The interview was published on Bilibili by 张小珺商业访谈录 (Zhang Xiaojun’s Business Interview series, produced by the 语言即世界 studio) on 25 July 2026. All credit for the interview belongs to them. I simply struggled to find an English translation and so produced my own which I’m sharing here.

The method flowed entirely through claude code, with it downloading the video, transcribing with Whisper large-v3 and then cleaning, translating and lightly editing for readability. I had it cross-check against another independent Chinese transcription and press coverage, and found no issues, but keep in mind this is pure AI transcription and translation so some errors are likely.

Prologue

Zhang Xiaojun: Hello everyone, I’m Xiaojun. Today our guest is Dr. Liao Heng (廖恒), Huawei Fellow and Chief Scientist for semiconductors. This should be the first time since Huawei’s ordeal of 2020 that a Huawei executive has come out to talk about how Huawei’s Ascend (昇腾) chips climbed step by step out of the trough. At the same time, this is also my own first time learning about the chip and semiconductor industry. So on one hand we’ll talk about the history of Ascend and China’s choices and story; on the other, from a more macro perspective, we’ll talk about the history, patterns, and overall picture of the entire global semiconductor industry. If you like our video, please give it a like, coin, and favorite. What follows is my interview with Dr. Liao Heng.

What changed in your state of mind? The difference when designing the 910 versus the 950 — the two generations of chips before and after the supply cut-off.

Liao Heng: This change in state of mind — I think putting it into words may sound a bit... What I want to say is: try to imagine, if you were Dong Cunrui, about to hold up that explosive charge — right? — ordinary people like us simply cannot imagine what his state of mind was. Or the soldiers at Shangganling (Triangle Hill) — what was their state of mind? Or — I once saw an interview with a Chinese soldier during the war, which is also a story from Huawei’s publicity posters. A reporter asked him: what do you want once the fighting is over? He said: I don’t think about that, because my parents are dead too... no point thinking about those things. So I’d say the so-called change in state of mind no longer has much meaning. It’s just: you face a difficulty, and you want to solve it.

Zhang Xiaojun: Today Dr. Liao made one request of me: he doesn’t want the focus on him personally. But he has witnessed more than 30, even 40 years of the global chip and semiconductor industry’s development, so I’d like to enter from that angle — through your eyes, take us traveling into this stretch of semiconductor industry history. Setting personal experience aside and talking about the industry itself: over these past 30–40 years, into which major eras would you divide it, and what was the core question of each?

Chip History: The Long Sunset Under Monopoly

Liao Heng: That’s a long topic. But before we get into it, let me first describe how this came about. In the past, HiSilicon (海思) very rarely appeared before the media, so I feel fortunate to have received this invitation. I did think it over at the time — there were probably three reasons that led me to make up my mind to accept this interview. First, having such a chance to look back on this engineer’s story might leave a bit of inspiration for other people in this world.

The second reason: about half a year or more ago, at Tsinghua, at the invitation of Professor Lu, I taught one session of the Computer Organization course in the computer science department. After teaching that class I was left with a deep sense of frustration — and Professor Lu, who teaches the course, also felt deeply helpless when we spoke afterwards. In this era, young students are far more willing to chase hot topics — AI algorithms, model training, even inference acceleration. Computer Organization is a hard course to study; the big assignment is to design a complete CPU, and everyone regards it as a “gate-of-hell” course. So it’s very hard to attract students to take an interest in hardware — computer processor hardware. On the one hand I felt extreme surprise; on the other, a kind of anxiety: if people aren’t even willing to study this anymore, will the industry lose its follow-on inflow of talent — could a gap open up in the human pipeline?

And of course the third: we feel that people’s understanding of this industry — even my own subordinates or colleagues — is lacking. They’re buried in hard work every day and really don’t have many opportunities to be shown a more complete view of the industry — what the full story actually is — so it’s easy to feel lost or to waver. So I thought: if there were such an opportunity — not just for industry peers, but for young students, or for my own colleagues — to present a relatively complete, decades-long perspective on the industry’s development and its underlying patterns, that would be a very good opportunity too. So these three opportunities — three origins, let’s say — taken together helped me make up my mind to accept this interview.

Now, the topic you just raised — the broad arc of semiconductors over the past thirty years. I started my undergraduate degree in 1987; then around ‘97 — I went to the United States in ‘96, and in ‘97 I joined a semiconductor company. And of course the field I studied — during graduate school and undergrad I actually worked on processor-related topics. If we summarize the industry over these past 30 years, it has really had some very large rises and falls. Looking at it longitudinally, I think there are two main threads interwoven with each other that form the through-line of chips, or hardware. One main thread is obviously the processor, right?

The CPU — from about ‘86, or ‘84, slightly earlier than the IBM PC. But from around ‘84–’85 you had the IBM PC, and then the CPU gradually moved from a desktop office terminal into enterprise IT servers, and then into the entire global infrastructure that followed — with digitization, humanity entered a digital world, right? In all of that, the CPU is clearly the most important main thread. And of course, on this thread, starting around 2005–2006, this AI wave is a new wave again — we’ll probably unpack that in more detail in a moment. So that’s one thread: the processor-centric thread.

Then there is a second, very important thread: the chips brought by communications infrastructure. Over these past 30 years, humanity went from an unconnected state — a physical world relying on physical connection — into a connected one: everybody connected, every device connected. That course of development is the development of the internet, and it represents another longitudinal thread. The world went from no internet, no broadband, no mobile phones, to every person and every device being connected together — and of course that required enormous quantities of chips and enormous infrastructure. And this infrastructure went from very slow dial-up over phone lines and modems, to broadband, to optical fiber — that’s the wired side — plus, behind it, the trunk lines between cities and between countries. That is what built up the entire internet’s infrastructure.

Then, with the emergence of mobile networks — and later things like Starlink — that is, wireless infrastructure, this amounted to the internet’s second wave. In fact we are right now in that second wave of the internet: the mobile internet. On one side it spawned massive infrastructure; at the same time, on the device side, the most representative product is the smartphone, which has basically become a companion that we humans use for perhaps six or seven hours or more every day — it has become a necessary part of our lives. And this thing is also a huge, massive pillar of semiconductors: the semiconductors it consumed at one point may have exceeded 60 to 70 percent — that is, if you count all the phones. So those two threads are the longitudinal threads.

But if we then look by points in time, there is also a horizontal thread. What we see, dividing things up by time, is that semiconductors went through a phase of extreme, feverish ascent and prosperity — and then rapidly entered a state of withering and decline, even a sunset industry. The most emblematic symbol of this “sunset industry”: before this AI wave, roughly before 2016, for a full ten years or more — even fifteen — Silicon Valley’s, America’s VCs made no investment whatsoever in chip startups, OK? Because everyone believed this was already a finished, sunset industry. And the most representative case was Broadcom, which invented a “Broadcom model”: acquire relatively mature, profitable semiconductor companies, merge and restructure them, and cut costs — that is, raise operating efficiency. That is entirely a sunset-industry harvesting model. Then after 2015–2016 came this AI boom, right? So now, you could say: the internet bubble, then the mobile internet as the second wave, and AI as the third wave of these most recent years.

This third wave — the enthusiasm it has stirred up, the volume of capital, and people’s expectations for the industry — may be even higher than the previous two waves. So we should think about a question: why, after that previous wave — after the internet, say — was there a 10-to-15-year sunset period, a dusk? Roughly 2005 to 2015 — really it had already begun after 2000, yes. But from 2005 to 2015, almost a full decade, VCs invested nothing, because nobody was optimistic; people felt chips were a hopeless industry. I think here — and what I’m saying may not be politically correct — I think it was precisely the extreme success of certain sectors that caused the withering at another layer. I don’t know — maybe that’s a bit counterintuitive, isn’t it, Teacher Xiaojun?

Zhang Xiaojun: Yes — when you told me that, I found it quite counterintuitive.

Liao Heng: The logic here is actually quite simple. Of course, China has now begun to move into an industrial model of its own, a kind of self-sustaining cycle. But taking America before that — if you rewind ten years, global tech was still basically led by the American model, leading the world’s currents. And the American model has one striking feature: when a fairly important, great invention or a new business appears, and it gets the backing of the capital markets, they can rapidly build a near-monopoly advantage. For example Google, or Meta — that is, Facebook — in social, Google in search, and before that Windows in terminal operating systems, plus Apple’s phones. In their respective fields, each of them — in perhaps just three or four years — could complete the establishment of a monopolistic advantage. That advantage is not only a technological advantage; more important are capital, infrastructure, and the number of users — in effect, people get used to one thing and won’t easily switch platforms, right? And this kind of monopoly, this monopolistic advantage, in fact has an extremely negative effect on innovation. The effect is this: if a customer is the only merchant in the world purchasing a certain class of product, then as its seller — the supplier — the supplier’s lot is miserable. Miserable — well, you can’t quite call it miserable; it’s just what the supply-demand relationship inherently determines.

When a customer accounts for the vast majority of purchase volume, it has enormous pricing power, and it will inevitably leave the supplier with nothing to earn — gross margins become very meager. Second, they come to dominate demand and the direction of technology. That dominance looks perfectly reasonable, but let me try a small example: if you listen entirely to your customer, there is no future. Because in chips, it usually takes at least three years from the main architecture or product-definition decisions to the product actually landing, in real users’ hands: the chip design cycle alone is a dozen-plus months, the manufacturing cycle now is maybe nine to ten months, then you still have to build it into a complete system, and then test. So three years is unavoidable — maybe two and a half if you’re fast, four if you’re slow. And so — people often lack that foresight — the calendars don’t line up: I live on a calendar set in 2030, while my customer lives in 2026. There’s a four-year gap in between. If I can only blindly obey my customer, then I will lag by four years. Therefore, if this so-called technical leadership — the power to define the future — is placed entirely in the hands of one dominant party with an overwhelming advantage, it will inevitably suppress new innovative forces. Because even if you have the right idea, you get no chance to put it into practice.

And this series of causes is what drove the traditional wave — that is, the pre-AI wave — of semiconductor companies, the vast majority of them, rapidly into a state of decline: first, operations were hard and there was no profit to be made; second, the things they had hopes for in the future rarely got a real chance to be built. Now of course, what I believe is that this latest wave — at least in the AI era, perhaps on the hardware side from around 2016–17, with AlexNet and the continual progress of these DNNs, roughly from 2016–17 — up to now still has not entered an unfortunate monopoly, or a single dominant pattern like that withering period I just described. And there is a very interesting phenomenon here: this field is extremely, extremely active, and its progress every month, every quarter, exceeds everyone’s imagination. Frankly, up to now — and of course there is also our China, an important participant in the world, a player that has stepped onto the field — you could say that, at least among the people we can reach, it is a sky full of brilliant stars; this field is intensely active. Nobody can say that Sam Altman, or Anthropic, will be the final winner. Because every day our peers — especially those super-smart, ambitious young people, perhaps in Wudaokou, perhaps in Hangzhou — keep bringing us surprises, preventing the monopoly from ever forming. And so more of this vitality keeps appearing at every single layer of this industry.

Zhang Xiaojun: The dominant players you’re talking about — you mean Google and those big guys, right?

Liao Heng: I think we in China also have plenty of dominant players. For example, if you buy things — shop online — you can identify who the dominant players are, right? If you say short video, you also know who those dominant players are.

Zhang Xiaojun: That is, the dominant players of the internet-application wave.

Liao Heng: Right. They represent enormous scale, enormous purchasing power, an enormous consuming population — and they also sustain massive infrastructure.

Zhang Xiaojun: I said it was counterintuitive because in my imagination, the development of upper-layer applications and of the underlying chips should be a mutually reinforcing relationship. I hadn’t expected that once monopolists form at the top, it would instead leave the chip industry in a period of withering.

Liao Heng: Then let me ask you one question — you’re a hand in this industry too — when you picture Google, what kind of company do you think it is?

Zhang Xiaojun: It’s a search company.

Liao Heng: It’s an advertising company. To this day, the vast majority of its revenue still comes from advertising tied to this broad search business, right? I think the tech industry has one striking feature: nearly all of them — these entities we can identify — have a first axe-stroke, a formidable opening act, and it’s always impressive. Typically it goes: they invented something unprecedented, and brought humanity some enormous, wonderful experience, or some indispensable capability that they enabled. But it is very, very difficult for a company to keep remaking itself. Say OK — after my first-edition super trick comes out and I’ve established a monopoly position, do I still have the ability to invent again — to reinvent, to innovate again?

This is the classic innovator’s dilemma: someone who has already built a huge advantage in a certain field — he, or a company, an entity — will certainly reap rich profits there. But does he have the ability to reinvent himself in another field, to create the next, even greater thing? There are indeed quite a few brilliant people who can keep breaking through their own DNA, their own limits, and keep bringing humanity surprises. But for the great majority of organizations, the capacity to remake and recreate themselves — to build an entirely new self — is usually limited. If you don’t believe it, run every company you know well through your mind, and you’ll find they each have an advantage somewhere, don’t they? But an advantage is usually precisely a disadvantage in a new field. So we see that today in the AI field, the companies with the biggest breakthroughs are in fact not the big names we knew — the previous generation’s big names. That’s right. Of course the big names still want to — and of course they have resource advantages — they want to do this and that, they want to work hard at reinventing, and we very much look forward to them reinventing, since after all they have their own advantages in talent, resources, organization, and execution. But what I want to say is: the world is wondrous in just this way — it is often not the richest child who becomes the most successful. For example, look at the classmates we’ve known: was it the classmates from the best-off families who ultimately achieved the most in their careers, or contributed the most to society? Usually not.

Zhang Xiaojun: You entered Tsinghua in 1987, and through the Youth Class (少年班) at that. What was your computer science department studying back then?

Liao Heng: I’d say the coursework probably wasn’t lagging that much yet. Growing up — from my teenage years — what I got my hands on was an Apple II compatible; around the time I started university, the IBM PC appeared, right? And we were still on floppy disks — a hard disk was a novelty. Musk hadn’t founded his companies yet.

Zhang Xiaojun: Had Musk even been born?

Liao Heng: He should have been born by then. Right, but it wasn’t yet his time to start companies. So in that era, in terms of the atmosphere at school, there were a few striking differences. First, the teaching faculty of that time was still rather old-school and didn’t much care about publications. The research the professors did was all about building the real machine — at the very least a prototype. You could not graduate by writing a paper; it was always: I designed such-and-such, and I must see it — not merely conceive the thing, but actually produce an ugly circuit board that works. That’s quite interesting — it represented a culture of research. At that time, because China’s own industry — the research and R&D capability of its companies — was very weak (and going back earlier than my time, the gap was even bigger), universities and research institutes in effect took on the role that corporate R&D plays today, because companies had no R&D. That includes China’s earlier “Two Bombs, One Satellite” pioneers — the computers they built and used, everything they used. It happened to be a transitional period. The research system that came after me — including how students were trained and graduated — would shift into a publication-driven mode, right? More like the American, Westernized model, where everyone treats getting papers into top conferences as the first priority — and that was very much in step with the rhythm of the times.

Because when I left China, a company like Huawei probably had maybe only a few thousand people, and what they were building was honestly still fairly rudimentary. That is, the system for organizing thousands upon thousands of people to develop something extremely complex in an organized way had not yet formed — it was still taking shape. But if you fast-forward to today, many Chinese companies, big and small, have development capabilities — developing products, knowing how to organize people and resources, how to build a relatively complete human-resource system to accomplish that kind of teamwork and organization work — that are already very strong, even world-leading. So universities no longer need to take on the kind of development-type work they did in the past.

But going back to the era I was just talking about — what was particularly interesting is that back then, besides going to class, we actually spent a huge amount of time knocking around in Zhongguancun (中关村). “Knocking around” meaning — it was actually very much like today’s startups: OK, you discover the world has lots of needs, and then we would try to develop a product, or at least a product prototype, to try to meet those needs.

Zhang Xiaojun: What did you make back then?

Liao Heng: We made countless failed things — laser printers, VCD players, electronic dictionaries — all ancient stuff. People today probably don’t even know what an electronic dictionary is, right?

Zhang Xiaojun: I know what that is.

Liao Heng: Right. But at that time, companies’ R&D capabilities were very weak, so university students essentially filled that gap, developing these things on behalf of those small companies.

Since I myself was a competition student, I was decent at programming — passable at software. At the time I found my best friend, named Wu Zhao (吴钊), also a Tsinghua classmate; we were close in age, less than a year apart. He was a hardware competition student. So for many years I really yearned — thinking, I want to be like him: I want to be able to build boards, design FPGAs, and I absolutely must figure out what is actually inside a processor — what is really going on inside the chip. Right? Because chips have now reached over a hundred billion transistors, an extremely complex system. So when I felt my software was decent, I would pour enormous effort into trying to pierce through that layer — that membrane, that boundary — as if it were the difference between the upper floor and the lower floor of a building.

Zhang Xiaojun: That was your yearning at the time?

Liao Heng: That was my yearning at the time. And so, in much of what came later, I had a very strong preference — I was especially eager to understand what the world on the other layer was like. So later, professionally, I became through and through — or at least primarily — someone doing hardware.

Zhang Xiaojun: So at university you leaned toward software, right?

Liao Heng: My starting point was a programming-competition background.

Zhang Xiaojun: How old were you when you entered the gifted youth class?

Liao Heng: The gifted youth class generally required you to be under 15; I was 14 at the time.

Zhang Xiaojun: Wow, so young.

Liao Heng: I don’t think that’s anything special. The vast majority of people get in through happenstance — various circumstances give them a particular kind of opportunity. Getting into that kind of school purely on your own ability is quite hard, so it’s not worth dwelling on.

Zhang Xiaojun: What you were just describing is crossing between different layers — we can probably come back to this topic later. On the way here you also showed me Jensen Huang’s five-layer cake, right?

Liao Heng: Well, in our minds it’s more like an eighteen-layer pagoda — eighteen layers of “confusion.” (A pun: 宝塔, “pagoda,” sounds like 糊涂, “muddle.”) We’ll focus on that later.

