts in optimizing everything. So that’s why I think those will be a very strong competitive edge down the road.

Grace Shao: Once you guys scale further, especially.

James Peng: Exactly, absolutely. Very interesting.

Grace Shao: The second one, I’ll put you on the spot again. What is one differentiative you hold?

James Peng: I think I’ll take the one related to the answer of my first question. Is that, again, people always put too much emphasis or give too much credit on zero to one and think about less for one to ten. Give a lot of examples, right? People always think an invention is so hard, but putting an invention to be a scaled application is equally hard or a lot harder. Because the scale involves cost optimization, involves user education, involves a regulatory approval, it involves making the things a lot easier to use. So many examples like this, right?

Grace Shao: Definitely. Say a rocket is put in the sky. Oh, it’s so hard. But having the rockets to always be able to safely take off and recycle, that’s extremely hard.

James Peng: So I think related with autonomous driving is we certainly crossed zero to one. I think we crossed one to five, maybe. But from five to ten, ten to a hundred, I think there will be still a lot of challenges ahead.

Grace Shao: That’s very insightful. I agree with you. When we look at the internet era and a lot of players that still stand today versus who are the actual ones that created a lot of the internet use cases we know of today. Thank you so much, James. It was a pleasure and an honor to learn more about your business, yourself, the man behind the company that is changing the future of autonomous mobility. Thank you again.

James Peng: Thank you for having me.

Have Chinese open-weight models collectively caught up with the frontier U.S. closed models? Not across every task, and not quite yet. Kimi K3 still trails the strongest OpenAI and Anthropic models overall, even as it appears to outperform them in some areas, particularly frontend coding. Its full open-weight release is also not expected until July 27, so some of the claims still need independent verification. For a detailed look at its performance and a closer look at the question “Have Chinese AI Models Caught Up to the US Frontier?” see here.

But can we stop calling every major Chinese model release another “DeepSeek moment”? When something keeps happening, it is no longer a moment.

Kimi K3 is the latest evidence that the AI frontier is becoming more contested, more global and potentially less proprietary. Chinese models may not consistently hold the absolute number-one position, but they are now close enough, cheap enough, and improving fast enough to change the economics of the entire market. And leaning into that shared R&D base, they’re able to iterate faster and learn from each other. But seroiusly, read this one below:

Kimi is not necessarily the cheapest Chinese model. But everything is relative. Moonshot’s reported pricing still comes in materially below Anthropic’s flagship Fable 5, which costs $10 per million input tokens and $50 per million output tokens.

That matters because the central risk to frontier labs is not that their models suddenly become useless. It is that frontier-level capability becomes increasingly difficult to monetize at premium prices when open-weight alternatives can perform most tasks at a fraction of the cost.

Now Kimi has received a lot of loving, and is now creating tiered pricing options and trying to scramble up more compute. But they’re not the only ones who have had to face this problem. At one point, GLM was also selectively choosing who they serve first; I'm pretty sure DeepSeek had this constraint too. Ultimately, enterprise users probably have the first pick (because they pay the bills), and consumers are just going to be prioritized until there is compute abundance. This is also really putting pressure on the domestic chip makers to step up and fill in this gap.

Token-maxxing meets budget backlash

It has been widely reported recently that OpenAI is exploring possible price cuts and facing growing pressure on model pricing.

Enterprises and startups are increasingly routing workloads toward cheaper open-source and Chinese models, including DeepSeek, Kimi, Zhipu GLM, Alibaba’s Qwen and, increasingly, Nvidia’s Nemotron.

Customers are no longer treating Anthropic and OpenAI as the default infrastructure for every task. Instead, they are building orchestration layers that select the cheapest model capable of doing the job and adopting the hybrid model approach we have written about extensively before. The most expensive frontier models are increasingly being reserved for difficult reasoning, coding, and other high-value workflows.

DeepSeek has been the most-used AI company on OpenRouter since mid-May. Among OpenRouter’s highest-spending customers, open-source token usage grew four times faster than closed-source usage between autumn 2025 and spring 2026. OpenRouter has also seen more than 500 organizations switch from proprietary to open-source models.

But I still do not think this spells the end for frontier labs at all.

Anthropic and OpenAI could retain the highest value share of tokens because their expensive models will continue to be used for the most complex and economically valuable work. Open-source and Chinese models could take the highest volume share, handling the much larger number of routine tasks.

Think Apple versus Android, rather than a winner-takes-all outcome.

For AI infrastructure suppliers, the read-through could still be positive. Lower inference costs should expand consumption. As AI becomes cheap enough to embed into every workflow, the number of tokens generated could rise much faster than the cost per token falls.

That is the Jevons paradox argument again.

Where is the moat?

For enterprises, the long-term use case may be less about renting access to the single smartest general-purpose model.

It may be about owning what makes the company special: its proprietary data, industry knowledge and workflows, and then choosing, customizing or fine-tuning whichever models best support them.

We wrote about this recently: who should own what makes a company unique and where value could migrate in a future of self-hosted, fine-tuned open-source models.

Why Didn’t Yang Zhilin Stay in the U.S.?

Finally, I want to address some of the commentary asking: why didn’t Yang Zhilin simply stay in the U.S.?

Honestly, it is a pretty ridiculous question.

Maybe it was personal reasons. Maybe it was lifestyle, family, familiarity, or simply a desire to build something in China. There is also something to be said for wanting to make it in your home market, in your native language, and within a culture you understand deeply.

For more on his personal philosophy and outlook on AGI, you can revisit this piece from a year ago.

Anyway, I read a good thread on X making this point. It is strange to suggest that China is somehow the easier place to build a leading AI company, or to assume that he must have left because he could not get a visa.

China’s venture-capital pool is much smaller. Competition is cut-throat. Monetization is difficult. The market is hardly an easy option.

At the same time, it would be naïve to pretend that rising Sinophobia does not create additional discomfort for Chinese researchers and founders living in the U.S. That may not be the main reason someone chooses to return, but it probably does not help.

Sometimes people simply want to build at home. That should not be so difficult to understand.

Ultimately, training these models remains part of a capital-intensive, energy-intensive, and increasingly politically contested infrastructure cycle.

Kimi K3 does not kill the AI trade. It does, however, make the distribution of value much less straightforward.

Frontier labs may capture fewer dollars per token. Hyperscalers may capture more of the orchestration layer. Enterprises may retain more value through proprietary data and workflows. Open-weight models may dominate token volumes. And the suppliers of chips, power, and physical infrastructure may continue benefiting as cheaper intelligence produces much more demand.

The frontier race is not disappearing. It is simply adding a few more players.

Coming up on the podcast is an exclusive sit-down interview with Pony.ai’s CEO, James Peng. Thanks for all the positive feedback on the episode with Meitu’s CFO, Gary Ngan. More to come in this series! CHECK IT OUT.

I went on Bloomberg TV on Friday morning to talk about the WAIC 2026 preview (you can see more.

As I write this, I’ve been watching the monsoon downpour in Hong Kong this past week and missing the conference, but I sent someone I trust who will give us a very detailed and informative report.

It was my own decision, given how tired I was from recent travel, though my teenage inner self can’t help wondering if I’m missing out after receiving at least 20 messages asking if I’m going.

Instead, I’m locked in on a Satya essay that I believe marks a pivotal point in how the industry is thinking about model commoditization. But seriously, this could be a pivotal moment.

Now, Satya Nadella wrote over the last weekend:

“You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.”

This echoed what some people called another episode of Alex Karp “losing it again” when he went on CNBC arguing that frontier labs could eventually take the data and alpha of the businesses using their models.

But as much as his presentation can seem chaotic, Karp has some logic to it:

“What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it’s not being transferred to someone else.”

It’s all colliding in this moment. Mira Murati’s Thinking Machines similarly made an argument from a more philosophical angle just this week. Closed models today are trained centrally and then largely frozen in time. They are not continuously shaped by the local knowledge of the people and companies using them. Thinking Machines argues that organizations should be able to fine-tune model weights with their own knowledge and keep adapting those models as that knowledge evolves.

The implication points in the same direction: a more distributed ecosystem of customizable and often smaller models that companies can fine-tune and potentially self-host, rather than relying on one centralized intelligence for everything.

Of course, everyone here is talking in their own favor. Palantir sells sovereignty and control. Microsoft benefits when enterprises use multiple models and build the orchestration layer on Azure. Thinking Machines is building tools that let customers train and customize models.

But when three such different parties arrive at a similar conclusion, perhaps there really is an underlying concern.

What I’m trying to work out is whether it now makes sense for companies to use more open and smaller models suited to their own industries, whether that will lead to more niche models, and whether the big labs will simply eat everyone’s lunch anyway.

Are the labs eating everything?

A recent LatePost (leading Chinese tech blog) article described the author attending ICML in Seoul and meeting several old friends who were building AI application startups.

Collectively, they seemed resigned to the idea that there was not much of a future for them. For now, they could innovate on product and collect user memory, context, and industry data. Eventually, they expected to become data companies and sell that data to the labs.

“We know our destiny.”

The fear is not simply that Claude will add a legal product or ChatGPT will move into healthcare and education. It is that application companies will spend years collecting the proprietary data and industry know-how needed to make a general model useful in a specific vertical. Once the labs understand those use cases well enough, they could build the niche applications themselves, and the only moat left is the data.

In other words, the labs may not begin with the knowledge required to serve a law firm, hospital, or design team. But the companies using their models could end up teaching them. So that is why people are realizing they shouldn’t be giving the labs their data anymore…

This does not require believing that labs are secretly training on protected enterprise data. It is more a question of incentives and bargaining power. To make a general model useful, a company has to show it what good work looks like, how its experts make decisions, which exceptions matter, and what its customers actually value.

That is the proprietary know-how that makes the company special.

So this is not really a question about whether companies should own “AI-generated intelligence.” Nor is it exactly a question about distillation, although distillation may be one technical response. It is about whether companies can use external intelligence without gradually handing over the data, context, and workflows that differentiate their businesses.

The thin-wrapper companies probably should be worried. If the application is mainly a cleaner interface around Claude or ChatGPT, the lab can eventually add the feature itself.

Models will also commoditize at different speeds across verticals, wrote Bernstein in its June 12 report. Consumer applications and lower-stakes generalist business uses are likely to go first, while highly specialized domains that demand extreme accuracy or have near-unlimited willingness to pay — drug discovery, advanced science, certain creative fields — should remain differentiated for longer.

But a company that owns a deep workflow, proprietary data, and a customer relationship is in a different position. The model may help deliver the product, but it does not automatically own the knowledge required to make that product work.

The solution, as of now, is probably hybrid

The answer is not for every company to become a frontier lab, nor is it realistic to even try to fine-tune their own models.

For much of the first phase of generative AI, US software companies paid for the best closed models because that was the quickest way to launch. Speed mattered more than cost, and almost everyone was still experimenting.

Now the tokenmaxxing phase is running into economic reality. And just as Bernstein wrote that the cost of Fable 5 tokens has driven a rethink of AI model usage by intelligence and cost…” people are rethinking what is ‘good enough.'

Once usage reaches production scale, companies have to ask whether every customer-service request, internal search query, or product recommendation really requires the most powerful and expensive model available.

The answer will increasingly be no.

A frontier model can still handle the hardest tasks or serve as a teacher. Smaller open or first-party models can handle high-volume, repetitive work where the company’s own data and workflow knowledge matter more than broad general intelligence.

That is why the eventual architecture is probably hybrid, which is what we’re seeing with AI-native firms already. Companies will route different tasks to different models, using closed frontier models where they are genuinely better, open models where they are good enough, and fine-tuned internal models where control, cost or proprietary data matter most.

Open models, fine-tuning and self-hosting are not the same thing, and not every company needs all three. But together, they give companies more choice over how much of their know-how they expose and how dependent they become on one lab.

This would also move more value toward training, fine-tuning, orchestration and inference. Cloud providers will likely benefit, but instead of simply reselling frontier-model tokens, they may increasingly provide the infrastructure on which companies build their own intelligence layers.

Did Chinese companies reach this conclusion earlier?

Remember how Xiaomi, Meituan, and a series of Chinese internet companies all seemed to be rushing to release models? Our initial thinking was that 1/ China’s internet companies' expansionary nature, 2/ ego, 3/ open-source lowered the barrier to entry for R&D, BUT MAYBE they figured out something we didn’t.

At the time, some of these efforts felt slightly baffling. Why did a smartphone company or food-delivery platform need its own foundation model?

It is starting to make more sense now.

For much of the first phase of generative AI, American application companies could simply plug into OpenAI or Anthropic. Chinese companies rarely had that luxury. Access to the best US models was uncertain or simply blocked, while Chinese internet companies have a long history of not trusting competitors to provide critical infrastructure. That probably played a role in this 大家卷起来~*juan*-ness too.

China’s open-source ecosystem also offered something closer to a shared R&D pool. Companies did not need to rediscover every architecture or training technique themselves. They could build on DeepSeek, Qwen and other open releases, then adapt them to their own needs.

So Xiaomi, Meituan, and Meitu began treating models as part of their own technology stack, even when building a model did not initially seem central to the business.

Xiaomi does not need MiMo to beat every frontier model. It needs models that can operate across phones, cars, and smart-home devices while giving Xiaomi more control over latency, personalization, cost, and the data moving through that ecosystem.

And that same logic applies to Meituan’s LongCat. Restaurant discovery, delivery, travel booking, and merchant operations do not necessarily require a model capable of PhD-level physics or mathematics. They require reliable and efficient intelligence embedded inside workflows that Meituan already understands.

Meitu is probably one of the clearest cases. Its advantage was never simply having a photo-editing app. It came from years of understanding how users, particularly Asian users, wanted to edit faces, images, and videos. It knows which outputs people save, which they reject, how aesthetic preferences differ, and what quality level persuades someone to pay. You can say its niche, but it is the best at doing exactly that.

CFO Gary Ngan told me in the interview linked above that more than 90% of Meitu’s AI outputs already come from its own models. The company uses its own fine-tuned models where they perform better and calls third-party APIs when it encounters edge cases its models do not cover. When it sees enough users repeatedly asking for the same thing, it can then train its own model to handle that scenario.

He was also quite clear that Meitu is not trying to build a general-purpose model. Visual standards are subjective; different verticals interpret the same prompt differently, and every vertical has its own workflow and standards. That is why he does not believe that a single general model can simply destroy every visual application. It comes down to aesthetics and taste, but that takes not just human taste but actual experience and data.

That is a much more grounded version of the hybrid strategy we can probably expect to see more — build your own models where your proprietary knowledge matters, use third parties where they are stronger, and keep improving based on what your own users actually do.

In some ways, constraints again simply pushed many Chinese models toward this architecture earlier. What once looked like needless model-building now looks more like an effort to avoid outsourcing a critical part of the business.

Alibaba and Tencent complicate the story

There is still a big counterargument.

Perhaps value is not moving away from model providers. Perhaps it is moving toward model providers that also own infrastructure, applications, proprietary data, and distribution.

Alibaba owns Qwen, Alibaba Cloud, and a huge consumer and merchant ecosystem. Tencent owns Hunyuan alongside WeChat, games, advertising, video and cloud infrastructure.

They are not standalone labs selling intelligence through an API. Their models can be deployed across internal products, improved through real use cases, and distributed through platforms that already have enormous customer bases.

In that world, Alibaba and Tencent may keep a lot of value inside the model layer precisely because they control everything around the model too.

So the real divide is probably not model providers versus application companies. It is between companies that own a strong combination of proprietary data, workflow, and distribution, and those that own only one thin layer. And we’ve written extensively on the labs’ need for applications for consumer distribution.

The labs will eat plenty of applications, especially products that do little more than package capabilities that the model can soon offer itself. More niche models will emerge, but many will probably live inside companies like Meitu, Xiaomi, and Meituan rather than become standalone labs.

And global companies will increasingly adopt hybrid model strategies, not because frontier models are going away, but because using the best model for everything is expensive and potentially gives one external provider too much knowledge and control.

The model may be rented. What makes your company special should not be.

We spent a lot of time on Meitu’s business edge: why visual AI is not just a foundation-model race, and why aesthetic judgment, controllability, and vertical context matter. Gary argues that visual creation is highly subjective. The same prompt can mean very different things across countries, cultures, product categories, and commercial goals. That is why Meitu is building verticalized products such as Picchi, DesignKit, Kaipai, Vmake, and RoboNeo, instead of relying only on one general-purpose AI model.

We also discussed the business model. Consumer subscriptions have become Meitu’s main revenue engine, while advertising is no longer the strategic growth driver it once was. Gary explained the shift in Chinese consumer willingness to pay for apps, the higher ARPU potential in overseas markets, and how new AI-native products like Picchi could introduce additional monetization through personalized models and AI credits. He also addressed AI compute cost, why more than 90% of Meitu’s AI outputs come from its own models, and why the company sees AI as a TAM-expanding opportunity rather than simply a margin risk.

Finally, we covered competition and globalization. Gary explained how Meitu thinks about competing with ByteDance, Kuaishou, Canva, Adobe, Shopify, Alibaba, and other AI-native visual tools, and why Meitu’s approach is more vertical-driven than general design-platform driven. Lastly, we touched localization, from different beauty preferences across markets to why true globalization requires understanding culture at a much deeper level than translation or marketing campaigns.

CHECK OUT THIS CONVERSATION. Gary’s so cool.

Share

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from analysts, VC investors, builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Chapters:

00:00 What is Meitu today? Mapping Meitu’s product portfolio
04:14 Why vertical focus still matters in the age of AI agents
06:36 Aesthetic standards, subjective prompts, and visual AI nuance
11:31 How AI changes art and creative expression
15:02 MeituHub and MiracleVision as visual AI infrastructure
17:01 Why Meitu needs its own models
20:55 How Meitu chooses models and the role of designers
25:09 Meitu’s AI legacy and generative AI strategy
28:27 AI compute cost, ROI, and gross margin
38:22 Subscription growth and advertising dependence
40:47 Partnerships with consumer chatbots and platforms
43:27 Competition with ByteDance, Kuaishou, Canva, Adobe, and others
49:49 Deepfakes, misuse, and AI safety safeguards
52:42 Globalization, localization, and cultural differences
59:16 The biggest investor misconception about Meitu

Transcript (AI-generated, for reference only)

Grace Shao:
Gary, thank you so much for joining us today.

Gary Ngan:
Hi Grace. Good to be here.

Grace Shao:
I’m really excited to have this conversation. To start, tell us what Meitu is up to these days. For a lot of investors and users, when they think of Meitu, they still think of the selfie and beauty-editing app. How would you define Meitu today? Is it still simply a consumer AI company, or is it much more than that now?

Gary Ngan:
Meitu is no longer just a selfie or beauty-editing company. I would define Meitu today as an AI application company specializing in photo, video, and design.

We focus on very high-value verticals where we can leverage AI to deliver high-quality results to users. We often refer to these users as prosumers: people who have strong design needs, but no prior formal design training.

So that is how I would define Meitu today.

Grace Shao:
That makes sense. Tell us about the products, because you have quite an array of them. Some are more consumer-facing, some are more prosumer-facing, and some may even be a bit more enterprise-facing. There is Meitu, BeautyCam, Wink, Picchi, DesignKit, Kaipai, Vmake, RoboNeo. Help us map out the ecosystem.

Gary Ngan:
We think about Meitu’s product portfolio in two main categories: applications for leisure and applications for productivity.

Applications for leisure include the Meitu app, BeautyCam, Wink, and Picchi. They serve use cases such as photo-taking, photo editing, and video editing, usually for sharing on social media.

The Meitu app and BeautyCam are our core consumer applications. Wink extends our capability from photo to video editing. Picchi is our latest portrait-retouching agent, focused on personalizing editing styles.

The second bucket is applications for productivity, which includes DesignKit, Kaipai, Vmake, and RoboNeo. These products serve professional and commercial content creation needs.

DesignKit focuses on e-commerce product-listing design. It helps merchants and creators produce product images, model images, and marketing materials much more efficiently.

Kaipai and Vmake focus on talking-video and marketing-video creation. Kaipai is more focused on the domestic Chinese market in verticals such as insurance and real estate, while Vmake is seeing strong traction in the U.S. fitness and wellness market.

To give you a sense, as of May this year, Kaipai had about three million monthly active creators, and cumulative content creation exceeded 400 million pieces. For Vmake, ARR in the first quarter of 2026 was about US$4 million.

Then there is RoboNeo, our AI-native agent product launched in July 2025. It is currently targeting the AI short-drama vertical. Its agent workflows can support scriptwriting, characters, storyboards, visual generation, and asset management.

So that gives you a rough idea of the different vertical products. But the ecosystem logic is very important, because many new products come from user insights we observe in existing products.

For example, DesignKit came from the poster-design function within the Meitu app. Kaipai came from the AI teleprompter feature in BeautyCam. Picchi came from new user behaviors we observed in the Meitu app.

So our portfolio is not a random collection of apps. It is a structured expansion from consumer imaging into AI-native workflows across photo, video, and design.

Grace Shao:
That makes a lot of sense. But in the age of AI agents, would it make sense for Meitu to consolidate a lot of these apps? Or do you still think it is better to keep them separate for different types of users and workflows?

Gary Ngan:
In the age of agents, we still believe we should focus on high-value verticals, because different verticals have many differences.

First of all, aesthetic standards are very different. I’ll give you an example. The phrase “handsome guy” would be interpreted very differently in an application serving the U.S. market versus an Asian market.

Even within the Asian market, if you are addressing e-commerce merchants selling gym products versus formal apparel, the word “handsome” will also be interpreted very differently across those verticals.

So being able to separate these different verticals gives you a very good head start in focusing on the aesthetic standards that each vertical needs.

Also, users in different verticals have very different behaviors and workflows. It is very important to build those workflows and that know-how into each vertical in order to create the right products.

With agents, you can cover a slightly bigger boundary. But I still think you want to focus on different verticals to maximize the output for the user, and also make it more efficient and easier to market within each vertical.

Grace Shao:
That is really interesting. You touched on something that a lot of people discuss when they think about visual AI, which is how to ensure consistency and accuracy when translating language into visuals, especially when text can be in different languages and words can be subjective.

As you said, if you say “handsome” and I say “handsome,” that could mean very different things in our heads. How do you ensure that identity, description, and nuance are not lost? You mentioned vertical focus, but what is the technical side of that?

Gary Ngan:
Instead of calling it fragmentation, I would say vertical focus is very important. That sets the tone.

Behind that, we also have a large team of designers who control different points in the model fine-tuning process. They help set the right direction for the aesthetic standards within each vertical. That is something differentiated in our product offerings.

Then, if you move one step forward, the data flywheel is also very important. Users within a vertical give us data through their behavior: which photos they use, which photos they edit, which ones they throw away. That is very important for us to improve image creation.

We are also in the camp that believes controllability in visual applications should not just come from AI. You still need manual touch-ups at the end for users to make last-mile improvements, because aesthetic judgment is very subjective.

Even if an AI model works with you every day, you will always have subjective comments and small edits you want to make.

One other interesting point is that when we talk about aesthetic standards, in the case of leisure products, the face is usually yours. So you have a strong say and a strong sense of what is good for you.

The way we learn that is by studying the trend in your geographic location, giving recommendations, letting you try them, and then as you use the application, you tell us what is most suitable for you.

Picchi is a newly launched app where you can upload three to five sets of original photos and edited photos that you have done yourself. We are then able to learn that pattern and create a specialized model for you. The next time you want to edit a photo, you can call up the model that you trained yourself and apply your own aesthetic standard to your photos.

On the productivity side, however, aesthetic is not the ultimate holy grail of an image. It is very important, but whether that photo or video is effective in driving conversion, likes, or comments is also very important.

When we deliver images and videos to users, we take into account key data from that particular vertical and the metrics that matter for results. It is not just whether someone is subjectively handsome. It is whether this person, image, or video can help sell your product.

So the two camps are quite different.

Grace Shao:
That is really interesting. From a pure consumer point of view, you pointed out an important nuance. If I upload my own face to a Meitu product, it might give me very smooth, pale skin and a more angular jawline or chin. But if I use an American fine-tuned product, it might give me more contouring. It is a very different aesthetic.

But to your point, if you are a prosumer, a content creator, influencer, or e-commerce seller, then it is not only about whether the image looks good. It is about whether it drives sales.

I want to ask something slightly more philosophical before getting into the businesses. Technology often changes art. Photography changed painting. Software like Final Cut Pro and Photoshop changed photography and video. How is AI now changing how visual artists approach their vision and craft?

I have also spoken to companies like Kuaishou, where they have Kling and are partnering with AI-native film studios. These people are creatives, but they do not view AI as disrupting their work. They use AI as a tool to create their work. What do you think about this at a high level?

Gary Ngan:
If you look at our core value proposition, our mission statement is uniting art and technology.

One step down from that, we are trying to democratize design, art, and creative expression.

What AI changes is that it enables people who have creative ideas, but not the actual training or skills, to express those ideas.

A lot of the time, we have creative ideas that we want to express, but our motor skills are not refined, or we do not know how to put colors together. AI can help us deliver those ideas.

That is fundamentally changing the artistic landscape to a certain extent.

Other companies may say that existing professional filmmakers and designers can use AI to make things more efficient or create things in a different way. That is great. But I think the bigger impact on the world is enabling many people who previously could not create anything. They had ideas, but could not express them. Now they are able to express them.

That is what is fundamentally changing the industry.

Grace Shao:
We have talked about how you have many different products, and you explained that they are targeted at different verticals. How should we understand MeituHub and MiracleVision? Are they the operating system underneath everything?

Gary Ngan:
MiracleVision and MeituHub are the visual, image, and video infrastructure that we have. Our applications are built on top of these things, so they go hand in hand.

We are still fundamentally an AI application company, but we also need visual infrastructure.

To give you an example, over 90% of our AI outputs come from our own models. There are many situations where we think the models we fine-tune ourselves perform better than third-party models. There are things that other people do not necessarily focus on, so we have to invest in R&D and create that infrastructure ourselves.

MeituHub is also a way for us to export that technology. People can use our APIs and skills to build their own applications or integrate them into their own systems. That also reinforces our vision of democratizing design.

Grace Shao:
That is a perfect segue to my next question. I understand your team fine-tunes your own models, but you also build on various open-source models. Why does Meitu need its own model?

Traditional application companies often did not need to own the foundation layer. So why does Meitu need that? And more broadly, why are so many Chinese consumer internet companies pushing out models? You see even companies in food delivery, ride-hailing, and other consumer internet sectors releasing models. Is this a cultural push, or something else?

First, how does Meitu think about it at the company level? And if you can comment, how do you view this competition across the China ecosystem?

Gary Ngan:
It is harder to comment on the overall market, because what we do, visual image and video models, is quite different from language models. So I will focus on why we do our own models.

Our belief is that one general model will have difficulty performing well across all verticals, because context is so important.

If we do not have our own models, then aesthetic standards will be set by third parties. When a third party creates a model, they have their own idea of what aesthetic standards should be. They have their own idea of what should be generally good given a certain prompt word.

But that may not be applicable to the verticals we are working on. That is why we need our own models to serve those purposes.

At the same time, we integrate third-party models because even within a vertical, there are corner cases or edge cases that our core model may not be optimized for. In those situations, we call on third-party APIs to serve users.

As an AI application company, the only point of optimization is user satisfaction. We use a combination of our own models and third-party models to serve that purpose.

Sometimes we see more and more users calling third-party APIs for similar prompts or similar creation scenarios. Then we will augment our models to cover those scenarios as well.

Our model is continuously growing, but we make it very vertical-driven. We have told the market that we are not in the business of creating a general-purpose model. We are creating vertical models. But that does not mean we are giving up model training altogether.

Grace Shao:
So there is a lot of industry know-how in each vertical that you have.

When it comes to which foundation model you choose for each task, how do you make that decision? I spoke to one of your colleagues at SuperAI, Rocky, your VP of R&D. We discussed the fact that you use a series of open-source models and also work with different model providers. What is the main factor in choosing which model to build on for which vertical? How do you delegate tasks across models?

Gary Ngan:
At a high level, there are two main forces behind that.

One is user behavior. If a user uses Model A to create a certain task, and many users do not press save or do not continue working on it, then we probably need to serve that task with a different model. It is a data flywheel type of operation.

The other factor is our large team of designers, who are very involved in training these models. Designers help set the standard for what the right model should be for a particular task.

This is a very important differentiation for our company versus most technology companies.

I am not sure if you are aware, but our founder and CEO was an art student by training. In his day, he was the top student in the Tsinghua Arts Academy entrance exam for oil painting.

In his mind, aesthetic standards are always very important. Because of that, designers in our company have a very strong say in every product and every feature we launch.

Over more than a decade of designer training, the rest of the company has also developed stronger aesthetic standards. Product managers and R&D engineers also have quite high aesthetic standards now.

Our company is organized toward delivering the best aesthetic standards for users. That is very differentiated from most tech companies.

Most model companies may think: We solved this problem, the photo is done, the video is generated, the main character is stable throughout three minutes, so the mission is accomplished.

But for our designers, apart from the stability of the main character, they also look at whether the lighting is realistic, whether the color fits that vertical, and whether anything feels wrong from an artistic point of view.

Those are the things we really focus on when fine-tuning. That is something we are very proud of, and I think it is a major differentiator.

Grace Shao:
Even as a consumer user, I can say your products have that extra last-mile touch-up tool that others often do not offer. It is meticulous and accurate. You can zoom into pores or details in the background. It is interesting to hear about your founder’s background and that artistic legacy, because that culture really shines through the products.

Speaking of legacy, I want to understand Meitu’s AI legacy and strategy. You have been working in image and video for over a decade, so you obviously have a vast database and deep know-how in visuals. How does that industry expertise translate in the age of generative AI and in the future agentic world?

Gary Ngan:
Generative AI has changed the speed, scope, and value of what we can deliver.

First, speed. New AI capabilities can now be translated into user-facing features much faster, helping us launch popular effects globally and drive overseas growth.

Second, user experience. Generative AI enables effects that traditional computer vision technology could not fully achieve.

For example, facial and body retouching is no longer just manual adjustment. AI can reconstruct details, lighting, and texture in a much more natural way.

Third, target addressable market expansion. AI helps us broaden into productivity workflows like DesignKit and Kaipai, which were things we traditionally could not do.

Overall, AI is very empowering in helping us get to where we want to go.

Before AI, all we could deliver was better tools. But in order to use those tools, you still needed pretty good aesthetic standards or some understanding of the basics. Otherwise, giving you those tools did not really help much.

With AI, you still need maybe 10% or 20% of that understanding, but the requirement is reduced massively. AI can give you many choices to choose from, and then you can start building from there.

That helps us move from leisure applications to productivity applications. That is really what the strategy is about today.

Grace Shao:
AI can act like a guide or mentor if you are new to a certain craft or sector.

Let me ask the spicy question. AI compute cost is obviously extremely high. Image and visual generation are expensive. How does the economics work right now? Does AI compute affect your gross margins, or are you seeing ROI already?

Gary Ngan:
As I said, currently over 90% of our generative outputs come from our own models. As long as we are using our own models, the cost is very manageable. Our gross margin is still over 70%.