Zhang Xiaojun: Back to 1987. When I was studying chip history I found it very interesting, because TSMC (台积电) was founded in 1987, I believe — Morris Chang (张忠谋) was 56 that year. TSMC pioneered the foundry model, which later rewrote the global chip industry’s division of labor: before, it was vertically integrated, and afterwards design companies became one category and chip foundry companies another.

Liao Heng: Yes.

Zhang Xiaojun: You lived through that era.

Liao Heng: By the time I experienced it, TSMC was already fairly powerful. Or rather — although it had not yet formed today’s kind of monopoly, that overwhelming advantage — it was already quite strong.

Zhang Xiaojun: How did this model come about? What is its root cause? Why was it a better fit for that era?

Liao Heng: I think the most fundamental reason is economics — it’s a principle of economics. Because a wafer fab’s capital expenditure keeps growing, right? Each generation it increases at — I don’t know — certainly something close to a Moore’s-law rate. That is, a process node that used to have maybe 10 mask layers now has 100 mask layers; each machine that used to cost maybe $100,000 or $1 million now costs $10 million or $100 million. You can think of the foundry — wafer manufacturing — as a business with extremely large capital investment and extremely long R&D cycles. This corresponds to what I’ll describe later as floors five and six of the pagoda — the basement levels. As a business it is extremely high-risk, with terrible ROI, and an extremely long payback period on investment — and as process technology evolves, it gets worse and worse. So at that point, as a design company, you simply don’t have that kind of capital capacity. It’s like how nowadays every home uses air conditioning and electricity — you wouldn’t build your own power plant just because you need electricity, because the outlay for a power plant is too enormous. So it became a kind of — I think Morris Chang was the first to realize this: that everyone would want to publicly share this infrastructure. Of course, on Bilibili you can probably find many interviews with him where he looks back on this history. But he was the first to realize it, and to explicitly turn it into a business model — that is a great achievement, a major contribution to humanity.

Of course, this contribution — countless companies have been built on this model and found success. I can put it this way: without the fabless model, even a giant like Google would have been unlikely to do the TPU. Because building a fab of his own might take three years; developing the process technology, another three years; getting the fab running properly would also take a long time. And its scale probably couldn’t sustain the economies of scale that a leading-edge fab requires. But once this model was established, it naturally enabled everything that followed — in any case, my entire career has been under this model; until about five years ago, it was all under this model, right? The model has its rationale for existing, but it is not everything. AMD founder Jerry Sanders famously said that only those with wafer fabs are real men — “real men have fabs” — but in the end even he spun off his own chip manufacturing division. This model has its strengths, but that doesn’t mean everything in the future can only follow this model. Because the rationale we’re talking about rests on certain big premises, and once those premises change, you have no choice but to adapt.

Zhang Xiaojun: What was the big premise at the time?

Liao Heng: The big premise at the time was global liberalization — “the world is flat.” Everyone specialized in their own part of the division of labor: at every layer, I want to procure the best capability in the world. So at HiSilicon, before 2019, nearly all — 95%, even 99% — of our designs were built on the world’s best suppliers: the “downstairs,” that is — I would pick the best fifth and sixth floors to build my product on. But that big premise has clearly been completely overturned by competition between nations, or by the influence of various other geopolitical factors. That’s the first point. The second is that when something develops too rapidly, this layered division of labor may not be able to meet the demands. For example, the extreme memory shortage that has appeared in the past half year has driven memory prices up more than tenfold — DRAM has risen more than tenfold in price. Vertical division of labor, to some degree, cannot cope well with such rapid — excessively rapid — change. It’s like the stock market, right? When there’s an explosive crash — a Black Friday — nobody can react in time; or when there’s an extreme surge, your normal trading operations can’t keep up with the speed of the rise.

Zhang Xiaojun: When did you realize you would probably spend your whole career in the chip and semiconductor industry? Was it at Tsinghua, or after you went to America?

Liao Heng: At Tsinghua I already really wanted to do this. Because of what I mentioned earlier — that yearning for an understanding of the “downstairs,” right? And I think the upstairs is more comfortable: the view is good, you can see farther. For instance, if you build an application, as long as you have the right concept, within maybe half a year you can rack up tens of millions of daily active users, right? The vast majority of super apps emerged that way. But the downstairs represents a more patient kind of player — a long-distance runner, so to speak.

Zhang Xiaojun: Are you that kind of long-distance runner?

Liao Heng: I don’t know. Throughout my career, most of my colleagues have said: you’re not a chip engineer, you’re a wolf in sheep’s clothing wearing a chip engineer’s skin, because you’re not really doing chips — you do software. And then when I’m with software colleagues, they either say I belong downstairs with the chip people, or they’ll suddenly say I do algorithms. So I think this polyhedral, multifaceted character — actually, doing chips requires knowing those things; or put another way, you need to be able to spar with others in those other domains without losing.

Zhang Xiaojun: I’m curious — in 1996 you went from China to America, to do a postdoc. At that moment, what was your imagined picture of the future of technology? What was the prevailing mood?

Liao Heng: That mood is a bit complicated to describe. I held two thoughts at the time. The first was that I felt I had to go, OK? Otherwise — even now, many excellent candidates I interview, say people who have already done PhDs at top schools and proven their abilities are second to none, still feel — many of my “Genius Youth” (天才少年) subordinates say: I still have to make a trip to MIT, otherwise I’ll feel I’m lacking. I had the same mentality back then. Because America as the technology leader — from roughly 1945 to now, I’d say, all these years — people would have a kind of regret: if I’ve never seen it, I won’t know, and I’ll forever feel something is missing. That was the first thought.

The second thought was this: I felt that in that startup-like mode I described earlier, I had made a lot of miscellaneous things that looked like products, but nothing I made had ever — say, sold a million units, or had countless people raving “I’m using the thing you built” — nothing had truly reached the market. At the time I didn’t know what my problem was — I genuinely didn’t — but after working for a year or two, I quickly found the answer. When I went over, I still carried that puzzlement: why is what I make never good enough, or never reaching the market? Actually the answer is very simple. After a year and a half of working, I quickly understood what the missing piece was.

Even if you’re making an electric kettle, you must ensure that even if one component inside fails, it won’t keep dry-boiling — otherwise it will cause a fire, or even kill someone. If you make an electric kettle, it absolutely must pass safety testing: a bit of water splashed on the side must not electrocute anyone; if the relay on top fails, it must not dry-boil; it must fail safe, not fail dangerously, right? That kind of process — for a thing to actually be usable, it’s not enough to have a clever enough mind to imagine how to design an electric kettle, or how to design a SpaceX rocket; you also have to execute it, execute it reliably, and you need sufficient verification and testing, before it can finally meet the standard of going to market and being used by ordinary people. As a student I hadn’t recognized this at all — I thought, I have a clever idea, I build it. In fact all we could claim was reaching a prototype. And that prototype — you can’t imagine: take a phone like the one you use — it may take tens of thousands of people in R&D, over ten thousand person-years of investment per generation, to get it to the point where it doesn’t, under some circumstance — black-screen the moment it gets somewhere hot, shut down instantly at a ski resort in the Northeast, or fail to charge. There are countless things that require a rigorous process, a strict process. Only after going through such a process, such a cycle, can the product be polished to the point of being usable.

For example, our intelligent driving — it was also first incubated in our team, but that process took seven years: from building a prototype that was already driving all around our campus, to finally launching and selling to the first consumer. Getting it to land was a long road — it took a full seven years. So as a student I had absolutely no awareness of this, and very naively thought that being fairly clever, or being able to program, or able to build a board, meant I could make a product.

Another interesting, related topic: only after a year of working did I realize that if you want to organize — never mind ten thousand people — even just 20 people to build a product, you can’t have every one of them be an Olympiad competitor. You need ordinary, conscientious staff — people whose intelligence or skills have simply come through an average education and your hiring screening process — to be able to work effectively, and you need to combine them together. You can’t pick a genius for every slot, and if they were all geniuses that would create problems too.

Zhang Xiaojun: Right, with 20 geniuses you’d definitely never manage it.

Liao Heng: All kinds of conflicts would arise among them, right? That’s another issue related to companies, or teams. So everyone has their own function. This is actually very healthy for us — whether for a company or for a society. Otherwise, only the strong survive — the ones just a tiny bit better than others, or with a slight edge in one particular respect, would leave everyone else no room to survive. But the reality is exactly the opposite: you need large numbers of conscientious, responsible people to finally get things done.

Zhang Xiaojun: In 1996 you arrived in America, at Princeton for a postdoc. What was your first impression of America?

Liao Heng: Well first, I had already been to America before that — I had gone in 1987. From then to 1996, another nine years had passed. On that first trip, I thought America was a fantasy land. I was only 14 then, and everything seemed like a dream, because the gap between China and America was just so vast. How silly was I? After coming back, I told others — or told the woman who later became my wife — America is so wonderful, they even have air conditioning out on the streets. Later, when I tried to recall why I had held such a stupid notion — in fact, I must have visited a shopping mall in Palo Alto, near Stanford. At that time China had no shopping malls, right? Back then we only had farmers’ markets and the shops lining both sides of the street. Well, a shopping mall is basically a street with a glass roof built over it. So my mistaken impression at the time was that America was so advanced they even air-conditioned the streets.

But by 1996 I was of course an adult. The first thing was a degree of disappointment. Looking at the world as an adult rather than a child, I felt a deep sense of not belonging. Because the Princeton I went to is an American elite-style institution — its temperament is rather different from other universities. Later I would actually learn that Princeton has many excellent qualities worth learning from, but at the time I didn’t feel that. I only felt, profoundly: this is a place where America’s elite aristocracy lives, and I do not belong to this class — so I need to leave this place quickly.

Zhang Xiaojun: So you stayed only one year and left right away.

Liao Heng: Right.

Of course I had a second disappointment as well. I had originally thought that maybe I could do a postdoc and perhaps have a chance at a faculty position. But by then I already understood — understood fully — that if I wanted a faculty post, I would have to do a PhD all over again, because the Tsinghua brand wasn’t strong enough. Today the world may be different — in academic progress, including integration with the world, Tsinghua’s gap versus that era has narrowed dramatically, and in places it even leads. Back then there was still a lot of inferiority in my heart. I wouldn’t say it was pure inferiority — perhaps a mix of arrogance and inferiority.

Zhang Xiaojun: Did you think about returning to China at that time?

Liao Heng: I didn’t think about it then — not at that time. I felt that going back then would be — I can only say it was probably a mistaken perception: the belief that returning meant you couldn’t make it, couldn’t hang on there, right? I didn’t want that sense of defeat. But later, the facts proved that this perception was very foolish. There’s no helping it — when you’re young you make some stupid mistakes.

Zhang Xiaojun: That probably doesn’t rise to the level of a mistake.

Liao Heng: I think it was simply a mistake. If I look back at classmates around my age, or ten years younger, who pursued this profession in China — even the average, ordinary people I described earlier, conscientious, positive, willing, not “lying flat” — they have all done very well. Why? A person’s growth depends partly on themselves, but even more on the larger environment, right? If the whole environment is rising, you naturally rise with it — like standing on an escalator that is itself moving upward — that kind of collective progress. Whereas in America, from my visits in ‘87 and ‘96, I felt they were not progressing quickly — they were even regressing. But China over these past 30 years has progressed very fast. And of course I was fortunate to return to China for the second half, and to take part in a piece of it.

Zhang Xiaojun: When I was reading chip history, I found that 1997 also had a big event in the chip industry: in April 1997, Nvidia launched its third-generation NV3. Only after the failures of its first two generations of chips did it finally gain a firm footing. Did you notice this company at the time?

Liao Heng: I didn’t. I’d say they were nobody at that time. In 1997, in the history of chip development, there was actually a far more spectacular drama playing out — the main storyline was not what you just described; that was the peripheral of peripherals, it wasn’t important. The main drama being staged at the time was actually an extremely important juncture for the CPU. Because at Princeton, I shared an office with some fellow students — and for those working on processors then, the company they yearned for most was one called DEC, Digital Equipment Corporation. DEC may have vanished from history by now, but at the time it was second only to IBM. DEC made minicomputers and mid-range machines, and was also in Boston. This was DEC’s final period. Digital had a flagship program at the time called Alpha — the Alpha processor. That processor represented the highest level of CPU design of the day, so everyone doing processor research wanted to join that team, to have the privilege of taking part — just as today everyone wants to work on Nvidia’s Rubin or Feynman, to be part of that program, which represents the highest level in the world right now.

Let me rewind this story a little — it actually relates to what I said earlier about how, once a monopoly forms, it suppresses innovation. The CPU started out single-core and single-issue — that is, executing one instruction per cycle, or even taking many cycles per instruction. Then gradually came the reduced instruction set processor (RISC), including its inventor, Professor David Patterson, who is still very active today, and Professor John Hennessy, who once served as president of Stanford. They led a processor-simplification movement: make every instruction very simple, so one instruction can execute per cycle. Then in the 1980s — by ‘87, perhaps starting from ‘84 or ‘85, and on through ‘87, ‘89, ‘90 — the era of multiple issue arrived, meaning multiple instructions could execute in a single cycle; multi-core had not yet appeared. And at that time there was an extremely important competition — probably still the most critical one in processors to this day: if a processor can execute multiple instructions per cycle, how do I schedule them? How do I generate a piece of software, how do I compile it into code, such that it executes several instructions every cycle? That way it runs faster. Because, as mentioned, at that time Windows — or rather DOS plus Windows — was already, on the desktop — on the CPU side a monopoly had taken shape, but in the server field it was still a hundred flowers blooming.

The hottest company back then was Sun Microsystems, plus SGI and a whole series of others — Digital was still around too — and the operating system was UNIX, in its various flavors. That was the most important golden age of processor development, and it lasted about ten years. The main battle of technical routes at the time: on one side, statically scheduled processors, called VLIW, where the compiler arranges the instructions in advance and, at execution time, each instruction stands for multiple instructions executing simultaneously; on the other, the superscalar approach, which is fully dynamic — that is, nobody pre-arranges which instruction executes in which cycle; the processor itself fetches the instructions, places them in a scheduling window, and schedules them dynamically. Why do I bring this up specifically? Because this important topic will come up later in our discussion of AI processors: all of today’s AI processors are statically scheduled — including the world-leading Nvidia GPUs. Everyone is in fact still walking the VLIW road, not the superscalar road.

That debate went on for ten years. Because software and IT systems had not formed a monopoly, besides Sun and SGI there were also HP, IBM, and Digital — a contest of many powers, like China’s Spring and Autumn and Warring States period: every school of thought was present, and everyone was actively experimenting, trying to prove their own doctrine right. But this story — let me fast-forward — ended in an epic disaster; it all came crashing down. First Digital went bankrupt; when it sold off its assets, the Alpha processor was acquired by Intel, probably for very little money. Digital’s biggest asset was a search engine called AltaVista — a forerunner that came about six or seven years before Google. AltaVista essentially pioneered the search engine. It just didn’t know how to make money at the time: it fetched a lot of money when sold off as an asset, but nobody yet knew how to turn a search engine into a business that earns money; technically it was a pioneer. Later, of course, very clever people invented better search-engine algorithms, and Eric Schmidt found the business model, so Google became a giant and AltaVista vanished into history. But what I want to say is that this debate still carries very deep lessons for us even today — many lessons to learn from. And as we keep moving forward with the evolution of AI processors, we may very well draw useful directions from this history.

Zhang Xiaojun: What was your own view at the time, in that debate?

Liao Heng: My view at the time was hardly worth mentioning — because what I was doing at Princeton was also VLIW. At that point I had never built a chip, never done a full-scale chip that genuinely had to tape out and go into manufacturing. So as a so-called PhD student I had no judgment — very naive. To have judgment, you need to know things, you need cognition of reality: first know what the situation is, be aware of your surroundings — OK, it’s hot today so I wear short sleeves; when it’s cold... You have to perceive the real world. That perception of reality, back then, I completely lacked. Later, after going through so much—

Zhang Xiaojun: After feeling that Princeton wasn’t quite the right fit, you chose to go into industry.

Liao Heng: I basically just went and found a job — whoever wanted me, I went. So I didn’t have many choices either. We all belonged to that cohort — I brought a thousand US dollars over, which probably already counted as being a rich man; plenty of classmates dared to go to America with a hundred dollars, some without even enough money in their pockets to pay the taxi fare from the airport to the school. So we had a fairly strong dose of reality back then: the feeling was, first of all, survive.

Zhang Xiaojun: In 1997 you joined PMC-Sierra. It had also just gone through a restructuring: a Canadian company and a Silicon Valley company had just merged, and you happened to join right at the point of their merger.

Liao Heng: Right. It wasn’t my active choice at first — they chose me. But actually it was a small company; at its very peak it probably never exceeded 2,000 people. What does that mean? The department I belong to today may have 5,000 — 4,000-plus people, with revenue easily 100 times theirs. So PMC was probably a small entity. But it lived through a magnificent, surf-riding industry pattern of the kind I described earlier. It happened to be standing right on the first wave of the internet, because its business was transport networks — several of the key devices in the internet’s backbone infrastructure, metro networks, including the long-distance transport networks between cities. That was exactly what it made. That wave was really just like the capital-market frenzy we see today, right? At one point PMC’s market cap became perhaps the No. 1 — at least the No. 2 — highest in all of Canada; it could easily have bought the Bank of Montreal. But that illusion vanished very quickly — a few good years, a few bad years. I arrived in 1997, and by 2000 it burst.

Zhang Xiaojun: Burst in 2000.

Liao Heng: In 2000 — you may not know; you look very young, maybe you were a child then, or not even born yet — but in 2000 there was a fantastic... anyway, it was the first bubble-and-burst cycle I lived through. The internet had become the world’s number-one hot topic; people believed everything in the world — that the internet would change everything. So on one front, Microsoft was in a fight to the death over the browser of the day, called Netscape: Microsoft built IE and was determined to kill Netscape — over the entry point, the portal. Just like now, aren’t all the internet applications fighting over the user entry point? It was the same then. At the application layer, people believed the browser was the first entry point, because everything had to go in through the browser, so whoever controlled the browser had the advantage. And at the infrastructure level, people believed everything in the world would pass through the internet, so the whole world was frantically, extravagantly overbuilding backbone and access networks, so that every city would have internet coverage, and the fiber between cities had to be laid with plenty of bandwidth. So PMC’s share price back then shot up just like Nvidia’s or Cambricon’s today — very, very high, right? But very quickly you discover — of course I’m not saying these present-day companies are the same — there is one biggest economic measurement in all this: whoever invests has to be able to earn it back. The internet bubble lasted about three years, and then people rapidly found that those investments... For example, the fiber America uses even today may still be what was laid in 2000, because so much was suddenly deployed and then — oh no — nobody’s buying, nobody’s leasing; you find it cannot be monetized. I think this problem does not necessarily exist in China’s AI today, but in the internet of that era, in America, it genuinely existed.

Zhang Xiaojun: Right — that was when you went through an internet bubble burst. But later the internet companies actually rose again, didn’t they? Around 2005 the internet application companies took off.

Liao Heng: Yes, the application companies rose again — but the people doing infrastructure never came back up.

Zhang Xiaojun: Then why didn’t you ever think of changing jobs, jumping ship? Was it that you had come to know people there — or, to put it another way, was it a question of them abandoning you versus you abandoning them, and you weren’t willing to abandon them?

Liao Heng: I never interrogated myself quite that explicitly. I just felt I still wanted to find another product — something that would let these chip-making colleagues, or myself, have the chance to build another product and bring in new revenue.

Zhang Xiaojun: And did you find one?

Liao Heng: Later we entered the IT industry — we started doing hard-disk, i.e. storage-related things. And the customers we served switched from telecom-side companies like Cisco, Ericsson, and Alcatel to IT-industry companies like HP, IBM, Dell, EMC, and Network Appliance. So I got the chance to learn: OK, this is what the telecom industry is like, and this is what the IT industry is like.

[1:00:00]

IT was once a very profitable industry. The equipment sold by companies like EMC and IBM carried very high margins. And the traditional IT OEMs — enterprises like HP — began to wither precisely after the hyperscalers appeared. Why? Because the so-called hyperscalers — those super-large internet companies — all need enormous infrastructure. In the global market today, I reckon that for purchases of servers, or of AI infrastructure, these big players — our term is OTT, over-the-top; the foreign term is hyperscaler, or call them super-large internet enterprises — should account for more than half of the global total, even 60–70 percent, even 80 percent. So when enterprises with that much purchasing power appear, they put enormous pressure on the brand-name OEMs. The old brand-name OEMs’ advantage was that they could make fairly high-quality products, but the customers they served were first the Fortune 500, then spreading out to the Fortune 1000 and 2000 — or, say, the tens of millions, hundreds of millions of giant, large, medium, and small enterprises worldwide. So an OEM needed a broad-coverage capability.

Right — say you want to sell servers into China: you probably need a service point in every county town, because that county might have a thousand enterprises as your customers, each buying only two or three physical servers; but when something breaks, you have to have someone make a house call. The hyperscalers are fundamentally different from that profile. First, there are very few of them — worldwide, maybe [unclear] — in the US perhaps the “seven fairies” (a colloquialism for the handful of US giants, akin to the “Magnificent Seven”) or however many; in China likewise roughly seven or ten, right? They are extremely concentrated. Second, their geography is concentrated too: all the machines sit in gigantic data centers, and you don’t need to set up a repair shop in some county town in Shaanxi, right? Third, when their purchase volume is 70 percent of what you produce, their pricing power is very strong. That is the other industry-changing driver I spoke of just now.

Zhang Xiaojun: And in that final year — the year you left PMC and joined Huawei, 2016 — the company itself was sold, for 2.5 billion US dollars. You happened to live through its entire cycle.

Liao Heng: That cycle was in fact precisely the sunset I described earlier. As I was saying, by around that time it more or less represented the entire American tech sector’s complete disappointment in, and abandonment of, semiconductors. First came Wall Street’s abandonment: their PE (price-earnings) ratios were maybe just three — on average about 3 to 5. What does that mean? Your profit this year is one billion dollars, and your market cap is only three billion. For reference: look at China’s AI chip companies today — their price-earnings ratios can reach numbers like six thousand or eight thousand. So first, the capital markets completely abandoned the semiconductor industry, because it had nothing new and no reason left to make anyone feel excited, right?

So what emerged at the time was the buyout model. Hock Tan — a genius of a leader, a business leader — invented swallowing the big with the small: first borrow a sum of money from private equity, then buy a company ten times your own size, buy it in at a low price, then restructure it — delete all the duplicated departments and personnel, cut the products with poor margins — and the financial statements turn beautiful. Operating that way — it is really restructuring-by-M&A, something that belongs to the late stage of an industry; effectively, certain people were liquidating the industry. Of course, the Nvidia you mentioned earlier belongs to the other category — injecting a brand-new hope into this industry — which is why we practitioners all deeply respect the contribution Nvidia has made.

Zhang Xiaojun: That sunset was truly long. From 2000 — three years after you joined the company — the dozen-plus years that followed were one long sunset period. What did it feel like to be inside it? I imagine anyone else might have left long before.

Liao Heng: I think the first thing was fear. Every day, colleagues would gather — over lunch, or while heating their lunchboxes in the microwave — and everyone would be discussing when the next round of layoffs would come and who it would be. But for me it was a — well, my wife also gave me a great deal of support at the time, or rather talked me out of giving up.

Because I think everything has two sides. First, it gave me ample opportunity to think and to observe: to think about technology — what kind of technology, what kind of product definition, has a chance to monetize; or to draw up a plan that contains the technology, the business, and a vision for the industry. And I had so much experience of failure, countless attempts. Also, although I was never laid off — right up to the last day — it gave me very much the role of an observer. For instance, I paid many visits to the IT giants — IBM, HP — dealing with them over the long run, going to Texas once or twice every month, and to IBM in North Carolina. I learned a great deal that way, including from the white-haired EMC engineers in Boston: from them I learned what this industry really was about, right — gaining more of a historical perspective. And I saw many failures, and came to know what kinds of things instantly trip the alarm — telling you this thing definitely will not work.

Zhang Xiaojun: Across that long sunset, what do you think were your biggest learnings? Looking back today, that stretch of experience may have been precious — because without it, the later Huawei chapter of your story might never have happened.

Liao Heng: What I learned — when I went to interview at HiSilicon (海思), the leader interviewing me asked: what can you bring? Because Huawei at the time wanted to hire “mingbai ren” — people who get it: they wanted to enter a certain field, but that field was new to them, so they needed to hire someone who had already worked in it for quite a while; that is what “someone who gets it” means. And I said: I don’t have many experiences of success — I have far too many experiences of failure, so I can tell you what won’t work. Of course I also often “kill the innocent”, right — my judgment is not always correct. But I have an especially large stock of knowing how things fail, and so the task is to work hard to steer around those failure factors.

Zhang Xiaojun: Could you name a few? Say three — failure factors, things you could tell at a glance would not work.

Liao Heng: For example, at the very beginning we set out to do ARM servers: we wanted to build a CPU to replace the x86 processor. My gut reaction — number one: failure. That was, at the time, a number-one product of what is now our Turing business unit. Back then I told the leadership every single time: whatever you do, don’t do this, because there is no market for it — or rather, no customer demand. Though look, I should stay humble here and admit my mistake: my judgment then was completely wrong. Because at the time CPU supply was plentiful, and as for the x86 ecosystem — AMD was not yet especially prominent, but everyone was used to Intel’s ecosystem, right? To change people’s habits you typically need a several-fold advantage: either you are cheap — a third of their price — or your performance is three times better than theirs. And at the time we possessed no such value advantage — that was my logic. But that logic, as I say, was wrong; the facts later proved I was totally wrong.

So my advice to everyone at the time was: we absolutely must find where the customers are. The only possible customers I could think of — because China’s cloud, China’s public cloud, did not yet have much scale; the cloud companies had all been founded, but they had not yet formed a good positive cycle — so I said the only opportunity was in America; we must immediately send people to Microsoft and Amazon. So we rushed straight off to Seattle. But unfortunately — in fact the action itself was not wrong; we were nearly on the verge of getting in, of having the chance to deploy this ARM CPU on Microsoft’s or Amazon’s cloud. That was 2016, 2017 — it started arriving around 2017; just as we had the opportunity to get into the large-scale mainstream cloud service providers as a CPU supplier, the US–China tech war broke out, and that story was terminated.

Zhang Xiaojun: That was still the previous era.

Liao Heng: Yes, that was that era. But after the tech war, when this story came back around, today we see that ARM servers really have — there are even predictions that within two or three years they may overtake x86 in the cloud, in the so-called data center, right? Because all sorts of people at these enterprises kept pushing — the hyperscalers even built ARM CPUs of their own. In China, too, we deploy on the order of a million-plus every year. And the rationale is not a pure value rationale — not purely “I have a threefold competitive advantage” — but involves many macro political factors. For example, China’s critical infrastructure can no longer run the risk of being penetrated through IT the way Iran’s power stations and nuclear infrastructure were, right? We must have something of our own. And our products have also improved very fast: no threefold advantage, admittedly, but no worse than the alternative — they have reached the point of being entirely usable. So society as a whole was fairly ready, and the product is also quite mature. That process took another ten years — but how many decades does a person get in one life? So I say my assertion back then was wrong — and yet, within that short time window, it was also not wrong. That is the first one.

Zhang Xiaojun: And the other two?

Liao Heng: The other two — well, as I said, it is about knowing what will not work and what will, right, and basically everything I can tell you is an example of my being wrong. At the time we wanted to build the Ascend (昇腾) chip, and the leadership said we absolutely must build a flagship product capable of training, capable of forming large clusters. I went and tried to persuade the leadership: whatever you do, don’t build this — it won’t sell. Because I had not yet seen the day when China’s sky would fill with brilliant stars — in 2016 I genuinely had not seen it. All the important algorithms and inventions were in America, right? Chinese researchers had not yet distinguished themselves — today it may just be Chinese in America competing with Chinese in China, but back then it was nowhere near so obvious. So I felt that a training-oriented chip — and Chinese internet companies had not yet begun training models themselves — I felt that product could never sell, and that we should build a small Ascend instead, right? When we first released the Ascend architecture, we hoped it would cover everything from one dollar to a hundred million dollars — six or eight orders of magnitude of broad coverage. And we did in fact achieve that: the smallest product sits inside our earbuds — those clip earbuds — and the biggest product is today’s training clusters of several hundred thousand cards, right? The spread we actually achieved is probably even more than eight orders of magnitude — whether in power consumption or in selling price. But at the time I was deeply pessimistic about the data-center product, and very bullish on the edge products. As the facts show, looking at it today, the data center—

I kept wanting to tell everyone: come on, let’s be realistic first and build something we can actually sell. But the facts proved me wrong yet again, right? So as I said: even when you hold that many prior assumptions (先验的认知), they are not reliable. And yet they still serve some purpose.

Zhang Xiaojun: Look, the several mistakes you just described were really all choices made on the basis of a kind of pessimistic expectation. Is that actually connected to the decline of the American chip industry that you lived through?

Liao Heng: Maybe there is. Maybe that experience trained me into a rather pessimistic habit — of first seeing the side of things that won’t work — and then working hard to escape that expectation of certain death, so to speak.

Zhang Xiaojun: And the examples you just gave all took place before the tech war.

Liao Heng: Yes.

Zhang Xiaojun: Including our autonomous driving — you mentioned just now the seven years, that we started doing autonomous-driving chips in 2019 — all before the tech war, when the external environment was still relatively good. So really, the choices you were making then were all small choices; you didn’t dare to bet big.

Liao Heng: At that time, I have to admit, some of the main leaders at Huawei — including the leadership of HiSilicon, or certain leaders at group level — had a much bigger frame of vision than I did. I was like a loach from a small pond suddenly thrown into this vast ocean, so I hadn’t yet learned that larger perspective. But after 2019, when we were placed in such a difficult situation, we had no choice but to learn to look at problems from a bigger vantage point. That is — take the examples I gave just now: at the micro level they all look fine, but at the macro level — I said my judgment was wrong — I was wrong precisely at the macro level. I didn’t see the bigger picture—

Zhang Xiaojun: Right.

Liao Heng: —didn’t see the expectations for the future, what the world would become. Really this is a limitation you fall into more easily when your experience is shallower, or when you’ve been a frog at the bottom of a well in a small pond for too long. Or, put another way, the younger you are the more easily you make this kind of mistake. I’m now in my fifties; back then I was only in my forties.

Moore’s Law

Zhang Xiaojun: I have a technical question I’d like to ask: Moore’s law in the chip industry — at what point did people feel it started to slow down? And what are the reasons behind that?

Liao Heng: First, Moore’s law has been slowing regardless of whether there’s a tech war — that’s an objective fact. Moore’s law has three aspects: its economics, its performance, and its energy efficiency — what we call PPA. As the area shrinks smaller and smaller, the cost averaged over each transistor comes down. That is what made electronics the only product in the world that is anti-inflationary — that is the greatest contribution Moore’s law made, right? If you buy anything else, what you bought ten years ago was certainly cheaper than it is now. Only with electronics is it the case that a ten-year-old electronic product today has no value at all, except perhaps some collector’s value. That is the economic improvement Moore’s law delivered: anti-inflation.

The second is performance: it used to be thought that the smaller the transistor, the faster it switches, so performance improved. The third is energy efficiency — energy cost: flipping a smaller bucket consumes less energy than flipping a big bucket. But at roughly 7 nanometers — or even as early as 16 nanometers — the economics of Moore’s law had already stalled: from then on, the cost per transistor not only stopped getting cheaper, it got more expensive. So the economic benefit was gone. The performance benefit is very small — yes, it is still slowly improving, but not by large margins. Where Moore’s law still genuinely delivers a benefit today is in energy cost: because the bucket has gotten smaller, the energy consumed each time you dump out that bucket of water is lower — that is, the picojoule- or femtojoule-level energy cost per switching event gets smaller.

Actually, this topic gets at something very fundamental. I can put it to you this way: the gap between China’s semiconductors and TSMC’s semiconductors — the main gap — is precisely in this energy cost. For example, if you ask me to supply the compute of a hundred thousand GPUs, I can supply it easily; it’s just that compared with the world’s number-one product, my power consumption is somewhat higher. But we have abundant energy.

Zhang Xiaojun: Right.

Liao Heng: So for us, for China, that is not necessarily a — well, first, this answer already has many facets. The first facet: nobody needs to worry that if we were completely cut off, China would have no compute. The answer to that is clearly no — we absolutely do have compute. The second question: am I more expensive? Not necessarily. The third: do I consume more electricity? Yes, I consume more electricity — there’s no way around it.

Zhang Xiaojun: But as you just said, China has energy, right?

Liao Heng: Our electricity supply is perhaps three times that of the United States, and we have a large amount of spare energy. Electricity for data centers in China may cost one-fifth or one-quarter of what it costs in the US; or compared with Singapore and other places around the world, on average our energy cost is about a quarter of theirs — roughly that level, a quarter or less. So first of all, we do not want to use this as the argument to persuade everyone to adopt something that consumes more power. All I can say is that as a floor — a fallback — this is not something to worry about too much, and China’s energy cost and energy supply are completely fine.

Zhang Xiaojun: So Moore’s law is still going?

Liao Heng: Moore’s law is of course still progressing. But to describe it more precisely, I think we can no longer use nanometers, a length scale — I think we should describe it by the number of atoms. Because an angstrom is one-tenth of a nanometer — that is, if you look at it on the angstrom, atomic scale, it has already reached the atomic scale. Therefore we know Moore’s law must have an end one day. And now, within the dimensions of a transistor, the smallest feature may already be at the angstrom level — at that level you’re counting in numbers of atoms, right? Going forward we might as well measure this thing in how many atoms, because in the end you can’t shrink below a single atom, can you? It becomes a matter of there being one or none — that is where its ultimate endpoint lies. But I think for now we should still hope Moore’s law can continue. Its economics, however, will certainly get worse — it will become more and more expensive.

And a second point, touching on what I described just now: if everyone has grown used to measuring how advanced a transistor or a semiconductor is with one particular yardstick, why not switch yardsticks? Ask instead whether a thing’s operation time can get faster — switch from a spatial scale to a time scale. Because what we actually need is the time scale: I need to do more operations per second. And then, can we switch to an energy scale: each time I flip a transistor, how much energy do I consume — at the picojoule level? You’ll find that when you switch scales, I think an essential difference emerges. For instance, when you evaluate a person — is it by their looks, or by how smart they are, or by whether they score well in math or in Chinese? Change the subject, and the same person may get a different exam score, right? So which subject you test becomes very important. This Tau Scaling Law (Tau定律) we’re talking about — Huawei’s time-scale counterpart to Moore’s law, unveiled in late May 2026 — tests the time scale: it says I just want the circuit to be faster. We could also switch to another one — Carnot-something, Joule’s law — saying what I test is: doing the same task while consuming less energy.

Zhang Xiaojun: A different exam question.

Liao Heng: A different question — if we change the question, perhaps the answer changes too, doesn’t it?

Actually, although these answers differ, they have strong factors linking them to one another. Any engineering guy has basically studied basic electrical principles: what governs the switching speed of a circuit is capacitance times resistance. Under Moore’s law there are two factors at play. Resistance is certain to get worse: the thinner a wire, the higher its resistance must be; and as resistance rises, the circuit slows down. The second factor, capacitance: the closer two plates are to each other, the larger the capacitance — so in that respect it gets worse; but as things shrink — as the plates themselves get smaller — the capacitance gets smaller. I can tell you why, past 16 nanometers or 7 nanometers, there has been little progress in speed: the essential reason is that resistance got worse, and capacitance did not improve significantly either. One factor makes the plates smaller, so capacitance should fall; but they also move closer together, which makes it rise again — the two factors cancel each other out. In fact, all semiconductor progress has come from capacitance shrinking, while resistance keeps growing — that is the essence of Moore’s law.