Also, when you are editing your own face or editing a product photo, these things are not purely AI-generated. You may want AI to edit a little bit, remove someone from the background, or create a new background for a product, but the entire photo is not purely AI-generated.

It is true that AI inference has a cost, but it is not as if every photo now incurs a lot of cost. We need to make that distinction first.

As we move into new verticals, like music videos and AI short dramas with RoboNeo, those are more experimental. We are using more third-party models, so margins on those new applications will be much lower than something like Meitu Xiuxiu.

But as we continue to progress, we will develop our own models to replace some of the third-party costs. Over the longer term, we also believe API costs will come down.

So we do not see this as a threat. In fact, the integration of AI has expanded the addressable market so much that it is a much bigger opportunity than threat.

Grace Shao:
I appreciate that nuance. You are explaining that the first type of usage does not use as much AI or token cost as people might expect from the headlines. The second part may be more expensive, but we are still in very early stages.

Let’s take a step back. For some of our American or Western audience, they may not be as familiar with Meitu. How do you fundamentally make money?

In your public disclosures, consumer paid subscribers grew more than 30% year-on-year. What is driving that growth? Is it that the AI features are much better now? Is it global expansion? Help us understand the business model and what is driving growth.

Gary Ngan:
Our main revenue source is subscriptions, mostly on the leisure side.

That is our second growth curve. The first one was advertising, but that business has matured.

The second growth curve, which is still growing quickly, is subscription on the leisure side. The main driver has several parts.

The first is China user behavior. Paying for apps really started after COVID. Before COVID, virtually all applications were free. They competed through free usage, advertising, or redirecting traffic to other applications to generate money.

After COVID, many user-facing applications realized advertising was under pressure, and they wanted new revenue sources. Without colluding, many of them started charging users. That kickstarted the user subscription process.

What is less understood is that users then began realizing that applications have to be paid for. As time goes by, the behavior of paying for applications grew on them.

Now there is much less of the issue of, “This app has to be paid, so I am not using it.” That was a real mindset before. Now it is more like, “This app costs 15 RMB a month. Is it worth it?”

That is what I would describe as the beta factor, meaning the overall market. Users are becoming more and more used to paying for mobile products.

That is one reason we are confident that paying subscribers and the paying subscription rate of our leisure applications can continue to grow.

To give you a sense, we have done surveys. The global paying percentage for photo and video applications is about 20%. If you benchmark music and video apps globally versus Chinese equivalents, the Chinese equivalent is usually around half. For example, if Spotify is around 40%, the Chinese equivalent might be around 20%.

So if global photo and video applications are at about 20%, China should at least achieve about 10%. Right now, we are around 5% to 6%. So there is still another 80% to 100% growth headroom there.

The second growth potential is international expansion. In high-ARPU areas like the U.S., Europe, and East Asia, including Japan and Korea, the base ARPU is already much higher than China, anywhere from 100% to 200% higher. The paying percentage can also be much higher.

To give you a sense, one of our applications called AirBrush has over 50% paying percentage in the U.S.

As we launch stronger operations in these high-ARPU countries, we expect our blended paying percentage to grow further.

One final point about monetization is that we are integrating more generative AI capabilities into these applications. For example, Picchi is an application for leisure, but it uses an agent for editing, and that has a completely new business model.

On top of regular subscription, if you want to create your own model to apply your own editing skills, you have to pay for that model separately. That is another monetization test we are currently working on.

Grace Shao:
When I was reading your earnings reports, I was a bit surprised that your highest revenue generator is consumer subscription, because the default mindset is that people have very little willingness to pay.

But as you said, whether it is the change in behavior in China, or people having more appetite for premium add-ons or AI-plus features, willingness to pay is changing.

There is also the fact that advertising can be annoying to sit through. You do have a lot of advertising, I have to say. Spotify does too, and I think that drives people to pay to get rid of advertising.

On that note, do you think you will gradually reduce your dependency on advertising? It is still your second-largest revenue model.

Gary Ngan:
We have not relied on advertising since 2022. At the corporate level, we made the point that we are no longer strategically trying to drive advertising.

You have seen our advertising business grow at low single digits over the past few years. Advertising is not what we are fundamentally trying to drive.

However, we are experimenting with advertisers on fun and engaging AI-infused campaigns.

It is hard to describe with words, but you can imagine users generating viral photos with a brand advertiser’s branding that fits the brand image. That gives the uploader a lot of likes and gives the advertiser a lot of exposure.

So we continue to experiment with those things. But in any case, we are not relying on advertising for business growth.

Grace Shao:
On partnerships, I had this idea and I do not know if you are doing anything like this. Would you partner with some consumer-facing chatbots in China to help them with video and visual capabilities?

For example, could someone go into a consumer chatbot and call up Meitu’s capabilities? There may also be competition there. How do you view your relationship with these players?

Gary Ngan:
We are open. In fact, we are already an official partner with WeChat, not on the Xiaochang side, but in another area. I do not remember the exact English name, but basically when you are using the chatbot, you can call up Meitu.

Right now, it is still a lighter relationship, almost like traffic redirection. But our goal is to democratize design. Being able to work with more people and enable more people to access that power to express themselves is something we are open to.

Grace Shao:
That makes sense. It feels like they may not want to put as many resources into this specific use case, and you have the know-how in doing the best video and image editing.

Gary Ngan:
I would not say they do not have the edge. I think they may just not want to focus on that.

Creating these applications requires a lot of focus. It requires the right organizational structure and a laser-sharp focus on trial and error, and on creating the best aesthetic output for users.

These may not be the things that larger companies want to invest in. It is important relative to our size, but to them it may be something they do not want to focus on. If they wanted to do it, I think they could.

Grace Shao:
Let me challenge you a little bit on big tech. In China, ByteDance and Kuaishou clearly have a lot of edge and moat from massive pools of image, visual, and video data. In the West, we have Canva, Adobe, and other global applications. Even Shopify and Alibaba are creating e-commerce staging and design tools.

In this big world of competition, or peers if we put it more nicely, how do you see Meitu’s strength? Who are the most relevant competitors that are similar to what you do? And who may look similar on the surface but are not actually doing the same thing?

Gary Ngan:
We have to separate it into two categories.

On the leisure editing side, with the exception of one business unit within ByteDance, there are not many large companies doing that globally. I do not think there is any real large company doing that in the U.S.

There are smaller companies, but they are much smaller compared to us.

On that side, our edge is really continuing to follow and set the trend for the latest aesthetic standards and what helps users stand out on social media. These are the things we have been doing for more than a decade, and we will continue to excel in them.

On the productivity side, there are many companies doing similar things, but taking a much more general approach.

For example, Canva and Adobe use one product to satisfy different verticals. Adobe is organized around media: photos, vector diagrams, video, effects, and so on. Canva is one editor trying to fit many situations. Figma is also one app serving many applications.

They are design-driven. We, on the other hand, are much more vertical-driven.

We are not restricting ourselves to a specific media type. We are saying, within e-commerce, what do you need?

You need product photos. You need very short product videos. You need the ability to generate batches and batches of photos. You need to know the latest trend on the e-commerce platform you are selling on. For that particular product, you need to know the selling points. You also need to know the rules of Amazon or Temu and what you need to abide by when selling those products with pictures.

All these things are baked into DesignKit.

If you are using Canva, I highly doubt it will have a red flag saying, “You should not be using minors in this product photo.” Canva may not even know that you are creating a product photo in the first place.

So these are the different focuses we have.

In terms of competitors, it is hard to say who is a direct competitor, because at the end of the day, you can use Photoshop, Canva, or our products to create an e-commerce photo. They are all peers, but we take different approaches.

If we take a step back, generative AI is still very early. It is 2026 now, but generative AI really only started in earnest late last year for visual use cases. Before then, a lot of generative AI photos still looked AI-generated.

Grace Shao:
They were quite bad. There might be six toes, or the face was disproportionate.

Gary Ngan:
Even if there was nothing obviously wrong, you would look at the photo and know it was AI-generated. It did not feel real.

Now we are just starting to see things that are harder to distinguish between human-made and AI-made. This is how we can empower the industry and increase efficiency.

We are still very early in this market. That is why we are very optimistic and see a lot of opportunities.

Grace Shao:
A little side note: in 2019, when I was still with CNBC, I covered deepfakes. At the time, there were a lot of deepfake videos of Obama or Zuckerberg. A startup even made a deepfake of me. It was literally just plugging someone else’s face onto my head and body, and nothing really worked.

But now, fake images and videos are getting very hard to distinguish with human eyes. How do you view the ethical side? How do you stop misuse of the technology?

Gary Ngan:
First, we have put in safeguards.

For example, on Picchi, if you generate a model of yourself using your own photos, that model cannot be applied to anything other than your face. If we detect that it is not your face, we will not allow you to apply that model to another face.

In some of our generation applications, we have also put in safeguards around certain words, such as violence or pornographic images. You cannot generate those using our applications.

So there are safeguards that we put in place. Obviously, we can only do so much.

One thing that makes it slightly easier for us is that we organize our applications into different verticals. Users come into our applications with a very strong intent. They know they are creating e-commerce photos, for example.

Instead of giving them a general chatbot where any random person can come up with a random idea like putting their face onto the President of the United States, it is harder to imagine someone using DesignKit to run a prompt like that.

Organizing into different verticals also helps us mitigate the risk a little bit.

Grace Shao:
The last area I want to talk about is globalization and your global strategy. Meitu is globally available. It is interesting because, as you said, you focus on each vertical, and you have not done a big splashy general marketing push. It also feels like that is true geographically. You are in Southeast Asia, Japan, Korea, Europe, the U.S., and so on.

Help us understand global scaling. What have been the challenges? How have you done it successfully? And how do international users from different regions behave differently from users in China?

Gary Ngan:
I will answer the second part first. Users in different regions all behave very differently.

Grace Shao:
Give me all the stereotypes.

Gary Ngan:
Not stereotypes, but I will give you one example.

We were doing a user focus group in the UK and spoke to a male influencer. He said, “Your app can edit my jawline? That is incredible. I would totally pay for it. But I do not think it is a good idea to smooth out my skin.”

Grace Shao:
That is interesting. So it is not okay to pretend you have better skin, but it is totally okay to have a chiseled jawline?

Gary Ngan:
He did not mention whether it was ethical or not. That was just his feedback, word for word.

The challenge, or the interesting thing, is that we have to really listen to what users want in those markets. Different geographic locations need the right mix of features and marketing campaigns.

I will give you another simple example. A few years ago, we were looking at Lunar New Year. Koreans also celebrate Lunar New Year, and in China Lunar New Year is a festival where we get a lot of usage.

We had launched features in China that were very popular that year, but in Korea there was no uptick. Later, when we had local Korean colleagues helping us run marketing campaigns there, they told us that Koreans generally celebrate Lunar New Year with white clothing and a white theme, while Chinese people celebrate with red.

Our Chinese marketing team was surprised, because in China, white is usually associated with funerals. It did not register.

That example tells us there are many things we need to immerse ourselves in culturally to understand how people behave, what they care about, and what the standards are.

We cannot stereotype anything. Every place and every person behaves very differently.

That is the biggest challenge, but also the biggest opportunity.

Now we are setting up offices in different parts of the world. We are sending product managers overseas regularly to do more focus groups and, more importantly, to experience the lives of the users they are trying to serve.

In the past, we relied too much on consultants, reports, or reading online. That is not enough anymore. We are fixing that, and I think we are making progress in some countries.

Fingers crossed, we will continue to grow bigger in Western markets.

One other tailwind that has helped us is TikTok and K-pop. Back in the day, editing a photo seemed socially unacceptable to a certain extent. But with TikTok, people are more relaxed about filters being applied and playing around with your face. It is no longer as taboo in many Western countries.

The rise of K-pop is also influencing cosmetic styles, and that becomes a segue for us to try different things in Western markets.

There are very interesting things happening. But the most important thing is for us to really understand, live, and breathe those cultures so we can create things users want.

Grace Shao:
That is meaningful, and it feels important for a new generation of Chinese companies going global. Localization cannot just be reading headlines or high-level reports. You have to understand the culture, because culture influences the business.

To wrap up, I really appreciate your time. My last two questions: first, what is the biggest misconception investors currently have about Meitu’s business?

Gary Ngan:
One of the biggest misconceptions is that general models are going to destroy everything, and that there is no place for AI applications.

We think that is quite unlikely on the visual side. I am not sure about the language side, but on the visual side, aesthetic standards are very subjective and personalized, and a lot of controllability is needed.

Different verticals have different interpretations of the same words. So the way models are trained and organized, even with agents, makes it unlikely that a one-size-fits-all general model can satisfy all verticals.

Every vertical has its own workflow and standards. AI application companies are very important in making those adjustments and optimizing workflows for users.

That is the biggest misconception.

Grace Shao:
I agree with that. We are seeing more of that realization in the market now. You have strong vertical use-case AI-native companies coming through, like Harvey. I have also met companies where former investors are building equity analyst research tools.

You can say generic GPT can be used for research very easily. But to your point, these teams know the niche use case. They know the process, the standard, and the workflow better than anyone else. Even if the TAM is small, it can be big enough for their business.

The last question I ask everyone on the podcast is: what is one differentiated view you hold? Something that is a bit against consensus.

Gary Ngan:
Is it related to the company or the industry?

Grace Shao:
It could be anything. Usually people answer about their topic, but it can be anything.

Gary Ngan:
I think life expectancy will be a lot longer than we think today for our generation.

Grace Shao:
So we are going to live to 150, thanks to Bryan Johnson’s experiments?

Gary Ngan:
Possibly. Then there will be more time.

There is a lot of advancement in AI. It speeds up many pharmaceutical processes. You can run different trials much faster and understand the underlying issues more efficiently than before.

And with more time, there is more time for us to create more art.

Grace Shao:
And live a healthier life. Although right now, anyone working in AI knows AI never sleeps, and I think we are all working more than ever.

But thank you so much. That is definitely a differentiated view. I really appreciate your insights and your sharing today. Thanks again, Gary.

Gary Ngan:
Thank you so much, Grace, for this opportunity. Really nice talking to you.

In this episode, I speak with Paul Triolo, partner at DGA-Albright Stonebridge Group, about how to think clearly about China’s semiconductor ecosystem, Huawei’s role in the domestic AI stack, and whether U.S. export controls are actually working as intended.

We start with the latest debate around restricting foreign access to advanced Chinese AI models, before moving into the semiconductor stack itself: SMIC’s role, capacity bottlenecks, domestic GPU startups, hyperscaler chip efforts, Huawei’s vertical integration, and why software ecosystems like CUDA, CANN, and MindSpore matter just as much as hardware.

Paul argues that the usual framing- whether China can “catch up” to Nvidia or TSMC is simply too narrow. The more important story is that export controls have pushed China toward a broader systems-engineering response across chips, tools, packaging, memory, software, and cloud deployment. We also discuss HBM, rare earths, remote-access loopholes, the logic behind Huawei’s roadmap, and why the collateral effects of controls may be larger than policymakers expected.

We close on the bigger strategic question: whether the U.S. and China are drifting into an AI race dynamic that raises risks for everyone, and why more direct dialogue — not just more restrictions — may matter most from here. [Paul co-authored a piece here discussing how to navigate the complexities of the U.S.-China AI safety dialog] This is an extremely insight-dense episode, and I hope you enjoy it as much as I did. Thanks again .

Btw, coming up next are a few episodes featuring founders and execs from hot-listed AI, autonomous driving, and spatial intelligence companies.

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from analysts, VC investors, builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Chapters

00:00 Beijing, Chinese AI models, and why regulators are paying attention
04:43 Why AI labs are moving into chip design
08:12 SMIC’s role and the fight for domestic chip capacity
17:56 Huawei’s capabilities and why it became the center of the conversation
26:53 CUDA, CANN, and whether export controls really worked
38:24 Can Huawei’s software stack win developer mindshare?
51:10 Why China can still progress despite compute constraints
57:07 The AI race narrative and why Paul is skeptical of it
1:07:46 Why governments still lack the technical capacity to respond
1:08:42 Will frontier AI labs eventually be nationalized?
1:13:59 The risks of zero-sum U.S.-China AI policy
1:20:46 Why more direct U.S.-China dialogue matters

Transcript (AI-generated for reference only)

Grace Shao (00:00)

Hi, Paul. Thank you so much for joining. Really excited to have you on. And I feel like you’re the perfect person for a few of the questions I have prepared in the beginning of the podcast before we get into the actual topic. There are so many things happening, and I can’t keep up with like Twitter these days. So no, so supposedly Reuters reporting saying Beijing is restricting foreign users in accessing Chinese models. Like, what is that all about? What’s your view on that?

Paul Triolo (00:15)

It’s hard to keep up. Yeah, I think in the wake of the fable mythos fiasco, if you will, in the US and the capabilities of these advanced models getting really good and getting into areas like cybersecurity and biosecurity, I think it’s not surprising that the Chinese government is at least considering what to do about open-weight more properly models that are that are that are reaching sort of frontier level capabilities, particularly DeepSeek and Zhipu and then even more recently, you know, Meituan and others. I think my sense is this is just preliminary discussions with the labs about this issue because, you know, even in the US, there’s lots of confusion about what the government’s role should be here in determining when models are released and under what circumstances and how do you measure capabilities. even though the US has been thinking about this for a while, and I’m sure that’s to some degree that’s happening in China, there’s no general agreement on how do you do this. And these companies in China, as you know, are all commercial companies that just like in the US are under a lot of pressure to continue to put out models. and so I think I would not read too much into that. I think that’s it’s clear that the Chinese government is Trying to figure out what to do with about this, but I don’t think they’ve reached any conclusion about which models to control and how to control them. they’re they’re learning from the companies. They’re probably going out and saying, you know, how do companies themselves evaluate these models internally, in terms of capabilities that could be of concern? and what should the Chinese government eventually do? I mean, they’ll they’ll they’ll do something eventually, but I think we’re in the early stages still. as a result of the

Grace Shao (01:57)

Yeah, for sure. I think you know, Zhipu, Minimax, these companies are public listed, like they actually face, you know, just shareholder pressure as well. So it but the one thing, the nuance is Chinese companies usually are a bit more prepared or aware of potential regulatory, I guess, involvement. so yeah, let’s let’s wait and see. Because I read this and I was like, this seems bit counterintuitive, frankly, to the model’s going forward. I think like the deadline was a little out in front here. I think it’s

Paul Triolo (02:49)

Right. I think that the headline was a little out in front here. I think it’s clear that there’s concern as these models become more about what to do about tiered releasing. but this is a more general discussion I think that’s happening. it doesn’t surpr it to people who’ve been following this sector for a long time, the idea that we would be here at this moment, you know, was not surprising. The problem government gov the ability of governments to keep up with the pace of development of the technology is just clearly here, it’s it’s woefully inadequate to the moment because you know within these large AI labs, and I’m now calling DeepSeek and Zhipu, along with Anthropic and OpenAI, you know, the top four global frontier AI labs. you know, th researchers n understand these issues and they’re they’re really concerned about this because P particularly things like recursive self improvement, which is which means models are basically training the not training themselves, but they’re they’re optimizing some of the orchestration or the harnessing that the sort of platforms that these things operate on. You know, that’s been a that’s a growing concern because they’re you know, the models themselves now are able to improve the overall ecosystem without human intervention. Right. and that people miss that, I think, in the in the US with all the fable mythos kerfluffle. The Anthropic released a blog that talked about this, that recursive self-improvement is now sort of part of the landscape. And so that’s also I think part of the concern including in within the Chinese government about okay, you know, Chinese labs are getting pretty good. what should the government how should the government think about

Grace Shao (04:43)

Yeah, definitely. And I think, you know, in China, usually the regulators and the industry actually work pretty close together. so hopefully, you know, people can kind of regulators can keep up, will hopefully catch up on understanding technology a bit faster and better. Okay, so another quick commentary on what is happening in the news. Supposedly DeepSeek and Kai AI are going out and making their own chips. What is happening? What is your high level view on this?

Paul Triolo (05:08)

Everybody’s doing it. Now, you know, in it this is not surprising. US, of course, some of the hyperscalers and more of the hyperscalers like Google and AWS have long d determined that it would be useful have specially designed ASICs, application-specific integrated circuits that are optimized for running certain workloads in their cloud. and you know and optimi na and now optimized for models specif specifically for you know large advanced models for which general purpose GPUs, which is what NVIDIA and AMD produce, may be, you know, may be suboptimal. but this is a complicated issue because to do semiconductor design, you know, this is a whole nother thing than building models. And so for both Zhipu and DeepSeek, you know, this requires building a team of design so some semiconductor design engineers, right? Who now they’re not they’re not a lot of these guys laying around that are not gainfully employed, particularly for designing really sophisticated chips here. So I think it’s not surprising that they want to do this, but I would be again sort of a little bit skeptical that they’re gonna do be able to do this in the you within the next year even. You have to build a team. It’s expensive. use you have to get advanced semiconductor design tools, electronic design automation tools. and then you have to begin figuring out where you’re gonna manufacture these, right? And in China, of course, as we know, because of export controls, these companies are likely gonna have to use SMIC, the domestic foundry. So they’ll be competing with all the other players, all the other GP GPU designers, the general purpose GPU designers like Biren and More Threads, a of these companies that as you know have gone public recently, they’re all vying for this limited capacity at SMIC to manufacture these advanced designs at say let’s say seven nanometers, which is a feature size of these chips. so I think it’s really interesting that they’re gonna do this, but it’s gonna be a challenge on two fronts. One is you know, assembling a team Sustaining that team over time, you know, this is expensive. It’s very expensive. Lenovo tried to do this, for example, at one point. They were considering getting into semiconductor design, but they determined it was it was too expensive and it was going to be a long term drain potentially if you know, depending on the success of these teams. and so, you know, I think if anybody can do it, I mean DeepSeek seems to already have a lot of expertise on hardware and understand the hardware really well. But it’s it but semiconductor design is another whole nother discipline, if you will. And then in China it comes with the added constraint that you have to you’re sort of stuck with SMIC and you know, maybe Huahong down the road will have sub some kind of seven nanometer process. but it’s you know it’s it’s a tricky thing. But again, the trend in the industry is to do this. so in the US you have the open AI, I think just recently Anthropic are both also considering you know designing their own semi

Grace Shao (08:12)

Yeah, they have partnerships as well with other vendors. So help me understand. I think twofolds. One is what is SMIC’s role in China and what is their biggest bottleneck? For them to actually churn out better chips. One thing you said is capacity, other thing you said is access to certain technology, you know, instruments, machinery. The other part of the question, if you can fold into it, is who are the actual players? So, you know, DeepSeek and Zhipu wants to build their own create their own chips, but Baba’s in this, you know, Baidu’s in this. There’s a lot of big tech trying to create their own chips as well. How do we understand all their relationships in the ecosystem?

Paul Triolo (09:06)

Wow, okay, you got a lot a lot in there to so let’s just let’s just look at the demand side, then we can look at the supply side. So the demand side, as I started to allude to there, is pretty heavy, right? So you have Huawei, right? They’re designing the ascends, and those are all at the at least the most advanced node, which is this horrible technical term for the process that is used to manufacture these at SMIC. For example, here. And so Huawei has a lot of demand. And past, Huawei was given most of the capacity at SMIC, just because a year ago or two years ago there weren’t as many other players. Now there are. Now it’s more complicated. But certainly the Chinese government probably also heavily weighed in to have SMIC prioritize Huawei production, both for their smartphone, their Kirin smartphone, and for the Ascend processors for AI. But now in the last year, we have all we have a we have two s two additional sets of companies vying for this capacity at SMIC. One is the G the sort of GPU makers in China, the startups. And this is Biren, Moore Threads, SoftGo, Inflame, Iluvatar. You know, there’s at least there’s a couple more too, but those five are sort of the top companies. And some of those companies were started by engineers from Nvidia and AMD. And so they their designs are really advanced and they are they’re more compatible with the NVIDIA ecosystem, et cetera, et cetera. So they’re all now producing GPUs at SMIC, and there’s allocation issues. I when I was in China recently, I heard that one of those companies had been promised certain number of wafers at SMIC, but then one of the big hyperscalers would come in and offered more money. And so SMIC had said, Okay, well, we’re gonna reduce your allocation, right? So there’s a lot of fighting for that. Th and then the other group that you’ve just mentioned here is are the companies like Alibaba, the hyperscalers, and then now the model developers like DeepSeek and Zhipu that are also doing their own designs. And I think again, as you noted, Baidu and Tencent and Alibaba are more are farther along on this. They have semiconductor design teams which are already producing, in the case of Alibaba, and in the case of Tencent, they have dedicated ASICs. They’ve been doing this for a while, sort of under the radar. Tencent’s a very capable company and has been they really good at this, right? It turns out, but they don’t they don’t they’re very low key on this. and then Baidu of course with Kunlunxin, which is also gonna go public on the Hong Kong market. So the other part of this also is that all these companies of course need capital To function and to do all these designs and to hire these engineers and to and to do all this stuff. So all the in the last year we’ve seen these companies tap into capital markets, particularly, you know, the GPU makers, the new GPU startups have gone public in Hong Kong and in Shanghai. And then you know, we’ve seen DeepSeek, of course, raise seven billion in funding from a variety of sources, and part of that will go to presumably the building up a design team and doing Designing their own chips. and then of course Zhipu has gone went public and its stock is crazy high on the if you look at the valuation on Hong Kong. So it’s a new a new game where Chinese companies are playing this game of d you know, designing their own chips. and then of course they’re all competing for this capacity at SMIC. So now we can turn SMIC, right? So SMIC is this crazy company, right, that for a long time was you know was sort of under radar. but they are under heavy US export controls. So that started in around twenty eighteen, twenty nineteen, when they were they had actually ordered a very advanced lithography machine from ASML, which produces all of the advanced lithography machines, extreme ultraviolet lithography, UV lithography. So they were denied that. They actually ordered it. And then they and then the Dutch government, under pressure from the US government and Wassenaar, which is this interagency or intercountry group, this multilateral group, pulled that license from them. So for the last whatever, six years, SMIC has been using its deep ultraviolet lithography, which is the second best, but pretty good. They’ve been pushing their suite of DUV machines to the utmost limits to try to get to these lower and more advanced nodes. And seven nanometers is sort of the limit, seven and maybe five It’s complicated. Some layers of these semiconductors can be but they don’t all have to be at the most advanced levels. but they’ve been doing something that nobody else in the world has been has done. Now in Taiwan, they did use some for example, that SMIC is using, but they that when they had a access to the more advanced lithography, they went to that because it’s it’s it the throughput is faster. And the yields are better. So SMIC is trying to do something that you know that nobody, no but no other company would have to do because of US export controls. And that means pushing these lithography machines to their limits. they’re having obviously there’s a they’re being successful, but there’s a limit to sort of the yield too that they can do. So for example, the AI semiconductors are much more complicated than for a smartphone handset. So for the smartphone handset, they yield. you know, say ninety over ninety percent. But for the GPUs or the NPUs as actually as Huawei is using, the yields are much lower because these are very complicated and dense chips, right? And so using some of these advanced techniques just it’s just hard to do, right? And it’s and you end up with not as many use useful chips at the end because the yields you know, you’re you can’t do it because you’re you’re you’re sort of pushing the machines to their to their the limit of capabilities. So but this neck but over the next six months it looks like they’re gonna bring online SMIC is gonna bring online more capacity at these advanced nodes, maybe double capacity, because obviously they there’s so much demand for this for this these chips in China that SMIC is responding to this demand and is putting in place more production lines in Shanghai, by the way, at SMIC South to produce to meet the demand for all of these. AI primarily not just AI but mostly AI optimized hardware because of these because of those three batches of companies. You know Huawei is sort of a batch in its in its own right. But then they the GPU makers, the startups and then the ASIC makers. So the challenge then is who decides who gets the capacity, right? and as I said, there I’ve heard a lot of anecdotal you know, chatter in China about that. It’s very complicated. The government obviously weighs in to favor certain companies, but then you know, there’s a tremendous amount of competition. And then finally, the other piece of this not just the logic, if you will, the sort process or die. It’s also memory, right? So for more for the for the AI hardware, high bandwidth memory is a really important part of that because These advanced GPUs are packaged with lots of memory on the on the actual package, in within the actual package, co-packaged, if you will, with the logic. and there, of course, US export controls again have affected CXMT in particular, this case ChangXin Memory, is in Hafei and other places in Beijing. I just saw a big fab in Beijing when was there. and so there but again, it turns out is it to of the export control restrictions on memory is not quite as hard as it is for logic, although it’s not easy. and so CXMT is also ramping up its production of high bandwidth memory, which would then be packaged with those with those dyes. Huawei in particular stockpiled a lot of HBM in twenty four before the US controls were put in place from what Samsung and SK Heino, some of the South Korean producers which are leading In the production of high-bandwidth memory. So anyway so it’s a combination of several bottlenecks that the Chinese domestic semiconductor industry is trying to overcome to meet this demand for these for these AI this AI optimized hardware, whether it’s the Ascend series from Huawei or these other GPUs or these other ASICs that are that are now being designed by the hypo hyperscalers and the model developers like Zhipu and

Grace Shao (17:56)

I appreciate that context. I think for me, like I’ve been reading about SMIC for a long time, but that just really clarifies exactly their role in the ecosystem. And it’s interesting to kind of see how they’re prioritizing certain companies over others. but anyway, I want to double click on Huawei. I think it gets the most heat, you know? It obviously is not And it’s a very interesting company because it’s not just such like it’s not only just like you know pushing out their own models, but they’re a foundry, they’re they’re a chipmaker, they’re a designer, but they’re a bit of everything, they their own hardware. And then they also have obviously then CANN and then MindSpore that trying to compete on the ecosystem side with NVIDIA. So walk us through just Huawei’s capabilities, Huawei’s competitive edge. And then why Huawei was put on the entity list and got so much heat over the last couple of years.