So we see that as it scales down, it doesn’t get faster — mainly because capacitance only shrinks slightly; with two counterbalancing factors, there’s no significant speedup. But because the whole thing has gotten smaller, fewer electrons have to be released, so there is still a gain in energy efficiency — just no gain in speed. As for cost: the cost of fabricating something this minuscule has certainly gone up. Those are the difficulties Moore’s law faces, as I described. But these difficulties don’t really matter, in a sense, because the whole world is counting on it — the scale of this industry is enormous — so everyone should keep pushing. Still, as I said just now: if we change the exam question, might the answer be different?

Let me give you a small example. If we want to reduce resistance, I should make the wire thicker, right? But if I want to reduce capacitance, I should minimize the overlapping area of the two electrodes. Now think about it: if two wires run parallel, their overlap area is large; if I make them orthogonal, the overlap is only that crossing point where they intersect, correct? So a structural change brings a gain. Plain scaling shrinks the original pattern: if I refuse to change the structure, that’s one answer; but if I say I can’t shrink any further and instead I turn two parallel wires into perpendicular ones — won’t the speed get much faster? The answer is: it certainly will.

This reminds me — you may have put a question to me before, saying one should ask a “cold question,” one whose answer is not so obvious. So let me give you this cold question — it actually speaks to the difference between Moore’s law and the Tau Scaling Law (Tau定律). Take a bold guess: in human history, how many patents have there been for mousetraps? Just guess.

Zhang Xiaojun: Three hundred.

Liao Heng: I don’t have the answer either — you could ask Doubao, or go look it up yourself on a patent website. I believe around 1900, the head of the US Patent Office submitted a proposal to Congress — wrote a letter — saying the US Patent Office should be abolished: the organization no longer needed to exist, because humanity was so clever that everything in the world that needed inventing had already been invented. One of the examples was that there were perhaps a thousand kinds of mousetraps — a thousand patents, of every variety, for exterminating that pest, the mouse — since humans basically all hate mice. So possibly by 1900 the US Patent Office already held over a thousand patents on how to kill mice.

Why is this an interesting question? Because back in 2020, I was making a serious effort to learn how lithography tools are built, and I found a textbook on electromechanical systems design. It left a very deep impression on me, because on the very first page of that textbook there was a diagram, and the diagram illustrated ten ways to kill a mouse. I can’t lay my hands on the book right now, but let me try to describe it for you. If you’re a chemical engineer, you’d say: to kill a mouse, I’ll concoct a poison, and then work through something the mouse loves to eat. Of course, killing mice always involves a lure — you need something the mouse loves, either cheese or some piece of meat, right? Put the poison on it, the mouse eats it and dies, correct? If I’m a mechanical engineer, I’ll build a trap — or a hundred kinds of traps: when the mouse comes for the bait, it trips the mechanism, and then I either snap it dead or shut it in a cage it can’t escape — in short, all sorts of methods, right? And if I’m an electrical engineer, I’ll rig up a high voltage: the moment the mouse passes my spot, it triggers a thousand volts and is killed instantly.

What I actually want to say is this: whether it’s Moore’s law or the Tau Scaling Law, it is simply human engineers solving problems with an Edison-style mindset. First: what is the problem? Second: have I found an effective way to solve it — to face the problem and find a good answer? Of course, if you’re making a product you also need economics — the solution has to be cheap enough. And fourth, the solution has to be repeatable, right? You can’t have quality failures — where this one unit solves the problem but the next one you build is unreliable, correct? All these factors stack on top of one another. And this is the big-premise issue I mentioned earlier: sitting here, if we start thinking about how I might reproduce TSMC’s or Nvidia’s capability — first, I don’t need to; second, however hard I think about it, I don’t have it in me to do so — my preconditions are simply different from theirs.

Zhang Xiaojun: You don’t have that context.

Liao Heng: Right. So a lot of what we’ve been saying has actually already drifted toward a metaphysical, or even philosophical, perspective.

Zhang Xiaojun: Just now we reviewed the chip history you lived through. For the second part, I’d like you to — imagine you’re a tour guide, taking us on a horizontal survey of this whole industry, the whole supply chain. Because you told me it’s an 18-layer pagoda — you don’t accept Jensen Huang’s five-layer cake; you think there are 18 layers. Take us on a tour of these 18 layers — a horizontal overview.

The 18-Layer Pagoda

Liao Heng: As we were describing — well, perhaps because our cultural backgrounds differ: Jensen Huang probably often eats French desserts, so he uses a cake as his analogy. As a Chinese person, when I picture a layered structure, the first thing that comes to mind is something like our Yingxian Wooden Pagoda (应县木塔) — right, the pagoda. But in substance it’s the same; we’re talking about the same thing, aren’t we?

That is to say — setting the pagoda itself aside for a moment — everyone has surely heard of the concept of co-design, so-called collaborative optimization: software-hardware co-optimization, or co-optimization between algorithms and chip infrastructure. The moment you discuss co-optimization, you’re already involving two layers, right? It means the classmates on the seventh floor and the classmates on the sixth floor know each other, know what each other’s difficulties are, work together, and make mutual concessions and accommodations: “Hey, your memory bandwidth is short here, so I’ll find a way in the algorithm to save some bandwidth; you have compute to spare, so I’ll spend — waste — compute to save bandwidth,” right?

Now let me point out a very intricate, very microscopic level of this. If you look at the Blackwell chip and today’s Ascend (昇腾), you’ll find a difference between us that perhaps nobody has ever described at this fine a grain. Let’s talk about Vector compute versus Cube compute. The Cube — they call it the Tensor Core, we call it the Cube — is the 3D matrix-computation unit we described: we name it Cube, they name it Tensor Core. Say their ratio is 32 to 1, and ours is 8 to 1. What does that represent? What important implications does it carry? I think the gap is perhaps a bit like this: imagine a family of four. When young, maybe they can only afford a 100-square-meter apartment. Later your career goes well, you’re a bit older, you’ve accumulated some wealth, and you buy a 200-square-meter luxury flat. Maybe five years later you get promoted again and buy a 400-square-meter villa. But aren’t you still the same family of four living there? You’ll find that when you live in the 400-square-meter villa, you probably pile up a lot of junk — say, delivery boxes, many never even opened, heaped in some corner nobody visits, right?

Zhang Xiaojun: Lots of redundancy.

Liao Heng: Lots of redundancy. What we’re getting at is: if you were suddenly told to move back into the 100-square-meter apartment, you could actually still live your life.

So 8-to-1 versus 32-to-1 means that when you go to use this compute, your model has to reckon with it: if my space is plentiful, I can waste it freely; if my space is scarce, I must find every way to use it fully — either build something “economical and practical,” or build storage shelving and clear out everything unneeded as much as possible. And you’ll find, if you look at the DeepSeek model, that they made some very deliberate, very conscious choices — I think their sense of direction is extremely clear. For instance, before 2025 they had already realized: I absolutely must economize on compute. Because of the scaling law, everyone believes more parameters bring stronger capability — but does every single parameter really have to be brute-force computed for the model to obtain that capability? Can I compute only one thirty-second of them? I think DeepSeek’s V1 — or V3, and R1 — amply proved the point: you don’t need to. That is sparse activation of parameters: if I just cleverly select one thirty-second of the parameters to participate in the computation, don’t I save thirty-two-fold on compute? Memory access is also cut thirty-two-fold. When people consciously confront this problem, they gain a 32x speedup.

Then of course, in 2025 to 2026, they did another thing with a very clear direction and goal: what about long sequences? As sequences get longer — a 4K sequence versus a 1M sequence differs by maybe 256x — must you pay a 256x computational cost to handle the long sequence? In fact, no. They did compression, sparsification: out of that 256-fold of tokens, I don’t need to compute every one; I select 1K, or 512, of them to compute, and I save that compute. But notice: such a selection process exacts a price in complexity. So you can see this model doing co-design very meticulously and very consciously.

Zhang Xiaojun: Right.

Liao Heng: They consciously said: we have only the 100-square-meter apartment, we still have to house a family of four, we still have to raise our child — and we want our child to be outstanding, no worse than Musk’s kids. But that requires pouring in more effort and taking on greater complexity.

In fact, that complexity comes precisely from the vector computation I mentioned. So from this angle we can say DeepSeek is extremely well matched to 8-to-1 compute: in order to pick out the items to send off for brute-force computation, it spends more vector computation on the selection. Whereas I believe Anthropic’s models — or OpenAI’s models — don’t need to do this, because the machines they buy come with 32x Cube compute. So let me put it another way: our current chip happens to be one-quarter of Blackwell’s spec, yet when running DeepSeek it seems to do fine.

Zhang Xiaojun: That’s co-design.

Liao Heng: That’s co-design — that is spanning two floors of the building: the algorithm folks realized where their model’s advantage ought to be built, right — that’s co-design. And this co-design has actually meant that with this kind of model, if I run inference on a DeepSeek V4 Pro, compared with the Claude — what is it, 4.8, or Fable — that I use every day, I don’t feel its token output is any slower, right? Of course we still need to respect what others do well — catch up where we should catch up, surpass where we should surpass, no? But — I can—

Whether we are the ones working on chips or on models, if you have this layer of consciousness, this understanding of the big premise, the direction of our efforts will differ enormously. I clearly have to give Liang Wenfeng (DeepSeek’s founder) a warm round of applause, because he consciously and proactively chose higher complexity — algorithms that are harder to make converge. With algorithms this complex, convergence can be difficult and all kinds of anomalies appear, and he worked hard to overcome those problems, because he had advance awareness. It is not that he waited until five years later, suddenly at a dead end with no compute at all — he had the foresight to choose to solve, ahead of time, a problem he anticipated. That is the value of someone with foresight: in effect he led the way, and others will keep pushing along that path.

Before 2019, we too were living in a happy world. HiSilicon (海思) back then was very hidden, very low-profile, but our scale was already very large — perhaps among the top one or two in the world. In terms of wafer purchase volume and product variety, we were already at least a top-3-scale semiconductor company in the world — it is just that all of its products served only Huawei’s own products. But what I mean by the “happy times” is that everything you procured was the world’s best, because you could afford to pay. For example with wafers, we would use TSMC’s most, most advanced wafers — even processes that nobody else in the world, including companies like Nvidia, would use yet. We were the first to use them, and if not first then second, because the position our products occupied could support that kind of spending, right? But once all of that was shattered, we had no choice but to face much harder things.

The second thing I have learned over these past years is that this, too, can be done. When you are pushed into a corner, you simply have to learn what is going on “downstairs”: how exactly to enable 7nm and 5nm process nodes, what efforts are required. I think perhaps the most valuable point is this: a problem that looks impossible, once you decompose it, becomes ten or a hundred more concrete problems; then you decompose those hundred problems one more layer, and it may become a thousand problems of physics, chemistry, and mathematics. This decomposition takes the macro question — say, today you ask: can you be as good as TSMC? The simple answer is no. But that answer does not help you, so what do you do? You start decomposing one layer down, and you find: if these particular places do not work, then which places do work, right? So an unsolvable problem becomes concrete, becomes tangible. Once it is tangible, out of 100 problems you can maybe solve 80, and you find you are apparently not doing so badly, right? Then you improve further on those eighty, and you find that on ten of them you are doing better than others — maybe you can play to your strengths and use certain advantages to compensate for your weaknesses.

Zhang Xiaojun: So what are the eighteen layers, exactly?

Liao Heng: The eighteen layers — first, the very top layer is applications: for example Doubao (豆包), or WeChat, or Ali’s Alipay. These are the application layer — the tip of the pyramid.

Zhang Xiaojun: The tip of the pyramid.

Liao Heng: Because it can monetize directly, and while it certainly has its technology, it is not necessarily an Einstein-level technical challenge, right — it is more about operations or business design. The bottom-most layer is very basic stuff: ores, mining, going to Nigeria to dig a mine — the bottom layers are essentially physics, chemistry, mines and the like, because a semiconductor is in essence a pile of sand plus certain rare metals. A bit further up is device design — say, whether you use a FinFET transistor or a GAA transistor, or what I mentioned earlier: if I can change the transistor design from two parallel lines to two perpendicular lines, does my speed not suddenly become ten times faster, because my capacitance shrinks a great deal? A bit further up again: how do you manufacture trillions of transistors on a single wafer, manufacture them reliably, so the chip can still work for five or ten years without failure? That is extremely hard.

But fortunately, the Chinese nation also has a great many very capable people — some Taiwanese, some mainlanders — and as a whole they have mastered these capabilities, right? So when they run into difficulty — constrained by machines: you do not have the most advanced lithography tool, you do not have the hundreds or even thousands of kinds of most advanced equipment — there are also countless people working hard to build that equipment, overcoming one obstacle after another. What I want to say is, this already brings us to roughly the fifth and sixth floors, right? The process node: do you have a reliable 5nm process? If you cannot do 3nm, what then? If you cannot shrink by another level, can you stack? What machines does stacking require? How do you stack these things reliably without creating heat-dissipation problems and power-delivery problems? These are the fifth- and sixth-floor problems downstairs.

Coming back to the seventh floor — that is roughly the level where chips sit: how best to exploit what is below, facing your own physical constraints, right? Because as we said, if our basic premise differs from others’, our process node is also different — where others need no stacking, I need stacking — you must face those physical constraints. And looking upward, we need to design the next chip’s architecture four years in advance so that software is relatively easy to program, right? So above the chip sit compilers, programming languages, the most sensible ways to partition for parallelism, reinforcement learning; then how to do KV cache, how to do model compression, how to do sparsification, how to invent a new data format so that algorithm people can use lower precision without harming model performance or preventing convergence. From the algorithm level upward, everyone is probably fairly familiar, right? Once you have a good slow-thinking model, or a model with agentic capability — how do you build a KV cache system, how do you build an agent framework, and finally construct a valuable application. Those layers are visible to the great majority of people. So above the seventh floor are the events everyone can see; below the seventh floor is the basement, where the crowd that ever looks is rather smaller.

Zhang Xiaojun: Which of these layers are the ones you know well?

Liao Heng: As I was saying — since I do not do any single layer well enough, basically all I can do is “muddle about”: roam between these layers, helping the colleagues who are more specialized at each layer to bridge the gap. Because it is very hard to be both deep and broad. It is like digging a hole: you can be a very sharp needle — stuck into a watermelon, you can pierce right through it; or you can be a knife and cut the watermelon in half. You cannot really be both a needle and a knife — you cannot be both sharp and broad. But I think this is exactly the point: we are actually not particularly short of experts at each layer — not just at Huawei, but looking across the whole world — because each layer is a profession, and once people have worked in a given role long enough, their understanding of it naturally keeps growing and they accumulate experience. What is relatively scarce is the ability to cross layers. As I said just now, maybe there are a hundred thousand AI algorithm researchers in the world, and maybe only Liang Wenfeng alone recognized deeply that he needed to break through that sparsification problem — to trade greater complexity for less compute. You see, that is a choice of direction; many people would not choose it. So this crossing of layers is the core of co-design. When a person can span those levels, they are like a thread stringing many pearls together, forming a necklace.

Zhang Xiaojun: Laying these 18 layers out, where does China have relative strengths, and where is it relatively weak?

Liao Heng: I think the strengths are obvious. First, the application layer is a clear strength — things like Alipay and WeChat Pay are just not as convenient anywhere else, right? These days, basically — I probably have not seen renminbi cash for several years now; I have not touched money, I have not touched cash in years. But if you travel anywhere else, you find you either still have to carry a credit card — otherwise you may not even be able to check in at your hotel — or you still have to carry some cash, worried that in Japan, say, you will not manage to buy a train ticket, right? So at the application layer we have a great advantage, and Chinese people are already very digitized, especially among consumers. And all the way down until you reach the chip, I think China’s strengths are at least enough to be no worse than anyone else’s. In algorithms it is even more a sky full of stars — I think perhaps 70% of the world’s outstanding algorithm people are Chinese. Why? Because a while ago I saw a Berkeley professor — a science-and-engineering professor at Berkeley — who wrote a piece saying that in his classes he now has to teach students the distributive law of multiplication: A times (B plus C) equals A times B plus A times C. So the gap is enormous. Chinese parents all take their children’s education very seriously; we have the imperial-examination (科举) tradition, the idea that “he who excels in learning becomes an official” (学而优则仕); and the upward channel for our people still exists — through your own effort, test into a good university, learn more, and you still have a chance, right? Therefore China’s talent supply is not lacking — indeed it is a far, overwhelming advantage. And of course our ability to monetize is not lacking either, because Chinese people pursue the good life — everyone wants to live better, earn more money, or achieve greater success, whatever the measure of success may be.

Zhang Xiaojun: Chips seem to sit right in the middle of this 18-layer pagoda — below it is the basement, above it the building; it occupies a middle zone.

Liao Heng: Yes.

Zhang Xiaojun: A connecting point.

Liao Heng: Yes.

Zhang Xiaojun: You came back to China in 2016 — between then and 2019, did your feeling about the whole chip industry also change a great deal, even though it was only three years?

Liao Heng: I think in 2016, I did not believe.

Zhang Xiaojun: Did not believe what?

Liao Heng: I did not believe that we, China, had this capability.

Zhang Xiaojun: Then why did you come back—

Liao Heng: I came back because of my wife. My wife said firmly that once we had a child, we absolutely must not let him grow up as an ABC (American-born Chinese), without an identity of his own, or trying to find his personal sense of self amid a kind of contradiction.

Zhang Xiaojun: How old was your child then?

Liao Heng: I came back as early as 2008, because my child was three.

Zhang Xiaojun: You came back in 2008—

Liao Heng: In 2008, although I was still employed at the same company—

Zhang Xiaojun: But you were physically back in China.

Liao Heng: I was already physically in China. Because I think my wife came to this realization earlier than I did; she said even more firmly: do not stay abroad, we must start over in China, especially considering the next generation. She was utterly determined to return to China, so I did too — at that time returning to China was a passive choice.

But by 2016, when I joined HiSilicon — I think perhaps the bigger shift in what you would call professional perspective came with that move. Before 2016, I completely did not believe. That is — please note, foreigners have a kind of arrogance, a blind self-confidence. The first element of this blind confidence is believing they are the best. Of course, once you have worked in that environment long enough, you become infected with this mistaken perception too — believing that a certain thing, only they can do. And “they” included myself — I was one of them, right? Although Chinese, having joined such an environment, I was part of that team — so you come to believe that for the harder things, only “we can, nobody else can.” But as for how things in this world actually work — as I said just now: take a very hard problem, and when you decompose it into 100 sub-problems, you find every one of them becomes more concrete and solvable. If someone is willing to solve those 100 problems, perhaps they will solve them better than you.

That is exactly the transformation I underwent starting in 2016. My first shift was this: things I had believed completely impossible for China’s industry at the time — in fact, our capability far exceeded my expectations. And I find this very interesting: I believe the great majority of Americans, American practitioners, still hold my 2016-era view. That gives us a huge opportunity. Because if you regard them as a competitor, and your opponent severely underestimates your capability, you hold an enormous advantage. They innately believe they can compete with you while living comfortably — clocking off at five every day, having a coffee, then going surfing. In reality, at that point, if you just push one notch harder, you can easily overtake them. That is my biggest realization: whether in 2016 or for many years before it, our Chinese counterparts — the effort they put in, their understanding of the problems — so long as their identification of the problem is at the same standard as everyone else’s, their ability to solve it far exceeds that of practitioners abroad, because they work harder. Put simply: you reap what you sow. But the precondition is that you need the right people to lead them to the right definition of the problem. Because if the problem is defined wrongly, all the effort is wasted.

Just now I gave the example of Liang Wenfeng. He defined two problems. First: model parameters must be sparsified — and he solved it. Second: sequence length — attention must be sparsified — and he solved that quite well too, right? You see, the definition of the problem itself — if we are allocating credit — already takes 80% of the credit; actually finding a concrete method to solve the problem once defined perhaps takes twenty percent. It requires talent, requires intelligence — maybe that intelligence comes from an intern — but the definition of the problem comes from the team’s leader, because the leader must steer how others should go about solving problems.

The History of Ascend and China’s Path

Zhang Xiaojun: Now we come to the very heart of this interview — I would really like to hear you tell the history of Ascend (昇腾), because these past few years your development has also been very low-profile. If you used three words to describe the past decade, 2016 to 2026, which words come to mind?

Liao Heng: I think it has just been a tribulation.

Zhang Xiaojun: A tribulation.

Liao Heng: Yes — or rather, a tribulation with hope in it.

Zhang Xiaojun: From the 910 to the 950, the environment you faced was actually very different. Standing at those design points, do you think the problems you defined for the 910 and the 950 were different?

Liao Heng: Yes.

Zhang Xiaojun: What were they, respectively?

Liao Heng: At the time of the 910, the basis we worked from was possessing the world’s most, most advanced logic process. But at that time, regarding the AI ecosystem, the development of models, and what their applications would actually be, we were fairly lost. In fact the whole world was somewhat lost then, right? Although I believe Nvidia was probably the earliest to sense that this area would explode, I believe that worldwide at that time — because model capabilities were not there yet, and no enormous scale had emerged — the confusion was more about how to guess which direction the future would take.

At that time we had already realized we needed to do model training; the inference market simply did not exist at all, right? So on edge devices — the earbuds and phones I mentioned earlier — we tried to find a path to monetization on the device side. That is, we thought data centers were for training, and elsewhere, trained models would be deployed onto all kinds of small devices. We did not foresee this kind of large language model inference — including things like AI coding; none of it appeared in our imagination. So at that time, it was merely that the conditions were very good, but we did not know where our own biggest future monetization opportunity lay.

[2:00:00]

Then by the 950, things had reversed, right? The conditions were extremely harsh: the opponent you faced was already Silicon Valley’s — the universe’s — biggest company, the one with the highest market capitalization and the strongest capabilities. So at this point, how could we survive, and then how could we serve — perhaps we harbored no over-lofty ambition of seizing the world’s number one position; it was simply this: step one, stay alive; step one, serve the customers who need us — satisfy that basic demand of being “customer-centric”.

Zhang Xiaojun: And what would have happened if you simply had not done it?

Liao Heng: We could also have not done it. What I want to say is that in reality, whoever this world loses, the Earth keeps turning just the same, right? It is only that, on the one hand, we still felt there was something to hope for — OK, we did not want to “lie flat”, so we still wanted to make an effort; and on the other hand, many industry peers probably had expectations of us, so we felt we did not want to give up so easily.

Zhang Xiaojun: What was the change in your state of mind? The difference when designing the 910 versus the 950 — the two generations of chips before and after the supply cutoff — that change in state of mind.

Liao Heng: I think putting it into words may sound a bit — what I want to say is: try to imagine, if you were Dong Cunrui (董存瑞) — the PLA war hero who died holding a demolition charge in place — about to hold up that satchel charge, right — we ordinary people cannot imagine what his state of mind was. Or a soldier at Shanggan Ridge (上甘岭) — what was his state of mind? Or — I once saw an interview from the Second World War with a Chinese soldier, also a story that appeared on one of Huawei’s promotional posters. A reporter asked him: what do you want once the war is over? He said: I don’t think about it. Because my parents are dead too... there is no point thinking about those things. So I would say the so-called change in state of mind no longer carries much meaning. It is simply: faced with a difficulty, you just want to solve it.

Zhang Xiaojun: You said that when you first joined Huawei in 2016 you did not believe either. When did you go from not believing to believing?

Liao Heng: That actually happened quite quickly. Because my disbelief was based on a foreigner’s arrogance — blind conceit — the belief that the capabilities they possessed could not possibly be possessed by anyone else. Reality educated me very quickly. After I joined HiSilicon (海思), I very rapidly came to understand that the colleagues around me, although different from me — for example, HiSilicon might newly tape out perhaps 100 chips a year. In any case, our President He (He Tingbo of HiSilicon) said at the Tau Scaling Law (Tau定律) launch event that over five years we had done 381 chips — that averages about 70 a year, right, 70 new tape-outs per year. But I have never once heard of a chip coming back “smoking” — that is, failing test and in the end proving impossible to mass-produce.

What does that show? Such a high hit rate shows that our colleagues — every single team — are highly qualified. Regardless of their origins, or whether their résumés look glamorous, in the domains they are responsible for they are all extremely responsible. As I said earlier about being an engineer — this is not about myself — it is something I did not grasp as a student, and grasped after a year of working: to be an engineer, you must first of all be dependable, and you must be responsible. That is the foundation of everything. And therefore, a hit rate like that would rank among the best in the world — and the outcome gave me a rapid education and transformation.

Zhang Xiaojun: I believe that no matter from what period you look back on Huawei’s history, the supply cutoff (断供) will count as a hugely significant event. When it happened, what was it like in your department? What was the reaction in that first moment?

Liao Heng: I think President He’s letter must still have been written before the cutoff. (Her internal letter at the time of the US action became famous.) But once the cutoff truly came, there was never again a chance to write that kind of letter. As with the example I gave earlier — you cannot imagine what Dong Cunrui felt. I can only say that everyone was in a different role, and their feelings were not all the same; I have no way to describe it.

I would just say, first, there were perhaps two sides to it. On one side, you needed to reassure the colleagues below you, tell them not to be afraid — because enormous fear arose in everyone’s hearts. Second — speaking for myself, and my memory is a bit blurry now — the deepest feeling I remember is that it ignited an enormous competitive drive in me, an enormous passion: this absolutely had to be solved. First, I had to know exactly what problems we had run into; then I had to break those problems down one by one and solve them one by one.

For an engineer, this was a godsend. Because under normal circumstances you never need to understand how a motion stage achieves one-nanometer precision, or how to measure something’s position. If you want to control a stage to move with one-nanometer precision, you also have to locate it during high-speed motion, know exactly where it has moved to; then you have to work out whether to measure it with a laser or with something else, right? These things are all interlocked, and you run into a huge number of interesting technical problems of the kind engineers need to solve.

Zhang Xiaojun: Did the team atmosphere change before and after?

Liao Heng: I think the team atmosphere probably did change quite a lot, but overall it was all right. Because precisely in the most difficult period, I think maybe only 10% — 5% — of the people around us quickly went looking for a new way out. But maybe eighty or ninety percent — especially the backbone, the most capable colleagues — I believe that to varying degrees, perhaps like me, each had their own problem-solving passion ignited. So many people — the great majority — chose not to overthink it: the problems, once broken down, were fascinating, so they just went and solved them.

That describes just that one phase. If we want to put it this way: every human heart has two sides, good and bad — the side that wants to survive and get more benefits, and the side with certain visions, or greater ambitions, a greater sense of mission; it is only a question of which side gains the upper hand. It’s like Gollum in The Lord of the Rings, right — that creature who on one hand wants to seize the Ring for himself, to gain greater power, but on the other hand feels that this thing is not quite right.

Zhang Xiaojun: The most difficult period — roughly from which year to which year?

Liao Heng: I’d say it was difficult at every stage — it’s just that the difficulties each stage faced were different.

Zhang Xiaojun: And the very lowest point?

Liao Heng: The lowest point was when there was completely nothing — that is, people felt that within maybe a few months we would be completely unable to manufacture chips, and the entire commercial loop would therefore be broken. But fortunately, those difficulties were essentially carried by a few exceptionally resilient leaders, who locked them tightly within a very small circle of their own. Those people bore the great majority of the pressure from these difficulties; everyone else remained more or less in the dark.

Zhang Xiaojun: And presumably there were also some landmark events?

Liao Heng: Perhaps there was no single event. But Huawei — take China’s 5G network, which is now everywhere — throughout that process we never once broke supply, and never caused construction of the entire communications network infrastructure to stall, or its schedule to slip noticeably. Behind that were the tireless efforts of countless people that made these things possible. I can say that perhaps upwards of five thousand, maybe as many as ten thousand boards were all redone — every item had components swapped out, and all of them rapidly reached mass-production capability again. That was the result of the joint effort of an extraordinarily large number of people.

It’s just that these things have no single marker — I mean, when something does not get cut off, what does that even count as, right? It’s what everyone takes for granted in daily life: no downpour today, no earthquake today. What I really want to say is that behind this outcome were very many people’s efforts, day and night.

Zhang Xiaojun: Why is Ascend (昇腾) called Ascend? And when was the name settled?

Liao Heng: I may not remember this very clearly anymore. It should have drawn on the Chinese phrase ‘ri sheng yue heng’ (日升月恒 — ‘as the sun rises, as the moon waxes’) — a kind of beautiful hope for the future.

Zhang Xiaojun: You’ve just described the overall line of thinking in the architecture design. For each chip generation — from the 910 A, B, C through to the 950 — what were the architectural changes and evolution in each generation?

Liao Heng: With the architectural changes, I think there were first some fairly clear and important realizations. The first: as I said, during those four to five years when we were stalled, our peers working on algorithms kept pushing forward — large models exploded, and they achieved rich results. Those results gave us a great deal of grounding for our designs.

The second very important event: some models — Llama being the first — chose the open-source route. As a result, roughly from its first generation, Llama became something everyone could analyze — it was like cutting open the body of a large model, letting people truly see what a large model looks like inside, down to every microscopic detail. That gave us excellent grounding for design. Later, of course, this series of Chinese models — whether DeepSeek or Qwen (千问) — kept open-sourcing, and in capability they were also close to the first tier, serving as that kind of reference point.

Beyond that, there was a lot in relatively non-language domains — for example China’s video generation and multimodal work, including Alibaba’s Wan series, and other internet companies, who also open-sourced a great deal of valuable work in those areas. Those works, plus our own closed business loops in autonomous-driving systems and in phones themselves, gave us a lot of real performance feedback, which also kept improving the models themselves. Simply put, by the time of the 950, the reference and evaluation basis for our designs was much richer, much more plentiful.

Second: after last year’s Spring Festival, when DeepSeek’s R1 came out, it directly triggered an explosion in model inference that we had never anticipated, and its scale was already quite considerable. So we gained a much more realistic understanding of the capabilities inference requires. Because inference doesn’t just have to be functional — it also has to be fast, right, the latency has to be low; and with low latency, the number of tokens each card produces — that is, the throughput — also has to be high. Balancing those several factors to the extreme gave us a sense of reality. Because before large-scale inference systems were deployed, we simply didn’t know these requirements existed.

Very quickly, once we started serving inference customers and genuinely deploying inference systems, we ran into a huge variety of engineering problems. One of the most visible: once you enter inference-system deployment — within maybe one or two months — you discover that although we used to think compute mattered most, in an inference system, especially in the decode part, compute actually doesn’t need to be that high; instead, memory bandwidth is especially important. As for the ratio between the two — if you take compute divided by memory bandwidth — you find, like the analogy I used earlier about the coordinated proportions of the human body at various scales, that the ratio needed for inference decode is markedly different from the ratio for training, for prefill.

And of course this process also, fortunately, confirmed one thing: what we call the supernode (超节点) — combining a very large number of chips into one tightly coupled, very closely cooperating whole — that capability is an important reason our 910C, and the 950, have been able to survive. Even though beforehand we had only a hazy vision of combining more things together, it was validated in actual business deployment. And of course we also discovered the places where we hadn’t done well.

Zhang Xiaojun: Which were?

Liao Heng: Namely: when you do communication — say, moving one megabyte of data at a time versus moving 7 kilobytes — the difference in granularity is an utterly fundamental issue. And the granularity of what we built in the past was all on the large side. What do I mean by granularity? If you fill a bucket or a wooden box with a pile of cobblestones, there are lots of gaps in between, because the stones’ granularity is large. If you fill the same box with sand, you find it packs much fuller, because each grain of sand is small. If you pour in a bucket of water — as long as the box doesn’t leak — it fills even fuller: that’s what I mean, because water is individual molecules, right, finer than any grain of sand.

So, for example, we originally thought this tightly coupled network would be used to move big blocks of data; later I found it was used to move very small blocks — we thought one megabyte, and it turned out to be 7 kilobytes. That’s a big difference, right? If the design isn’t good, you may be very efficient moving one megabyte of data but inefficient moving 7 kilobytes. So all these places need continuous polishing in practice.

So I can say why the 950 is significantly better than the 910C: first, that one is simply too old; second, that one was essentially designed by blind guessing, whereas by the time we designed the 950 there were a great many real-world things against which to test whether our guesses were right — and wherever they were wrong, we could quickly polish, optimize, and adjust. Moreover, the products have differentiated: as I said, the chips needed for inference and the chips needed for training require different ratios, and so may need different types of memory.

Of course, the last half year has brought an even deeper lesson. The lesson is — well, in fact we realized seven years ago that memory is especially important: rather than saying we sell an AI compute chip, you could better say we sell high-bandwidth memory. That realization directly drove us, seven years ago, to set up a dedicated department to build design capability for high-bandwidth memory. Yet even with seven years of preparation, when the storm arrived we found our preparation still wasn’t enough. In this current memory wave, you’ve seen prices rise more than tenfold, delivering an enormous shock to the entire electronics industry. What I want to say is: even with that much advance understanding and preparation, on the day the storm came — the day the tsunami came — there were still many regrets: OK, why didn’t we do more back then, put in more painstaking work, or prepare better in some way. Before people truly face the storm, they inevitably harbor some wishful thinking.

Zhang Xiaojun: Between chips needed for training and chips needed for inference, which do you think will be more numerous in the long run? What will the ratio look like?

Liao Heng: I think it should already be true today — this has already happened: inference will certainly be more. If inference isn’t the larger share, it means your economics are bad. Because inference is a revenue-producing process, while training is an investment in building capability: one is pure expense — a cost center — and the other is a profit center. And the profit center must be bigger than the cost center.

Zhang Xiaojun: In which chip generation will your thinking about these inference forms be reflected?

Liao Heng: I think it is strongly reflected in this very generation. For example, we split it into two tiers: one tier with very high bandwidth, and one tier with relatively lower bandwidth but better economics. For training, say, you might be able to use the more economical one.

Zhang Xiaojun: At the 2021 STW — the System Technology Workshop — you proposed that computing architecture had a new trend: moving from the old heterogeneous computing architecture, with CPU, GPU and NPU separated, toward a new type of homogeneous computing architecture. Could you explain what heterogeneous and homogeneous mean, and what distinguishes this new homogeneity from traditional homogeneity?

Liao Heng: To some degree I’d say that was something of a gimmick. A gimmick meant to push back against certain voices of doubt.

Zhang Xiaojun: Internal doubts, or external ones?

Liao Heng: There were doubting voices both inside and outside. First — hold on, this topic goes fairly deep. The first thing I want to say is that the most common argument in the industry is the SIMD-versus-SIMT fight. Some people say the GPU is SIMT — Single Instruction, Multiple Threads — that’s the GPU architecture; while Ascend currently — in the 910B and 910C — is a SIMD architecture. What does that mean? When you process an operation, do you express the computation as single elements, or do you compute on a big block of data at once? Put another way, it’s like the difference between a truck and a private car, right? If you drive yourself to work, each person takes one car; if you ride a bus, each bus carries a whole load of people — maybe dozens of people sharing one vehicle.

On the surface, this SIMD thing — what I want to say is, why is it a gimmick? In reality, today’s SIMT processors are internally SIMD as well; and today’s SIMD processors themselves also have multiple cores running almost identical code — that SPMD kind of pattern. But some people love to split hairs — or rather, people especially keen to exploit the CUDA ecosystem will feel SIMT is wonderful, because a huge amount of original research work has been developed on the CUDA ecosystem. So if your processor looks exactly like a GPU, you can download the code straight from GitHub and it runs without you doing anything — the ecosystem’s first-mover advantage.

But as I said, SIMT itself has SIMD elements inside it. For example, you’re processing an 8-bit floating-point number, but the compute unit may be 32-bit, so you have to pack four floats together to do the computation — that’s a little bus too. Nobody is fundamentalist about computing only one number at a time; if you computed only one number at a time, you’d be wasting the compute unit’s capacity. So even in code written for SIMT, people pack four 8-bit numbers together for computation. That’s an analogy.