Paul Triolo (18:44)

My God, great question. Great question. You’re really asking good questions here. But you know, there’s some, and I’ve written so much on this. So look, Huawei, if you remember, was primarily a telecommunications equipment company. the in the 2000, 20, that, you know, up until say 2018, 2019. And then Huawei got into the handset business too, right? They built they started building smartphones, they were ramping up. you know, to they were competing and out competing in some sense Samsung and Apple. And then the US put Huawei on the entity list in twenty nineteen, May of twenty nineteen. I still remember where I was when they were when I heard they were put on the entity list. It’s like when I when I remember where I was when shot. and then they were and then they were the even more critical in twenty, they the US added this so called foreign direct product rule, which meant that not manufacture its designs at TSMC basically. And I remember being in at Huawei in 2019 when I toured the their headquarters in Shenzhen and they showed very proudly all of the semiconductor designs they were doing at the most very advanced notes for all of their product lines. But at the time those were you know there was more on telecommunications equipment and also on cloud and on you know servers for clouds, the Kunpeng series of chips, for example. So anyway, so Huawei and the secret there that still important is that Huawei had spent a lot of effort to develop their that design team for those semiconductors, right? So I mentioned earlier how hard it is to do that. Huawei had over the about a ten year period had developed this arm of the company which was designing by the by the twenty time frame, was designing cutting edge chips, you know, on par with like Qualcomm and Nvidia even and some of the other areas. And it was getting better and better before the US basically you know w by putting Huawei on the entity list, HiSilicon was also there. So then High Silicon could not use TSMC. But in the process, Hi Silicon of doing some generations of chips at TSMC, Hi Silicon gained a lot of knowledge about you know both semiconductor design and how to use those complicated tools, those EDA tools, and how to do manufacturing. Because when you work with when you’re doing design and you’re working with a fab like TSMC, you’ll learn a lot about how that happens. So high silicon is sort of the secret weapon and if you will of Huawei. And so after the US controls, the Huawei kept high silicon designing. The designers kept designing Even though they didn’t have a place to manufacture yet, right? And eventually SMIC, working with Huawei figured out how to do all these optimizations of their existing equipment, this DUV equipment, to allow Huawei to manufacture SMIC, even though they weren’t the most advanced process, they were still pretty good, right? And Huawei made all these innovations in terms of overcoming some of the that not being not having access to the latest and greatest tools meant, which was, you know, for things like power consumption and other things. They did they designed around this, right? So that’s the other sort of this is a theme that you’ll that I’m sure you’re familiar with is, you know, the force Chinese companies to do different things and optimize things. This is what happened with DeepSeek, right? And other Chinese companies that don’t have access to all the latest and greatest Nvidia chips. So same thing with Huawei, they figured out how to design around this. Now the other thing they did, of course, which I don’t think mentioned was software, right? Because the US controls, for example, restricted access to Google Mobile Services for their handsets, Huawei invent had to invent Harmony, the HarmonyOS, right? So Huawei had to become a software company, right? Which they didn’t really want to do. Arguably, I talked to the senior Huawei officials and they were like, wow, you know, if we could use Android, why would we go to all the trouble inventing having to invest in and build A whole new operating system for which it doesn’t generate any revenue, right? It’s it’s just it’s sort of a cost center for Huawei. But they had to do it because they were under pressure. And so as part of that process, it’s important to understand that because now we get to the AI stack that you mentioned, the AI era. So Huawei has as a result of US export controls and having to develop Harmony, they have now known sort of the process of how to develop a software ecosystem. and how to get developers to use it, right? Although again with AI it’s it’s harder, right? So you mentioned CANN, the Compute Architecture for Neural Networks, which is Huawei’s equivalent of CUDA, which is the sort of developer and software developer environment that’s critical for NVIDIA. So yes, Huawei is trying to now develop as it did with HarmonyOS. And that was a long process by the way, a long and hard process. It wasn’t easy to do. They had to they and still, right, still China s smartphone companies still use Android, right? They don’t all use Harmony. But Harmony is sort of a cross device thing that you know goes for automobiles. You can use it on your car for the Huawei the Huawei invested EVs. So it’s so it’s a it’s a pretty good system. I’ve seen it when you go with your phone to your car, it’s all seamless, et cetera, et cetera. So anyway, so Huawei knows how to do software development in a complex hardware environment now. And so for AI, which is harder, they’re doing yes, they’re doing can is something that they’re working with. And then also MindSpore is sort of the equivalent of like PyTorch and it’s an it’s a development environment that AI developers use. And so they’ve gone a come a long way on that. In 2020, 2022, 2023, you know, because Chinese AI companies could still use Nvidia, they you know nobody was using Can and Huawei. But now DeepSeek and Zhipu and even Meituan, which looks like they have their million perimeter trillion parameter model on 50,000 ascends. This is something that Meituan has complained. And then now we have Minimax saying they’re going to do a 2.7 trillion parameter model. They’re all working with Huawei to optimize some of the aspects of that software development environment. Now it’s complicated because there’s training and there’s inference. And so there’s different needs for each of those in terms of development. But the again, the US export controls have forced the Chinese model developers to work very closely with Huawei because Huawei is the main alternative, right? And the Chinese government, of course, has been encouraging this. And so, you know, now we’re in the situation where the development environment around the Huawei and the SENS and CAN and MindSpore is better, arguably, than it was even a year ago. When I was at the World AI conference. last year in Shanghai, I talked to a lot of hosting, you know, various capabilities using Huawei Sense and they said that the Huawei system was hard to work with. They would tell me, they wouldn’t tell me this, you know, they I didn’t want to be quoted on this, but they said, you know, they were being told to use Huawei hardware, but it was hard for them to offer the kinds of services they were offering with Huawei hardware the same the same caliber of as with NVIDIA. But now I think that gap is closing. It’s not like Everybody in China’s all the AI developers are rushing to Huawei. but there’s still a complicated mix of both NVIDIA hardware and as we’ll see now, they’ve they’re the Chinese government is allowing I think 10 companies to buy these H200 GPUs, which we should talk about. anyway, so it’s a very complicated and heterogeneous compute environment in China. But one thing we can say with some certainty is the Huawei because of the export controls and because they’re closely with. DeepSeek which is very good at programming the hardware, for example, the that environment is at a stage where probably it wouldn’t have been without the export controls. And so we’re in you know, it’s it’s it’s getting better and better. and because companies are gonna have to use it at some point, they’re sort of d deciding, like DeepSeek is deciding, well, we better put a lot of effort into helping optimize that development.

Grace Shao (26:53)

I think I agree with you and what I’ve been hearing on the ground as well. A lot of the developers saying like if they had a choice, they wouldn’t really leave CUDA just so much better. But if they don’t have a choice, it’s kind of like damn it, I’ll have to just try to learn how to use this. And even if it’s not as good, like we’ll try it’s also like a chicken egg thing, the more developers on it. The better the s the software and the system. But I want to play devil’s advocate here. Like obviously, we all heard the Dario and Jensen like interview. Right. So like we don’t have to get into the details of that. But part of the argument. But part of, you know, what you just talked about was like, you know, high silicon kind of got shafted. They couldn’t get access to certain machinery. You know, obviously, right now we can say that. Objectively, factually, Chinese chips are probably not as good as the leading chips globally. So thus some may argue exp export control worked, right? Like so what’s your view on that?

Paul Triolo (28:02)

Wow. Okay, that’s a that’s a rather large topic. So it sort of depends on what you mean by work. so look the original goal of the export controls as sort of articulated in the you know federal register notice in October twenty two was originally related to s you know to sort of military other sort of nefarious end uses of right? but the real driving force, if you will, was really the this idea slowing Chinese companies’ ability to develop frontier models down so that the US would get to some advanced level of AI first, right? And so if you just look at that and you look at say Zhipu releasing GLM five point two, that’s not as quite as good as fable or mythos, but it’s pretty good, right? And it’s the gap between le the le the leading models from anthropic and openai and the leading models from Zhipu and DeepSeek and other Chinese companies and Alibaba in particular and now you know Meituan and even Xiaomi and Minimax, you know, that gap is still there, but it’s not really is it months, is it a couple of months? So if you’re going to argue that the export controls worked, then you know, is the does a two or three month gap even matter now, right? so that’s that’s one way to look at it, right? Now, if you look at it in terms it make did it reduce the ability of Chinese companies in the semiconductor industry to manufacture advanced GPUs, for example, on par with NVIDIA, of course it worked, right? Nobody would argue. that it did that didn’t happen because if you’re gonna restrict exports of GPUs and you’re gonna exp r you know restrict exports of critical tools that are used to manufacture those GPUs, of course you’re gonna you’re gonna slow them down. But then they then you have to say well how has Chinese how has Chinese industry responded to that, right? And what are and what are the costs of that for US companies, right, for example. And then you have to look at what is the retaliation from China to those export controls. So you have to at least look at it, look at the picture more broadly than just, you know, did the US slow down Chinese model development? Arguably there, the jury’s still out on that, right? Because I would argue that they haven’t the slowdown hasn’t really been that significant. If a company like Zhipu, like who had heard of Zhipu like even a year ago, right? if they can release a model like GLM 5.2, okay, wow, that’s a that’s a frontier model, right? It’s it’s matching fable in some benchmarks. Okay, so it’s clearly somewhere near the frontier. How close we can argue about and a lot of that is complicated, depends on the benchmarks you’re using. But in the in the semiconductor industry, then the you have to look at that in a little more depth. One thing that I’ve written quite a bit on of course it’s forced Chinese toolmakers to work With the with SMIC and Huahung and some of the other manufacturers, CXMT and YMTC. And so the overall level of capability, for example, of Chinese, the Chinese semiconductor industry to do stuff domestically has gone way up. So those toolmakers, for example, NARA and AMEC and Piotech, they’re now competing outside China with US companies in a way that was inconceivable in 2022. And so what happened, of course, is As a result of the controls, the US companies had to pull all their people out of those fabs in China. guess what? US competitors, US company competitors from Japan in particular, and also Chinese domestic companies had got access to that equipment. And they learned things from that equipment that they wouldn’t have learned if the US companies had been in control of that equipment. And so that’s one just one of many examples of sort of the way you have to look at this if you’re gonna say, did they work? Because as a result of the controls, the US now US companies now have competitors globally for the in the tool making sector. So for example, NARA and other companies in China have been qualified for TSMC to provide tools to TSMC and to Intel and Micron, right? and so now US companies face bigger competition. And the ability of China’s semiconductor industry to pr to eventually produce more advanced chips has gone way up, right? So because the that semiconductor part is complicated. You know, the idea that the US, example, could use controls to forever keep Chinese companies from developing n capabilities sort of, you know, it’s unrealistic, right? Because th this is a this is an applied science. And so the ar the US argument was this is a choke point that we can stop China from doing, but no, China’s designing around that because there’s many ways to do things, right? There’s there’s more than one way to do to develop a tool, for example. And The industry has pursued many different paths and over the years, some have more commercially viable. Those have been the ones that dominated, but now China is pursuing other ways to do things. And so you know, that’s that so like hu like Huawei, just a quick example. So Huawei just you probably saw a couple those last month, I think. They came out with this Tao scaling idea. And so is the re reduction of feature size is to increase the speed and the and reduce the power consumption of these chips. And so that’s Moore’s law has held for a long time. But now we’re run, you know, the industry is running up against just the limits of physics in that in that regard. We’re down to you know one nanometer, you know, really small feature sizes. And so Huawei is saying, okay, well there’s maybe another way to do that. We can we can use a sort of three dimensional structure here to also to move the components closer together and to reduce the time, the latency between signals going to those components. And so that’s not new in industry. This approach has been used before or you know, people have been looking at this. But Huawei is now putting a lot of effort into the actual tools and the technologies to actually do that at some kind of scale. Now the jury’s still out on when that will happen. They’re saying by 2030, for example, they’ll have a s a feature size that will be a system that will be of like a 1.5 nanometer system. And again, the other thing to remember here is that it’s not now just about feature sizes, it’s about sort of the entire package of the system, right? It’s about the memory, it’s about the interconnections, the optical interconnections between the GPUs, where Huawei, for example, has a lot of knowledge of optical interconnectivity. And so you can’t just you can no longer just look at the individual sort of feature size die to die and then and then determine that you know China is ahead of the US or US is ahead of China. So it’s a more much more complicated calculus. And I’ve written about this quite a bit. But you know, I wrote the I think two years ago I noted that you know that now we were in a different ballgame. It was really systems engineering at a at a higher level that’s going to determine you know the capabilities. And here again, Huawei has some significant advantages. so anyway, so the long worded answer to whether the export controls worked is well, yes, of course they worked at some degree to stop and slow down Chinese industry. But at the same time, they’ve accelerated key parts of that industry. And then finally, I would argue the rare earth issue, which by the way I live every day because we’re trying to help companies overcome some of the issues around many licensing and other things, you know, that has been a huge thing because that was directly responsible in response Gallium, graphite. And then of course in April last year, the controls on heavy rare earths and magnets, right? And so those are still, you know, with us. Just today, you may have seen and yesterday, and Nikkei had a story about Japan, Japanese companies who which have been cut off from rare earths in Jan starting in January. They’re filing with the Tokyo stock exchanges are indicating they’re because they’re running out of these critical materials. And all of that result of the of the US export control regime and China’s response, which is to put in place this very strict licensing regime around rare earths, the way, are key inputs for the semiconductor industry too, right? So yttrium, for example, is used to line etching chambers. and most of all those machines I mentioned, DUV, UV, they all use lots of rare magnets for various purposes. every ch semiconductor produced in the world, virtually every one, is touched by a plasma, which is this gaseous, you know, material that’s controlled by Chinese rare earth magnets, and the chambers where that plasma is contained are lined with Chinese rare earths materials. So in other words, the export the US ha put in place have resulted in this very serious response China. That we still are in the middle of. We don’t know how it’s going to come out, but it’s already had a huge impact on the entire supply chain for the semiconductor industry. And not just semiconductors, but of course and power tools and any industry that uses these materials. So anyway, so the disruptions that caused by that are huge. And so when you’re looking at the cost, so we’re looking at the costs and benefits. Did the US slow China’s AI development? Yes, degree, but Jury’s still out on how much. And then if you look at all the collateral damage that those controls cost, you know, those are stacking up and there’s no end in sight right now. So that’s but my view is always, you know, you can you can I agree that the controls work to some degree, but then the question is, you know, what was China’s response both from an from an industrial point of view in terms of working around the controls, and then what was the collateral damage created by the controls and that is that is considerable, I can tell you.

Grace Shao (38:24)

Paul, I love interviewing guests like you because I was gonna follow up with like HBM in the whole picture, how that affects it. You already answered. I was gonna ask you about inferencing versus training on Huawei chips. You answered it. I love you give the full picture. but I wanna ask, what is it like you talked about collateral damage and how these like industries kind of came out because of export controls? Now, how do we understand actually potentially CUDA? I sorry, not CUDA can. Taking some market share, I wouldn’t say lead at all, but some market share away from CUDA and potentially courting more developers globally, maybe beyond just China. How does that new ecosystem and operating system meet work? Because I would challenge and say harmony at this point is still nowhere close to being a dominant operating system, right? So despite you making the point that they obviously had to go around it and create harmony and it does exist and suffice for their own hardware ecosystem. It’s not a leader. How do I understand that?

Paul Triolo (39:23)

Yeah, yeah, that’s a great question. That’s a great, great question. So, you know, this is a this is an older question remember, you know, the China didn’t the Chin there was no Chinese operating system, you know, for j like for PCs back in the day, remember? so we had things like Red Hat Linux, you know, or Red Flag Linux, right? Which was the which was a sort of source Chinese version of Linux that was touted as gonna you know, that was gonna be sort of the Chinese version of Windows because of course China has been dependent on Windows for a long time and still is to some degree, right? And so this big the big this issue of sort does how does how does China how does China develop alternatives to existing dominant software ecosystems like Windows or like Android CUDA. You know, is a is a is a really good question. And it’s and it’s sort of it’s a complicated issue because each of those has a different a different dynamic there. And it’s and it turns out to be really hard, right? Because developers and I know this from installing CUDA software environment on my home computer where I have a I run an RTX 4090 NVIDIA GPU, which is export controlled to China, but I wanted Seek. a deep seek model on my home my home system and I had to install all of this development environment which is very complicated which included you know PyTorch and CUDA and all these things, right? And so when you’re a developer and you’ve been working with all of these things for many years, the idea and somebody tells you, you’re gonna have to now switch over to this other system, which you don’t know, and you don’t know the limitations and the and the strengths and the weaknesses of that be like, Like really? Do I have to do that? You’re not gonna want to do that. You’re gonna resist, right? and so this and same with Nope when Chinese companies were using Windows and somebody said, Hey, here’s red flag Linux, which of course wasn’t very good. and you by the way, you can’t run all your Windows applications under Red Flag Linux. so you’re gonna have to run, you know, weird open source versions of all of all your favorite programs. Again, you know, I did that for a while. I actually Linux exclusively for a while, but then I ended up coming back to because you know, there was certain things I couldn’t do. so same thing here and same thing with Android and Harmony. So it’s a it’s a but as I said, when like when Huawei first started on Harmony, you know, they had a hard time convincing developers in China to use Harmony. But now you know I think that process is pretty far along and other I re just recently you know other companies are starting to use Harmony and so Event it depend and again it depends on their business model. If you’re s a Xiaomi and you want to sell handsets outside of China, you’re probably gonna go with Android because you can still use Google Mobile Services, right? I mean it was really a d a devilishly clever thing for the for the administr for the for the Trump administration to control access to Google Mobile Services because that really killed Huawei’s business China. And that was a key source of revenue, by the way. So that was not an accident, right? Like wh at one level it was like why should they do that, right? It’s not military technology. It’s it’s it’s you know, YouTube and Gmail, right? But the reason was they really wanted to kill Huawei’s handset business. And so that’s why they targeted that. But other Chinese companies can still use Google Mobile Services. So Huawei in that case is operating in a in a in an environment where Android is still out there, they haven’t Android is still available in China. So they have a they’re competing against they’re still competing against Android. Now CAN, it’s tricky here because in addition to Can, as I mentioned, those other GPU companies like Biren and others, their develop they have their own development environments. And those development environments are more compatible with And then in addition to that, Huawei is trying to make CAN and the whole environment more compatible. with CUDA. So the idea is that you know the difference between the two. If it was here three years ago, now the difference you know, is less. And so it that willingness of the developers to move to environment easier as you sort of reduce the differences. And so and it’s hard to gauge exactly where that is, right? Because each and each company is different. So DeepSeek, example, I is different in the sense that those guys were programming the hardware directly, right? So if you don’t if you if you are really good and not that many have engineers that can do this, you don’t need CUDA. You can program the hardware directly, right? You CUDA is sort of this intermediate layer that makes it easier. It’s a bunch of libraries and it makes it easier developers to train you know use the training environment. But if you know how to program the hardware directly you don’t need CUDA. So anyway, DeepSeek is sort of unique in that they were really good at the hardware. and so that’s why it’s important that they’re working with Huawei because they understand you know the sort low-level way that these systems all work together so they can help Huawei to improve the ability of the capability of Huawei’s hardware development environment to more to be with CUDA. And so I think we’re in the process of having that happen What’s probably gonna happen in China is gonna there’s gonna be a sort of s system where Huawei will be and CANN will be used more for inference on the inference side to inference and some and CUDA and NVIDIA will still be used to some degree on the training side. Because remember, it’s complicated. Right now, Chinese companies can still, for example, use remote access to services like in Japan and Southeast other places. to train their models. And so they can continue to use the NVIDIA development environment for that, right? And then and then when you get to inference, they can then use Ascend and they can run that on there and they can they can optimize using CAN to run on the on those on the on the for on the inference side. So we’re in this sort of a weird world where you know there’s the developers haven’t all switched over to Huawei and they still don’t really want to. But more and are there’s more effort And ease that transition CANN with CUDA. And so where we exactly we are in that is hard is hard on any given day is hard to say. But clearly, as you noted earlier and I and I tried to stress, the problem is that for the long term, the Chinese government and these companies don’t know what the US policy is here. So we just saw that hundreds, you know, maybe two hundred thousand H200s will probably be approved by government. For companies like ByteDance and Alibaba and Tencent and others to buy, right? Okay, so they buy those. They can use those. Those are really good for inference. They can just, know, they can they have a lot of demand for their for their services. They can they can throw those in and they can be used for inference. They can also be used for training if you know what you’re doing, right? You can tie a lot of those together. but what next, right? So what is the US government’s policy? basically, under the influence of Jensen Huang and agreed to stop. Forcing NVIDIA to downgrade their chips for a set sale to China. And so Trump said, okay, you can we’ll allow them to sell, you know, not the cutting edge, but something a couple generations behind the cutting edge. So hence the H200 class GPUs. But what’s next? So if you’re if you’re a Chinese company, you can’t count. Any other in the world doing AI design can say, okay, I’m gonna, I’m gonna, I’m gonna upgrade my cluster from H200s to Blackwell, and then I’m gonna upgrade to Vera Rubin, which is the next one, and then I’m gonna upgrade to Feynman, right? So there’s a roadmap of updating your hardware cluster. China, you know, what’s what comes after the H200s? So therefore the pressure is to and this is why Huawei eventually issued a roadmap, right? Huawei had never done a roadmap for any of this, but now Huawei, because dynamic, had to come up with a roadmap. And so that’s why they have the Ascend 950 you know, the nine the nine twenty and nine fifty. So now they have a roadmap out to twenty thirty or tw twenty thirty one that’s their roadmap for upgrading their the domestic processors. So if you’re a Chinese company, like ByteDance or like you know Alibaba or Tencent, all the leading players, you have to figure out a very complicated equation which is how do I keep my core developers who are using CUDA happy, using some hardware, either in China and then how do I gradually transition to using domestic hardware for some workloads, right? And again, these companies can they can run different workloads on different systems depending on what the need is. and so they’re and then at the same time, you know, maybe they can get some GPUs from REN or you know Sofco or some of the other smaller players and run those are primarily inference workloads. And so but they can but those are those are really good, you know, those are very, very capable. GPUs. So they can run some stuff on those and experiment with those. And those are gonna be easier because those are more compatible with the Nvidia ecosystem. So anyway, so have a very complicated hardware environment to navigate compared to Western companies. You know, like OpenAI can just keep its clusters depending on you know how many GPUs it can Nvidia and AMD. so it’s it’s it’s a very interesting and heterogeneous situation here. Where it’s different than Harmony because Harmony is, you know, it’s still developing the developers develop apps to run on Harmony, right? And so you have you have that piece. it’s like a it’s there’s inference and there’s training and there’s a lot of different things going on here. there’s runtime stuff that you’re doing, there’s harnesses, which are the ecosystem around these models that make them capable. so AI development environment is much more complicated than har than for a mo just a mobile operating system. Will. But again, the export controls and the uncertainty of policy really they’re there, right? And you know nobody has said that eventually the US will allow black wells to be exported to China, for But we’re still in this weird Chinese companies can and access those restricted semiconductors outside of China. They can they can run training workloads in Japan, right? And so that loophole may be may or may not be closed over the next year or so. but in the meantime, you know, Chinese companies have options and each company’s different, right? Because DeepSeek, for example, wants to have its hands on the hardware. So they don’t they’re they’re probably not gonna use anything overseas. They wanna have the actual hardware because that’s what that’s what they do. But Alibaba and ByteDance and companies, the hyperscalers that have data centers China. You know, they’re probably gonna they’re they have more options, right? And some of those H200s I think will probably go into could go into data centers outside China too. and then NVIDIA is selling the CPUs now are also really important for some of this. And so there’s no controls. It’s a weird loophole, but the Vera CPU, which is used with the Vera Rubin GPU architecture, can now be sold to China NVIDIA just this in the last couple of weeks is marketing that to China. That’s a very capable CPU, which could paired with other accelerators and used for AI training and other things, right? So that so the compute environment in China is really complicated by the because the are there, but they don’t cover everything. They don’t cover remote access. They don’t cover CPUs. and so Chinese companies now have some options here, but they’re, you know, but again, it’s like what is Here, right? It’s much more complicated for a Chinese AI developer than it is for OpenAI or Anthropic, and that’s that’s sort of the

Grace Shao (51:10)

Absolutely right. I’m really glad you brought up the H200s because I was gonna ask you about that. And it’s very interesting to learn that the CANN like system is trying to become more like CUDA and it makes sense if you’re trying to entice people to move over. Okay, I have a question, I don’t know how to ask it because I’ve heard it asked in both ways. Some people are saying, therefore, why is China still lagging behind if they’re capable of still getting access to certain ships and they’re so talented, right? Or the question could be asked in different kind of framing, which is why like wait, I just asked why are they so behind, right? Others are saying, why are they be able to catch up, play catch up if they’re so limited to, you know, generations ago. So, like, you know, the same question is basically being asked with different framing. At a high level, how do you view this right now? Because frankly, going back to your commentary even on open AI being able to just keep spending and keep purchasing. The most frontier GPUs, then the question is, is that price even justified if you can get almost frontier near frontier with, you know, four generations ago GPUs, then why do you need to spend so much on the latest, right? Like there’s a lot of discussion around that. Is the CapEx kind of justified? I guess this is a big question. See how you want to answer it. It’s complicated. Yeah no

Paul Triolo (52:25)

We should probably do a whole show just on that, because it’s complicated. Yeah, no, that’s a great question. And you’re you’re as you’re you’re really good at asking, you know, the really tough questions here. So I mean the and I think the d the difference then is that in the US, this idea of scaling, right? The scaling laws still hold. So the more GPUs you throw at training, the better the models will be. You know, that’s still sort of the view in the US. And it turns of that may be true to some degree, but there’s a lot of factors, for example, besides scaling that make models capable. There are these, there’s the harness thing, right? Which is the which is the ecosystem, the orchestration around the model. That’s really important in terms of the performance of the model. What tools can the model call, right? There’s a whole huge effort, you know, to standardize the calling of MCP protocol. which is used to connect the model to other applications. I just hooked up, for Claude. I gave Claude access to one of my brokerage accounts. And it can go in there and pull all the data on all my investments and analyze it, right? and so it does and it’s really good. And it learns, you know, more about you know certain other topics. A little bit well I have I had to sign away I had to tell the brokerage made me like you know sign away all the rights to any you know that happened.

Grace Shao (53:38)

That sounds so risky, Paul, and you’re so brave.

Paul Triolo (53:50)

So anyway, but the point is that the raw model and the scaling and the GPUs, it turns out that there’s more the scaling does still hold, right? I mean, Dario from MADA, where of course the CEO of Enthropic, you know, he was like discoverer of the scaling laws. And so you the major US labs like OpenAI and Enthropic and Google, and you know, there’s still the sense that the that the more GPUs throw at it and the more training you’re doing, you’re gonna get better models. But it turns out that like DeepSeek and others, there’s you know, through optimizations, because they don’t have access to unlimited compute, they figured out ways to optimize the and to enable them to run more cheaply. and that and that’s that’s affected the diffusion of the models. So if you’re when you talk about you know who’s ahead, there’s sort of raw model capability is one thing. And then there’s like who’s using the models, and everybody, it turns out that everybody doesn’t need the most advanced models. To run to run really useful applications, right? And so that’s where the Mabel fable and mythos thing, it turns out that you know people are now worried that the US government will, for example, cut off access to these models. And so why wouldn’t you use an open s a really capable open source model from China like GLM or Moonshot? Kimmy is very popular in the US. And you see in these recent reports that you know Coinbase and Microsoft and all these US companies are considering or using Chinese open source models in production, right? And so the question there is, you know, those models are exactly as good as the US models, but they’re pretty good, right? They’re good enough. And so it’s is the question of who’s winning the sort race, you know, is sort of less material in some sense because of these other factors, right? And so it turns out that the that both the model capability, once it’s near frontier, it’s good enough to run most things, right, that need. Some companies will still want to have the most cutting edge model. And also they’ll wanna have the issue of like their where their data is, right? They’ll wanna trust the company that’s that’s running their data whether it’s through an API. they’ll wanna trust that company their data. and maybe they don’t wanna they don’t wanna do that you Chinese model cloud, but they might be willing to run it on premises, a Chinese open source model and build on top of that. And that’s that’s that’s also what Fable and is sort of forced issue. Now companies are thinking, why do I want to give all my data to Anthropic or OpenAI when I can run a very good Chinese open source model on my own infrastructure and I can control the data and the and the security of that of that of my you know my business model. you may have seen Alex Carp’s sort of rant, people some people called it a ramp a couple days ago where he was talking about that issue where he was basically saying that, you know, don’t want to give all their data over to these model developers because then those model developers will compete with them for certain things. Which what’s happened like what

Grace Shao (56:41)

Yeah, they will eat their lunch instead. And people also have a misconception that like when you use a Chinese model, it’s not like you’re giving the data to an open source Chinese model because you actually can self host that model I in your home countries. Anyway, I’m gonna start wrapping the conversation. I want to go big picture. Last okay. Last question on this. The dominant narrative, DC, is that, you know, AI policy circles, you know, circles often talk about, you know, whoever achieves AGI first, whatever that means these days. Will gain a decisive strategic advantage and ability to reshape global power. So it’s very, very scary. You know, how do you view this? Because through our conversation, what I’m hearing is that both sides are, you know, cautious, both sides are healthy and skeptical, but both are putting regulatory pressure, whether domestically on the companies or, you know, on protecting them from, I guess, foreign actors. Is this actually conducive for the future? Like how should we kind of view this? Because It also feels like from our conversation, a lot of these export controls, protective measures are not actually working or actually good for the industries domestically. So just a high level, like how do we understand this? Yeah.