So it comes down to nothing more than whether the bus is big or small. Undeniably, a small bus means small granularity, which helps with filling — as in my earlier example: filling a box with cobblestones versus fine sand, the sand packs fuller, right? So all we are really doing is judging whether a data block’s size should be large or small. The difference is 128 bytes versus 512 bytes: in effect, the GPU’s natural width is 128B, while our Ascend — especially the early products — had a width of 512B, which is a bit too wide. We have to admit that, to be honest and factual: our cobblestones were a bit too big. Although in the great majority of networks — especially Transformer-type networks, which are themselves very large-block, so a slightly bigger cobblestone doesn’t actually matter much — in certain networks — recommendation networks, or some early vision networks, especially on-device ones — smaller granularity may be needed.

In the 950 generation, what we call the new homogeneity means the processor supports both SIMD and SIMT, so the problem is greatly eased. Rather than carrying on this religious-style argument, better to use both — or to use different modes of computation in different situations. Because apart from the most essential point — the granularity question I just described — by today perhaps 90% of this issue can be set aside, since the technology as a whole has evolved to this stage; arguing about it further is no longer a core, critical question.

What matters instead is what I said just now: how much memory bandwidth to provision; how the interconnect should be built into a supernode; how to achieve extreme fusion — mega kernels — in kernel computation; how to reduce the synchronization overhead among multiple cores, even multiple chips. Those are the questions of the present — we consider the earlier debates to be already-solved problems, but these are the more important problems of today.

Zhang Xiaojun: These more important present-day problems — do you have answers now, or some interim thinking?

Liao Heng: Of course we have our own answers — otherwise the product couldn’t be carried forward. In our pipeline we have the 960, 970, and 980, all in the development pipeline, and we keep producing more answers of our own.

But let me try to tell you a more serious fact. In the past, AI algorithm developers were all used to PyTorch: to describe an algorithm, I use a pile of Torch operators — each operator is a function, a basic computational function — and stitching these operators together represents the whole model’s computation. But the more important present-day problem I mean is this: suppose you want to develop a modern model that way and also achieve, say, one-millisecond inference latency. One millisecond means: when conversing with the model — whether it’s a human or an agent talking to the model — submitting a prompt and getting output, one millisecond implies outputting 1,000 tokens per second. Today our typical chatbots — the great majority — output 20 or 30 tokens per second. Going from 20 to 1,000 is a 50-fold time compression.

Fifty-fold time compression — or, from a latency of maybe 20 milliseconds down to 1 millisecond, call it 20-fold — right, twenty or thirty tokens means forty or fifty milliseconds down to one millisecond, so a compression of several tens of times. At that point, the more important problem is: in almost every case where you are chasing performance and latency, you can no longer make calls operator by operator — because every call carries launch overhead and latency, plus the data-transfer overhead of the launch. So the more important question becomes: when these processors write the entire model as one enormous operator, what should that style of writing look like? That style is what we call a mega kernel, or kernel fusion: fusing tens or hundreds of operations into one larger operation, to reduce the launch and data-transfer overhead.

This change has been happening rapidly over the past year or two. The most famous work is Flash Attention; and the more important work after Flash Attention is the series of operators DeepSeek has open-sourced, and so on — every one of them an outstanding exemplar of fusion. This fusion process in turn delivers a big shock to processor design. In the past, when people designed a core, they only cared about one thing: my computation runs fast, and that’s enough. Now the question is: if I want to express a very complex computational process, the programming has to be reasonably easy to write, and I have to be able to modify the program on your processor well and quickly, so that it achieves maximum performance. This has given rise to some modern programming approaches. People aren’t quite — especially when the algorithms are changing very fast — people can no longer bear expressing such complex computations through low-level programming.

So naturally, new programming languages and compiler technologies emerge. These include, as mentioned earlier, Triton, which is a piece of OpenAI’s work; then TileLang, built by Professor Yang Zhi of Peking University and his student Wang Lei, which is currently also DeepSeek’s main development approach; and PyPTO, which we built ourselves. All of these try to express extremely complex computation in a higher-level way — to the point that an entire model is written as a single operator. And that has a fairly far-reaching influence, from the software side, on compilation and on processor design.

Because if you look at the classic Hennessy and Patterson textbook, Computer Architecture — the first time I saw this sentence, I was genuinely, deeply struck. It says computer architecture is ‘the interface between software and hardware’ — it is that interface, right —

Zhang Xiaojun: The communication interface between the two.

Liao Heng: The communication interface between the two. Did most of the people who used to work on processors, on the hardware core side, really have a particularly deep appreciation of this? Because they only saw their own side, the core: I need to design a compute unit — how do I make it efficient, how do I save die area. But once you raise your understanding to the level of this interface, you can see that changes in programming languages affect this interface — it is two sides of one thing; the interface necessarily has its yin side and its yang side — and where the dividing line falls is determined not only by the core, but also by the way you program and the way you compile.

So for the design of today’s AI cores — of Ascend processors — I think we, or rather our peers, everyone who designs processors, are facing the fact that this interface is being reopened. Or to put it another way, the protective wall of the CUDA ecosystem is rapidly disappearing, because it is no longer the primary interface. If all of DeepSeek’s work is built on top of TileLang, then TileLang becomes a critically important interface.

Zhang Xiaojun: So that’s also the point of what you’re building now?

Liao Heng: Right — we ourselves of course had this understanding years ago. Despite facing a great deal of skepticism, we have also been doing our own open-source compiler work, the PyPTO line of work. This work is purely first-principles based; there isn’t much that insists on... that is to say, the methods and design philosophy we use should be universally applicable, not applicable only to Ascend. In other words, although it was we who open-sourced this work, I believe all AI processors can benefit from the same compiler stack — that is, what’s called cross-platform.

Zhang Xiaojun: CANN — what are your expectations for it?

Liao Heng: CANN is our software stack — essentially our entire suite built on top of Ascend: the runtime environment, compilers, libraries, plus the models we have already ported.

Zhang Xiaojun: Do you expect it to be the next CUDA-type product?

Liao Heng: It is already comprehensive — on one hand, it is comprehensively protecting the legacy. By ‘legacy’ I mean: whatever capabilities CUDA has, our CANN must have too — even one-to-one at every single layer. For example, you can run PyTorch, I can also run PyTorch; you can run vLLM, I must run vLLM too. But we genuinely still have a certain gap — that’s the legacy part. Then there’s the part that is advancing by leaps and bounds.

First, legacy serves the masses. What do I mean by the masses? If you’re a university student just starting to study AI — the deep neural networks course — you’ll most likely use PyTorch, use Torch operators. If the professor teaching the course uses Ascend hardware as the environment for the coursework, then you’ll need CANN. All of that belongs to legacy; it’s the long tail, and it may persist for a very long time.

But the part serving the few is this: when I want to deploy a GLM 5.2 and I want the highest economic return — the most tokens produced per card per second — and I also want the lowest latency, that’s where the ‘advanced’ part comes in, right? First we need to use more advanced methods, in the shortest possible time, so that people can quickly achieve this high-throughput, low-latency tuning. And this tuning is tuning taken to the extreme, because every additional 10% of throughput is an additional 10% of revenue, right? So it’s real money. And that advanced side may not involve ten thousand PhD students using the thing — maybe only a hundred people working day and night to crack the problem, raising the throughput of the production system as early as possible, because it is a production system. I think we are making continuous efforts on both fronts.

The reason Ascend can sell at all is, first of all, that both sides need to show progress. If the performance of the production system doesn’t go up, nobody will buy it, right? Because if you buy it, you lose money — you spend the same amount of CAPEX and produce fewer tokens; or you train a model and it won’t converge, or it takes longer — converging five times slower than others — nobody can accept that, right? And the ecosystem part I just mentioned serves the broader community. For this community, we have in the past gone fully open source and worked with countless university professors, hoping they will build this ecosystem together with us. The CANN community has now become the hottest open-source community in China, because so many people follow AI and there is a great deal of enthusiasm around it. So, riding this ‘many hands make the flames rise higher’ trend, we have made great strides over the past year.

Zhang Xiaojun: Which do you think is harder — this extreme physical pursuit at the nanometer scale, or building an ecosystem that surpasses CUDA? Which is harder?

Liao Heng: I would say surpassing CUDA is harder — even though the extreme pursuit at the nanometer scale, or making breakthroughs in physics, is also hard. Why? One of them mainly rests on your own effort. The other requires changing the shared habits of a whole group — of countless people. For example, if today you invented an instant-messaging app and tried to persuade everyone to stop using WeChat and use your new app instead — that’s hard, extremely hard, right? Because it’s a group habit, and that group habit needs to keep receiving positive reasons before it will make such a shift. Whereas developing the next, more advanced chip, as I said — as long as we ourselves put in the effort, maybe 70% of the factors are in our own hands, and perhaps 30% are objective physical constraints, and we just have to find every possible way to break through those physical constraints. But these two problems are, in essence, solved by colleagues from different disciplines.

Zhang Xiaojun: Solved by them.

Liao Heng: Right.

Zhang Xiaojun: Was open-sourcing CANN a difficult decision for you?

Liao Heng: It was not a difficult decision. Why? Well, I have my own personal view — of course everyone has their own perspective. First, communication between humans: the human brain is highly developed, but the connection capacity between people is very poor. For example, sitting here doing this interview with you, my speaking rate is maybe just three to five characters per second — and that already counts as fast, right? So the communication bandwidth I give you may be only 1 kbps. But if you were an Ascend chip and I were an Ascend chip, our communication bandwidth would be 7.2T — 7.2 TB per second. TB per second is a terrifying number — who even knows what power of two that is, right? So the orders of magnitude between the two differ enormously.

So why do I use this metaphor to describe the importance of open source? Because when you deliver a very complex technical system, what’s actually needed is for the other side — the people using the thing — to quickly understand: what are this system’s interfaces, how is it designed, how should it be used? So this communication bandwidth — when we go and restrict such a communication bandwidth, you realize that this communication was already an enormous bottleneck to begin with; and when on top of that you control when a given document may be sent to whom, and even viewing it requires signing an NDA — that inherently adds unnecessary security gates onto an already very constrained communication channel, right? So-called open source is really openness taken to the extreme. Meaning: once I have shown you all the code, I actually no longer need to communicate with you, because you can go look for yourself at the countless files in the countless directories, right — no more need to waste breath on it. So it solves the biggest bottleneck in this kind of exchange — sharing ideas and getting aligned.

Liao Heng: So we feel that from the moment we open-sourced, both on our side — the developers — and on the customers’ side, there was an immediate sense of a weight being lifted. A problem that used to be difficult — every time you had to make phone calls, file issue tickets — now that man-made bottleneck has disappeared, right, and the speed of solving problems is much faster. And that’s not even counting the many more people making their own contributions on top of it. There’s also the fact that AI-assisted programming is now becoming an important — even a dominant — way of working for us. After open-sourcing, those models and agents can much more quickly get an ample corpus, giving them training data for programming on CANN. So those models are also playing a very good role in CANN development.

Zhang Xiaojun: I feel the early Ascend story had something of a backed-into-a-corner, rising-up-in-resistance flavor to it. Coming to today, do you think Ascend and Nvidia are becoming more and more alike, or less and less alike? Have you taken the same path or a different one?

Liao Heng: I would say that if there were more similarities early on, part of it is that they have become more and more like us, right — the Cube I mentioned earlier: their tensor core has come to look more and more like our Cube, getting bigger —

Zhang Xiaojun: Bigger.

Liao Heng: — to achieve higher compute efficiency. But apart from such details, at the macro level things were fairly similar back then, because everyone was centered on the single chip, and deployment approaches and so on were all fairly alike. The bigger dissimilarity, though — I think we are now becoming less and less alike. The most fundamental reason lies in what I described earlier: the difference of small versus large, the difference in granularity, the difference of using the small to fight the large, the difference of the supernode — including various system-level differences — less and less alike. At the micro level, we may still have plenty to learn from them — for example, as I said, realizing our granularity had become too coarse and should be made smaller, roughly their size, so that adaptation efficiency across more algorithms is higher. But at the system level, we are less and less alike.

Let me offer one small angle to try to illustrate this. At least — this is our bold speculation, or judging from the roadmap they announced at GTC — I think Nvidia has been pursuing extreme increases in rack-level density. The increases are astonishing: almost every generation is a doubling — twice the compute stuffed into the same rack. Personally, I think there has to be a limit; you cannot push it to the extreme, and you should not increase rack density without bound. Why? There are several dimensions to this. First, I think the specs of a single chip should not be pushed to the extreme — or simply cannot be pushed further. The single chip — OK, the single package. Why shouldn’t a single rack push its specs to the extreme? It does need to improve, but it should not exceed the limits of physics. Because once you exceed the physical limits, you run into catastrophic consequences.

Let me start with the single chip. Everyone has seen that one of the biggest predicaments the AI industry has created is the HBM shortage — because demand has exploded, everyone is using bigger, more numerous, higher-bandwidth HBM. Let me try to explain, with a fairly vivid example, why this kind of 2.5D scaling has a ceiling. Take the compute die — there are actually multiple dies these days, but let’s simplify the problem — suppose I treat the compute die as a square with side length N; then its area is N squared, right? A square with side N — isn’t its area N squared? Then its compute power is N squared — because each compute unit requires some number of logic gates, and if you have N-squared gates, compute power is proportional to N squared. But when you have N-squared compute power, the memory bandwidth and interconnect bandwidth you consume scale in the same proportion.

But if I place one huge compute die in the middle and keep making N bigger, while my HBM, all my IO, and all my power delivery are arranged around its perimeter — then doesn’t every single bit of memory access have to cross the edges of this square’s four sides? The perimeter of the four sides is 4N. So you find that as N grows bigger and bigger, your compute power, your bandwidth, your interconnect, your power consumption all scale as N squared — but your four sides only give you 4N. Between that quadratic curve and a straight line, won’t there be an ever-widening gap? That opening — the contradiction between the two — will explode more and more, right? So this shows that the current path of relying on ever-bigger compute dies plus ever-more HBM has a ceiling. Of course, before we reach that physical boundary, we will keep making it bigger; but once you cross that boundary, you will see catastrophic consequences — an avalanche.

Zhang Xiaojun: How big might that boundary be?

Liao Heng: The boundary might be four reticles, each reticle being 800 square millimeters; maybe six reticles, maybe eight. I trust our whole upstream — and this gets into the lower levels, the manufacturing chain I described earlier, the colleagues on the fifth, fourth, and third floors, the people in the basement — to see how far they can extend that capability boundary. But I already know with certainty that this physical gap between N squared and N is irreconcilable. So this kind of architecture can perhaps run for one or two more generations, maybe three, but most likely by the fourth generation it will no longer work.

Zhang Xiaojun: Because they are still building along this architecture?

Liao Heng: Currently, yes — the roadmap they have published so far is like that. So I don’t think a single chip can keep doubling every year. If you double every year, you make this gap ever more contradictory, ever bigger, right?

The second question is: why can’t a single rack be scaled up without limit — and why is it also unnecessary? Because if a rack’s power draw is 100 kilowatts today, 200 kilowatts next year, 400 kilowatts the year after, and the year after that — reaching a megawatt might take only four years — two to the fourth power, already many times over. Now why shouldn’t you push it that high? Because as you stuff in more and more chips, they have countless interconnects — everything — there are chip-to-chip links, and they need bandwidth. The more you stuff in — it’s not just the chips’ compute that doubles: their interconnect doubles too, their power draw doubles, their heat dissipation doubles. Let’s look at it in mundane, practical terms: for a one-megawatt rack, if it used natural ambient-air cooling, we can do a back-of-the-envelope calculation of how much cooling-tower space is needed. Say it takes 500 square meters to dissipate one megawatt — then you find that a gigawatt data center might contain only 1,000 racks, yet it might need a full square kilometer of land just to house the cooling towers.

Between those two — a rack’s footprint may be under about two square meters, but if it needs 500 square meters of space for cooling, that’s a 250-fold spatial amplification. When that ratio becomes seriously out of proportion, you find that the water pipe may have to run a kilometer to reach the cooling tower, because the space the cooling towers occupy is vastly larger than the space your machines occupy, right? Everything I just described actually goes far beyond the elements of processor architecture design itself — but it belongs to what I called another layer of the problem, right? In terms of the 18-layer pagoda: if you build an AI gigawatt data center, someone always has to construct the buildings, lay the water pipes, design the cooling system. The point is that this ratio must not become seriously distorted. We can only say — it’s really not our place to comment on how others do it — we can only say that in our own system design, perhaps as early as three years ago, we had already realized: I do not want to stuff too many chips into the same rack.

That then means the interconnect distance between my chips — if others stuff 100 chips into one rack while I spread them across 4 racks — the physical distance between those 100 chips becomes longer. Of course, longer physical distance brings latency and all sorts of overheads. But if I recognize that this is a fundamental problem — that one day I will have to spread things out — then I need a wiring approach whose connection cost is very low, and whose transmission distance won’t get to the point where the signal can no longer get through. So we were in fact proactive: when we designed the first-generation UB — the UnifiedBus (灵衢) I mentioned earlier — we designed it so that it can run both over electrical wiring and over optical cable. And you should understand that the thickness of an optical cable and of an electrical cable differ greatly — the diameters may differ by a factor of ten. When you’re connecting a single wire, there isn’t much difference; but when a rack has to connect five thousand wires, you see the physical reality — the size of five thousand electrical cables bundled together may already be as thick as an elephant’s leg.

I think Nvidia’s approach is rather brute-force. On one hand they keep raising speeds — pushing up the SerDes rate, the signaling rate on every single wire. Everyone is raising speeds; everyone wants to carry more signal over fewer wires. But on the other hand, they have used the most advanced board materials in the world, which has directly caused a shortage in high-end PCB capacity too; and they have also used some very advanced physical-layer technologies, right? All of this is fantastic — these are areas where they are extremely good. But I, as a system designer, do not want to take on the world’s very hardest problems in every single spot. Why? Because with every generation, I have to deliver a new, highly reliable product to customers, on time and at quality. If I have to break through 20 physical limits simultaneously, then if there is even one I fail to break through, my product is dead. I would rather put my energy into breaking through maybe five especially hard — and high-value — problems, and leave some of the other hard problems for others to solve. That way, my probability of failure — of the product dying — drops substantially, right? Because even if each one has a 99% probability of success, 99% to the 20th power is a number approaching zero; whereas 99% to the 5th power may still be tolerable. So this is a design philosophy — a philosophical question: play to your strengths and avoid your weaknesses, bring what you are good at into full play, and do not try to contend with others on every front.