Paul Triolo (57:47)

Yeah, great question. I think level, my concern, and I’ve written quite a about that, you know, if we race argument that the US is indeed to prolong the gap between US models and Chinese models so that when we reach something and I think AGI, I think we’re already at We’re already the models already passed the Turing test, right? But we’re talking about like artificial superintelligence where models are you know self they’re self-improving and they’re they’re coming up with really amazing new designs or weapons designs. If you look at the AI 2027 scenario, that’s sort of that’s sort of how people are thinking. Now I’m skeptical of that scenario because I think you know we’re still away from these models being to do you know the design super weapons and take over everything and so that one side who gets there first wins And then can kind of lord it over the other side. You know, this is essentially what Dario saying in his essays, like the machines of grace, and the adolescence of technology. He has said basically like AI, d democratic AI needs to win so that then that can be used to sort of force regime change in China, right? Or force authoritarians to sort of, you know, kawtow, if you will, to Western the Western governments. But I think that’s a That’s a I don’t like that framing because I think you know we’re not gonna wake up one morning and have that capability. It’s gonna be a gradual thing. and then the real issue is how governments deal with this? Like this whole issue of fable and mythos has forced the issue to the fore of how do governments deal with even just capabilities? This isn’t super intelligence, but this is like really good capability to detect vulnerabilities and software that exploited. And we don’t even have a framework for that, let alone for something more advanced intelligence. How would the government and industry work together on that, right? So there’s a couple things. One is, and I’ve written I just wrote a piece in Cairo Review about when does the government think about nationalizing the AI labs, right? Because people are using these analogies like these are nuclear weapons, right? Even though AI is not nuclear weapons. And I think it’s a very dangerous analogy. But people are saying, you know, this is a technology developed in the private sector, previously, weapon systems that were very capable were developed by government. And here we have a private sector ca capacity that’s that’s starting to edge towards weapons systems with cyber capabilities, for example. What is the gu how does the government do fit think about that? And the mythos thing, frankly, showed how unprepared the US government was for this, right? If you’re in the industry, you know that you knew that this was coming. I we talked about this last year at the Paris AI conference, right? You knew that this capability was coming, but nobody in the US government, the Trump administration was just saying innovation, innovation. China’s in the same way, right? How do you balance regulation and innovation? They want the companies to compete. So there’s no regulation. now Mythos and Fable have forced the issue of like, my God, well now we need to have some government role in determining, you know, how to test the models for certain capabilities and how to determine what’s a covered model and what should the conditions be around w how that model is released. And at least with Fable, we saw company, in this case Anthropic, have to, you know, put guardrails around the cyber capabilities of that model. And now they’ve finally the government has allowed them to release Fable. But that’s not there’s still a lot of questions around that, right? So the problem is if we’re in this, if we accept this race idea, then we’re never going to get collaboration between the US and China here, which I think is really dangerous because then, you know, malicious non state actors are gonna get access to this capability. and then, you know, the implications of that are really, really serious. And so the problem with the race idea is that it forces everything is it that then becomes distrust. The US distrusts China, China distrusts the US, you know, the US is gonna ban open could ban open source Chinese models. China could you know r restrict the release of open source models because they don’t want to contribute to The US developing capabilities. So we’re gonna get if we get into this race, then it’s a bad thing for everybody, I think. So we have the US China AI dialogue, which is coming up hopefully after the World AI conference in Shanghai, which I’ll be attending next week. and you know that’s gonna I think that’s the last chance. It’s the last chance for the US and Chinese governments to say, okay, we understand we don’t trust each other, but this is a threat, the threat of you know uncontrolled access to these models. It’s a exactly. It’s so it’s the last chance for the for governance to recognize that. And I think we I think that’s the mythos fable thing. The good news is that really drove I think this agreement in Beijing and Mart and May to between the t the two presidents to start talking about this. But it’s a complicated issue because ha you know, you got what are what is the goal here? How are you gonna agree on both sides to some limitations on this, right? And then how do you how does this translate into eventually a global agreement on putting guardrails around frontier AI models. It’s it’s it’s the problem is the technology is developing so fast that the ability of governments to keep up with this and come up with, you know, credible and viable structures to put some controls around this, you know, it’s really tough because there’s just not enough expertise in government. It’s going to probably have to be an independent, private sector led effort to do this. And this is This these are the kind things that are going to be discussed in week. I’m on a couple of panels, including some closed-door panels, that where these issues will be discussed. Now nothing’s gonna be decided next week in Shanghai, but I think the level of the discussion will be much higher because of fable and mythos and because the Chinese government is kind of freaked out about this. and you know, and the good news is that at least at some point the US and China will eventually sit down and try to begin this. Scott Bessent is gonna head up the US side and Vice Premier He I just did a piece with Alvin Graylin that you’ve probably seen in on the ASPI website, which tries to look at who’s gonna participate in this from both sides because there’s lots of lots of equities and we saw the that whole issue become complicated just in terms of deciding what to do about Fable. it was good in the sense that the governments to have a serious discussion of what are we going to do about this, right? So that’s the good news. But when I was at the way the finally, when I was at the World AI conference last year, I think it was Stuart Russell, my good friend Stuart Russell, who said, you know, he had talked leading CEO of a of a US lab, and he had said the best thing we can hope for in the next two years is a Chernobyl style event, right? Now think of what that means, right? This is the head of a lab. admitting that you know AI could lead to a very bad outcome here, right? and so this is where we are here in the summer of twenty six, where US and China both have leading labs. Governments don’t seem to know how to get a handle on what to do about that. But the US and China have to talk about it. Because if we get into this race, if we if we if we if we basically give up sort of say, okay, it’s gonna be a race, you know, and then th what will happen is something bad will happen and then and then people will say, my God, now we this. and you know, some people think that Fable Mythos thing is good because there were there was no bad event, you know, no loss of life or no nothing. But it did sort of force people to realize, okay, now we need to do something, you know, that’s a good thing. But still the political pressures and all these things we’ve been talking about, the export controls and everything, you know, there’s no there’s just no trust on either side. we’ve dug a deep hole in terms of and so Digging out of that, as you say, you know, to for the benefit of humanity is gonna be r a real challenge now. but you know, hopefully, as I say, in the next couple months, we’ll know better where that dialogue is gonna go and where China’s gonna be, for example, coming out of the World AI conference. They might announce the World AI Cooperation Organization, for example. I suspect they will announce that. Xi Jinping is coming, just to show you how important this issue is. Xi Jinping is coming to Shanghai. So the security arrangements around this conference are nuts. I’ve just been trying to figure out where I’m gonna be on different days. so that shows you how important it is. If Xi Jinping is coming, it’s important. and if Xi Jinping has agreed with President Trump to discuss this at some level, that’s that’s good. That’s good news. But as I say, I think this is like humanity’s last chance to get a handle on this because you know it’s the US and China have to agree. Everybody else matters, you know, there’s a big safety community, there’s other capable model developers in other places, but really the US and China are where ninety percent of the action is. and so if there is no agreement between the US and China, beginnings of an agreement around this, then you know, then all bets are off. And so I think this is a really the next couple months are really critical in this in this arena. And you know, the and that’s the technology continues advance and recursive self-improvement kick in. And as, you know, these things, you know, that’s not gonna stop. There have been all these efforts to say, hey, let’s let’s stop until we figure out what to do, right? Let’s pause, right? Who’s gonna pause at this point, right? I mean, those a year a year and a half ago that all these scientists, including it from China, signed on, like, we gotta pause, six month pause. The hard part is if you pause, you know, how do you decide when to start up again? Right? it’s it would be you know impossible. So nobody’s gonna agree to a pause. So therefore we need to agree on A minimally viable framework around governing these advanced models. And that’s what the goal is going to be in the next couple months. But you know, it’s it’s it’s really going to be hard because of the government the lag in capability in government you know so the mythos thing highlighted both need to do something but also wow like it’s the people who really understand these issues are still limited in number

Grace Shao (1:07:46)

Yeah, we need more technical people in the government. Like actually every government. That’s the thing. Because this technology is not for the laymen to understand, frankly. Like you need someone who’s technical to understand it. But I think, okay, on that, like I’m feeling serious FOMO. I was planning on not going, but maybe I’ll go up. It seems like everyone is going. I was speaking to Alvin Graylin this morning actually. We’re working on a piece together. So it’s very interesting. I’m glad he’s he’s potentially going, you know, Ray Ma’s going, a bunch of people in the circles going. So You know what, like I think you’re right. Like I really do hope that something positive comes out of this. It’s just seems like it’s really hard to regulate something when regulation takes so much time. There’s so much bureaucracy that comes with it. And then on the other hand, like exactly to your point, AI doesn’t sleep. I was joking with my husband, I was like, I just want to summer. And he’s like, AI doesn’t summer, you can’t summer. And then he like being a tiger husband there. But you know, the reality is no one’s gonna stop right now, right? And like

Paul Triolo (1:08:28)

Right, right. I want three months off from all this, but catch up.

Grace Shao (1:08:42)

There’s a commercial interest and there’s also the com competition competitor like I guess even spirit in these researchers at this point. So it is gonna be incredibly hard. Yeah. So are these things gonna be nationalized? Do you think these like I mean, the irony and all this is like a look at my Cairo Review article? Yeah came out last month when I tried to lay

Paul Triolo (1:08:50)

Right, and we also have massive IPOs coming up, right? We have anthropic that’s the other complicated Well, take a look at my Cairo Review article that just came out last month, and I tried to lay out how both the US and China view this. And I think y arguably already, you know, there is some there isn’t national you know, is gonna happen in different ways, but some level of nationalization is gonna have to happen, right? We’re already talking about open AI g you know, pr that the government taking a taking a investment or taking a share. OpenAI. So that’s kind of a there, right? so I you know it’ll be it’ll be different than nationalizing other industries, right? But yes, I think at some point it’s it’s hard to see the government leaving this capability in the hands of the private sector fully, right? because a important capability. And as we get closer and closer to more advanced you know AI and The idea of like loss of control, what happens, what happens if we lose control of the AI? all these things are out there. And so I think, but again, we can’t even figure out basic government role in, you know, how do we how do we determine what is a covered model and who is who is equipped to test that model, right? there are very there’s some efforts going on that I’m aware of to try to figure that out. Right. And it’s gonna but it’s gonna it’s not gonna be just the government. It’s gonna have to complicated, you know, body outside the government that’s that’s that’s that’s plugged into the government, kinda like the IEA, right, for n for nuclear for nuclear technology. But it’s right, there needs to be standards and there needs to be there needs to be a sort of neutral international body. But you know, we’re still quite a ways from that too. So we first have to get US and China to at least agree, Then building on that There could be the some new body. Now, again, I my the cynical view in the AI safety community I is that there has to be first a Chernobyl style event, right? That hopefully won’t be too serious before people get concentrated. Yeah, yeah, it does. It does. It’s ac absolutely Right. Right, right. But that’s sort of this that’s the worst case sort of cynical view within the AI safety community. But upcoming US China dialogue, it’s gonna be really important to see who’s participating, how serious it is, and you know, how quickly something can happen, right? Because the safety community has been arguing, we’re getting closer, we’re getting closer to artificial and it’s gonna take time. and so we need to start the serious discussion now. And that You know, that it started happening under the they were very serious about this, very thoughtful people. But then when Trump came in, it was basically let’s let her rip, right? Like US innovation is gonna dominate AI, and there was really a downplaying of governance. And then, you know, mythos sort of punctured that optimism in some sense and was like, okay, now we have to do But, you know, having not thought about that for a long time. You know, there were thoughtful people like Dean Ball and others who contributed to the AI Action Plan. And who’s now jumped up and out? Dean’s a great Dean’s great. I love Dean. He’s a great thinker on all these issues. So there are people out there who’ve been thinking about these, but it still turns out to be really, to, for example, set up a new organization. Late in the Biden administration, there was a discussion that Frontier AI is so different, right? It’s a different technology. You need a different regulatory structure around this. But it’s really hard to do that, to set up a whole new body and fund it and find but now I think people realize no, we this as another technology that we can just fit into our existing regulatory structure. It’s a different problem, it needs different capabilities, and so we need to rethink how to do this. I think that could happen too on both sides, both in China and the US, is okay, we need a we need to figure out a new structure here, an organization with the right authorities and the right capabilities and the right technical expertise to actually manage this problem, right? and I have I have a paper coming out with Alban’s not part of this paper, but I have a paper coming out with ASPE that’s looking at you know how one potential structure that could down this road having and both between both the US and China, right? A structure that includes the key players on both sides that would allow this to happen. But again, very tough you know, we’re it requires a sort of trust and concessions on both sides to figure out how to do this right. and, you know, the bilateral tensions that you see every day, right? are still a real impediment to this, right? Because AI in Washington, as you know, has become such a charged issue. You know, I mean we I mean people are talking about, you know,

Grace Shao (1:13:59)

But some of it’s talking point and some of it’s reality. I feel like at least in the business world, right? Sh maybe the policy world should have a little bit of that too. You know, what happens on the surface, what happens under understanding the reality and the realistic consequences that these talking points may lead to.

Paul Triolo (1:14:01)

Chinese. Right. Right, right. But Right. Well Right, no, that’s a great point. My the is that the during the later Trump first administration and the administration, this constituency developed around the AI issue, right? This sort of this weird sort of consensus that AI China, you know, we had to slow China down, we have to restrict these things. And that there’s a you know, that was in the in government, in the in the media, in think tanks. So there’s this huge sort of constituency of people. Who are wedded to the idea that we have to win over China at all costs on AI, right? and then there’s a group, the AI safety groups and others, and then people like me who were saying, well, no, that’s not the way that’s the sort of that zero sum thinking is gonna lead to disaster, right? and that we need to fig figure out a bet a better way. Like that the better way is how can we collaborate with China in these areas where we do respect national security concerns, but we don’t over index on them to the point where we can’t collaborate with China and then, you know, it’s a free for all and you know, bad things happen. and so that’s sort of where we are now. And I think the good thing is the Trump administration isn’t wedded necessarily totally to the previous administration’s approach to this, but it’s still it y you need smart people in D C like David Sachs and others who understand the industry and where the industry’s going and the technology who can kind of who are outsiders. outside the beltway who can actually look at this more holistically and say, okay, wow, we can we can work with China here, we can compete with them there, we can we can, you know, control certain things, but we need to figure out a way to skin this cat. We need to figure out a way to get to some basic level of agreement here. Otherwise, you know, we’re all in for a world of hurt, as that CEO of the of the lab admitted in a private setting I mean Chernobyl style events sounds pretty serious, right? And avoiding that needs to be something that focuses people, in DC and Beijing on, you know, how to how to how

Grace Shao (1:16:23)

Paul, I agree with you. You are full of knowledge and insights and you’re full of differentiated views, but there is one question I ask every single guest as we wrap up the conversation. what is one differentiated view you hold? You think that’s something just very against consensus?

Paul Triolo (1:16:41)

Well, I just think that technology controls and the idea of choke points really bad idea because we’re in a world as an interconnected world, right? And in fact, like when I’m working with clients across the AI stack every day. And when I look at US China, you know, y the degree of inner of interdependence and interconnectivity here is much deeper than people think, right? People are like, decoupling here and there. No. I mean if you really if you really look at the at what’s happening on a day-to-day basis and the complexity of supply chains, for example, the idea that we can simply decouple in AI or elsewhere is just, I mean, yes, we could do that, but the cost of that to the to the to the and to the companies and to global supply chains is just, you know, really fully accounted for that. So my I of like always coming back to the reality of okay, what is what is the what’s what’s what’s happening with businesses on the ground on a day to day basis and how are they being affected by this, right? And when you look at that level, you the sort of, you know, the thinking and comments that I hear that are just are very divorced from sort of that day-to-day reality of how interconnected the US and China have become over the last thirty years. And, you know, if we’re gonna indiscriminately, you know, pursue policies that where that collateral damage and the sort of the full cost benefit analysis done, you know, then it’s like, what’s you know, what we’re what are we doing here, so I’m always just I’m always just sort of arguing for a thinking about policy that’s based on a sort of a really deep understanding of the reality on the ground and you know how innovation happens and companies de-risk supply chains and how there are certain dependencies. For example, like rare earths turns out to be a real choke point, right? In the way that semiconductor technology is not, right? There’s just way around China’s Chinese company’s dominance of say samarium cobalt magnet production, right? That’s a real choke point. that will take ten years to you know to unravel. whereas other choke points that have been used on the US side are not really choke points. So I think they’re just my differentiation is the need to step back and look at this and think about like what is the point particular policy? Is that does it make sense? And is it is it having you know is the cost benefit sort Clearly on the side of, you know, too much cost and not enough benefit. and so that’s what I keep coming back to. And then we didn’t even talk about Taiwan. My also my sort of nobody few other people talk to is the impact of all this potentially on Taiwan and the risks around Taiwan. We work with companies every day and we do exercises, for example, about around a risk around Taiwan to their supply chains. You know, something short of a military exchange, which then there’s no de-risking. but you know, it turns out that, you know, Taiwan and supply chains and Asia in general are so intertwined that when you start pushing on some of these buttons, the worry I have the worry that you’re gonna, you know, you’re increase the potential for disaster there, you know, unintended or intended or whatever. and I think not enough people are thinking about that. they’re only thinking about you know, deterrence and arming Taiwan, think is frankly a sort of a mistaken way to problem. so anyway, so I think that out-of-the-box thinking and sort of getting a getting a getting away from the standard view which has developed over the last 30 years, you know, is necessary. And I just don’t see enough of that in Washington or Beijing. how do we rethink some of these things in the age of AI? So what we need a we need China policy for the age of AI and we need a technology policy for the age of AI, And I think

Grace Shao (1:20:46)

We need more dialogues. The thing is like I hear from so many people, whether they’re working on policy side or the actual researchers and developers, they’re like people aren’t talking officially because of all the geopolitical headwinds and noise. And then but actually, you know, obviously people talk, you know, behind the scenes, but we need more official dialogues to get to more fruitful results. I think more Yeah. So Paul, I’m glad you’re going to WAIC. You’re gonna be leading these dialogues.

Paul Triolo (1:20:47)

That’s where governments and we need more dialogue. Yeah. Mm-hmm. I’m with you, Grace. you. And I really appreciate your perspective. Well, I’m gonna try to contribute a little bit here and there. I mean, I love the people that I’ve met with a lot of the Chinese AI safety people, for example. They’re very thoughtful. they’re very good. and you know, in some areas ch China’s China’s Chinese you know, organizations and individuals are leading. But also there’ll be a lot of really good people from the broader AI safety community right? From the Future of Life Institute and from Concordia AI and s you know, really good Players who are really have smart people that are thinking about these problems also. So it’ll be a really good effort. Unfortunately, because it’s in China, you know, some of the leading US AI labs and some people and the US government, you know, will not be participating in this. It’s it’s seen as a sort of Chinese thing. but there will be the a the APEC meeting is happening just after this. And so I think there may be some there’ll be a US presence at the APEC meeting. And then as I said, you know, eventually. Probably shortly after this, I imagine that the US and China will kick off this AI dialogue. And so you’re right, dialogue is really critical here. And this is such a complicated issue that, you know, the sooner this dialogue gets kicked off and the sooner that they can they can, you know, feel each other both sides can feel each other out and get to the real issues. Yeah, exactly. Exactly.

Grace Shao (1:22:36)

Paul, I’ve taken up so much of your time today. I really appreciate it. I’ve learned so much. Can we please do this again sometime? I have more questions for you. You have you’re so knowledgeable, but thank you so much today. Thank you for your time.

Paul Triolo (1:22:36)

Yep. And thank you, Grace. I really appreciate your thoughtfulness and the and the thoughtfulness of your questions on these complicated issues. You really bring a lot to the conversation.

As I was prepping for my day trip from Shanghai, I was dreading the hour-long taxi ride from the bullet train station to Alibaba’s Xixi campus that I used to have to take during my stay there. But to my surprise, a former colleague told me that there is now a new bullet train station, Hangzhou West, conveniently located near Alibaba’s main campus. Trimming down that hour-long taxi ride to 10 mins.

It takes about an hour on the train from Shanghai Hongqiao to Hangzhou West, and when the doors opened, I couldn’t help but chuckle. Is this the uniform or the epitome of the Chinese techfit? Everyone on the platform was wearing a Tumi backpack or laptop case. It was quite clear that everyone looked frankly more “expensive” in their outfits getting off at that station, and there was a clear urgency in their footsteps.

I got into a taxi and started chatting with the uncle driving. If I hadn’t been back in five years, he told me, then too much has changed. New highway bridges, rows and rows of new development, and of course, the completed new campuses for Alibaba.

So I asked him, “Does that mean housing prices have gone back up with the AI boom?”

“Oh no, no”, he said, only a few are buying up luxury apartments, but they work for semi companies. He added that he heard kids working in AI are making 2 to 3 million RMB a year, which seemed like a crazy number to him. But those kids working in AI are so tired that they get into his car and pass out, and he has to wake them up at their stop. In awe of their talent, but also worried about their health, as one would expect a typical Chinese uncle to be.

Now, I was flabbergasted at the response from this uncle, frankly, who could barely speak Putonghua clearly, but had the sophistication to know that only the semi companies, their investors, and everyone along that supply chain have made a lot of money this year, and that the rest of tech and AI is still mostly waiting for its turn.

In some ways, that taxi driver in his 50s in Hangzhou had a more accurate understanding of the cycle than a lot of professional analysts.

Before, I felt like Hangzhou hosted Alibaba. As in you know by default, the dominant payment system is Alipay and not WeChat in the city, but it still felt like Alibaba was part of it, rather than the other way around (at least in Hangzhou East).

But getting in from Hangzhou West, it doesn’t feel like a city that hosts Alibaba anymore; it feels like an Alibaba campus with a city attached to it.

You know how tech companies have sprawling campuses? This whole city is an Alibaba campus. The train station itself was reportedly built in part to make the Shanghai commute more convenient for Alibaba staff, and apartment complexes went up around the new developments Alibaba established. The spending is crazy, and it cuts both ways. On the one hand, it shows confidence in the company’s direction; on the other, when you speak to investors, there is a level of skepticism because they are wondering how much of that money should have been spent versus how much should have been returned to shareholders.

With that, I will guide you through my day, my observations, and thoughts in chronological order. This is from a day spent meeting with four (five) business units: the Qwen App team, Accio, the Qoder team, and a brief hello with Banma and (light touch) Alibaba Cloud.

Qwen: The closed activity loop is real, but who is willing to pay for it?

The Qwen app is extremely impressive in how it connects all the commerce activity, even though we wrote about how it hasn’t fundamentally changed how users engage with the idea of shopping yet.

But this is Alibaba’s edge if tokenomics eventually comes down. Think about it, everything can be completed within its own platform. E-commerce, same-day delivery, grocery, maps, ride hailing, bookings, payments, all integrated, agentic, agent to agent. This is something we’ve already seen OpenAI couldn’t really figure out with its reliance on third-party partners. It’s easy to push a user to DoorDash and then to Stripe after they’ve shown intent to buy, but it’s much harder to have the agents link up so the full loop, from car booking to airplane to hotel, actually completes.

In this case, Alibaba owns the rails, so the loop closes all under its control. There is good and bad to that. The good is that this activity loop could even exist; the bad is that you bear all the liability, right?

The representative I met repeatedly reminded me that any purchases or transactions still require human verification to prevent mistakes, but if we are to push for an agentic commerce future, maybe that step will eventually be removed?

Anyway, what I found is that the Qwen app is more nuanced than a checkout bot. The app is memory-based, so when I want a dress, it already knows my size, my body type, and my usual preferred style, and unlike GUI-driven agents, it doesn’t need to operate a screen because it calls directly on stored memory. Again, this is something unique to Alibaba, having the Taobao purchasing data.

It behaves like a shopping guide rather than a fulfillment engine. The PR said, if you want to buy a jade bed because influencers on RED are saying it helps with chronic illness, they will tell you not to. It has built-in authenticity, and I guess, in some sense, integrity.

Mention at night that your cat isn’t sleeping, and it will suggest it might be an illness or a calcium deficiency and ask if you want to buy something for it. This agentic shopping companion is its future vision.

The team’s framing, which honestly I buy, is that China already app-ified every booking and ordering need over the past fifteen years, and it is much easier to AI-fy a mature internet infrastructure than to build agentic commerce on top of fragmented rails.

The payment step stays deliberately human, so the final purchase has to be personally confirmed, except for utility goods like the weekly cat kibble or toilet paper, for which you can choose to authorize a recurring order. Shopping has been live since May, and consumer usage is reportedly growing 2x month-on-month; overall data is not publicly disclosed yet as it’s still in its relative infancy stage.

One example that really impressed me, and that I would personally pay for, is planning a day trip from Shanghai to Hangzhou. All you need to do is tell the agent the time you want to arrive and your preferred transportation, and it will plan out your time, your snacks, your coffee, and everything in between, frankly, even more meticulously than a human secretary.

But right as I was getting excited about this, it dawned on me: how can you justify the cost of token usage for such simple activities?

I would pay for an app to manage my itinerary, but how much would that token cost, and would it be worth it? I guess it really depends on how expensive your own hourly rate is.

And there is a second problem that pricing can’t solve, which is liability. If there is traffic and you miss your train, will you blame the AI or your own judgment? Who bears that risk, Alibaba, the airline partner, or you?

The bigger question here isn’t really an Alibaba question at all. It’s a question I have for the whole industry: how much is a consumer actually willing to pay for agentic support?

Just as Azeem’s team wrote in the recent report The State of the AI Economy, “The AI economy is bigger and faster than any technology wave before it, yet still small enough to be early.”

That is how it felt. Qwen, in many ways, felt ahead of the curve. But maybe the infrastructure cost-benefit just isn’t mature enough to justify such a use case yet? Because the stock price has not been a friendly reflection of this vision. That same report shows that AI is scaling three times faster than any prior IT wave, but much of that is from the enterprise level. Who is bearing the cost for consumer usage?

This is why OpenAI’s pivot last year back to enterprise rather than consumer shows a bigger revelation in the industry. Right now, there is only willingness to pay in enterprise because AI is fundamentally a productivity-gain technology, and only businesses really prioritize productivity, plus a small percentage of people whose hourly rate justifies it. Look at Doubao; its MAU dropped the moment it started charging.

The rest of the consumers are still looking for value for money, really, which is exactly why PDD and Temu, despite the hate, are doing so well. How much is the consumer agent really going to be contributing to GMV? And if Alibaba does want to monetize the transactions themselves by charging a premium on each agent-completed service, how much is justifiable before the consumer just does it themselves?

For now…I think there is still a lot to grow, iterate, and unpack as tokenomics matures and AI user behavior evolves. This is a watch-and-see app for me… no conclusive verdicts yet.

An interlude at Hupan

Between meetings, I visited a newly opened exhibition hall where they have installed a replica of the Hupan apartment, which is where Jack Ma and the original team started. The PR shared a funny anecdote: because there were so many people crammed into this little apartment and only one squatty potty, people joked it was “Lundun,” which means “taking turns squatting” and obviously sounds just like London.

Jokes aside, it was very nice to see how much respect is paid to the heritage of the company’s founding. And a special Chinese corporate thing is to have showrooms open to guests.

Accio: AI UX on a twenty-year supplier data moat

TBH, my real goal in going to Alibaba this time was to learn more about Accio, and it was the highlight of the day, a business I find very undervalued. I read about it a year ago, but at that time it was still merely described as a merchant support AI tool.

I explained on the ~100-investor JPM call on Friday that it really undersells it, especially now after a year of iterations. On the surface, it looks just like any other agent service for merchants, and the interface even looks like Claude. But it relies heavily on Alibaba’s sprawling supply network, which is its ultimate advantage.

Accio is less an “AI product” and more an AI UX on a twenty-year supplier data moat. This is the clearest enterprise 2B use case Alibaba can push out right now, and frankly I don’t even think it’s getting enough attention internally at this point.

The value of Accio isn’t just the AI agent. It’s that the AI sits on top of 1.5 million verified suppliers, 400M+ SKUs, and decades of transaction and rating data that Alibaba owns because it is the B2B marketplace. There are already millions of MAUs.

And there is also a reason that there is no Accio equivalent in the West, because the supplier graph simply doesn’t exist there for anyone to build on. No Western company has that asset. Even Amazon can’t do this.

For a merchant, it is essentially everything in one; it shocked me how it removed the barriers for business/ drop shipping and all the tedious search and due diligence work that is needed to start a brand. Look, I even toyed around with the idea during COVID because I was bored, but it was much more work than expected to source the right product and manage the distribution channels. The best way to understand it is to walk through what a merchant actually does all day, because Accio covers nearly every step of the workflow. Most users are startup entrepreneurs, brand owners and e-commerce sellers, brick-and-mortar retailers, and service providers, with 30 to 40% coming over from alibaba.com.

It starts with supplier finding. The agent sits on the supplier graph and can find, filter, and talk to vendors for you. Autochat can speak to vendors through one interface, say, for sourcing just the lid of a bottle, run the first round of filtering against your core requirements, and then hand off to you to speak to the shortlist directly. Sometimes it is literally Accio talking to Accio on both sides of the negotiation. And the sourcing guidance is real know-how, not chatbot filler. One buyer didn’t know the measurements or the requirements for their product, and the agent guided them through the safety requirements, the weather conditions, and what those mean for raw material choices- the kind of professional judgment that lives in the network and twenty years of transaction data, frankly. It’s not more intelligence but more know-how. [Industry-specific use case: AI tools are the future is what we’ve been writing about!!]

From sourcing, it moves into pricing and margin. The agent negotiates pricing with vendors behind the scenes over multiple rounds, then benchmarks against industry comparisons and hands the merchant a rough margin before they’ve committed to anything. This is the part I keep coming back to, because it means a small merchant gets the pricing intelligence that used to require a sourcing office in Guangzhou.

Btw, I know I’m like super positive about this product. Swear to god, I got zero dollars from any of them. I self-paid for my trip and all.

Then the operational layer: supply and inventory management, plus the back-office work around it. Accio Work claims to handle VAT filings across 100+ markets, multi-round supplier negotiation, and logistics ops, though how much of that is truly AI-executed end-to-end versus AI-assisted with a human in the loop is one of my open diligence questions. On the storefront side, it covers CRM and platform management, lead generation, content management, and SEO and GEO optimization. It connects to Mailchimp, WordPress, and the like, where some opened up access to Accio, and some skills were built directly by the Accio team, so the experience is smoother, and it is opening up WhatsApp and Telegram so your vendor conversations get summarized straight back into Accio instead of dying in a chat thread.

And then distribution. Like HELLO. Plugins are already live on Amazon, Shopify, Etsy, and Shopline, and you can swap the 1688 agent for an Amazon seller toolkit depending on where you sell. The data runs both ways, too, internal Alibaba data from Alibaba.com and 1688 alongside external data from Amazon. A marketplace’s self-run toolkit only works inside its own marketplace; Accio pushes to every channel the merchant sells on.

The dropshipping honestly would be perfect, and it is incredible. If I were without kids and had more mental bandwidth, I would SOO create my own loungewear brand with this. (During COVID, I had ideas of setting up my own Shopify shop, but the process was actually much more tedious and consuming than I expected. I toyed with doing dropshipping and creating a pet product storefront, opening up a bagel shop in HK but realized you can’t mess with that here, there’s enough drama already, and/ or a luxury loungewear brand inspired by my West Coast upbringing living in leggings and Birkenstocks)

Anyway, back to our main topic. So you can give a verbal prompt. Accio provides the sourcing, then provides visuals, then technically can communicate the pricing behind the scenes given that you’ve given it permission, then provides an industry comparison and a rough margin, and then it creates the Shopify store link and can run the whole website with product descriptions and everything, and helps with social media and image banners. It is sourcing to storefront to marketing in one conversation box.

While I was going through the demo, what really stood out wasn’t the chatbox or the interface or anything. It was the supply chain layer where the real differentiation was, built on their own network and sourcing know-how, while store management, marketing, and operational support were more similar to what other agents could do, helpful but not as unique.

This is also where the willingness-to-pay math finally works. A consumer won’t pay for twenty minutes of saved itinerary planning, but a merchant will absolutely pay for an agent that compresses sourcing, negotiation, storefront creation, and marketing into one workflow, because the ROI is measurable and the alternative is headcount.

I would love to have someone from the team join us on the podcast to talk more sometime.

Models, Tokens, Bureaucracy

The most candid conversation of the day was with someone in the cloud/Qwen unit. Qwen, he admitted, still lags well behind GLM on coding — awkward when you’re selling Qoder, because in practice many Qoder users are paying Alibaba for a managed service to run GLM. In effect, Alibaba is taking a margin on its rival’s model.

Domestically, this doesn’t matter much. There’s no dominant Cursor or Codex equivalent in China. I mean, ByteDance’s Trae competes, but Alibaba’s enterprise reach carries the product. Outside China, it is a different story, with competition obviously led by Anthropic and OpenAI.

However, I’ve heard a few voices saying that they feel startups have an advantage in leading innovation because of the lack of layers and bureaucracy, which makes big tech feel slower. In general, a lot of these big tech companies are really good at integrating AI into their existing businesses, but they are spending way too much and still having a hard time pushing out new technology. And this is not just a China thing; the same voices pointed to the U.S. peers too, look at Google and Microsoft. The bureaucracy and the management are sometimes the ones holding new innovations back.

I’ve been getting mixed responses on internal token usage, but in general it seems like Alibaba has not officially capped it for employees.

Developers can use whatever they want, while purely writing or document-focused work gets a 3,000-credit limit, though people can apply for more. They deliberately switched from tokens to credits because explicitly showing how many tokens each employee uses is too transparent, and they don’t want to disclose that. With the cost per token coming down so fast, the point is really about managing efficiency rather than rationing consumption. Even inside the company that owns the models, nobody wants the unit economics to be THAT transparent.