Zhang Xiaojun: So on the question of whether the single chip keeps getting bigger and bigger — is your answer that...

Liao Heng: Of course we are making it bigger and bigger, but we do it within our own comfort zone.

Zhang Xiaojun: Not unboundedly big.

Liao Heng: Not unboundedly big. We will not go and take on the kind of design that is simply asking to get itself killed — we won’t do that.

Zhang Xiaojun: And single-rack density — should it be pushed to a megawatt?

Liao Heng: We don’t need it — the answer is no. We know with certainty that we don’t need it. Because our space is dirt cheap — each square meter of a machine hall may cost only 2,000 RMB. Compared with a 10-million-RMB rack, that 2,000 RMB is completely negligible. So we have space in abundance. Through optical interconnect, we let these chips sit comfortably spread out across a larger space, and thereby obtain data centers of 100,000 or even 500,000 cards — we can build all of that. A single machine hall can probably reach 100,000 — 100,000 chips is already no problem at all today.

Zhang Xiaojun: In the large-model industry, people still say we’re roughly a few months behind. In the chip industry, what do you think our real generational gap is?

Liao Heng: I think on this question — people have this worry, or...

The publication of the Tau Scaling Law (Tau定律) I mentioned earlier was, to some degree, a way of letting more people know: shrinking (微缩) is not the only path. When you can no longer shrink, you can stack — and stacking can be done at the chip level or at the system level. I believe that in the second half of this year, when people see the newest phones —

It should be released in the autumn. Because there are plenty of outfits like SemiAnalysis and TechInsights that will certainly do reverse engineering, you will see a lot of analysis reports come out, and you will be able to see whether the so-called Tau Scaling Law, stacking, and logic folding can actually solve — or narrow — the generational gap. My answer is: to a large extent, yes. And people may have consistently overlooked the most critical point of all: when people buy a GPU or NPU, they are actually buying memory. Because calculated at today’s costs, 80% of a GPU or NPU is the cost of the HBM; only around 10% of the money is spent on the advanced logic die. So the main competitiveness of an AI chip arguably comes from how much bandwidth it has, not from how large its compute specs are — or rather, when the compute specs grow, the bandwidth must scale up in proportion. After all, HBM cannot be sold on its own; it has to live parasitically inside some GPU or NPU, so its monetization depends on the final integrator. The reason I bring this topic up again is that the biggest core competitiveness does not necessarily come from how advanced that logic die is — it comes more from how advanced the aggregate bandwidth and the memory matched to it are, right?

Zhang Xiaojun: What you have just given are a lot of conclusions. In the actual process of feeling your way forward, when did you find this path?

Liao Heng: I would have to say our capability in this respect is still somewhat lacking. The first thing is instinct. If you are lost deep in the mountains and forests, a good hunter has a kind of innate intuition about which direction might lead to water to drink, which direction leads back home — maybe he looks at the stars in the sky to judge his bearings, then picks a direction and goes. So-called instinct means making a choice without sufficient evidence — intuition, plus some judgments of your own. And those judgments are not necessarily rigorous conclusions from reasoning that could be mathematically proven, right?

The second thing, of course, is that you need to keep seeking verification in a given direction, and run some experiments yourself. And even more — this is where I say we are still lacking — we need to find the people building models, or working at the scientific frontier on all kinds of quantitative work, and form an effective dialogue with them: our instinct is such-and-such, your work is in this field, let us see whether we can align. So this is a very important part of the four-year lead time I mentioned earlier: first, you yourself need enough credibility that others are willing to engage with you and form an effective exchange; and then, at that level, people need a kind of exchange of ideas, and continuous alignment.

Zhang Xiaojun: A lot of capability comes from those model pioneers. Can you talk about some stories from the past two years of mutual support, adaptation and optimization with some Chinese large models — for example, DeepSeek?

Liao Heng: The specifics I am not in a position to discuss. But what I want to say is that in the so-called AI ecosystem there are perhaps three different levels, with the difficulty and the requirements rising step by step. The first level is: a model arrives, and I can run inference on it. That is relatively simple. When a model arrives, its service life usually lasts — maybe half a year, or perhaps a year. Once you have adapted it and tuned its performance well, people will copy and deploy it over and over, right? That is inference: one-off work that can be replicated.

The second level: taking a model that a research team has already finalized into first-version production training — the workload is perhaps three times that of adapting an inference system. All the operators already used in inference are needed in training too, but there are also backward operators, and backward operators are usually a bit more complex. On the other hand, the demands for extreme performance efficiency are relatively lower than for a production inference system. Either way, large-scale training is also one-off — or rather, you do it once and it runs for three months, then you do the next version. This still counts as a difficulty that can be solved, and crossed, with manpower.

The third stage is research. An advanced research team may have dozens of excellent algorithm people, each running different experiments every day, constantly modifying their code. We consider the inference breakthrough basically solved, and the training breakthrough entirely within reach — it will not take very long. But breaking through at the research level can only be accomplished with a more advanced compiler stack. Researchers have now switched to TileLang or other higher-level languages, because they are easy to modify. No matter whose processor you use, everyone actually faces the same problem: change the algorithm a little and you have to rewrite the CUDA operators — nobody can stand that. So we are also working very hard on this with compiler stacks — no matter which company leads them — we are striving to get this right. Once that is done, we will be in a position to expect that within the next year or so we may partially reach that research level I just described. But it genuinely takes time.

[3:00:00]

Zhang Xiaojun: In your view, as of today, is there original innovation in your work? Where does it show?

Liao Heng: Perhaps I can talk about it on a few levels. One is the chip itself. I personally am a staunch member of the ‘don’t copy the homework’ camp. Not copying, of course, creates enormous difficulty, because not copying means you cannot borrow someone else’s ecosystem — you have to pay that price. But I have a deeply rooted conviction: we are a technology company, and no technology company in the world has ever achieved great success by copying homework. Because technology itself is innovation — your primary means of value creation is building a differentiated advantage. Even if you have a hundred flaws, you must have one strength. That is my understanding after working in this industry for so long. I am especially afraid of making products that are the same as other people’s. Because we are not the Haining Leather City or the Wenzhou small-commodities market — where every stall’s Christmas trees look exactly the same, and then it is just a contest of who has the lower cost. That model may work in low-end industries, but in an industry that leads the world it is absolutely unacceptable. You must have your own strength, because only that strength converts into the value of the product.

Zhang Xiaojun: When was ‘no copying homework’ decided? From day one?

Liao Heng: From day one. Or rather, in every product I have ever worked on, we have never copied anyone else’s homework. Put another way, the moment I hear we are supposed to be the same as someone else, I feel that definitely will not do.

Zhang Xiaojun: What is the price of not copying the homework?

Liao Heng: The price of not copying is, for example, that your ecosystem gap is enormous at the start. Everyone is used to Windows, and you suddenly want to make a non-Windows PC — you have to change those people’s habits. Or, like now, everyone uses iOS or Android, and you suddenly want to build HarmonyOS (鸿蒙) — you have to put in enormous effort to establish a unique advantage, or give people a reason to switch. That is the price. As for these problems, I think we are now crossing the so-called painful plateau, and may be entering a better phase.

Zhang Xiaojun: For you, would copying the homework not be easier than not copying?

Liao Heng: Of course it would.

Zhang Xiaojun: Then why insist on not copying — from the standpoint of your position?

Liao Heng: Because our mission is not to win some praise — or take a scolding — in an exchange with some customer this year. Our mission is to build a sustainably developing product system. And within that product system — as I mentioned from that Hennessy and Patterson book, the interface between software and hardware is two sides of one coin, right? Copying the homework means: I want to make my hardware sufficiently similar to someone else’s, and then for software I entirely use the ready-made software others have built, so I do not need to put in much effort building my own ecosystem. But I am deeply convinced this has several fatal problems. The first problem: you can never — suppose today I want to transform myself to become exactly like you — I can never look completely identical to you, and there will inevitably be consequences from not being exactly the same.

Second: such a technical system is completely incapable of evolving. Because if you copy step by step, you must copy at every step; you have entirely lost the possibility of independent design innovation. The other side’s software is already there, so you have to cut the foot to fit the shoe — whittle your own shoe down until it fits into, what, Snow White’s crystal slipper (as spoken — the glass slipper is of course Cinderella’s), right? Once you have whittled once, next time you have to whittle again. So I think this is the follower’s inevitable backwardness: the very process of copying along has already determined that he will be behind. And we do not want to be behind — we want independence and self-reliance. Moreover, our problem, our constraint — say even if I copied to be the same as him, his interconnect capability is different from mine. What do I do then? There will inevitably be a communications part I cannot copy. And once the communications part cannot be copied, the operators that fuse communication and compute cannot be used. So with this kind of problem, it is not enough to copy part of it — you would have to copy all of it.

Zhang Xiaojun: We were just discussing innovation — you covered chip architecture; what about other aspects?

Liao Heng: As I said, the important part is the system level. As mentioned, we have completely taken two different roads, with different design philosophies: they go dense, I go sparse — right, I go loose. I want everyone to have their own bedroom and sleep comfortably; he wants twenty people sleeping on one kang (heated brick bed).

Zhang Xiaojun: For the next few chips that have not yet been officially released — including the 950DT, and the future 960, 970, 990 — what should we look forward to?

Liao Heng: I think the foreseeable next three generations are all about the same — roughly on par with the competitor — at a cadence of doubling every year. Beyond that, there is the system level. What I find most worth looking forward to is this: we have been pondering a question — do you build one highly integrated building block, or do you build smaller building blocks and, by wiring them together, let different users in different scenarios obtain more flexible combinations? That is a rather interesting system-level question.

Look at the equipment people used in data centers in the past: earliest were servers, each a 2U server; later, AI servers were maybe 6U or 8U — taller, the box is bigger; and then the Pod appeared. A Pod is basically the typical all-in-one: the whole rack frame is one device, with different switch boards, compute boards and a backplane on the front and back. These two forms obviously each have their strengths. First, the all-in-one device is highly integrated and can perhaps achieve extreme density. For instance, Nvidia’s NVL72 is an all-in-one big machine, right — and on the later roadmap there are bigger and bigger machines of this kind. That is an important form factor, and we have similar forms too. But I think while this form’s advantage is high density and the ability to pre-integrate, the challenge it brings is: if you want to make any adjustment to the proportions of what goes inside, you cannot. Because the machine has effectively hard-fixed all the parts that can be plugged in, the ratios between them, and their connection relationships.

And what we see is precisely that, because of the business itself, models went from 600-odd billion parameters to 1.6T in a single year, and next year it may be 5T to 10T. The resources they need — the KV cache used to sit in the CPU’s DRAM, and now everyone has migrated it to SSDs; and model inference speed, or latency, has to drop from 20 milliseconds to 1 millisecond. Whether in size, speed or latency requirements, these are drastic changes of half an order of magnitude to a full order of magnitude. Such changes may well be hard to predict at the time you design the Pod — two years later, when the machine reaches the market, the original proportions have already shifted a great deal. What then? So going forward we will quite actively consider the need for a more reconfigurable kind of building block, where the blocks can be recombined between themselves in the simplest possible way. This actually ties in with the relatively loose design philosophy I described earlier, and with using all-optical interconnect — these ideas are all connected, or mutually reinforcing. Put another way, once I have all-optical interconnect, I can build smaller boxes, and these boxes can be freely combined — just by connecting a few optical fibers — into the optimal configuration. So that is a system-level innovation, or something to look forward to. And our supernode domain will also be made very large, because we have found that if you have a 20T model requiring ultra-low latency, the communication domain becomes quite large — hundreds of nodes, even a thousand-plus nodes.

Zhang Xiaojun: In the chip field, would you say China has walked out a Chinese path — in your view, as of now?

Liao Heng: I do not think my saying so actually counts for much. I think you can go look at these companies’ financial reports. Their financials to a large degree represent an aggregation, because the great majority of the companies are fabless — everyone goes to certain fabs for manufacturing — so it is roughly equivalent to a sum of the whole industry. The financial reports are very telling, because they represent, first, a scale of economics, and second, whether they can be positively profitable. Because whether a Chinese path has been walked out requires not only technological breakthroughs but also an economically closed loop. You cannot subsidize forever, or bleed indefinitely — that is an unsustainable state, right? So I think if you look at those places, you will get a very good answer. And the financials are not just the logic fabs alone — you should also look at the memory fabs, the NAND fabs, and the packaging houses. My direct or indirect impression is that, especially on a five-year timeline, they are all in a state of order-of-magnitude surge.

The next question is whether that curve is sustainable, right — that it does not fall back down. I personally believe it will not. Why? Because first, in many fields we are actually not behind: advanced packaging is not behind; the NAND field, for instance, is actually not behind, right? The DRAM gap is not that large either, nor is the HBM gap — there is some gap. For logic, as long as the loop can close positively, that is already very good. And then of course there is the macro question: will China and the US ever return to that past so-called strategic-partner relationship of mutual trust and mutual willingness to depend on each other? I think it is quite clear the problems lie more on that side, right — the other side is not willing.

Zhang Xiaojun: Everyone says compute is scarce, compute is scarce — every model company is severely short of compute. So what is the true state of China’s compute today?

Liao Heng: I do not know what the actual situation is, but I can guess from some tangential data — and it is probably a very inaccurate guess. Reportedly, last year China deployed something over one gigawatt of data centers, while the US deployed roughly seven to eight gigawatts. So we are at roughly one-fifth — or somewhere between one-fifth and one-seventh or one-eighth — in relative terms. Of course, they have also deployed a lot in many other odd places owing to particular factors — Southeast Asia, for example.

Zhang Xiaojun: Can your supply volumes be ramped up?

Liao Heng: We are working hard on it — probably yes.

Zhang Xiaojun: Say by this time next year — how much more than this year?

Liao Heng: That is hard to say. All I can say is that this problem should still be solved rapidly.

Zhang Xiaojun: When did you realize you might have made it through the hardest days? Which year?

Liao Heng: I felt I had made it through the hardest days when I started receiving more and more criticism — especially internal criticism.

Zhang Xiaojun: Why?

Liao Heng: Because at the most difficult time, no one comes to criticize you.

Zhang Xiaojun: Oh.

Liao Heng: Last year, the year before — I would say by around last year or the year before, that point had already passed.

Zhang Xiaojun: So you were still being challenged internally?

Liao Heng: Why do I call this a way of drawing the line? Because if you are walking in a desert and have not drunk water for seven days, then any water you see, no matter how dirty, you will feel it is saving your life — you can drink it, and you will not be picky about how the water tastes. After the first bottle of water goes down, everyone’s body immediately recovers; then by maybe the second bottle, people start to feel the water tastes bad — and that is the moment of criticism I am talking about. So I think the most difficult time is when nobody criticizes you.

Zhang Xiaojun: Does being criticized feel unfair?

Liao Heng: I think, for me — perhaps my virtue is insufficient, or it is a flaw of temperament — there is still some pushback in me. But afterwards I come to think that this too is one of life’s inevitabilities, so there is nothing really to feel wronged about.

Zhang Xiaojun: You described those dozen-plus years at that overseas semiconductor company as living through a long period of decline and shrinkage — the sunset of a sunset industry. What about these last ten years? What have these ten years felt like?

Liao Heng: These ten years, for China — or for the environment I am in — even though we went through a nine-deaths-one-life calamity, have absolutely been an explosive period for capability.

Zhang Xiaojun: What has it felt like?

Liao Heng: The feeling is that, apart from process technology being a tiny bit behind — a little behind, specs a little behind — the vast majority of the chips America can make, China can now make itself, from design through manufacturing. And not just our own team — this includes the overall capability outside Huawei too. For example, you will see countless companies making autonomous-driving chips, or embodied-intelligence chips, or WiFi chips. You will find this capability has become fairly widespread — that is, the ability to design relatively complex SoCs with a certain barrier to entry is now blooming everywhere, which is to say this capability has become relatively less scarce. The second interesting point is that the age of the people in the field is a good twenty years younger than the Silicon Valley crowd.

Zhang Xiaojun: Twenty years younger? Than people in the Silicon Valley semiconductor industry?

Liao Heng: Yes. In Silicon Valley semiconductors I would count as at least not old — on the younger side; but in this industry in China, I am at an age that should probably soon be phased out. So this place represents a kind of hope.

Or put differently — young people have more explosive energy, and they have many more years ahead of them in this profession.

Zhang Xiaojun: So here is the question: so many people are now entering the chip industry — including the large-model companies. How do you view this domestic competition?

Liao Heng: In any case, I’m not particularly worried. I think having competition is quite healthy. I hope they are all very capable — but I am not afraid of them.

Zhang Xiaojun: During this ordeal, was there a so-called near-death moment? When was the most desperate moment?

Liao Heng: I didn’t experience one personally, so I don’t want to put myself into other people’s roles and recount those events on their behalf. For example, perhaps the closest thing to a near-death moment — one I was only indirectly involved in — was our Kirin (麒麟) chip, that is, the phone SoC. But fortunately there were some very resolute people who absolutely refused to give up, and they brought it back to life.

Talent and Compute

Zhang Xiaojun: Most of the companies I interview lean toward software. In a hardware-leaning industry like yours, is the design of the organization and its culture very different from those software-leaning companies? Do you have any distinctive culture?

Liao Heng: This topic is a bit deep. I can only say that, since I’m not someone responsible for talent management, or for managing a very large organization on the people side, whatever I say about talent from an organizational angle may not be very accurate. But I can, from a personal angle, describe how a talented person grows within a chip team — some points that are perhaps especially valuable and important, and how to grow. Maybe that’s more meaningful.