Beyond the technology side of things, he sees the potential issues stemming from the fact that the last generation of Chinese tech people are not international enough, in both language and mindset, whereas the AI-native companies and the next generation of AI companies are often very international and focus on going global first. Domestically, they obviously have the mindshare and dominance, so whatever they push out for B2B is a lot easier to accept because of the credibility and trust already built into the brand, but that credibility doesn’t travel.

As for what comes next, he thinks the next wave of products might be general-purpose usage agents that find vertical use cases (!!), maybe, say, building agents for content creators, with specified, tailored plugins for each use case. And within Alibaba, its Cursor-like product Qoder is already showing signs of this pivot, integrating more industry know-how and running smaller models to help with the token usage constraints and worries, much like what we wrote about before. It echoes the epiphany from the Accio demo.

Wrapping up my rant

While it’s much easier to get carried away by narratives about the consumer story, the real question is how they can monetize it.

The moat is the rails and the data, not just the model. Qwen commerce works because Alibaba controls Alipay, Taobao, and Amap, and Accio is defensible because of the supplier graph, while Qwen the model lags GLM, and it barely matters domestically.

Similarly, if you think about it, whether it’s the smart cockpit business Banma or Dingtalk, B2B monetizes before B2C, because businesses buy productivity and consumers buy value for money, and until token costs fall far enough, or someone solves the liability question, the consumer agent is a moat-defending feature rather than a revenue line.

And the reason why I’m so gung-ho about Accio is that the merchant use case is obviously chargeable and defended by a real edge.

And despite some concerns around the company’s heavy spending, one small detail that gives me some confidence in it is that the mgmt is now taking a page from ByteDance and running A/B teams apparently on various products - including the foundational model layer- to see who can outcompete and win out. Maybe that kind of “wolfness” or hunger is back at Alibaba?

Last but not least, I’m super happy to say we’ve been approached by some super interesting mid-size cap companies that span AI software to spatial intelligence and physical AI, and we should be releasing some interviews with them in the coming weeks. Stay tuned :) I’m really excited to share them.

Haha, ok now that I’ve made that disclaimer…

Joining me today is the ever-so-knowledgeable Tu Le. He is the founder and managing director of Sino Auto Insights, author of the SAI Weekly Substack, and co-host of the China EVs and More & At The Wheel podcasts.

He has worked across Detroit, Silicon Valley, and China, so he views the industry from the inside, through the traditional auto industry, the tech industry, and the Chinese market.

I wanted to do this episode almost as an educational primer, not just for you all but for myself as well. Most people now understand that Chinese EVs are competitive. But very few people understand why and how that is translating into the Physical AI space.

We talked through the Chinese EV landscape, why traditional OEMs struggled to make good EVs, how autonomous driving fits in, how these carmakers are integrating AI, and why home appliance and smartphone companies like Huawei, Xiaomi, and Dreame are suddenly making cars.

Follow Sino Auto Insights here: https://x.com/SinoAutoInsight

For consulting inquiries, go DM Tu Le on LinkedIn!

Website: https://www.sinoautoinsights.com/

Btw, I’m rebranding Differentiated Understanding to AI Proem Podcast.

Subscribe now

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, and product managers. For more information on the podcast series, see here.

Chapters

00:00 Introduction to Tu Le and Sino Auto Insights

04:28 Mapping the Chinese EV Industry

09:19 The Rise of Xiaomi in the EV Market

14:17 Understanding BYD’s Market Position

17:54 Challenges for Traditional OEMs in EV Production

31:29 The Role of Government Subsidies and Policies

36:12 AI Integration in EVs and the Future of Mobility

45:13 The Evolution of Brand Experience in EVs

46:53 The Future of Manufacturing and Market Dynamics

50:46 Safety Concerns in Rapid Development

52:51 Current Landscape of Autonomous Driving in China

57:51 Challenges in Deploying Autonomous Vehicles

01:05:00 The Future of Mobility and Urban Planning

AI-generated transcript (for reference only)

Grace Shao (00:00)

Hi Tu. Thank you so much for joining us today. I’m really excited to have you on.

Tu Le - Sino Auto Insights (00:04)

Thanks for having me on, Grace.

Grace Shao (00:06)

Yeah, to start, why don’t you tell us a bit about yourself? We were just having this conversation right before recording. I find your background really fascinating.You know, you can talk to a very diverse group of kind of people. You run a successful consulting gig, a consulting company. Tell us about everything that you do.

Tu Le - Sino Auto Insights (00:24)

So name is Tu Le I’m the managing director at Sino Auto Insights. I also create content. I run or I co-host two podcasts, China EVs and more, and at the wheel with my co-hosts that are very, very good at what they do as well. And then I write a weekly newsletter, almost weekly anyways, called Sino Auto Insights Weekly that just kind of goes over my thoughts on what’s happening in the industry now globally every week. And I’m actually not Chinese. I’m Vietnamese. And I was born in Vietnam and moved to the United States when I was a year old and grew up right outside of Detroit. My whole family, youngest of eight, whole family’s automotive. So grew up car kid and did that for a few years before going back to grad school and moving to Silicon Valley to work for seven years. So that’s where the knowledge of the tech comes in, especially the hard tech, where hardware software integration is such an important part of creating a great user experience. And then I met a girl and in San Francisco. my girlfriend, who’s now my wife, was transferred by her company over to Beijing, where she was born. And I decided to pull the ripcord and and follow her over. And what we thought was going to be a three- or four-year assignment ended up being thirteen. And during this time I worked in automotive; I worked at a few Chinese EV e-commerce startups. And so that’s when I learned and experienced nine nine six myself for about two years. And yeah, it’s not fun, super intense, but again I wouldn’t trade those experiences for the world because it gives me the perspective that I have now. about eight years ago I saw this huge disconnect because EVs were becoming a thing because of Tesla. Companies like NIO and XPeng had just been founded.

And you know, in Beijing, as you know, Grace, there’s a lot of the German OEMs, and so there was a bit of arrogance about how hard or how simple they thought software was and really, really being consumer focused as opposed to product focused. So I saw this opportunity, and I started this consultancy, Sino Auto Insights, and you know we’ve been growing since we’ve done traditional work. We’ve worked with the UK government, US government on things. And then also when I moved back four years ago from Beijing, I left during COVID. So August of 2022 and then November, December timeframe, China opens its border and says, What COVID? Come on in. So we didn’t know that was going to be the case. so we decided to move back. And we opened an office here in just outside of Detroit. And we’ve been helping more on the investment side, looking for investment opportunities, what’s around the corner, but also giving our clients a better understanding of the Chinese EV players and the battery players and what they’re doing outside of China. So it’s a very, very interesting time. The mobility space, as you know, Grace, involves now AI, silicon, data centers, data privacy, data security, batteries. So it’s just, just a tremendously unique sector that I get to be a part of.

Grace Shao (03:55)

Thank you so much for sharing your life story. First of all, kudos to your mother. Eight kids. Like, I don’t know how she did that. Like I have two and I’m already dying. And also, I love your personal touch, you know, why you moved to Beijing and just learning about your background. I think it’s super fascinating. You pointed one thing out. Like when I was living in Beijing in Shanghai as well, I met a lot of German OEM like employees and people kind of low key don’t know this, that there’s a huge German community i it in the huge like and and French as well. A lot of Europeans are actually working in China for these, especially like luxury vehicle companies. and a lot of them did relocate out of China during COVID times. And a lot of them I’ve even heard anecdotally from two friends who say they’re dying to get back because they were born in Munich or Frankfurt. just because they’re so bored.

Tu Le - Sino Auto Insights (04:23)

Huge. We hear those stories a lot, don’t we, Grace? We hear those stories a lot.

Grace Shao (04:48)

it’s just because it’s just the fast-paced energy in China. However, okay, COVID was crazy. China’s fast paced. Let’s get to that actual topic today. I wanna talk about EVs. I wanna learn everything from you. so before we get started, when we think of Chinese EVs, most people outside of China think of BYD. think of maybe like the few other ones you mentioned, like NIO X Peng. Now Xiaomi Dreame, which is crazy; essentially, these home appliance companies are going into the space as well. they are there are state-owned companies, there are old independent automakers, there are startups that we just talked about. And then some of them also produce batteries; some of them are, like I said, home appliance and phone companies. Basically, all of these different moving parts, they’re all coming into the same arena. Help us map out the industry first. Like to start with, who are the main players? What are the buckets? what does each group bring to the table or what’s their differentiating kind of offering? I know this is a very big question, but start with a big picture.

Tu Le - Sino Auto Insights (05:45)

So well, let me press rewind and kind of frame it and create more context as opposed to just we’ll we’ll zoom out and then we’ll zoom into the China market. So last year, twenty twenty-five, Toyota was the number one global automaker, eleven million units, around eleven, just over eleven million units. Volkswagen was number two at eight million. To give you a sense of scale, Tesla was one point six. million units and BYD was about 4.6, which makes them a top 10 automaker. The other top 10 automaker for the Chinese was Geely. Geely and everybody else outside of BYD and Tesla build ICEs and EVs. Or in China, they call them NEVs, new energy vehicles, which means that they’re battery electric vehicles plus plug-in hybrids, E Revs, and then fuel cells. So fuel cells, for our intents and purposes, are rounding error. So when we talk NEVs, we’re talking battery electric and plug-in hybrids and extended range electric vehicles. So Toyota’s been number one for a long, long time. And you know, the China market has been the number one passenger vehicle market since 2009, overtaking the United States. And now the China market is almost twice as big as the US market. If we add the European market, which is around 12 and a half, 13 million units, and the US market, which is around 15 and a half, 16 million units, it’s almost the same as China. And so the scale of the China market is enormous. And so to talk about EV specifically, I would create different sets of buckets. And I would look at BYD, Chery, Geely, Great Wall as separate companies, SAIC because they produce in the millions of units. Okay. And then this lower tier, I won’t say lower, but this other tier of EV makers, the NIO, the XPengs, the Li Autos, the Zekers, these are companies specifically that Western investors pay attention to because they’re traded publicly in the US. They’re in the hundreds of thousands of units. Go ahead. Yep.

Grace Shao (07:58)

I want to comment on this. So you’re actually separating them by the number of units versus their technology, because it seems like, just not like SAIC, they’re actually a traditional OEM company, but they’re also producing EVs, whatnot. So you’re actually categorized by number of sales versus, I guess, I don’t know, EV native like NIO and XPeng. Just help us understand what why that industry does that.

Tu Le - Sino Auto Insights (08:18)

so the automotive industry is very capital intensive. And so scale creates cost efficiencies. Okay, so if I buy 10 of something versus one of something, I’m gonna get a better price, generally speaking. And that’s why it’s so important that BYD has this enormous scale of 4.6 million units. And that’s through the traditional lens. Now It’s multi-layered as you’d mentioned. You know, they’re they’re EV only companies that we can talk about. But from the standpoint of scale and global reach, that’s where I’m really creating these separations because scale also creates flexibility because it’s gonna be harder for a company that only sells 300,000 units of anything to go global as opposed to someone that sells four point six million units of something.

Because these companies, the BYDs, the Geelys, the Leap Motors, they all already ship and build and manufacture outside of China. Big, big steps. And so those are really kind of the uniqueness to some of those top-level guys that have the sales volume. They have the ability to go abroad. Because think of it just from a number standpoint, Grace, if we have capacity of half a million units. You and I run a car company. Building a hundred and fifty thousand unit factory is a huge consideration for us because we have to find demand somewhere for that hundred and fifty thousand units. Whereas if you have millions of units, 150,000 units isn’t that huge in a grand s in the grand scheme of things of sales and distribution. And so that’s where it’s a little bit easier for these larger companies to really command, you know, the pricing scale that they can negotiate over some of those smaller players.

But you know, back to kind of how I would look at this. You know, the NIOs, the XPeng, the Leottos, they’re publicly traded in the US. That’s why there’s a lot of attention paid to them. But they’re still puppies. And you know, all of them shipped less than half a million units last year. Now Xiaomi is one of the newer players, but it is making a huge, huge impact because their automotive division, and you and I, I think, were there when Xiaomi first was founded in 2010 in Beijing. But their automotive division is less than six years old. And this year they’ll ship over half a million units. Contrast that with the NIO and XPeng, who’ve been around since 2014, 2015. They are barely shipping 500,000 units. So that tells you kind of the impact Xiaomi has had. Xiaomi only has two products. NIO and XPeng have four, five, six products. And so, now with Huawei, I would pull them into a very unique budget because they don’t build any cars. What they do is partner with other OEMs, whether they’re state-owned or non-state-owned, and offer the hardware and software stack. Okay. They call it HEMA, which is the Harmony in Mobility Alliance. And so there are companies like when you hear about Ito, Micstro, Stelato, Luxe; I won’t talk about who their partners are, but these are all using Huawei technology. And more and more, there are foreign companies that are using Huawei technology. So, Audi, I think Mercedes is using some of the Huawei technology in the China market. So we get into a lot of crossover when it comes to Huawei, because in order for Huawei to really, really legitimize their tech stack, they probably need more foreign automakers to sign up to it. And they’re really, really, really aggressive on trying to scale that part of the business because, you know, once they lost some of that handset business from North America. They try they’re trying to quickly find revenues to replace that. And I think focusing on the automotive sector was one of the ways they were thinking of doing that.

Grace Shao (12:33)

That’s super, super helpful just to get an understanding of the different buckets. so why is Xiaomi doing so well then actually? Because, you know, I saw a Xiaomi car, I think, maybe last year when I visited Beijing. It’s very sleek. It feels very nice, but just in comparison to Li Auto, XPeng, and NIO, which, as you said, they’ve been around for like more than a decade. They’re very, very they’re actually beautiful cars. I was quite like shocked when I first saw some of them during COVID time. Yeah, why is it that they’re just outbeating them? Is it because of the ecosystem? Is it because of what people are talking about with the control, the operating system? Because Xiaomi’s operating system is just significantly better because their software is better? Or what is it that’s driving consumers by them over others?

Tu Le - Sino Auto Insights (13:20)

So there’s a few things going on with Xiaomi. So the Su 7, which is their Sedan, if you squint, it kind of looks like a Porsche Cayenne. And then the U7, which is the yeah, so that helps, I think. And then the U7, which if you squint, looks a little bit like a Ferrari Purosangue, which is their Ferrari UV or FUV. So they

Tu Le - Sino Auto Insights (13:46)

Have borrowed design language from these two amazing automotive brands, and that’s helped them. But they have offered these vehicles at less than $40,000, starting at $40,000, with features that are very, very technology-forward. And one of the big reasons they’re successful is I bet. Grace, if I looked around your apartment, I would probably see a Xiaomi product, one or two at least. And in China, I didn’t know anyone who did not have some sort of Xiaomi air purifier, rice cooker, TV, computer, mobile phone. So the brand is ubiquitous in China. And that really helped Xiaomi when the vehicles launched, create this automatic instant demand because the brand is there: complete awareness of the brand, and there’s a lot of trust amongst Chinese consumers. And one of the important things about the China market versus the rest of the world is that anyone born after 1990 in China is a digital native. So they grew up with WeChat, Didi, Meituan, and Alibaba, you know, and Xiaomi. So, and these are all Chinese brands. So, they trust Chinese brands, not like their parents who only bought foreign brands.

And I think that’s the larger shift in the Chinese Chinese market across sectors, consumers goods, you know, technology, high, you know, consumers products, now automotive. But also Xiaomi is very well connected across different, you know, product segments. They have an app that controls everything; it extends into the vehicle, and you would call that a consumer-focused product company. I think that resonates with a lot of Chinese consumers. So the impact that they’ve made is enormous. I haven’t even talked about how cool the cars are or what they can do. They’re breaking records at one of the most historic racetracks in Germany, the Nürburg Ring, and they’re beating the pants off of Porsche. So if I am Porsche and I know that Xiaomi is entering the European markets, Germany in 2027, I’m pretty worried. Because Porsche’s not doing well in the China market. And I think that’s the canary in the coal mine for a lot of companies in Europe, that Xiaomi’s gonna be a major player as long as they can continue to build these cool cars.

Grace Shao (16:17)

Super interesting. I would wanna look I wanna talk about Chinese vehicles, Chinese EV companies going global and going to Europe later. But to start, I wanna talk about the hype that BYD gets as well. Some of them cost less than even a hundred thousand RMB. You know, what is it about BYD, and how to understand them from a business perspective as well? Like they create their own batteries. They are getting into semis. They have their, you know, their whole whole integrated industrial platform. Like help us understand BYD.

Tu Le - Sino Auto Insights (16:51)

BYD’s story is amazing. In 2009, my first day in Beijing, there were BYDs. I got into a I I want to say a BYD cab, and it wasn’t great. It was not good. You could hear the exterior, the outside world, pretty, pretty clearly. And I I wasn’t feeling that safe in the vehicle, to be honest with you. But fast-forward to 2026, and Let’s just look at the last seven or eight years. Before COVID, BYD was shipping less than a million units. They’re at four point six in twenty twenty five. We’ll likely get to five million. They’re gonna be exporting about a million of those, a little over a million of those. So they’re currently in over 100 markets. So they are uber aggressive. And to your point, they’re very vertically integrated. BYD started out as a technology company that supplied batteries and other components to companies like Apple, and I want to say like Intel, but they got their experience and their scar tissue from working with some of the toughest technology companies. And that’s really kind of created s this resilience and ability to grow and scale and stay aggressive. Now they created a monster because companies like Geely and companies like Leap Motor are right on their heels, and they’re actually not doing that well. Their growth is flattening out, and the competition in China is super, super intense, hence the importance of them to export. Wang Chuang Fu is the founder CEO, and Stella Lee is the head of international. They’ve made it clear that they’re targeting Toyota to become the largest automaker in the world. And three years ago, they were going to be number one with a bullet, but now we’re seeing, as they scale and that denominator gets much bigger, double-digit growth is much harder to come by. Competition is catching up. They created the competition. They really, really were catalysts for all these other companies to build these sub $100,000 or sub-100 RB cars that are. Pretty amazing. Now, you make a great point because in China, BYD is the mass-market value car. Okay. But it creates an entry point for Chinese consumers that wouldn’t otherwise be able to purchase a vehicle, especially in big cities like Beijing and and Shanghai, where it’s, you know, getting a license plate is a challenge. And then finding parking and being able to pay for that is also very challenging, especially if you live right in the city center. And BYD has really, I think, in 10 years, 15 years, we’ll point to BYD as a democratizer of many, many, many things, not just mass market clean energy vehicles. And their focus on going international really, really put them behind a little bit on the technology curve. And Companies like XPeng are leaning into the technology. And Wang Chuang Fu has acknowledged that they’re a little bit behind on some of the features that other Chinese automakers are providing in the China market. But he’s determined, with his over a hundred thousand engineers, to really push that envelope to catch up and surpass some of their domestic competitors. Examples, recent examples of that. They’re launching a megawatt charger, or they’ve launched a megawatt charger in the China market. And that’s basically charging as fast as gas. You can charge fully within six, seven, eight minutes. And then also recently, just last week, I want to say, their intelligent driving system is called God’s Eye. And they are now providing a year’s worth of insurance. If there are any accidents while you’re using God’s Eye in China. So they’re putting their money where their mouth is. You don’t see Tesla doing that. And so they are willing to take that next step that everyone else begrudgingly has to follow them on. And, you know, one of the important things in, you know, I’d mentioned ten years ago that in 10 years we’ll say they’ve democratized things. The other thing is they’re offering God’s Eye as standard on many of their vehicles. And so think of it from the standpoint of not only in China, but to your point, now in Hong Kong, in Thailand, in Latin America, these people that have a 10,000, 12,000, 15,000 US dollar car, they might be able to drive themselves or at least have portions of the road where the vehicle has intelligent driving capabilities. And if BYD doesn’t do that. Intelligent driving doesn’t happen in those emerging markets for 10, 15 years at least. So they’re really, really pulling, I think begrudgingly, a lot of their competitors forward on that technology curve. And I applaud them for that. Now, at 4.6 million to get to 11 million, I think Wang Chuang Fu and Stella Lee appreciate more the level of management capability and the operational efficiency needed, like by a Toyota, to get to eleven million units. And so, but they’ve doubled down, and it sounds like they’re still determined to be a top two, top three player in the next five to seven years.

Grace Shao (22:32)

That’s really interesting that you pointed something out, which I didn’t notice at all. It’s the fact that they’re democratizing the technology, so it’ll be very interesting that they will actually introduce the kind of next-generation technology to these markets. So I guess bringing down their price right now is a long term strategy because once you capture that. you know, mind share, then they could always increase their prices later on.

Tu Le - Sino Auto Insights (22:53)

One of the most important things, Grace, is that there are two emotional buys in a person’s life generally, and that’s a house and a car. And in China, I think there’s around 60 hours, I wanna say, of research being done online, at the retailer, test driving, before you actually pull the trigger and buy something. And so

Grace Shao (23:02) Yeah.

Tu Le - Sino Auto Insights (23:17)

If you and I- I don’t know if you have an iPhone or an Android phone, but if you lost your iPhone tomorrow, you’d be upset, but you’d walk into an Apple store and buy another one right away. But, you know, a car, people take consideration because it’s a reflection of who you are, who you want to be, you know, and really outside of your home and office, you spend most of your time in the car, especially if you live in Asia. And so That’s why it’s so important for them to be in these markets as a first mover. They create that awareness first, they build that trust first, and it helps them elbow out other players that might have equally impressive products, but because they came two or three years later, BYD already is in the mindset of a lot of these international consumers, especially in the emerging markets where a lot of times they’ll enter and six or seven months later, they’re the number one brand. In this segment, in that segment, or these segments that they enter.

Grace Shao (24:19) Definitely.

So let’s bring it back to traditional OEMs. You worked with some of them or you worked at some of them. Why did traditional automakers struggle to build compelling EVs now? Especially, you know, in China. The factories were there, the people were there. As you said in the beginning, you said some of them became a little bit complacent about the idea that, you know, Chinese consumers maybe just really liked luxury cars from Europe; you know, the default was buying Japanese cars for families. Why is it that almost every traditional OEM has produced a kind of crappy EV version?

Tu Le - Sino Auto Insights (24:55)

A lot of it has to do with, so let me first qualify this by saying if you ever talk to someone that tells you they predicted that the market was gonna move quickly over to clean energy vehicles in China, do not believe them because no one could have predicted how fast the market moved over in twenty twenty. We were in 2019, we’re at like one point two million units of NEVs relative to a twenty-two million unit base. Okay. And then in twenty twenty we were like one point three. Then we got to three point five, we got to six point five, we got to nine million, and last year we were close to eleven and a half, twelve, thirteen million units. And so we are currently over one of the inflection points, over 50 percent. So every one of every two cars is an NEV sold in China. Okay. And no one could have predicted that. So everyone was caught flat-footed if you’re a foreign automaker. Now let’s add in the fact that they’re analog companies, right? They’re not software companies. They’re not technology companies. And I bet today if you and I were to go into a boardroom or any meeting room in Detroit or Dearborn or Auburn Hills or Stuttgart, They’re still talking about the product. Okay. If we look at NIO, XPeng, Li Auto, their founders come from tech. Okay. They iterate. You know, they don’t, they, they ship product that’s good enough, and then they figure out what the bugs are, and then they create over there updates to fix those bugs. And whereas traditional automakers, they try to wring out as much profit as they can over a five-year period. And that’s because traditional product development cycles are a five-year period or a four-year period. The best company on the legacy side is Toyota, and they’re at around 30 months. Most good Chinese EV companies, like BYD or Zeekr, can go from clean sheet to job one. And job one means the first sellable vehicle off the production line. They can do that in about 15 months, which is absolutely insane. Now they don’t have the blinders on that say we can only do it a certain way. They’ve challenged everything all the way through. Simple things like where you might be you know a product engineer or component engineer and you you throw it over to me as a manufacturing engineer. Everything’s in serial. That’s why it takes so long sometimes. The Chinese EV makers, they do a lot of things in parallel; they simulate. A lot of safety tests, you know, validation to engineering validation tests, you know, manufacturing validation tests. They simulate a lot of that stuff, and it shrinks the timelines. And they treat manufacturing like a technology as opposed to, you know, an analog product. And if I’m being frank and honest, and I know that you want me to be that, the Volkswagens, the GMs, the Mercedes, the BM, they’re busy back in their home markets counting their money for a long time. You know, these narratives that they they steal IP. You know, is there IP theft? I think you and I would agree yes there is IP theft in China. But let’s qualify that: if there was IP theft in this instance, the Chinese automakers would make great ICE engines, right? Like gas engines, you would think. Exactly. So in this particular case, it it wasn’t because they stole this IP and it wasn’t because of subsidies. Because there have been subsidies since 2009, and it has been a substantial dollar figure, right? Tens of billions of dollars, if not hundreds. But if we look at Tesla, Tesla has not sold a vehicle.

Tu Le - Sino Auto Insights (28:59)

Globally without some form of subsidy. Okay. Full stop. They got their factory in Fremont for next to nothing. Shanghai Giga, tons of incentives by the local government. Li Chiang put that deal together. Guess what? He’s now the vice premier of China. So that was probably one of the reasons he got that promotion. So the idea that Chinese subsidies are bad, US subsidies are good, European subsidies are good. That needs to just stop. And at the end of the day, if you have those subsidies, it doesn’t automatically mean that you’re going to win. Because you probably remember this, Grace. X Peng Li Auto NIO in 2016, 2017, 2018, they were struggling. They’re almost bankrupt. And it wasn’t until December of 2019 that the first Model 3 rolled off the line in Shanghai Giga. That you really saw that hockey stick inflection point. Okay. So despite all the subsidies, despite all these promising EV startups, it took Tesla to really bring excitement to that market. And so you can credit Tesla for being the catalyst for EVs globally, not only in the United States, but China. I think we should acknowledge that they’re a huge part of why EVs became a thing in China. Now, because of the competition, all these competitors came in because of the subsidies, because of the the the RB that was being given out by local governments, you had a ton of players come in. And last year it got so competitive that the Chinese government was like, no mass, no mass, no involution. And so we have to acknowledge that the China market acr across a number of sectors, not just automotive, is likely the most brutal automotive market in the world. And if you’re able to survive out of this mess, you’re gonna be a very, very formidable competitor outside of the China market.

Grace Shao (31:15)

Definitely there’s the Musk effect, I think, on you know, Tesla bringing out EVs, and now we’re seeing that with humanoids again. Cause with Elon Musk obsessing with humanoids, we’re seeing the world obsessing with humanoids. We’ll talk a bit about that later. I appreciate you giving kind of the backdrop of the history of China’s EV space. Tell us about the industrial policy push though, because you mentioned subsidies as a very vague term. But what kind of subsidies or industrial push do you think China actually gave this industry to bolster it?

Tu Le - Sino Auto Insights (31:48)

The Chinese government has the ability to long-term plan. That doesn’t happen in the United States. Unfortunately, and it’s really costing us in a lot of areas and sectors. And we’re seeing these sectors that would normally be pretty strong be a pretty competitive struggle against the Chinese. But l let’s say in 2009, the Chinese government looked at manufacturing as a pillar industry. That supported jobs. And within the manufacturing sector, they wanted to become number one in batteries, number one in silicon. And silicon fabrication. Not only silicon design, but silicon fabrication, automotive EV manufacturing, and then, you know, and that all supports this notion that they’re the world’s factory. Okay. And they doubled down on that. Now The types of subsidies, whether it’s tax abatements for land, whether it’s discounts on building factories, whether it’s purchase subsidies or you know, tax abatements for the consumer who don’t have to pay taxes on buying EVs. To your point, in Hong Kong, same thing, right? Similar things going on. I I wanna I wanna stress that anywhere in the world, emerging technologies, in order for them to become really ubiquitous and blossom, governments need to put their thumbs on the scale. Okay. So it’s not just a Chinese thing. The United States was ready to give consumers $7,500 for every car, every EV that they bought. Okay. So it’s not just a China thing. Norway, which has a ton of oil money, used that oil money to subsidize EVs. And now the take rate in Norway is over 90%. Now their market is tiny, 400, 500,000, 600,000 units a year, but nonetheless, it’s another example of the government putting the thumb on the scale. In Germany and other parts of Europe, there are also subsidies for EV purchases. And that’s also when you saw growth rates a little bit higher than they are now because they’ve taken away a lot of those subsidies. So again, emerging technologies need help. And normally the governments in any one of these countries need to step in. The exception is kind of the UK, which has done pretty well on EV adoption, the growth of EV adoption, despite not having a ton of subsidies on the consumer side. But that being said, it these subsidies, this focus, this diligence, this perseverance, this investment over time. It didn’t look like, let’s say, 2010, 2012, 2014; it didn’t look like it was going to really, really work out that well. And then all of a sudden COVID happens, and you get this huge spike in demand for electric vehicles or clean energy vehicles or new energy vehicles. Now, my quick story. You know, it’s it’s a

Grace Shao (34:51)

Why is that?

It’s like we’re locked in our homes. So why do we need EVs?

Tu Le - Sino Auto Insights (34:58)

It’s a weird phenomenon and and I haven’t read anything or talked to anyone that can I you know, I th we can try to theorize. For me, it’s like, yes, we’re locked in. I wouldn’t say we’re locked in our homes, but we’re locked in the country, and we don’t travel, so maybe, you know, we spend money on something that we think might make us happy or something like that, right? Like, I can’t tell you why outside of, okay, Tesla starts building in 2020. And and and a quick story about my COVID experience. My family went back to Michigan. So I was sending my kids to a local school. So they had basically the month of January off in 2020. We were going to go back to the United States for three weeks to visit family. And the return flight was canceled. After two weeks in the US because China was starting to have COVID. And we bought two two one-way plane tickets that were canceled. And then a third one where we finally got to fly into Narita, stay there a night, fly to Shanghai, and then to Beijing, and then quarantine for two weeks. And we didn’t leave China for two and a half years at that point. The day we got back, the day after we got back, the Chinese government closed the border. And during that time, Grace, you saw more and more and more of these green plates. And I I was just amazed because think of all of the things that need to happen for demand and supply to work together. Because it’s not just, okay, we can produce these things, no problem. There needs to be charging infrastructure; batteries need to be scaled up to support, you know, the EVs. And then creating this awareness, creating this excitement. The NIOs and XPengs and Li Autos weren’t able to do it on their own. The BYDs weren’t able to do it on their own. But on the backs of the the the Chinese consumers really, really trusting the Tesla brand. And we’ve seen that the Tesla brand in the China market is extremely resilient because they’re still selling quite a few, despite not upgrading their vehicle or updating their vehicle in a number of years. And so it is a strange phenomenon that I can’t definitively tell you why. I just know I saw more and more green plates when I was in Beijing and Shanghai over twenty-twenty through twenty-twenty-four or twenty-twenty-two, so.

Grace Shao (37:41)

Yeah, it’s an interesting phenomenon. And like every single tech founder or AI founder I’ve ever met basically says they drive a Tesla. So contrary to what you were saying earlier, you’re like, a lot of Chinese younger generation actually prefer the Chinese consumer brands. There’s something about Elon Musk and his brand in China. People like love it, worship it, wanna become him, whatever. So all of all the founders and tech people still drive Tesla. I I wanna bring it to the next point, which is on AI and tech, actually. So we talked about EVs, and I know I can ask you like 3,000 more questions on this, but I wanna understand the AI element to all of this now, because essentially EVs are not like old cars; they are tech-first cars, right? Like you said. But because they're tech-first cars, they seem to have integrated. AI or hardware plus software much more seamlessly. We’re seeing like Li Xiang push out like even wearables like glasses that can control their cars. we all know all the mentioned brands just now have voice control. They all have some sort of AI already embedded in them. Now it’s just like giving them a more formalized name. Autonomous driving obviously is a form of AI. We can talk about that as well, but help us understand: for people, for these cars, are these kind of like software-hardware integrated really that seamlessly, or are they struggling as well? And then on top of that, you’ve said something I think in one of the podcasts before is that you push back on the phrase software-defined vehicle.

Tu Le - Sino Auto Insights (39:22)

So the terms mobile phone on wheels and software-defined vehicles, to me, and I would love your opinion on this, Grace. That tells me that these people don’t know technology or understand technology or have worked in the technology space because software doesn’t define anything. Software, AI, you know, silicon, these are all tools that create a compelling user experience. Now, you put them together, you design them well, you combine them with the right hardware and the software that instructs the hardware what to do, when to do it, how fast to do it, that creates the user experience. And I’m an Apple alum, so I learned that early on. It was really, really drilled in my head that the user experience, the stickiness, that creates the brand, that creates the brand loyalty. Okay. And what the Chinese automakers are doing right now is like throwing spaghetti on the wall to see what sticks. And because of the enormous pressure from competition, they just try to be first. Okay. What they likely need to do, and I don’t know if they’re going to be able to do this in the next 18, 24 months, just because I don’t see competition really, really slowing down in the China market. Is to take a step back, you know, I had a conversation with Sam Livingstone and Matt Mechelvoigue talking about the Ferrari Luce and some of the missteps. And part of that is, what does this mean to the NIO brand? What does this mean to the XPeng brand? Because that is what creates the awareness, the trust, and the loyalty; like everything just kind of makes sense. Simple design is really, really, really hard. Johnny I’ve has said that many, many times. In I’m sure. I would again love your opinion because I’m a Westerner, so I’m not used to Chinese apps that have a million things popping up at me or Chinese websites that have all these windows and all these lights blinking and stuff like that. It’s it’s a little intimidating to me. And I feel like front consoles of Chinese EVs are still a little bit overwhelming. Now, if we look at Xiaomi, I think they do it pretty well. They could They can improve, but they have years and years of experience among their teams to build out consumer experiences. Okay. And to me, like the reason I don’t like mobile phone on wheels is because a mobile phone can’t run you over and kill you. Okay. So anything automotive grade is a completely next level thing. So a big thing on the battery side is energy storage systems. Okay. A battery for an energy storage system is not automotive grade. It needs to be bulletproof if it’s automotive grade. So the level of engineering and manufacturing quality needs to be much, much higher. That’s why when we oversimplify it like that, I don’t think people appreciate the amount of effort and level of detail that needs to be had for putting something on automotive grade. And then software-defined again, it just tells me that these guys. think software is the end-all, be-all. Software, if software, AI, and hardware is running great, you don’t notice it. You just experience, you just have a great experience. If I have to say this hardware is not working or that hardware is not hard, then the automotive designers and engineers need to go back to the drawing board because something’s wrong. and and that’s what I want to emphasize being an Apple alum, that that hardware and software integration is is is what is going to create and differentiate you in the market long term. And one thing I will point out about Apple is that because they have a closed system and they don’t have to Frankenstein a bunch of disparate, you know, firmware together because they control the design of the hardware and the software and integrating, well. They might integrate third-party AI now because they’re super behind. But that’s kind of the only third-party thing that they’re doing. but eventually I’m sure they’re gonna look to to create a native AI support system for their ecosystem. But I I think that’s the huge differentiator. Now, is it realistic for an automaker to have a closed system and control so much? If you’re Tesla, maybe, but you started. on day one as a closed system. Okay. It’s gonna be extremely difficult, extremely, extremely difficult for most other automakers to do this. Now, you know, one of the areas I think you wanted to talk about was partnerships. And this is where the Chinese automakers are getting a lot of credit through the announcements of all these partnerships with their Western counterparts. And the important thing is that there’ve always been partnerships in the China market. You know, it’s been a requirement of the Chinese government historically, you know, with the exception of the the the Tesla factory in Shanghai recently. But now these partnerships are bleeding into Europe, they’re bleeding into North America, and we’ll continue to see that as long as the Chinese are pushing the envelope on innovation. But again, Grace, can the legacy automakers use somebody else’s tools? Again, they’re tools. Can they use somebody else’s tools? To create a Volkswagen experience, a Volkswagen brand experience that you know historically is this way or that way. Okay, like you trust Volvo, right? You don’t like their new cars, but can if Volvo is using Geely software, Geely AI, Geely hardware, or you know, like Geely qualified hardware, is it still gonna feel like a Volvo to you? And I think those are kind of the important things moving forward. that’s gonna differentiate some of the legacy automakers and some of the better Chinese EV makers from the rest of the field.

Grace Shao (45:28)

I feel like listening to you explain this actually makes me feel like there’s gonna be two fragments of the market where the hardware people, like the people who still want the best hardware experience, will still go with a traditional OEM because you will still get the best craftsmanship, get the best kind of hardware experience, right? But if you are gonna go for an AI-native experience as we go forward with this, you know, you want the best voice control, you want the best. Whatever interaction with your AI agent within your car, then it would make, as you said, it would be extremely hard for a Volvo to use someone else’s software. So wouldn’t a company with their own software actually have the advantage of building that? So like a Xiaomi or Tesla, right? Like your point, you have your closed ecosystem, you have your existing s software, you have everything you need to make the experience better. But that said, the car might not be as sleek as an Audi, Mercedes, whatnot, right? I don’t. What do you think?

Tu Le - Sino Auto Insights (46:32)

Well, I think that when you get to clean energy vehicles, so especially battery electric vehicles, manufacturing is is simplified by orders of magnitude. So the GMs, the Volkswagens, the Porsche’s, they all Mercedes, they all have entire powertrain divisions that only work on the engine. Imagine those entire departments effectively going away. Okay. Electric motors I’m oversimplifying this, but they’re fairly commoditized. Okay. They’re super fast, super efficient, generally speaking. Automotive grade is something different than everything else again. But that is really going to be the differentiator moving forward. Grace, if I told you that in 15 to 17 years, maybe less than that, building a car is going to be commoditized. So there’s no value in that aspect or part of it. Now, with the world being as bifurcated as it is, especially in North America, that doesn’t want Chinese battery cells in North American vehicles for now. Maybe it’s a longer timeline, but effectively China is really, really creating or forcing other companies to rethink how they manufacture things, how they develop things in the spaghetti on the wall. There will be some spaghetti that sticks for the Chinese automakers. And you better believe that that copycatting is gonna be reversed now. The Europeans and the North American car companies are really, really going to create their own versions of X, Y, and Z. Now, can we say that it was innovated and perfected in the China market? Probably moving forward in the next three, five, seven years, but we also need to look at the demo. Right? Because you had mentioned some people want performance, some people want the digital experience. And I think a lot of that is gonna be is gonna correlate to what the demographic is, because a BMW or Mercedes owner in China is around twenty, twenty-five years old, younger than in Europe and North America. As a you know, as an old man, I have different needs than you do, as a as a young woman. So What I like in a car is going to be different than what you like, than what your husband likes, than what my wife likes. And I think that’s where the Chinese are gonna have to play the global game. Okay. Now they need a solid foundation of extremely high sales in their domestic market to create the flexibility to sell abroad and to sell at a premium abroad. But are all these digital features And technological advancements going to resonate with a 60 year old year old European man in Germany who is used to driving a BMW? Probably not. Okay. But one of the other big advantages is that the Chinese have is, and I’ll give you a quick example. Friend Nick Carey, who writes for Reuters, last year he wrote an article about Chery taking six weeks to change the suspension and steering system in an Omata 5 because they were shipping the China-spec version of the Omata 5 to Europe and the Europeans were like, yeah, this steering is way too mushy and the feel isn’t there. The Europeans will not like this. And so over six weeks they qualified new parts, they updated the software through OTAs and firmware and then shipped the new product And that would take a year in in at least a year at most legacy automakers because of the layers of bureaucracy, the approvals needed. But this was done in six weeks. Now that’s also a reflection of the nine-six in China.

Grace Shao (50:24)

Does that not frighten you a little though? Because like how fast, like you mentioned, how fast they ship, w I wanna ask something slightly sensitive. Then what about the safety of these cars, right?

Tu Le - Sino Auto Insights (50:36)

Well, if you asked the Chinese automakers, they will assure you that they’re not cutting corners. Okay. And I have driven many of these Chinese cars. Now I d I haven’t owned one for 10 years. So long-term efficiency and safety, I don’t know. Most people don’t, because a lot of these cars have been on the road for less than five years. But you know, when you talk to the automakers, and again, a lot of these people come from the automotive space. A lot of the leadership of some of these Chinese companies, they come from the Mercedes, the Volkswagen groups. And so they do have some visibility into how things are being done differently. But I would also counter what you just said with yes, there’s a little bit of risk with cutting so quickly the product development and so severely the product development cycle, but Also, how things traditionally work at large conglomerates and in even governments is there’s something new that’s happening. We’ll add a layer. There’s something new happening; we’ll add a layer. Technology changes, we’ll add a layer. So no one ever takes a step back and says, These 15 layers, does this still make sense? Because I mean, that’s kind of the definition of bureaucracy, right? So time will tell. I do I feel not safe in these cars? No. you know, and I’ve driven dozens of them, so

Grace Shao (52:01) No, I was playing devil’s advocate. Like

I’ve been in so many of these in China, and you know, especially across Asia. But I just think the it’s just people tend to ask questions like, it’s so short then. If you’re shipping them out within a year, what kind of corners are you cutting? And then thus the question is easy to say. The next question is, is it safe? Right. But I think I I I kind of feel you on the point. If it’s completely new, we treat it like a startup; it’s innovative. There’s a lack of bureaucracy, there’s a lack of this is how we do things. Then you actually can just get things done much faster. I’m mindful of time. I want to ask you some questions beyond the traditional car makers and whatnot. Help us understand where China is with autonomous driving right now. Who are the main players and just roughly understand, you know, who are the ones kind of competing with Waymo, who’s Pony AI, right? Like who’s We Ride? I know again, it’s a super big question, but let me just throw this to you like open-ended.

Tu Le - Sino Auto Insights (52:57)

Let me kind of close out that last topic that we were talking about with food for thought in something that I know you know as well, but maybe your audience isn’t quite aware of. If any Chinese company is found to be cutting corners in the China market, the Chinese government would not look kindly on that. And there would be severe consequences, right? So I think there’s this healthy fear of If we are cutting corners and we’re found out, we’re gonna be in a lot of trouble. There’s not gonna be years of litigation, there’s gonna be severe penalties right away. And I think that healthy fear motivates many of these Chinese automaker leaders to stay on the right path. Right now, again, time will tell, but to pivot towards your question about autonomous vehicles, so Waymo is the global standard. I think most people would acknowledge that. They’re in many, many markets. They’re entering foreign markets. But there to your point, there is WeRide, there’s Pony, and then there’s Baidu. These are the three largest players in China currently. But then in Apollo Go, yep. And and and unfortunately, when I was in Beijing last month.

Grace Shao (54:13) I do is call Apollo, right? Or yeah.

Tu Le - Sino Auto Insights (54:22)

Or two months ago, Apollo Go was not running because in Wuhan about a thousand of them or a hundred of them were on the roads and they just turned into bricks, right? On the roads. And so they had stopped the pilot programs. I don’t know if they’re running again, but I normally when I’m back in China will try out all these systems for the latest software to make sure to just kind of see, feel, understand, what’s going on and and what’s unique about China is that the feel is a little bit different in each of these cities just because how people drive is so different in in different cities.

Grace Shao (55:02)

This is something I feel like no one understand if you don’t live if you haven’t lived in China. Like people in Beijing are just aggressively wild. Like people don’t realize this.

Tu Le - Sino Auto Insights (55:11)

What? Aggressive? man. So I’ve driven in like Changsha, I’ve driven in these tier two cities, and you’re like,

Grace Shao (55:17)

Okay, I haven’t driven into your two cities. I just haha I usually just get a car there. Like, I I I already think Beijing is so terrifying. Like, I I start driving 16 years old in, and I refuse to drive in Beijing, and because you stop the car for someone to pass. Next thing you know, like 10 cars have passed, like 20 bikes passed you, 30 pedestrians, and you’re still there, and then there’s like 20 people behind you honking you, and you’re just like, yeah.

Tu Le - Sino Auto Insights (55:33)

my goodness. Yeah, so a fun, quick funny story. My wife used to be very worried because I would get super fired up when I’m driving in China. For some reason, I learned to compartmentalize it because I would get super upset. And then when I got out of the car, I would just not be upset. I don’t know how I did it, but because my wife didn’t want me to take yeah, I it was just.

Grace Shao (56:04)

Like you have to. Because you’re like constantly road raging. Anyway.

Tu Le - Sino Auto Insights (56:10)

And so what you’re talking about is called cutting in. And if you have a meter of space between you and the car ahead of you, someone will cut in. Someone will cut in for sure. And if you’re not used to that, someone will cut in, and then there will be a San Luncha delivery vehicle turning the opposite way. And your head needs to be on a swivel. And it is quite an experience. It’s similar, and I won’t say similar, but it has a similar feel as Southeast Asia because it’s so crazy. And it just kind of works in Southeast Asia. And it doesn’t work super well in China because there’s traffic jams all over the place. But Pony and WeRide are trying to help some of that stuff. Let me segue to that.

Grace Shao (56:57)

Yeah, so how does it work though? Like that’s my point. Like how do these autonomous driving cars work if, you know, people are so unpredictable? The whole idea is they’re supposed to predict what the car is gonna do, but you know, they just like zigzag and people just pop out of nowhere. Like i is it safe? Like what’s really a holding up, like what’s a bottleneck of deploying these at scale right now? Is it regulation, is technology, or is it just the craziness of the roads?

Tu Le - Sino Auto Insights (57:28)

Again, let’s do a 30,000-foot level. We ride, pony. They’re both publicly traded in the West. And so I think there are Western investors that know who they are. They’re not as large as Baidu Apollo Go, which has, I want to say, well over a few thousand cars on the road in China pilot programs in a dozen cities. But Pony and WeRide are also moving aggressively outside of China, partnering with Uber, partnering with other companies, ride-hailing companies, and there’s pressure because they’re publicly traded to really scale and create some profitability. Whereas Waymo is still owned by Alphabet. So I think they have pr internal pressure, but not pressure from external markets. in the difference between China and the US, because at the end of the day, these are really the only two players that have multiple horses in this race. Now in the UK, there’s a company called Wave. And I think there would be other European players that would argue that, hey, we’re also a major player, but let’s, for the intents and purposes, oversimplify this by saying there’s the US and the Chinese players. In China, there is this first wave of A V companies: the Werides, the Pony AIs, the Baidus, and then there’s this other wave, and we’re only talking about robotaxis. Because on the commercial trucking side, on the slow-moving delivery vehicles, we also have autonomous vehicle players. But for robotaxis, there’s this second wave. And they’re more of an asset-light company, autonomous vehicle startup. So like DeepRoute, Momenta, QCraft, you know, they’re what they’re doing is partnering with traditional OEMs in China to get their stacks onto these vehicles. DeepRoute, for instance, is working directly with Great Wall to integrate their hardware and software stack into the design of the vehicle. So even before, because what we normally see right now, Grace, is a car with lidar, sonar, radar bolted on as, you know, an afterthought. But once these companies are working with the traditional OEMs, they can design them, and it looks like part of the form of the vehicle as opposed to like this bolt-on after the fact. And the convergence between RoboTaxi Company and traditional OEM is blurring. And I’d mentioned earlier that Huawei also has a significant stack. So they’re a player as well. But Huawei, as far as I know, is not getting into the RoboTaxi space. But they’re going to move into level three intelligent driving, and they’re hoping to be in millions of vehicles in China within the next few years. And that’s where it’s very different in China because there’s not a lot of convergence going on in the US market. Now we know about Neuro, we know about Zoox, we know about Waymo, and with the exception of Neuro, who’s working with Lucid to put their stack on the Gravity for a premium experience. Most of these companies are not working with OEMs. And then the OEMs have their own systems as well. And that’s the big difference between the China market and the US market. And what we’ll likely see is a bifurcation of the US market being primarily North American autonomous vehicle providers working with the Ubers and the Lyfts to create that larger install base to try to reach a broader audience. And That’s one of the big reasons why these companies are working with the ride-hailing companies, because it’s gonna be hard for Pony to attract 100 million users. Whereas if I partner with Uber, my install base is 160 million global users. And so that creates an opportunity. And I think long term, Uber sees Robotaxis and Evital as their profit drivers. And you know, the delivery services and all these ancillary mobility services as a way to increase their install base but not make a ton of money. And and so

Grace Shao (1:01:45)

Wouldn’t that cannibalize our own business, existing business a little bit?

Tu Le - Sino Auto Insights (1:01:51)

Yeah, and you know that’s that’s the that’s that’s the million-dollar question because Uber is now also buying its own autonomous vehicles. So not only is it a ride hailing platform, but it’s a fleet manager now. And that changes the economics of their yeah, exactly. So that really changes the economics on their balance sheet. Okay. So I I don’t

Grace Shao (1:02:06)

hedging.

Tu Le - Sino Auto Insights (1:02:17)

To your point, I do think they are kind of hedging their bets a little bit. Because if to answer your question, we should separate autonomous vehicles into those that use lidar and those do not use lidar. Tesla does not use lidar. Wave, which is the UK company based out of London, does not use LIDAR now. We can have, and I’m sure you’re well- I think you’re talking to some expert in a couple of weeks, so maybe you can ask them to use lidar or not to use lidar. they are using lidar, so I’m sure he’ll he’ll say that lidar is necessary, but it’s more philosophical now, right? It’s more philosophical because Tesla can’t all of a sudden put lidar because it changes their whole system. Okay. But to me, LIDAR.

Tu Le - Sino Auto Insights (1:03:06)

Prices have gone down so significantly that creating another redundancy and, you know, kind of creating that sensor fusion with multiple sensors and lidar creates a safer environment, I would think. but again, it it it it’s an interesting thing that I don’t think a lot of people definitively can answer. I’m sorry?

Grace Shao (1:03:29)

It’s a philosophical choice. Or is it actually a design choice because the vehicle already can- like, you cannot put a LIDAR on top of it because Teslas cannot use LiDAR. Like, is there a reason?

Tu Le - Sino Auto Insights (1:03:42)

So, to me, it was an engineering choice at first, but now it’s more philosophical because now you’d need to change your system if you all of a sudden incorporate lidar into it. Right. And for Elon, I think it’s also a mienza thing because he’s been so strong against LIDAR that if he turns it around, now don’t get me wrong, he says things sometimes that

Grace Shao (1:04:03) Yeah.

Tu Le - Sino Auto Insights (1:04:09)

Never come true or haven’t come true yet. So that’s kind of the crazy thing. But I think LiDAR, that’s one of those things where I think he’s willing to die on that that doesn’t need lidar. But anyways, I don’t want to get into this this this discussion about lidar, but but but but to the bifurcation thing. The reason I say bifurcation is because

Grace Shao (1:04:24)
Okay. It’s gets getting too technical. All right.

Tu Le - Sino Auto Insights (1:04:34)

We know that Europe has a strong data security, data sovereignty policy, security, privacy, and sovereignty. So will Europe allow the Chinese autonomous vehicle makers to ship European data back to China to the servers to train the models? And/or so the United States, because we have a ton of allies, China has a ton of allies, that’s why. It’s likely that the Chinese allies would use the Chinese systems first, and then the US allies would incorporate the Waymos and the Zoox. That’s kind of, and I’m oversimplifying this, but because the Trump administration has kind of poked the eye of a lot of our traditional allies. So maybe they wouldn’t want our autonomous vehicles on their roads. But I’ve ridden in all of these systems.

Grace Shao (1:05:08) I see.

Yeah, like Canada might be saying no these days. Sorry, I’m just going. It’s like really late for me and I’m just thinking about yes, but like now Canada’s gonna have eight Chinese EVs. Now American cars aren’t gonna sell there now; American Waymo’s not gonna be in Canada.

Tu Le - Sino Auto Insights (1:05:28)

Yeah, yeah. No, you’re well, we could do an entire episode about all of that stuff just between North America, right? So I think that would be an interesting conversation. But you know, that’s exactly my point, right? Like, traditionally, prior to Trump, I think Waymo was going to rely on our allies to launch its services. And if we look at the Middle East and India, they

Tu Le - Sino Auto Insights (1:06:06)

kind of want to be Switzerland a little bit. So because they they they they don’t want to take one side over another. India’s the same thing. And these are potential markets where both of them will compete. Famously, in London by the end of this year, early next year, Waymo and Do will all be testing and rolling out pilots. So next time you’re in London, if it’s early next year, you might be able to try all three systems in the same place on the same streets, which I think is gonna be a very, very, very unique experience. because I don’t think there’s gonna be many cities that you’re gonna be able to do that in over the next three, four, five years. And, you know, at the end of the day, is it a data thing? Is it

Grace Shao (1:06:47) That’s pretty crazy.

Tu Le - Sino Auto Insights (1:07:02)

Who has the most data? Who has the most robust edge case data? Is that ultimately who wins? Or does AI really change the game? Because if it’s about data, if it’s about kind of real-world miles, then you would think the Toyotas and the Volkswagens would have a distinct advantage because guess what? They put 11 million cars a year on the road. Okay. So a company like a deep route, a company like a wave would love to work with these companies that high have high sales volume. But what is the great equalizer? If you talk to Elon, it’s his system because again, everything is closed, everything is native to the Tesla system. FSD is the best intelligent driving system. Is it better than Waymo? And will there ultimately be a convergence between level three and then level four? You know, can a Waymo system compete directly versus a Tesla because and and I don’t have answers to that. And that’s what makes the industry so interesting because the politics of things change, the technology changes, and then the commercialization opportunities change as well. And what I do know is that Waymo is backed by one of the most valuable companies in the world, which means that they have

Grace Shao (1:08:09) Mm-hmm.

Tu Le - Sino Auto Insights (1:08:28)

A huge check to help them get to scaling. And one th other differentiator with Waymo and the rest of the players is that they’re really moving into a lot of four-season cities, like Detroit. So next year, Waymo is going to be launching a service. And what we’ve seen so far is that a lot of these autonomous vehicle companies are launching in Arizona in the Middle East, where guess what?

Grace Shao (1:08:29) Resources.

Tu Le - Sino Auto Insights (1:08:55)

Weather is super predictable, and it’s pretty one-note. And where the edge cases are, you can look at it like an inverse normal distribution curve, and I’m oversimplifying this, Grace, but the edge cases- so they happen few and far between- but they’re likely where the most severe accidents happen. Okay. So so it’s like there’s the least amount of data available. Exactly.

Tu Le - Sino Auto Insights (1:09:21)

Right. Severe storms, blizzards, snow, whiteouts. And so what we’ll likely see the final frontier being Robotaxis, because we’ll see commercial trucking from companies like Kodak. Remember that company like Too Simple? Those competitors. We’ll see those Aurora. We’ll see commercial trucking happen sooner. And likely highway to highway. I I divide commercial trucking into like three segments, and that’s

Tu Le - Sino Auto Insights (1:09:49)

You know, intra-city, city to highway, and then highway to highway as three separate use cases. Yeah. And you know, robotaxis, I think it’ll be phased; it’ll be geofenced for a long period of time. There’ll be certain use cases that it makes a ton of sense for robotaxis, you know, but ultimately, in order for this service to become ubiquitous, there probably needs to be

Grace Shao (1:09:54) It’s just much more predictable. Yeah.

Tu Le - Sino Auto Insights (1:10:16)

More services that have multiple people than one person in a car. And so that’s where it’s interesting because Waymo just launched the OHI, which is the Zeker contract manufactured vehicle, which has multiple seats. And I think that’s how Waymo looks towards profitability to get more than one person in an autonomous vehicle and almost looking at it like a bus, you know, a smaller bus to get

Grace Shao (1:10:41)

I was just gonna ask actually, like, are we gonna see a redesign of what robotaxis should look like? Because you, you, I tried out Waymo in San Fran, and it’s kind of creepy because you still have the driver’s seat, but like no one’s sitting there, so you’re constantly freaked out, like, what, you know, if you’re not used to it. So, like, will we see more like these little boxes or something that will signal to other drivers as well? More obviously, this is a robotaxi versus like a normal car.

Tu Le - Sino Auto Insights (1:11:10)

For sure, for sure. Right now there are policies in place that say you have to have a steering wheel, you have to have brakes. But as the autonomous vehicle landscape evolves, we’ll probably start to see, and I have a theory, Grace, that more and more cities will limit private passenger vehicles coming into the city center. Okay. If we look at Paris, they’re investing 300 million euros to make all the boulevards that lead into the Champs-Élysées bike-friendly, and they’re gonna limit private passenger vehicles. And so especially in Asia, I could see that being, you know, maybe you park, or you take the train into the fourth ring road, and then you take an autonomous vehicle into the city center. And then to get to your office, you take a scooter. Right. So there are these scenarios where I think more and more cities will try to take back some of the streets, some of the roads that bleed into the city center in order to lighten up traffic and take back some of the land. Because if we think about, and we’re getting getting off topic here, but I think these are important kind of secondary and tertiary effects of autonomous vehicles. Look at these parking structures and like these parking lots. We use them from like 7 a.m. to 5 p.m. And then they’re not used for the entire rest of the day. It’s kind of a waste, to be honest with you. And then and I’ll right. And so if we can take that back, you know, ideally make more housing affordable, make more office buildings, or whatever, right? Like round out, make more green space as opposed to having so much so many parking structures. Hong Kong could use more green space because all they can do is build up. And so there’s a lot of opportunity.

Grace Shao (1:12:36)

Hong Kong has, like, only I think like less than ten percent of the population even have a car because the public transit is so good. To your point, like there’s minibus, there’s a double-decker bus, there’s like the T MTR. Everything is walkable. you it’s and it’s c really, really, really well planned. And I think a lot of Asian megacities are like that. Actually, like think of Singapore, think about like Shenzhen, obviously obvious Shenzhen where you have like

Tu Le - Sino Auto Insights (1:13:05) All right. Yes.

Grace Shao (1:13:26)

EV buses to EV cars to EV scooters, all for rent, all for access for the like average person on the street, right? So yeah, that does make sense.

Tu Le - Sino Auto Insights (1:13:34) I’m gonna drill down on that.

I’m gonna drill down on that because when I was living in Beijing, I was a mobility practitioner. I walked, I rode share bikes, I rode subways, I rode high-speed rail.

Tu Le - Sino Auto Insights (1:13:53) Yeah.

Yeah, well, I mean, rings around a road is a little weird, but now that I live in the United States, I’m just an advocate ’cause all I do is get in my car and drive everywhere I go. And on the weekends yeah, yeah, well

Grace Shao (1:14:07) But it’s also because they’re in Detroit. It it makes

a difference, right? If you’re in suburbia versus like the middle of like ring three Beijing.

Tu Le - Sino Auto Insights (1:14:14)

Yeah, and I think that’s a big difference between like North America and Asia. And I would lump North America and Europe a little bit into that because a lot of these cities aren’t very big. So to invest in subways and things like that would probably be dis a disproportionate expense for the city’s budget. Where when you’re in China, there are dozens of cities with over a million people, you know, hundreds of cities over a million people.

Grace Shao (1:14:41)

Over like twenty million people, like a couple cities are over, yeah.

Tu Le - Sino Auto Insights (1:14:43)

That’s why I think, and what’s important about China or distinct about China as well, is that the automotive sector didn’t build out this transportation system because there’s a balance of high-speed rail, to your point. There’s a balance of subways, intracity transportation. But I’m off topic.

Tu Le - Sino Auto Insights (1:15:07) The autonomous vehicle space is

gonna be very interesting because our is the US government going to restrict silicon? you know, does because right now NVIDIA basically supplies every automaker with a a high-end transport or intelligent driving feature. Okay. But you and I know that the Chinese government is really pushing for companies like Horizon and and and

Tu Le - Sino Auto Insights (1:15:35) Black Sesame and XPeng, NIO, their

Huawei, they’re all silicon design companies now, too. And eventually NVIDIA is gonna get pushed out. And that’s also another bifurcation point as well. So if the Chinese can’t catch up to NVIDIA and Qualcomm and some of these other silicon design, Western, more established Western. Silicon design companies, does that mean their AI is not as good? Does that mean it’s not as robust? I think these are really open ended questions that you probably have conversations with your other guests on. So, and and I listen to you because that’s important to me of about understanding other perspectives on that stuff, because although I understand the chip sector, not to the level where I’m not an AI expert. And so I, like I said, I use it. as a tool as opposed to the end-all be-all. But but yeah, so

Grace Shao (1:16:32)

No, appreciate your insights. Really, really appreciate your time. I’ve definitely taken up more than, you know, I asked for. So, to end, I wanna ask a question. what is one differentiative view or you think a misunderstanding the the world might have of on the topic of China EVs, mobility?

Tu Le - Sino Auto Insights (1:16:51)

I think in twenty, twenty-five years, we’re gonna look back at this time as a renaissance in mobility because of everything that’s happening so quickly and being driven by the competitiveness of the China market. And we’re gonna see BYD is definitely gonna be a player. And you know, the other thing that I think is really, really important is that. I don’t believe traditional automotive folks can think outside of their normal way of seeing how the world works through transportation. And the top 10 mobility providers, to me, in 15 years, there might be a handful of traditional automakers, but I see an Uber maybe being a top 10 player. I see a Baidu or a Waymo being a top 10 player, and they don’t. build cars, but we’re the the importance of building vehicles is not going to be it is going to be reduced over time very quickly because of China. And so if you’re not providing a value added service in the mobility space, which I think which I think China is going to be able to do at at a much more affordable price point, especially in the emerging markets. I think that’s where the important thing is because in the Western markets, the BYDs and the Geely’s still have challenges and customer acquisition costs are much higher in in those emerging or those established markets. But in the emerging markets, the Chinese are going to try to roll out not only passenger vehicle buy-sell and their brand, but they’ll probably try to sell a lot of services once they have that sale. And I think that’s really going to be that opportunity for the Chinese to really make a name for themselves. Because you know this, Grace. One of the coolest things about the United States for me as an American is that anywhere I go, and you can not like the food, you can not like the coffee, but it’s consistent. If I go to a Starbucks in Munich or Vancouver or Toronto, and that’s soft power. That’s American soft power, right? The Chinese would love to have four Chinese brands. Creating soft power for them, creating aspirational desires to have their products. And and so that is gonna be the priority for a lot of these entrepreneurs that you and I speak with because, you know, they’re as ambitious as Elon, you know, maybe maybe they don’t get covered as much by Western media because their English might not be fluent or whatever, but they shouldn’t be underestimated just because they’re in China and they’re not in the rest of the world yet. And and I think that we’re gonna look back at this time and point to a few Chinese people at the level of being close to Elon. So

Grace Shao (1:19:47)

Very interesting. Yeah, I think I think to your point, a lot of the entrepreneurs I speak to these days, they set their eyes on the global market in the first day and they want to set the industry standard. Like that is their goal. And it’s no longer about shipping out something cheaper, shipping out something just to make that quick buck anymore. So there’s definitely like a a sentiment shift and that confidence is different as well. to it Yeah. Well yeah. Thank you so much for your time today.

Tu Le - Sino Auto Insights (1:20:09)

It’s off the charts. Confidence is off the charts.

Grace Shao (1:20:15)

Really appreciate your time. I feel like I need to invite you back for another conversation because, you know, for what I prepared, we can go on for another two hours, I feel like. But it is late tonight, for me. So I’m gonna call it a day. Thank you so much.

Tu Le - Sino Auto Insights (1:20:30) Thanks for having me, Grace.

What a whirlwind of a week. I recorded an episode with Bloomberg Odd Lots on China AI on Monday, flew to Singapore on Tuesday for three days of back-to-back moderating and meetings across the ecosystem during Singapore AI Week, then came back to Hong Kong on Friday to be with bean 1 and bean 2. I’m only now recovering enough to think it all through.

Before anything else, I want to thank the SuperAI team, Lightspeed, Liminal, Vertex, HubSpot, and the various local friends who extended invitations and hospitality. As usual, I’ve tried to comb through everything rattling around my head all at once in one long-form piece.

Share

Token-Minimizing: Notes From a Whirlwind Week at SuperAI

So, you know what’s actually happening?

People are reversing out of token-maxxing and into token-minimizing — or at least rationing. Qianer at The Information made the point well in her newsletter: “While Silicon Valley is moving away from tokenmaxxing—the deliberate act of consuming tokens to demonstrate AI nativeness—in Southeast Asia, home to nearly a dozen developing countries, developers and users never went through that stage and have always been cost-conscious.”

Her two takeaways from SuperAI rhymed with mine: one, it’s getting harder for startups to wow users because AI has already shown it can do so much; and two, even with token costs falling, it’s still not cheap enough to scale.

Let’s sit with that for a second. From Stripe’s customers to the largest enterprises — think Uber, Microsoft, Tencent, and ByteDance (all have publicly declared some reprioritizing token spend) — companies are, one by one, starting to ration the tokens they burn. Token-minimizing is really just this: once tokens are a managed cost rather than a free good, traffic flows to whatever model is cheapest for the job. For some that means a hybrid setup, but for another group, a growing number, it means routing to China’s efficient open-weight models.

So then you ask: 1) where does value accrue once the models themselves are commoditizing? Benedict Evans pressed exactly this from the main stage, and nearly every panel circled back to it. 2) the one MIT’s Max Tegmark reminded us all of: who are we even building AI for? How do we build it safely, and what does “pro-human AI” actually mean?

Let me unpack this step by step, starting with the SuperAI stage.

The visual AI stack

I moderated the visual AI stack panel, which was a treat because it spanned the whole stack. Alibaba Cloud held down the infrastructure layer, leaning into its sprawling ecosystem and courting developers with open-source models and tools — the Qwen models, the Wan video model, and its video-generation tool Happy Horse. Meitu sat much further up, in applications; it runs a proprietary R&D team focused on facial recognition, but it largely builds on top of other people’s models.

What we kept coming back to is that the ecosystem is vibrant, but the real use case for visual models still hasn’t scaled. Today it clusters in gaming, marketing, and prosumer work and e-commerce vendors making product ads, indie filmmakers, that crowd. Which creates a funny tension: the market feels small even though it’s big, if that makes sense. And for startups like Video Rebirth, their vision is not just to build frontier video models but push that technical capability into world models, leveraging their know-how in 3D and spatial intelligence.

And the main bottleneck faced by everyone?

It’s the thing nobody fully controls: data. Data is both the edge and the ceiling here. You also can’t think about training image and visual models the way you’d think about training an LLM; a visual is so subjective, so hard to pin down and keep consistent, that the whole exercise is different.

On the business model, my instinct has always been: why not build for enterprise? That’s where the money is, no?

But raw APIs carry almost no margin. Wrap that same capability into a consumer product, and margins can reach ~73%. Usage is lower, and it’s harder to scale, sure, but the value per generation — per token — is far higher, as one investor put it to me.

And for visuals specifically, so much of the moat is taste. How you tune color and effect isn’t a black-and-white question of fact; it’s a matter of taste. For companies like Meitu, though niche, they are made for the aesthetics; that is where they lead.

That’s also why I thought maybe they should build a Stripe for enterprise visual tools, like a clean plug-in layer. But when I sat down with Meitu’s IR, they pushed back: it doesn’t work like that, because enterprises want their visual tools deeply personalized. It’s nothing like the streamlined simplicity of payments.

Which is the perfect segue to my next thought…

Stripe: when the cost of building collapses

Abhi Tiwari, Stripe’s head of global products, made the week’s core theme concrete: the cost to build and the cost to go global are both falling away. His headline number — a $200/month coding subscription can now replace $2 million of up-front spend. And Stripe sees it in its own data: agents are increasingly the ones reading its docs, and by the end of this year more agents than humans will be doing so.

The result is a flood of builders. iOS app releases went from declining month over month to up roughly 24% since agentic coding showed up, and more businesses are adopting it, too. They’re also monetizing faster since Stripe baked payments into tools like Lovable, Replit, and Bolt; recent Stripe Atlas cohorts have gone from incorporation to first payment in about six weeks.

His point is that the cost of intelligence is no longer the bottleneck. What separates winners now is knowing what to build and when, managing tech debt, and taste plus distribution and reach.

But here’s the tension Stripe sets up. If intelligence is no longer the bottleneck, then taste, timing, and distribution become the game — but so does the bill. Cheap per token still adds up fast at scale. And that’s where the heart of my week really begins.

Token-minimizing: rationing the thing that used to be scarce

So let’s actually sit with token-minimizing for a beat.

If intelligence is getting cheap, the flip side is that companies get deliberate about how much they consume. There’s even a tidy Chinese corporate phrase for it: 降本增效, “cut costs, raise efficiency,” which is essentially what every company wants to do to maximize profit… and right now it’s being aimed squarely at the AI bill.

And it’s not just the American names. I’m told Tencent has been nudging employees toward something like RMB 1,000 of token spend a month (a significant drop from like ~5000 USD last month), and ByteDance is finding its own ways to cap internal use. (I got into how the BAT are reorganizing and spending around tokens in my 2025 wrap.) It cuts both ways — some firms cap usage to control cost, others push staff to burn more tokens to prove they’re being productive, but what is clear is that token spend is now a managed line item, and no longer an afterthought or a free good.

Which brings up something I’ve written about before: cheaper tokens don’t shrink demand; they multiply use cases, the Jevons paradox, alive and well in AI infrastructure. So rationing and an explosion of usage are happening at the same time. That sharpens the real question: which model, or which layer, captures all that new demand? Although, honestly, I’m starting to ask a harder one: does demand actually multiply if no one’s proven the ROI, while the costs just keep climbing?

For now, the answer is that traffic flows to whatever’s most efficient per token, and guess what? That is increasingly looking like Chinese open-source models, a posture I’ve argued is an ecosystem strategy, not just a release choice. The open-source strategy was always a business decision first, a philosophical choice second. To bring people into its ecosystem, building trust was the first step; monetization can always happen, maybe in less per sale, but that doesn’t stop a sustainable business.

So anyway, this is a good moment to talk about the labs themselves.

Where China’s models sit now

First, the fun part: it was awesome to catch up with so many of these labs offline, and a real pleasure to introduce a bunch of them around SuperAI.

So where do China’s model labs actually sit today? I had a decent vantage point this week, given that I was able to catch up with reps from StepFun, Tencent, Alibaba, Z.ai, and many more on the panels and on the sidelines.

Ultimately, it isn’t about “winning a race.” It’s that, for once, the structure of the moment runs in their favor. Token-minimizing is a tailwind that plays straight to their strength — efficiency per token — and the models keep leaping. GLM-5.2 is genuinely good (Bernstein puts it close to Opus 4.8), and Kimi’s K2.7 Code is pushing hard on reasoning efficiency. For the first time, you could argue Chinese open-weights models hold the rhetorical high ground, too.

It’s not a clean run, though. China AI still faces two hard constraints: a compute shortage for inference, and an API price war that grinds margins down even as hardware costs — memory especially — keep climbing. New anti-distillation efforts and export controls pile on more catch-up hurdles, and no one’s hiding that. The timing is delicate, too: this is the same moment Anthropic is working through its Fable and Mythos debacle and has doubled its API pricing on Fable 5 (vs. Opus 4.8) — which only widens the gap the Chinese labs are there to exploit.

One tell from the sidelines: the labs wouldn’t talk business with me at all. StepFun and Z.ai are both midway through major capital-markets activity (StepFun filed for an IPO on the HKEX, Z.ai for a dual listing on the Shanghai STAR board), so they kept the conversation focused on tech. It’s now or never, with everyone rushing to gold, and my take is that the HKEX helped set a fair valuation, and that turning back to A-shares helps bring in more liquidity and cash riding that AI fever.

So the picture is two-sided. On the one hand, token-minimizing keeps pushing traffic toward the most efficient models, including Chinese open-source models. On the other, those models keep leaping — GLM-5.2, Kimi’s K2.7 — and narrowing the gap with the frontier.

Which raises its own questions: will it continue? And when is a model’s capability simply “enough”?

All of which makes model routing a much bigger conversation. Can a cheaper Chinese model be the default, with a hybrid setup that escalates to a frontier model only when the task genuinely requires it? More and more, that’s the architecture people are reaching for.

Bernstein put the momentum well in a recent note (Global Internet: Never interrupt your opponent while he is making a mistake):

“AI never sleeps. It probably wasn’t an accident that Kimi and Z.ai, two AI labs we consider frontier in China, announced new models over the weekend. K2.7 Code promised better reasoning efficiency and reasoning capability upgrades. Z.ai’s GLM-5.2 meanwhile positions itself against Claude Opus 4.8; both online developer feedback and our own locally-hosted benchmark tests returned encouraging initial results. Real counterarguments remain, but on the margin these releases support the idea that China’s leading labs can continue to keep pace with global peers. The idea that Chinese open-weights models might now occupy the rhetorical high ground too is a novel development.”

Where does the value go?

That keep-pace question is really a value question, and in some ways it is the one Benedict Evans built his keynote around. I got to meet him briefly, tell him about AI Proem (fan-girl), and his talk is the one I keep chewing on (here’s the full keynote; it expands on his 70-page deck, AI Eats the World, and he goes deeper in this a16z conversation).

He won’t dismiss AI as a fad, but he won’t buy the story of an immediate, sweeping economic transformation either. He opened with: “AI is whatever machines can’t do yet — the moment they can do it well, we stop calling it AI and start calling it software.”

Databases once had superhuman memory; image recognition once felt like magic. Each was “AI” right up until it dissolved into being just software. LLMs may well go the same way. So the real debate isn’t whether the technology is real, but where value gets captured as the models commoditize. Is generative AI the next platform shift after the PC, the web and the smartphone, or does it become another layer folded into everything until it’s invisible?

He even pushed back on the Sam Altman “AI as a metered utility, like electricity” framing, pointing to telecom: mobile data volumes exploded for a decade, and the value never followed. Volume and valuation don’t travel together in a commodity. Evans left the ending genuinely open, which felt more honest than most takes I heard all week.

So where is AI actually diffusing? It’s the thing we’ve been writing about for a while: didn’t internet innovation take decades to play out? When Cisco was founded, did anyone imagine that Uber would become an internet company? Will the models capture all the value this time — or does it pool somewhere else entirely?

Think about it,

• Businesses without network effects may lack a scalable model.

• How do commoditizing LLMs become sustainable businesses — or do they stay very expensive infrastructure?

• A capacity gap and a usage gap are not the same thing.

• New workflows take longer to build than anyone expects; for now we’re mostly using new tools for old tasks, not inventing new ways of working.

The frontier models panel turned all of this into theater, with Minimax and Z.ai taking turns telling each other what huge fans they are. Cute. So much for the involution, the cutthroat one-upmanship you’d expect.

Underneath the niceties, though, there was real substance. Mistral pitched itself as the neutral third option and drew the panel’s sharpest line, a potentially specialized tool versus the Swiss Army knife and Minimax with its M3 model pushing toward a one-million-token context window. It argues that distribution, not raw capability, determines how AI evolves, and without GPU distribution, that capacity remains bottlenecked.

Z.ai went back to their favorite car analogies; they’re selling a Mercedes, and Opus is the Rolls-Royce. How many people will need Rolls-Royce or even be able to afford one?

Max Tegmark: the pro-human path

Which is exactly when Max Tegmark widened the lens — from “who captures the value” to “who is any of this even for.” MIT’s Tegmark gave what I thought was the most balanced keynote of the event: neither anti-AI nor accelerationist-at-all-costs, but “pro-human AI,” as he put it. His frame is a fork in the road between a “race to replace” path and a pro-human one. We’ve been on the replacement path a while, he argued, and he even quoted Elon Musk’s own line that, in a likely scenario, none of us will have a job, and AI ends up in charge. But the tide is turning, and he pointed to a recent bipartisan set of 33 principles for pro-human AI built around keeping humans in charge and preserving agency.

We didn’t stop developing fire, he said; we just stayed mindful that it can grill a great barbecue or burn the house down. The point of pro-human AI isn’t to slow the fire. It’s to decide what we cook with it and not burn people with it…

A small proem to what’s next

And yet the most telling signal of my whole week didn’t come from any stage. It came from my WeChat moments, where someone posted that traffic in Yuhang, a quaint suburb of Hangzhou, was completely gridlocked.

Trivial on its face, but maybe the truest tell that China AI is genuinely on. Yuhang, specifically the Hangzhou Future Sci-Tech City and the China (Hangzhou) Artificial Intelligence Town, is the epicenter of the country’s AI buildout and home to global giants and open-source leaders, from Alibaba and Ant to the “Six Little Dragons.” Think Unitree, BrainCo, ManyCore and more. And guess what: I’m headed there in two weeks to visit a few of them.

True to this newsletter’s name — proem means preface — the whole week felt like a preface to the next chapter rather than any conclusion.

Three things I’ll be watching: whether token-rationing hardens into a real procurement discipline (a budget line with an owner, not a vibe); whether model routing will shift to cheaper Chinese options; whether frontier-on-escalation becomes the standard enterprise architecture.

Some shameless self-promo and announcements

It was a pleasure to be back on BBC’s Artificial Human podcast to talk about robotics and China’s positioning in it all. The last time we spoke was a year ago, about DeepSeek and the rise of Chinese open-source AI. And of course, please check out my Odd Lots interview, which should be out in the next week.

In the meantime, summer is kicking off, and the kids are out of school soon. A reminder that we get something like 75% of our time with our babies before age 10, so do slow down, make some lemonade, sit in the sun, kick a ball, and read a book with them. :)

I, for one, will be taking the next two weeks slower. Thus, apologies, I’m posting one more podcast ep this week and then taking a breather for a bit. I appreciate your understanding!

Btw, a few quick news updates

A few things broke while I was sitting on this piece — worth keeping on your radar:

1. DeepSeek’s fundraise closed, and honestly the structure is the interesting part. Read Jing Yang of The Information on how it’s actually put together.

2. Alibaba shipped a world model: the Qwen-Robot series. Which loops right back to the visual-stack conversation up top: everyone’s racing from flat generation toward world models.

3. Lin Junyang, the former head of Qwen, closed the raise for his new AI lab — a world-models shop, reportedly at a ~$2 billion valuation.

In this conversation, we explore how European companies are actually approaching AI implementation. Rather than racing to deploy the latest models, many are focused on organizational design, employee adoption, process changes, and measurable returns on investment. Alex explains why European firms tend to be more cautious than their Chinese counterparts, how concerns around AI sovereignty shape technology decisions, and why companies increasingly find themselves balancing U.S. frontier models, Chinese cost-efficient models, and European alternatives such as Mistral AI.

We also discuss the economics of AI adoption, including the emerging concept of “tokenmaxxing” or rather if that is even the wise path forward, whether AI is truly replacing jobs, how companies should think about ROI when AI introduces variable costs, and why the future may involve token budgets becoming as commonplace as mobile data plans. Finally, we explore Europe’s position in robotics, industrial AI, and regulation, and whether Europe’s strength may ultimately lie not in building the largest and best-performing models, but in defining how AI is deployed responsibly at scale.

Subscribe now

To find the previous episodes of Differentiated Understanding, see here.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

Season two will host a series of guests from early-stage investing, as well as builders, researchers, founders, early adopters, and product managers. For more information on the podcast series, see here.

AI-generated transcript (for reference only)

Grace Shao (00:01)

Hi Sheng Yun. Thank you so much for joining us today. Really excited to have you.

Alex Lu (00:05)

Yeah, thanks very thanks for inviting me. I’m also very excited to have this conversation with you.

Grace Shao (00:11)

Yeah, awesome. So tell us about your journey. I think you’re in a pretty unique position. You know, like I said in the intro, you know, a lot of the conversation about AI right now is often positioned between China versus US, But you actually work predominantly with European companies in adopting AI and their digital transformation. So tell us about your your background and how you got into this.

Alex Lu (00:31)

Yeah. so thanks a lot. So actually, I went to France. I spent more than 10 years in France. I went to France in 2004 and I studied in a school called Ecole Polytechnique. and then when I graduated from the school, I started my work in in Europe, mainly for automotive industry and afterwards for the consulting industry. And still when I was in the consulting industry, I worked mainly for for the auto sector. So

I have a very traditional background of automotive. That’s why some of the work I’m doing currently in the in the AI, we can come back on that, is in the automotive manufacturing sector and mainly for European companies. Because I started my career in Europe, so I know I don’t I know them pretty better, pretty good. And the the the other thing point I want to mention is the school I started actually the Ecole Polytechnique was

Let’s say it it was a famous school in France or in Europe, but it it’s not so famous in in the world. actually this is in France they have a different educational system. but still with the with the rising of of AI in Europe, especially the French large language model called Mistral AI, the school becomes famous because the founder of the of of of Mistral AI comes from the the same school. So basically it’s also a a little bit like

Tsinghua university in China is like the the Tsinghua in in France, having the best talents for for the AI. So nowadays, when I continue my work in the AI transformation for companies or AI implementation for the companies, I work a lot with European companies. Firstly, I know that my I as I said before, and secondly, is when we look into the global competition between China, US, and Europe.

In the AI landscape, it’s pretty clear. It’s like China and and US or US China being the tier one or first ranked models. And Europe is kind of lagged behind. So most of the European European companies, they have this kind of attitude of being a little bit complex, I would say. on one hand, they are kind of seeking for, of course, for the best technology in the in the world to enhance their company’s competitiveness.

There comes the question, how I can define my AI strategy for next year’s between Chinese and US tech stack in AI. And the second question they raised often is while we are European companies, we want to keep keep our AI sovereignty, which is a very important topic in AI. again, we can come back on that. So their question is: okay, between this US and China tech race.

Is there any place for European companies regarding the foundation model companies or application companies or even corporate clients? What could be the playground for European companies? So these are major two questions are often received from European companies and you will you can see the thinking angle is they European companies want to at the same time keep it keep the AI sovereignty and at the same time keeping their competitiveness. That makes the question a little bit complex. Yeah.

Grace Shao (03:49)

Actually why don’t we just double click on the unpack that a little bit? What’s your view on it? Like what what do you advise your clients to do then if if they are kind of cut caught in a pickle or unsure how to build out the next stage of their infrastructure kind of being caught in between China and the US?

Alex Lu (04:08)

Yeah. So the the first thing I I always shared is in in in this tag race actually China and US we are not I want to twist twist a little bit the angle saying this is a competition between China and US. Actually, if we look into details, actually China and US are taking different directions in terms of the AI development, if I can say, because let’s say if we look into the US,

AI ecosystem or the AI development. I think a lot of efforts are put on the foundation model or kind of foundational research regarding how AI can be become AGI can be bring beneficial benefits to the humanity, or how we can guide Rails AI so that okay, one day we will not go into the direction of science fiction movies. So this is a little bit the the push from the US AI companies. While in China, actually

the ecosystem or from the national perspective, China’s AI is more about applications and more about how we can have the s beneficial from the whole society from the AI and how I can combine AI with my traditional technologies or traditional business to to to to to grab more values. So if we think in this angle, actually it will give us two different pictures. One is we cannot say that it’s kind of from

front to front front competition, because these two nations are just take different angles. The second thing is if we look into details based on these assumptions, we will say one nation is pursuing having the most advanced AI technology and one nation is pursuing most kind of most beneficial AI for the society regarding cost effectiveness, et cetera, et cetera. So then it comes to the question that you raised for European companies is

We always brainstorm and conclude on the simple question is what kind of AI are we looking for for European companies? Are we looking for, let’s make it simple, I take some an analogy. Are we looking for kind of you need all the employees to be the PhD employees having the most intelligence in the world? Then that will be the US foundation models. Or if we want to say we have the most cost efficient and best performing employees, virtual employees in your company.

Then we might consider Chinese models, foundation models. Then this is the trade-off. I think the companies should figure out. And the answer will not be so simple like that, saying, tomorrow I will switch to all US tech stack or Chinese tech stack. I would say the two ecosystem, as in the past in the digital area, will still continue for European companies, meaning that they need to juggle with Chinese tech stack in certain markets.

maybe in Chinese market for sure, but for other markets, developing markets where the Chinese foundation model are taking influence and as well as with US models. So this is the thing. And I think the other angle answer to to to their question is I I usually take the statement from Jensen Jensen Huang saying the AI is kind of five layer cake.

So what we are talking about is only one layer, which is the foundation model. And if we go deeper, then we will have infrastructure like data centers, like powers, chips, and electricities. And if we go upper, we will have the applications. So I would tell European companies or I told European companies often is I think the use cases in Europe makes a lot of sense because the cost is there and the employee was pretty

much expensive than Chinese employees. So if we deploy the same model, let’s say, and it of cost the ROI return on investment, you make the business case very easily in Europe than in China because the labor cost is kind of lower. And the the advantage of Europe, one I would say one of the advantages is about power and electricity. I study in France and in France y you would see they have the most advanced nuclear nuclear power technology in the world, at least in the past. And

I think the French government is also think about how we can build more power plants in the in the country to support Mistro’s AI development. And I listened to the founder of Mistral AI, Arthur Mensch. he explained to European Commissions how we can keep the AI development in the Europe is just he make a very simple analogy, meaning that intelligence equals to token. So we all know that, and he said token equals to electricity.

So if we want to make our society more intelligent in Europe, then we need to build infrastructure and more efficiently and more sustainably. And and my last point is I looked into the report released by Stanford, the HAI index. And very interesting because US is far away beh in advance compared to other countries in terms of the number of data centers, it’s around

2000 or more than 2000. I I didn’t remember the exact numbers. And the second and third, it’s not China in terms of the number of data centers. Based on the data, it’s United Kingdom and Germany. So Europe, Europe has the capability to build power plants, but I I I tend to believe these power plants are not currently used to train US models, in my opinion. So again,

This is the European competitiveness if we want to talk about AI. So if we enlarge a little bit picture, we say that okay, it’s five or five layers cake, then again, China and maybe is better performing better in terms of electricity, maybe a little bit less performing regarding the chips. Same situation for Europe. So it’s not only about the most performing models, right?

Grace Shao (10:13)

That’s an a really interesting take. So help me understand like what do you actually advise on these companies for? So you gave me a big high level picture, right? Well give me some examples on the kind of work you’re working on. It’s because I think on this podcast, we often invite people who are builders, founders, investors, and they give us a lot of high level views, which is great, right? But but I want to hear from you, how are you actually helping companies go through this AI transition? And

What are the bottlenecks? Maybe further down we can talk about that. What are the challenges? What are the exciting areas? But just help us understand what are the day to day tasks that you’re working on.

Alex Lu (10:49)

Yeah, thanks. again, I I might share two different perspectives from my experience when it’s again with the European companies. It’s very interesting example because actually I build a product doing this kind of market intelligence, market research for European companies and the value proposition at the time it at that time is we can save time for your employees and they a and and we make your organization more efficient.

And basically we find when we when we s when we sell this kind of value proposition to different companies, I see very interesting different answers. One is on the European side, he would say, this is very interesting, but before implement implementing, we need to think about a kind of tomorrow’s process process, meaning that if we put your AI product into our organization, how our employees will work with the product together, what’s the process look like?

And how many new skills would my employees need to perform or to better use your products? So I think the European company’s mindset is they will need some time to conceptualize the AI products or AI use cases. And then they will need to conceptualize and project, especially once the product is in place, what my company will look like. So they will spend

A little bit more time than Chinese companies to figure out the the regarding the talents, regarding the organization, regarding the process. And in my opinion, it might be a right right approach, in my opinion, meaning that they put human before the techno technology. And this is what I observed when I implement AI for companies saying that, okay, I bring you the best technology, but often we might have improved efficiency by 10 times or five times.

In one single process, but actually there will be some bottlenecks in other organizations, in the in the rest of the organization, then you cannot you cannot increase the efficiency of the whole workflow, let’s say. So that’s why a lot of people in US they talk about AI native organization to kind of remove the bottlenecks in the in the organization. And I worked another example, I worked for a European company, and it’s very interesting. He said, I receive

high level management. He said, I received so many reports from my employees, I I don’t have enough time to review and to approve them. That’s the case because we increase efficiency of the working level of the people, then the bottleneck becomes suddenly the a the leadership. And then we need maybe an empowered leadership by AI in the future to make the whole organization more efficient.

Or we need to think about a new organization where we include AI agents and human beings together because the two natures are producing things on a different scale. So this is mo most of the time, this is the European companies. And for the Chinese companies, the mindset is totally different. if we implement the AI solution, the same product to a to a Chinese client.

The the the answer would be, that’s very interesting. You save 20 or 30% of my employee’s time, but you know I cannot let’s say lay off the employee and to make some savings. So just tell me for the time we saved where he can work to produce more. So it’s always in the mindset, okay, we have some some time saving, but you I cannot pay 80% of the salary to the same guy. So

In in order to so I I need to pay him hundred percent salary, so I y your business case doesn’t work for me. So where where we can grab more values. Yeah, so you see a di

Grace Shao (14:37)

That’s really interesting. That’s a really interesting approach.

Yeah, because it’s like the company is reflecting actually a broader, I think, social, even cultural and perspective on how they’re perceiving AI. And in the China and US, often the conversation is so fixated on improving efficiency and people who are utilizing AI are actually more burnt out because they are like 10Xing themselves or whatever these days.

Alex Lu (14:49)

Exactly.

Grace Shao (15:03)

But you know, in Europe, that conversation is so different. And you can say maybe much more humane. However, like you said, the bottleneck right now is then how do these companies become the next generation? Like still relevant in the future, once this becomes normalized. So it’s interesting. I wanna go back to that a little bit later as well. I I wanna touch on something before we get further into I guess the comparison of you know, the adoption and everything is

Alex Lu (15:17)

Yes.

Grace Shao (15:30)

You and I met each other essentially online because I found out about your work that you helped a lot of European companies adopt Chinese model. I found that was very, very fascinating, right? you advised them on how to basically integrate, say, the Minimax and C AI of the world. Now, a lot of these model companies, when I speak to them, they say their priority right now is to basically sell globally. And of course, the Western markets are some of the most lucrative markets. the US.

headwinds are mostly in geopolitics, compliance, Europe. How do you view that as a market for them or opportunity for them? Like, is it equally challenging for these companies to sell to enterprises in Europe or do you think there are more opportunities for them right now and they’re they’re kind of taking off a bit more?

Alex Lu (16:16)

Yeah. I I I I if I think the conclusion if I I can state at the very beginning is kind of in Europe definitely there are more opportunities than for Chinese large language model companies than in in Europe, than in US, sorry. I have two proofs for that. Firstly is I discussed with a a CTO who is also a schoolmate from my school Polytechnique.

And actually he was very interested also by my newsletters on LinkedIn and then when they he asked me the question so apart apart from Deep Seek, what are dip other models and Chinese models that we we that we can use to improve our efficiency? Because when we meet all of he said he told me when we meet most of the IT implementation companies, they came with the solution like Anthropic or ChatGPT or OpenAI or or Google Gemini.

So we don’t see so many options. He mentioned the word options of Chinese models. And we know that Chinese models are more cost efficient. And then we can talk about token mapping. I think it’s kind of related topic. So this is one thing. I think in Europe, actually for the companies from the business perspective, they are also looking for different variety of different models so that they can bring what I said before, a best cost performance ratio models in the in the organization.

So this is one thing. And then I I told him that most of the Chinese large link model companies, firstly they started their business in China and then they tried to inf have the global influence. Like the most advanced one is zero dot zero one dot AI, but the other ones they are trying to catch up, like that AI you mentioned also Minimax. so I I think the thing is what I see today is the ecosystem of Chinese models.

Are not currently penetrating into the European markets. But definitely there’s a a room for Chinese players. The second thing is I always take the comparison with the other industries like EV industries, like car industries. because you you will see Europe put a lot of tariffs on Chinese vehicles. because okay, you you see a lot of Chinese vehicles because of

Europe wants to protect their own industries, et cetera, et cetera. But at the end of the day, they are not putting hundred percent tariffs. They are putting somehow reasonable tariffs on the Chinese vehicles. So the bottom line I want to mention is I lived in Europe before and I know the mindset of Europe European people. The mainstream, of course, we have different views. I think the mainstream for European people, most open ones.

Are saying okay, we need fair competition. The EV cars is just because okay, the European commissions are claiming that okay, you produce in China, but we in Europe we produce in a more sustainable way, so our cost is higher, blah blah blah. So if we take this comparison, I think definitely there will be some places for Chinese companies in condition that we play fairly in the European market.

And then we might come back to the third point I mentioned before, of course, there’s a a point of AI sovereignty. the biggest, the biggest player of European AI ecosystem is still Mistrol, so it’s the biggest player in the foundation model. And of course, Mistrol should be one of the choices options when we suggest to European clients as the large language models.

So I would see that if tomorrow the Chinese model enter these European markets, they will face a fierce competition with Mistro because Mistro basically they have a government back, let’s say, from France, and they have a very good positioning in the ecosystem. you would see in in two or three weeks you’ll there will be a VivaTech in France and Mistrol for sure they will be on the stage and for sure they will be

French or German presidents, French president and German chan chancellors. And with their unique positioning, I think most of the European companies they were firstly considered Mistrol, but still if Chinese companies can bring something on the table, business wise, the European companies will not only limit to only one model, there will be some balance between different models. And today, the balance I see is Mistral versus other US models.

Grace Shao (21:11)

No, I just think it’s really interesting ‘cause I think it also totally makes sense when I to talk to people who are in the Korean market or, you know, covering the Middle Eastern markets. sovereign AI is just such a top of mind like conversation for companies, whether it’s for compliance reasons, or regulatory reasons, whatnot. So it makes a lot of sense that Mistral’s position very well in Europe. However, are there any other players that maybe we’re overlooking outside because we’re not that familiar with the European market? Any other

foundational model labs that we should know of coming out of Europe.

Alex Lu (21:45)

Yeah, there will be apart from Mistro there’s another large language model whose name is H, but it’s less famous. And then you’ll have it’s not if we can say it is kind of word model by Yen Laquen, the ex researcher in Metafair, and he just came back to France and lab raised raised a large amount of money for the for for his model. It’s called AMI, yeah, AMI.

Grace Shao (22:01)

Mm.

Interesting. Okay, so let’s talk about the token maxing thing you touched on just now. So offline we talked about this a little bit recently. There’s been getting some buzz. It’s quite funny, you know, whether I’m it’s like big tech in the US or big tech in China. When I talk to them, people are saying, Okay, our managers are pushing us to token max. If we don’t basically use AI in our job and figure out ways to essentially replace ourselves, we get replaced, which is the irony in all of this. It’s it’s all kind of sci fi. but

Grace Shao (22:41)

Then the joke’s kind of been played now on the companies because you know there was just headlines coming out saying, this one guy basically spent like more than half a million dollars on tokens in a month, and that’s obviously more than his salary. And then companies are realizing, wait, this token maxing strategy is not cost efficient at all. So from an operational standpoint, I know you are someone who work a lot with companies to implement AI and find

Alex Lu (22:54)

Yeah.

Grace Shao (23:10)

the most cost efficient way for their for their operations, right? And not just costs, like you mentioned, it’s like a balance of costs, you know, and and operational sustainability as well as obviously company morale and everything. So how do we view this trend? Where is this going? Is this sustainable? Like just just give us some high level views on this.

Alex Lu (23:32)

Yeah, there there’s a a lot to talk about this, because the token is becoming really a trendy topic for individuals and as well for companies. so to answer your first firstly to answer your questions, I don’t think that’s sustainable. My view is the token mapping is kind of marketing for infrastructure companies. and of course, as you say, there’s a lot of people burn a lot of tokens and more than their salaries.

Then the question would be if I pay your salary or if if I pay your tokens. We’ll come back to this point afterwards. I discussed with some some Chinese companies. Very cost cons cautious. I think the the the thing is today when we actually for for the tech companies in China, there’s also some ranking of token consumed. but it’s kind of indicator of how people are use AI. But

It’s not is it the right indicator? I don’t think so. basically I think in the in the in current status we didn’t we didn’t find a very good metric to measure the performance of a human being empowered by AI. that’s the thing. So we take a kind of proxy indicator, which is the token for and of course there’s a lot of waste of token in in in in in the usage and I’m I’m not sure that every single employees

would be the master of AI if we don’t provide the sufficient upscaling in terms of the AI. Because from individual perspective, sometimes we use by coding, but if we don’t master the basics of coding, then we might waste some time and as well as some money and tokens in the by coding. So this is my view. So the token maxing is kind of marketing stuff and and the the day when we find out

Again, for the AI organization or for the organization, how we can measure the performance of individuals with AI, then we might have a clear picture and no longer token max. And the other interesting thing you you mentioned already, but I read also is Microsoft they are kind of switched to their copilot because ever s if everyone used used the entropy cloud model then become too expensive for the whole organization. It’s just not just not cost efficient.

And brings me to my point is when discussed with some Chinese companies. So, you know, Chinese companies are very cost conscious. And they are thinking is I think that it was a joking, but this is right angle of thinking is show we in the salary of our employees to allocate a part of the tokens monthly for our employees. meaning that okay, if the

If in the in the in the past situations hundred percent of the salary tomorrow might be eighty-five percent of yesterday’s salary plus fifty percent by tokens. And the tokens you can you can use and if you don’t use tokens efficiently then it’s the savings for the company. So this is it’s

Grace Shao (26:43)

That’s really crazy. But I kind of see what you mean.

Like so essentially it helps you with your job. So that’s why it’s on you. But then what if you just don’t w but what if you don’t want to use AI? What if you just like I can do my job perfectly fine the way I did it before and I don’t want to token max and I want to keep my hundred percent?

Alex Lu (26:50)

Exactly. That’s the question that the Chinese company needs to answer, but you reflect on your point mentioned that the token consumption is sometimes much more expensive than the salary. So it causes Chinese company companies to think that okay, I spent salary, I spend tokens for the intelligence, I spent two times to hire employee. So why not combine them together and doing kind of tomorrow’s package is your basic salary plus tokens?

Grace Shao (27:32)

So actually on on that,how should companies think about it then? Because, you know, it’s really easy to say, okay, this is an AI native company. There’s 20 people in this company. Everyone’s token maxing because it does bring the 20 people’s efficiency to say like 400 people, whatever it is, right? However, what about the traditional companies, especially the ones that you work with? Like a lot of them are OEMs, manufacturers, you know.

It it doesn’t make that much sense for them to really jumping on this AI bandwagon as well then. Or how do you advise them then? Or how do you think how should they think about it?

Alex Lu (28:06)

Yeah. I I think for the for the traditional companies or European companies, it doesn’t make sense for everyone to give the token maxim because as I said, I’m pretty aligned with the European approach saying that okay, in order to release or unlash the value of AI, we need at least to upskill a little bit our employees. We cannot expect employees like with thirty ex years experience in the industry and tomorrow he switched to a kind of AI expert in the

in the in the in his company. So I I just want to combine our question with my previous comment saying that today if you look into the Chinese market today there are some big big traditional telecommunication companies like China Mobile they are proposing the token plan for individuals it’s like your smartphone monthly monthly plan yeah

Grace Shao (29:02)

Wow. Like data plan.

Alex Lu (29:05)

It’s a kind of data plan, exactly. So the token is becoming kind of infrastructure like electricity, like water, or like your smartphone, monthly subscription. So this might be the way the companies might pursue, saying that, okay, yesterday I might give you a kind of monthly plan for your telephone. So I can reach out to you and you can read the emails and you can use the telephone to walk with emails, work teams or with Zoom, etc. etc. And tomorrow it might be a com kind of monthly subscription.

For different employees, then you have a monthly token plan you can use for your personal, not for professional work in AI. I guess that might be the way that the China might be moving for individuals and for companies. And again, for European companies, they are not there yet, but when I discuss this vision and this kind of trend, and they are pretty interested, they might be moving in the same direction.

And for the companies, at least not at the national level, but at the company level, to provide kind of a monthly subscription to a limited number of people who master AI, and the first wave of people adopting AI is their coding team, their IT team, their digital team. So they will be the first employees to use this kind of concept of monthly subs subscription to tokens.

And of course, for manufacturing companies, there’s a lot of people working in the factories, in the plants, or or in the on the production lines, and they are not be impacted, they will not be impacted by this kind of AI wave. But still, I think the things are are moving slowly and it’s it’s changing so quickly. but this is currently my discussion with European companies.

Grace Shao (30:48)

That’s actually very interesting. I it makes a lot of sense actually to build it in in as like a infrastructure like 5G data. And then it’s really, it’s really like there’s a cap on how much the company will pay for, but then how you utilize it should be, and you’re more mindful of how you’re utilizing this, right? And not wasting the tokens and and buy the that thus you know, wasting your energy, compute everything. So

Grace Shao (31:13)

I want to bring it back to the Chinese pricing models really quickly. I know you work with a lot of European companies, they are the buyers essentially. You also help them connecting with the Chinese vendors, essentially, which are like the Chinese LLM labs, Minimax, Drupal, Moonshot, etc. Now, how should we understand the pricing model of these companies right now? Because it’s obvious that they are pricing themselves much cheaper to US peers.

Some might you know, obviously argue that their performance might not be as on par like on par or as at the frontier. however, even when they do play catch up, you know, the reflection of it is it just s seems like a complete different cost structure. Help us ex understand that, like they’re thinking, why they’re pricing it much lower and how that plays out in the long run.

Alex Lu (31:46)

Mm-hmm.

Yeah. Actually, there are two perspectives on that. maybe I will firstly talk from the client’s perspective and then I I might conclude with the recent price decrease by Deep Seek. maybe you you have already read about it. so f from the European companies as I said before, the thinking is w that the that that’s that’s the statement for the companies I met. we do not need entropic models for all the time.

That’s for sure. Because this is very expensive even for a company. So for sure they will need a kind of different options from different models, like the best U US models and the cost-performing, the best cost-performing models from Chinese models and the AI sovereignty models like Mistro. So basically there will be three combinations and then there will be engineering of technical issue that meaning that how to manage these models to

Perform the right tasks. So, meaning use cloud to perform the most complex tasks and use Chinese models to perform kind of less complex tasks. And I think European companies they understand this. And they, of course, they are looking for Chinese models for the cost effectiveness. And I would say this is also one of the bottleneck of US models because they are very

In a relative way, very expensive. Therefore, it it’s the bottleneck of the massive adoptions. Only the European big, big companies can afford like continuous use of US models. well there are a lot of SMEs in Europe. So this is this is the thing. And for the Chinese model suppliers, I think the the way I I see the the the price issue is if you ask me

Can Chinese companies increase their token prices? I would say surely, because if you look into the financial report of ZAI or Minimax, actually they are not they invest a lot in the research to develop these models. And the expectation from the industry AI industry is if you want to train or pre-train a next model, you will cut it will be more costly than the previous pre-trainings. so for sure.

Chinese companies can increase their token prices. And w that that’s what they are doing actually after the open cloak, if you read into the news. And the thing is, compared to the US model, still the Chinese model are very cheap. I think there’s one very strategic thinking thinking angle is if you think about the Chinese models, most of them they are open source models. And the the the the thinking angle is, I think, for the Chinese model players is

We want we open source these models because we want people to use these models. Because they can deploy it on their own infrastructure, they will have more freedom, or they can use our open source model to train their own models. and maybe they they will use our our tokens by or they will they will understand or know our models better by open sourcing. So if if we combine this thinking angle, I would say.

The Chinese model strategy might be to increase the influence in the world, maybe in the developing markets, where people are more cost conscious, and to help people to use this AI to adopt AI in a cheaper way. And then in the long term, in the future, that’s very Chinese, maybe again to increase the prices once we take the market positioning.

It’s like the price competition for the last decade regarding this digital sharing economy or digital era. Nothing has changed. So a very aggressive c pricing strategy to at least to to have the market share and then once we have the market share then we can establish our our our our position in the market and ca kinda do a lot of monetization stuff.

That’s the one thing regarding the increased influence globally and taking the lead in the AI industry for the developing countries, in my opinion. Of course, go going to Europe is is part of the their strategy. So this is from the Chinese model’s perspective perspective, and it’s a very special case, of course. It did this is Deep Seek. DeepSeek released just the before and right after the release, during one month, I think for the developers we enjoy the

75% of discount regarding the token price. So it’s very deep discount. And recently, I think one or two weeks ago, DeepSeague announced that they will keep this 75% discount for for for forever. So it’s kind of they they just discount their token prices by such huge amount of discount. I’m pretty surprised. and

Again it di it it launched a price war in the market and you see recently Xiang Mi decrease also their token prices and I don’t know if other players will will follow in Chinese market at least. But if we think about Deep Seek cases, it’s a very special case because Deep Seat this year it doesn’t create a lot of buzz in the AI community in the US. I think so. I I’m not living in US but I read some newses. news, sorry. I think the

nowadays DeepSeek, I’m not saying that we have the best performing model. and and and in terms of of the tok coding performance, DeepSeak is is not at the top top level compared to other models. But the interesting thing is DeepSeag this time is trained on the Huawei ASEAN chips. so again, I think the price decrease of DeepSeag combining with their recent news of raising money and hiring some harness engineering

Across the world, I would suspect that DeepSeek by decreasing their prices, they just want to break through the ecosystem established by NVIDIA. This is my thinking, and and that’s why after the President Trump visit to Beijing, there are 10 Chinese companies are not authorized to buy Nvidia chips, but up to now you see few others.

Grace Shao (38:09)

That’s interesting.

Alex Lu (38:23)

I think there’s a thinking from the national wise from from the nation thing that okay with Dipsy can we break through the Nvidia chips plus CUDA? And if because that’s so cheap, so most of people they might use Deepsi in the future and they might be used Huawei as ASN chips because Deepsi got trained on these chips and it’s best support DeepSeak’s performance. So this is another angle. Yeah, so you would see

Grace Shao (38:23)

Mm-hmm.

So the open source strategy. Sorry, go on. It’s basically a strategy

to get people in to get the developer into its ecosystem, its own community first, which is what Jensen’s been saying the whole time. Yeah. no, I I agree with you on that. I actually I I wanna and steer away from the chips today because I I am quite fascinated. So you work with companies, adopt AI, but how does that actually what are companies really using AI for? Like we hear about stories.

Alex Lu (38:54)

Exactly.

Grace Shao (39:16)

you know, companies are token maxing, whatnot. And obvious the obvious one, like you mentioned, is in coding capacity in IT, but no again, not every company is in tech, you know, not every company needs coping co coding capacity. sorry, let me just say that. Not every company needs coding capacity. So like what are we seeing actually on the ground, especially for maybe more brick and mortar stores or old school traditional industries? Why would people all want to adopt AI right now?

Alex Lu (39:47)

the the the adoption rate actually for European companies is pretty low, to be honest. most of companies, if we say at a large scale, they don’t adopt sufficiently AI and they just are afraid of missing out something. So this is a FOMO. they are just feared of missing out some opportunities, and if they don’t use AI today, they might be less competitive in the future. So the the f the most common use cases I see in

companies for coding and for it and sometimes it’s easier to measure the effective effective sorry effectiveness of ai that’s in the most most of the time in the sales marketing department so meaning that if you use ai you can produce produce more contents and with more contents you have more impressions with more impressions you might have more conversion rate you might have more conversions and you might have more sales revenues so

This chain is actually well formed. So by using AI, you can track the individual metrics on the chain, and then you can kind of monitor the results by using AI. And most of the time, I get a very simple question of European companies, and very difficult question actually to answer is: what’s the ROI of implementing AI? What’s my return? then it’s a very difficult question because in the

digital, 10 years ago in the digital era, I can tell the ROI, I can estimate why, because the incremental cost of using digital products is kind of almost zero. You just need your digital products and then it makes more efficient, it makes more automate. Well, in AI, that’s very difficult because if you think about it, if you use more AI, you will consume, as you say, more tokens. So, meaning that

An employees, you need to pay the salary. If he is a heavy AI user to produce more content, then you will need to pay his tokens bill. And then the ROI might not be so immediate. Or there might not be ROI actually for the individual use cases. Then we come back to the question: is okay, by using this AI, how we can make the whole organization more efficient and how we can generate more revenues for the whole organization.

While for the individual users, maybe there’s no business case. So I think again, the the the the the difference compared to 10 years ago is the people who use AI and who use heavily AI, then he will have a bill to to pay. That’s a variable cost. That’s very important. And secondly, is the variable cost will be really

The beneficial of the variable cost will really depend on the skills of each individual. You may pay $100 for employee A or employee B. If B master better AI, then you will have 10 times more results, financial results, compared to the first case. So again, I think you asked the right question. the ROI question is definitely a very good question. and most European companies they seek about ROI before investing. So they are very cautious.

While again, if we compare to the Chinese companies, we are more pragmatic. So let’s implement a POC. it costs a little bit, but let’s implement it. If it doesn’t work, never mind. We waste some money, but we we we continue, we iterate or we continue with another use cases.

Grace Shao (43:17)

So you think the Europeans are taking a more cautious approach, but actually more cautious on what the potential ROI is. Then I bring it to the question that is a bit more philosophical and like a societal, not so businessy, is then isn’t the headline or the mainstream discussion on AI is replacing our jobs completely overblown then? If companies are not even investing in the like, you know, buying tokens, I don’t think they’re replacing people and comp just replacing roles with. Like AI, are are they? How do I understand this?

Alex Lu (43:50)

For the tech companies, I think your statement or the statement is true for the tech companies because they’re traditionally there are a lot of coders, there are a lot of programmers, and and actually I see a lot of developers, individual developers in the market because they work for tech companies and now with the with AI. That that would be very challenging. And again, currently for European companies, if I would say

They’re still at very, very early stage compared to to China. the cost is one thing, and we can take at the other angle, causes equals to conservative. So they are a little bit conservative and they care a little bit more about their employees. So actually I I I will not see in European market AI replace a lot of human workers. It’s not happening today. Will will that happen tomorrow? I think so.

Grace Shao (44:46)

Mm-hmm.

Alex Lu (44:49)

but again we need to find another society structure or we need to find other job opportunities for the human beings when AI comes to the companies and replaces some of them. It we’re not like very aggressive like at the tech companies like Meta or other tech companies. it will happen slowly, but of course AI has impact on the on the employment on employment, even for European companies. and

Grace Shao (45:13)

Mm. The economy itself will evolve and and jobs will look different.

Alex Lu (45:21)

Exactly. it that that’s exactly what I I was in Europe ten years ago. It’s exactly the discussion around industry four point zero if people remember. We say that okay tomorrow we’ll have some automated machines in the plant. So it’s kept it’s not it’s happening currently in China. We call it a dark light factory. So it’s very automated. you can run the factory without turning the light on.

so basically at that time in Europe we had a very big debate on where the employees employ employers should go once industry four point zero is in place. And the answer was there were sorry, the answer was there will be some upscaling and new job opp opportunities created with industry four point zero, and we need more skilled people to master these machines. And that’s that’s the same thing for the AI.

Tomorrow we will need people who can orchestra, who can manage the agents, AI agents, instead of doing the same job as a simple agent.

Grace Shao (46:23)

Yeah, I see. So so on that, I wanna ask, you know, given Europe’s strength in industrial, like industrial strength manufacturing, where do we see opportunities for companies to really couple that with the development evolution of AI right now?

Alex Lu (46:41)

You mean the the use cases, right, for the companies?

Grace Shao (46:44)

Use cases, new opportunities, new potential businesses. where could we see p like, you know, new businesses come out or, you know, new business revenues for current industrial companies?

Alex Lu (46:55)

Yeah. for European companies currently the use cases we’re discussing is more around kind of efficiency use cases. So for example, they want AI to help them to do some root cost analysis because if you run a a plant and if the machine is kind of done, the production line is kind of stopped, and then you you you lose basically a lot of money because you missed up.

opportunity of producing X unit units of of your products of your cars. So basically people care a lot a lot about how I can analyze the root causes of of a machine being done. And this traditionally was a very heavy task. We need we need a lot of experts to be involved and because there’s a whole system of different machines in the same plant. And the machine is kind of the product production line is kind of

made in a industrial sequential. So every parameter on different machines might have an impact on the chain. So we need to involve a lot of experts and by using AI actually we can we can understand better. We can do some causality analysis and do some root cause analysis and find the root causes more easily and in the future to do some predictive predictive maintenance and to improve the efficiency of the companies. So this is currently happening.

People are asking for that. And some companies are also asking for these kind of knowledge management platforms. Like we we need knowledge management for new enrollment of employees, for HR policies, for reimbursement policies, for new employees onboarding, etc. etc. So a lot of around that. And if we look into the vision and into the future, I think European companies are start to think about it.

I’m talking a lot a lot about European companies, but that’s the same thing for for the companies in China, it’s just kind of more advanced. So sorry, I I’ll come back. So if we take into the vision of European companies, actually they are also thinking about the future, which is how I can use AI to increase my revenues and to make the pie a little bit bigger. And then it comes to the discussion of agentic economy.

Meaning that can I use my agent to kind of sourcing, to kind of sourcing for my company? Can I use my agent to do some business development, to write emails, to do some code calls, to reach out to potential clients? So these are the things that people will come to think in the next wave, saying that okay, if we have a very good engineering of our agents, guidelines of our agents, what an agent can say, what he cannot say.

what he should say in which context. So once this is done, again it’s very European, they need to use everything kind of under control. Then I think we are ready to to go for the athentic economy so meaning that agent can do business in in the place of the companies.

Grace Shao (50:04)

I see. And if I were to say I’m the founder of AI native company, how would you advise me other w because it would be very different from what you’ve been saying about advising more traditional industries?

Alex Lu (50:10)

Yeah, it it i if you are a AI native founder, I think I’m I’m doing the currently the same position. there are a lot of things to consider. For example, in terms of the technology, the foundation model is evolving very pretty quickly. So how I make sure that my AI agent idea or concept or business model will not be revolutionized or disrupted by this

Foundation models. This is something we need to think about. The second thing is I always tell myself and also people in the say same AI community is we we don’t start to build our products from scratch without discussing with the clients. So why not in in a more safer way, why not discuss with the clients, build products for certain clients, and then kind of

Conceptualize the products and build more standardized products that we can sell, we can say, we can sell to market and we can scale in the future. It means the build of the product comes always from a specific demand of the clients. And once if there’s a demand, then we can do something, we can build things. Why this? Because, in my opinion, all the AI native funders, I think we are pretty aligned is produce.

something or build a product in the future will be much easier in the past. And if we compete with AI in terms of the intelligence, there’s no way a human being can catch up with AI. And we should place our time where the AI cannot compete and where we still need a human being. I I I make very simple analogy to some friends of mine saying emotional intelligence, meaning that how we can establish relationships with the people, how we can build a trust.

So still I think if I’m a founder or if AI native founder, he should go out to meet clients, discuss with clients, build a trust and have some demands from the clients because building the process will be pretty easy and the cost of failing is pretty low. So build fast, fail fast, scale fast and it works even more in the in the future.

Grace Shao (52:36)

And then my question on that is how do we actually understand how to build guardrails and safety around this? Because you talked about how Chinese companies you work with are often a bit more like gung ho, let’s go, we’ll t we’ll fix it if after it’s broken, kind of mentality. Whereas the European companies maybe are seen as a bit of a slow adopter in many ways, you can say more cautious, more humane, and protecting their concurrent employees. But, right, like

End of day, if this is the future evolution of our economy, how do we go forward with this? And then how do we actually build more intentionally?

Alex Lu (53:13)

Yeah. technic technically, actually there are a lot of skills, there are a lot of technical stuff in the area to build the guardrails for the agents, like Anthropic, I they are doing doing a very great job, and also some Chinese foundation model companies and also agentic companies. So all of all of that they call that the harness engineering. So they put every concept into the harness saying that okay, we need to build a harness and to make the guardrails.

So this is the technical perspective. But still, this technical perspective is very from the developers or programmers. And if we bring the case into a real company case, then it really depends on each use cases on each company. I would say for any new human employees which is who is a new hire in the company, at least when I join European companies, there’s always a code of conduct.

You see, it’s it’s simply a a document that we need to learn. We need to we need to we need to be compliant in the future in in our work or professional work within the company. So I would say for the AI agents that the same thing. they are very important in the future, a kind of infrastructure to evaluate the performance of the AI agents, meaning that if the AI agents is delivering the performance as we wished before, so there’s a kind of benchmark evaluation.

And also the evaluation should include also is the AI agent performing correctly as we wished in terms of the code of conduct. And the code of conduct should in my opinion, be written by human human human beings. It’s like an extra bic team, they have they have written a a hundred-page of constitutional constitution for for for for cloud. And then each company should write their code of conduct for.

every agent in every department. And a lot of Chinese founders then they are entrepreneurs, they are also joking at okay, we develop an AI agent today for companies but the next question will come shortly is when should we retire our AI agent it if it doesn’t perform correctly or why when we should replace them. So you see the evaluation or benchmark of of the AI agents would

shortly become a a a pro a p a problem in the market when we adopt massively the agents.

Grace Shao (55:45)

So then each organization will have to institutionalize this, essentially you’re saying, and have their own standards of code of conduct, whatnot. That makes a lot of sense. Yeah. And right just like how companies right now regulate data usage, even company devices, whatnot, right? Like this will all just be part of the compliance that employees will have to learn. I want to ask you one last question, which is what’s one differentiative view you hold?

Alex Lu (55:54)

I think so. In terms of the AI?

Grace Shao (56:16)

In terms of everything, it’s a question I like to just kinda throw throw it at people when they come to the podcast. It’s a it’s a wild card.

Alex Lu (56:24)

Okay. I think one of the points I always mention, it comes back to my background, is today the AI race is between US and China. So we say that European is kind of lagged behind. but do not forget that actually technology is one thing and the usage of technology is another thing. And again, if we come back to our

my my statement saying that implementing AI is not about technology. It’s not it’s about process culture and organization and human being. So I think the placard of the Europe is they’re pretty good at regulations. And if you think about they issued GDPR before the Chinese PIPO, which is protection of personal data. And they have this kind of European AI Act. And then if I think about how anthropic

They penetrated these enterprise solutions versus ChatGPT and generate today more revenues than open AI in terms of AR, because of the simple concept of responsible AI. Then I would say tomorrow, if the AI comes to the enterprise level, enterprise implementation, and if everyone should be responsible in the company with their own agent or with their own developed AI, maybe Europe has a part to play in that.

in the in the AI in the in the world of AI, because their initial statement is kind of we want AI to be regulated, we want AI to be responsible. So this is my point of view.

Grace Shao (58:06)

Thank you so much. You know, today you’ve been really generous just explaining to me and and the audience just how AI is really being implemented into these big companies and the more European perspective. is there anything else you think we’re missing or any misconceptions we might have about the relationship between European companies and Chinese companies or how Europe is perceiving AI? Is there anything you think we’re missing or do you think we covered it all mostly?

Alex Lu (58:36)

Yeah, I I I think we covered most of them, but I just want to mention one thing is even though we say that okay, there’s two different nations in the world, US and China, competing AI, or in we we we take different directions of AI. And still I received a lot of recently questions from European companies, and they are really, really interested by Chinese tech companies. So you would see

they are pretty open and they come frequently nowadays to China and they have the mindset of learning what Chinese companies are doing, what Chinese foundation models are doing, and especially seeking their use cases. so one thing I would say is when I receive them, we show some very advanced Chinese use cases. They would say, you are in a different environment because we have different laws, we have different regulations compared to you Europe.

but they are quite interested about what’s happening in Hong Kong because the regulations in Hong Kong is pretty closer to European markets. So still, I I see we might have a lot of potential collaborations between China and Europe in terms of the AI, in terms of the physical OI. We didn’t mention the robotics, and definitely it’s an area where European can have more playground, not only

About the humanoid robots, they want also to have their places in the hardware value chain for the robot robots. Like a lot of

Grace Shao (1:00:10)

I’m sorry, it’s I know we’ve hit our time, but what what is your view on that? Because you know, European companies traditionally been the leaders in robotics, right? Industrial robotics, like machinery. where do they stand now in the world? You know, are are the Germans and the Japanese still leading the space or or how how are they gonna be kind of presenting themselves or positioning themselves on the supply chain right now?

Alex Lu (1:00:15)

No worries. Yeah. So for the very traditional industry robot robots that let’s say it’s like KUKA, you have a lot of robotic arms. So they are still kind of leading the world, so you have a lot of robotic solutions implemented in the in different car makers’ plans. but for the humanoid robots, actually Europe Europe is lagged behind again because it’s not all only about the value

About the not only about the supply chain of the robots itself, it’s also about again the software and the large language models behind the robots. So the mindset of European companies today is: okay, we understand China again has the most advanced humanoid robotic companies in the world. US has maybe advanced in software in large language models or word models. China is pretty good at the supply chain.

So again, the same question they ask themselves. But the the recent demands I receive from European companies are are two. The first one is as a traditional European companies, we know that they know that the value chain of making a car is quite similar. Let’s say it’s not hundred percent the same thing, but there’s sixty or fifty percent are common of making a car and making humanoid robots.

So their thinking is okay, can we participate in the wave of these kind of robots with the development of China? So like motors, like electric motors, like actuators. Yeah, German, German guys are pretty good at at this apply. So that’s the first thing. The second thing is demand is a lot of European companies saying that okay, we have the real use cases in Europe because we are lack of workforces in our plants.

It could be an aging population, it could be some strike of labor unions. So they in order to keep the plant working, as we said before, about the predictive maintenance, they are very welcome, the Chinese robotics in the European markets. Again, the robots need to be compliant with European regulations, conditions, and they are very welcome. So the most common demand I receive is hey, hey, I I want to do a kind of analysis about

how I can be part of the supply chain in China and how I can leverage Chinese supply chain to be more competitive. The second one is okay, I have a use cases, then we need to think about how I can implement the humanoid robots in the European markets. And then we we can discuss about the business model of the robotic companies like Unitree of AJ Boss, because it’s not only about putting their robots in the factory, it’s about calibrating the robots, it’s about capturing the data, it’s about think about a closed loop of robust training. It’s about

the again, the guardrails how make sure robots will not harm a human being if they cross each other in the plant. So yeah, this is quite common nowadays for physical AI for European companies also, yeah.

Grace Shao (1:03:35)

Mm-hmm.

But in fact, actually you mentioned CUKA and it was bought out by Matee, right, a couple of years ago. So you’re also seeing a lot of Chinese companies like in the embodied AI, physical AI space actually actively buying out traditional brands in in Europe. How is that received actually locally?

Alex Lu (1:03:47)

Yes. actually for the for the embedded robots humanoid robots, there are not so many MA of Chinese players acquiring European companies. So basically I think for the humanoid robots, let’s say the robots like AJ Bot or like Uni3, China is much more advanced. And there was one robotic company in France, but they are kind of in financial difficulty. And another robotic company

They were in they are invested by Renault in France, but still their technology if you look into that is not as advanced as Unitree or AJ Rob A Gi bots, for example.

Grace Shao (1:04:42)

I see. one last question is just do you think it’s fair that we’re overgeneralizing all the European companies into just one EU right now? Or do you think actually a lot of different com countries have different goals, ambitions, or even, you know, future tracks for them laid out?

Alex Lu (1:05:03)

very good question. So I can only when I see European companies, sorry, actually I’m thinking about French and German companies. So actually I cannot represent all the European countries and for different countries like Spa Spain, Italy. I’m I’m not familiar familiar with the country. I didn’t live there. I I didn’t receive enough clients from from these countries. So actually you are right.

when I talk European companies, I’m more thinking about French and German companies. And of course, they are pr pretty different.

Grace Shao (1:05:35)

Okay. Well, thank you so much. Yeah, thank you. I just think it’s such a unique perspective because, you know, it it’s it’s more it’s easy for me to find someone who tells me the pure European perspective. It’s easy for me to find someone in the China US, but it’s harder for someone to for me to find someone f you know, who straddle between Europe and the Chinese market. You know, it’s obviously not as mainstream. So I’m really appreciative of your time and your insights and your sharing. Thank you so much, Alex.

Alex Lu (1:06:04)

Thanks,

Grace. Yeah, thanks a lot again for i inviting me and accepting me for the podcast. And thanks a lot for your audience. And yeah, let’s keep in touch if any chance happens. we can have another talk if needed.

Grace Shao (1:06:17)

Definitely.