First: to be a chip engineer or architect, you absolutely must have actually done a chip. A genius PhD cannot do chips without that. You have to start from the module level — module design, subsystems — and go through this long process, typically at least 18 or 20 months, and you must go through it more than once — many times. Why? Because chip design is very different from writing software. When writing software, out of habit, after writing say ten minutes of code I will definitely run it: hit compile, then execute the code. With a chip, you only get to press the button once, after 18 months. So you have absolutely no room for trial and error — the moment you press that button, you have spent at least 200–300 million RMB.

Second: a chip involves a great many physical things. A fairly smart person is often strong in the digital world, in logical thinking — but you must go through the physical side: be off by one nanosecond, miss timing by one picosecond, and the chip immediately fails; it dies right in front of you. So there is this extremely rigorous, exacting engineering dimension, and school education cannot give you these things. You must have been through the process to deeply appreciate why you have to be so meticulous about this — because you cannot be off by a single picosecond, cannot route a single wire wrong, and a logic gate cannot contain any critical bug. Software, by contrast — you can just fix it a minute later, right? You find your bug when the run fails.

So I think this process is exactly why I am not especially worried about those seemingly star-studded teams — because perhaps their leaders or organizers have not realized this point. They put too much weight on how smart people are, how good their resumes look, or whether their education was under famous masters. It has nothing to do with famous masters — you must go through the experience, and that experience carries a cost. It is through a continual tape-out process that large numbers of people get forged. Someone unwilling to go through this process will never grow into the position of architect. That is the first point.

The second point: if we are doing chip design — say fabless design sits on the seventh floor of the stack — then do you have sixth-floor capability? What about the eighth and ninth floors? Can you reach a shared understanding with someone who works on algorithms? If you have no basic understanding, people won’t even bother engaging with you, right? If you cannot understand a word of what they are saying, the exchange simply cannot continue.

So I think, as an engineer, you need to open your field of view — your horn of curiosity — fairly wide: care not only about what is in your own hands, but also about what the people on the floor above are doing, and the floor below. And don’t just listen to others — you should Google things more; or nowadays, with large models, you can ask a lot of questions. For instance, I just asked one: how much cooling-tower area does one megawatt require? For a question like that, asking one model may not be enough — I will ask four in a row and compare whether their answers roughly agree in order of magnitude before I believe them. You see, these are not questions that someone focused solely on delivering their own assigned duties would ask — but it is a very meaningful question, isn’t it?

So what I want to say is: opening the horn wide means investing a lot of extra curiosity; and beyond curiosity, investing a lot else — reading a few more papers that seem to have nothing to do with you. Then, when you are able to hold a conversation with people working on the most frontier models, you will find you have entered another comfort zone: not only can you understand them, you can even predict what they will be thinking next, and you may even anticipate many problems they haven’t thought of — because you have a more low-level perspective, right? Then, given time, as your experience shuttling between different floors accumulates, you integrate more and more layers, and your ability to get a grip on a technical problem grows stronger.

Zhang Xiaojun: So when you value a candidate — setting the resume aside, someone who perhaps has not yet actually done a chip — what is it you look for in them?

Liao Heng: The first thing I look at is their values — or maybe using the word ‘values’ is putting it too strongly — basically, whether we can get along with one another. That said, maybe ‘values’ isn’t wrong either: whether we fit when discussing things, be it technical questions, expectations for the future, or expectations about the career. Because even a candidate whose other qualities all look extremely high — if they only spend half a year or a year, and leave before going through, without the patience to go through, those training cycles I just described, then for me it is a complete waste of time. So we increasingly look for a rough alignment of basic outlook — only then is it meaningful. Otherwise, we are just wasting each other’s lives.

Zhang Xiaojun: You don’t poach people with astronomical offers, do you?

Liao Heng: Astronomical offers to poach people — we don’t have that kind of capacity. We don’t have the capacity for sky-high prices, and it also wouldn’t really fit Huawei’s so-called ‘striver’ (奋斗者) culture.

Zhang Xiaojun: With a striver culture like Huawei’s, how does the organization create innovation? Or is it a militarized style of management?

Liao Heng: No, no — Huawei is not like that. I think the innovation lies entirely in things like this: an employee of mine who joined just a year ago can come to my office and argue an issue with me — and I very much welcome people like that. I can’t speak for the so-called organizational culture as a whole — I can only say that each of us sees only a very small range around ourselves, right? We work hard to help the people around us become self-driven — self-driving is more effective than being driven by others.

Zhang Xiaojun: There are also young people now facing the question of whether to build their careers overseas or return to China; a common concern of theirs is that resources at home are constrained.

Liao Heng: Which resources do you mean?

Zhang Xiaojun: Compute resources; the gap is large.

Liao Heng: I don’t think that is necessarily so. First, within universities, domestic compute resources are absolutely far more abundant than overseas — which is very surprising.

Zhang Xiaojun: Really?

Liao Heng: Of course. Look at Beijing and Shanghai — those national resource laboratories and the like all have scale of over ten thousand accelerator cards. That is an astonishing number. Abroad, at even the most famous university, a single department might have only 1,000 cards, or a few hundred. So in compute resources, China’s academia is absolutely far in the lead — abundant to an unimaginable degree. Of course, this is also thanks to certain senior leaders who, earlier in this process, recognized that compute had to be provided in order for academic research to be done well. I think this condition in China is unmatched. So as far as academia goes, China’s compute is very ample. As for industry, I have not seen excessive scarcity either — as long as a team is doing valuable work, they can still obtain a certain amount of compute. Is compute so over-abundant that it can be casually wasted? Certainly not.

Zhang Xiaojun: Do you think compute is a bottleneck for large models today?

Liao Heng: I don’t think it is the main bottleneck — no, it is one bottleneck among several. I can only say that there are some outstanding teams that achieved extremely high breakthroughs using very little compute. For example, the DeepSeek team mentioned earlier — their compute was quite limited, absolutely not on the scale of a hundred thousand or a million cards, but on the scale of a thousand cards.

And perhaps there is also Kunpeng (鲲鹏) — everyone puts the vast majority of their attention on AI compute, but general-purpose computing is also part of the infrastructure, along with its corresponding networking, optical modules, NICs, and SSDs. You could say we have assembled the full set of these ‘eight big items’ (八大件), which make up a complete data center. The most fundamental components, of course, are the CPU and the NPU or GPU; beyond that there is memory, the SSDs or related storage systems needed to hold the data, and then NICs and switches. The switch is also a component with a relatively high engineering threshold: the more ports a switch fans out, the stronger its interconnect capability — and it directly determines how many tiers the network ends up with. Say one switch can connect 512 things — then a single tier does the job. But if you need to connect 1,024 things, one tier cannot do it; you must have two. And once you have multiple tiers, there is latency — the extra latency between the two tiers and all sorts of extra costs — which double directly. And then there are the NICs too, right?

So when we build this system — especially these networking technologies — on one hand they need to be advanced, and on the other hand they need to interconnect. Interconnection means compatibility and inclusiveness, because within any system there is a great deal of old equipment; you cannot simply construct a brand-new world from scratch. So we have been trying to build — on the one hand, this LingQu (灵衢, UnifiedBus) system; and of course our Ethernet suite, built up over many years, is also very complete: from high-performance NICs to RoCE to switching equipment, in every lane we strive to be at least in the world’s first tier, with no generational gap. For example, if others are using 51.2T while I am still on 25.6T — at that point there is a generational gap. This factor matters quite a lot, because if we only made one single component, it would be very hard to form a complete system of one or two hundred thousand cards — there would inevitably be a large amount of legacy constraining your competitiveness. This is also the necessary condition for why we could step out and create an entirely new interconnect bus of our own — because as soon as you are doing interconnect, you are dealing with linking all the necessary components together, right? On one hand we have the complete Ethernet system, which ensures interworking with all legacy devices — and that interworking capability is not inferior; it is world-class.

But when we go to build this supernode (超节点) technology, it again involves those ‘eight big items’ I just mentioned — lacking any one of them creates a major defect; you could say every one is indispensable. This partly explains why the world has so many vendors — anyone who works in communications will tell you it is about standards: I have to go to IEEE, to IETF, and pass admission certification; even to do PCIe you have to go to compliance labs for interoperability testing, because it is a multi-vendor system — each component may come from a different vendor, and they have to fit together. So I think, especially when we started designing this LingQu (UnifiedBus) system in 2019 — on one hand we were forced into it, but even at that time we already knew that a technology system must be able to stand on its own, and the precondition for standing on your own is: for the things that need connecting, you must hold the full hand of cards, right? If there is even one component you have not assembled, you do not meet the condition — because that connection simply will not connect. So with these two factors combined — we happened to have fully independent capability in every key component, built to the point of being interchangeable with, on par with, world-class products — at that point we had what it takes to define a proprietary protocol of our own. And now we have in fact made the LingQu protocol free-license and published its spec, and we welcome vendors around the world to use it themselves. But from our own perspective, we had the ability to build a full-suite, complete system, and so we rather bravely took this step.

Zhang Xiaojun: Are there still any shortcomings?

Liao Heng: I think our biggest shortcoming is actually still at the software level. The hardware certainly has its deficiencies, as just discussed — we would like the specs to be bigger, right, with efforts over the coming years that may double every year; and memory — we hope to move faster ourselves so that it does not drag things back. But the biggest shortfall is still in the so-called software ecosystem. Our hope is that as our deployment volume keeps climbing, more and more people will have both the conditions and the necessity to join in contributing on this new hardware platform, working continuously on performance tuning. Because whenever a vendor or a customer deploys this system, they will necessarily have teams that need to pull their business up onto it. I think there may be a kind of threshold to cross here: once this population is perhaps double what it is now, what you might call the communal force — the strength of the group — will be large enough for the system to sustain itself in a virtuous, self-perpetuating way. So this is what we most look forward to.

Zhang Xiaojun: At the very beginning we talked about the ups and downs of the chip industry — you have been through its boom, then its sunset period, and then with the arrival of the AI wave it entered another boom. Now, in the application layer built on top of chips, the monopoly landscape is still unsettled. If one day they again form the kind of monopoly effect that those big American giants had back then, will the chip industry face another withering period? How do you see the future of the chip industry?

AI and the Chip Frontier

Liao Heng: I think… it is possible. It is possible. But there are certain factors here that keep piercing this — if that monopoly were a balloon, there are factors constantly puncturing the balloon. Because from Wall Street’s perspective: if I am an investor, and my entire pension is in OpenAI stock, or Anthropic stock, of course I want its returns maximized, right? It becoming the world’s sole provider of AGI models is most advantageous for me — from an investor’s standpoint. And if one company is not enough, what about two? Betting on two is not bad either, right — buy both companies’ stock at the same time.

Why do I think certain factors have kept this from happening so far? One reason — one important factor — is simply that AI technology has not yet reached its saturation point, its plateau; the curve has not yet entered its flattening phase. So anyone — even today’s leader — might be overtaken by another player six months, three months later, right? That possibility has kept alternating over at least the past few years. We once considered Llama the world’s best model, but now it clearly is — well, temporarily — not; maybe tomorrow it will stage a comeback, right? And within this, I have to say, the threshold of human IQ required is not actually that high. In other words, if you take a walk around Wudaokou (五道口), you will find perhaps 5,000 or even 10,000 students who can completely understand where the tricks in the latest models lie — and every day, in hands-on practice, they are running experiments at smaller scale, finding their next breakthrough point. So in other words, when something has not yet reached saturation, it is not easily monopolized — that is down to the technology itself.

The second important factor: I think behind this there may also be some people with real aspirations who do not want it — even if they had the capability, they do not want to become the world’s terminator; rather, they want, in a more inclusive way, to let more people join this field and make the next invention. In fact, among those I have been in contact with — at least the one or two leading teams in China — their entire vision of value is exactly this: they do not necessarily want to use their momentary lead to capture maximal short-term profit, but rather hope that, with this thing, smart people all over the world can keep chasing and overtaking one another, and so achieve the next breakthrough in capability. So I think there are two different value systems coexisting here — their visions are simply not singular.

And of course there are other factors as well. I believe that in an environment like China’s, there are a great many people — even our country itself — who do not want to see, absolutely do not want to see, all of AGI monopolized by a single American company. That could even produce a crisis for humanity, right? So: vision and values, plus drive — inner drive — plus China’s galaxy of brilliant talent and a very ample talent supply, with outstanding students emerging generation after generation — some of the most important inventions might even be made by an intern; that possibility has always existed. For this reason, I think open-source models are especially important — perhaps even the single most important needle puncturing this monopoly. On the hardware side, of course, we are making our own efforts too, right, to ensure it does not become a single, monolithic hardware system — and many of our industry peers are making this effort as well. So I think in the near term — at least for me — I remain full of hope, and believe this balloon is unlikely to —

Anyway, I’m firmly among those who hope to pierce that bubble. One more thing worth looking forward to: in the digital world — the virtual, digitized world — AI’s capabilities really do seem to be advancing very fast. Every aspect of my work now depends deeply on it, and I find that in many respects — almost every concrete task I give it — it does better than I do. That is what people mean by being very close to AGI in the digital world. Whatever your definition of AGI is, it is already useful enough, capable enough.

But for it to cross over into the physical world, there is still a considerable gulf. From our own seven years doing autonomous driving, my sense is — I already gave one reason just now for why legged robots may be hard to monetize for the time being, an energy-consumption reason, right? Maybe trailing a power cable solves that problem, but trailing a cable then restricts its working range. The bigger gulf, though, is that physical AI’s models have not yet reached their ChatGPT moment. That gap still needs — maybe it is the very next moment, maybe it is two or three years out — a major model breakthrough before physical AI can cross its zero-to-one moment.

And that domain will itself give rise to a fairly enormous, promising industry. Because once something is physical, it has to have a body — it must be a machine, right? And machines are inherently diverse. So what you can expect is that, as a result, once you enter the physical world, monopoly becomes much harder. Look at the car industry: it has been developing for over a hundred years, and there are still so many different brands, in different countries and regions, and new brands are still being born. Even people being fatter or thinner might call for two different sizes of car, right? So that is where the diversity lies. And what I find most promising here is China — I am very bullish on it.

I once heard that Buffett may have said something like “nobody wins shorting America” — meaning if you short America you are bound to lose. We don’t actually want to short America; I only want to — I want to long China. Because I think China itself, in terms of its conditions across the board, is very promising; it should be able to climb a very big step up. Moreover — including that interview you mentioned just now — many of these interviews seem to look at the question from a kill-or-be-killed perspective: if China wins, America loses; or that a Chinese superintelligence, a model of comparable capability, would be used as a weapon to attack American networks. I think that is a truly ridiculous perspective.

That perspective is basically a Satanic perspective. Because if I had such a good model, why would I use it to attack you? Why not use it to live my own life a bit better? Why not improve my own economy, my own healthcare, all those things — improving people’s lives, as they say — instead of using it to attack American networks? I think that very notion is a deeply wrong-headed motive.

Zhang Xiaojun: Is that a difference in values?

Liao Heng: I don’t know whether it’s a difference in values so much as a kind of pirate culture — they are accustomed to viewing the world through that lens. Even when China is relatively strong, it does not necessarily need to go and make others weak.

Zhang Xiaojun: Over the foreseeable next five or even ten years, do you think the global chip industry landscape will change? Will it go through some degree of reshuffling?

Liao Heng: The first question is whether it will. The factor with the biggest influence on this is whether we go back to the state before the US-China tech war and trade war — because that has very real short-term effects. If it continues another ten years, I think it will definitely split into two — even if there is no more serious conflict, everyone will need, simply in order to survive, their own complete manufacturing capability. Because chips, for modern life, more than oil — they have almost risen to the same level as water and air, or close to it. In terms of economic scale, it is certainly a bigger industry than oil. You can imagine: coming home to no rice cooker, no refrigerator, nothing at all with a power cord — that is unimaginable, isn’t it?

So I think it is a necessity. And that necessity leads to what you asked — if we’re asking whether there will be big changes, the first important macro factor is the impact caused by these factors between nations. I’m no expert in that area, so it’s hard for me to predict what that impact will be, but there will certainly be an impact. Only after that comes what you just raised — whether there will be monopoly, whether the withering-out phase I spoke of earlier will take shape. My answer is: not that fast. Maybe in ten years — hard to say, right? But right now, because it changes every day, you can’t even say who the leading vendor is in capability. So maybe you lead by three or six months, but in all likelihood someone else will reach a similar level — so the conditions for monopoly don’t yet exist.

Zhang Xiaojun: Looking at the floors above — the upper floors of this 18-layer pagoda — what would you want to say to these model companies or AI application companies? What expectations or outlook do you have for them?

Liao Heng: First of all, I think, number one: the best-endowed doesn’t necessarily win. It’s like when we all go to the same university with many classmates — you find that the classmate from the wealthiest family is not necessarily the one who ends up far ahead of everyone. Not that they’ll do badly, but they may just turn out ordinary. So the organization with the most money and the richest resources won’t necessarily — take America as the example: America’s top model companies are not Microsoft, not Google, not Meta, not Amazon. Why? They have plenty of money; if it’s GPUs you want, they have countless cards; their capacity to invest is strong; their research teams are large and formidable, right? Because — it’s a new thing, it’s not an old thing.

So I think a large, established company’s advantage lies in the enormous edge it has built in its own domain — otherwise it wouldn’t have become a giant, right? But the point is precisely that AI is a new thing. Whether in applications or in business models — the most valuable applications, the most valuable business models — they have yet to be invented, or are being invented right now. Look, I gave the AltaVista example earlier: Digital [DEC] was a giant in its day, and then it sold itself; AltaVista fetched a good price, but nobody ever figured out — such a great technology. When we were graduate students we thought it was wonderful: I could search for papers without going to the library. After it appeared I almost never set foot in the library again — perhaps only to photocopy one of the papers I had found, because back then PDF downloads were mostly not yet available.

So we have to look at the classmates upstairs — admittedly, each of those layers has countless giants in it. I expressed my hope just now: of course, if the giants can play to their strengths, provide really good services and keep improving them — for instance, I would love for AI coding to be a service provided by some Chinese giant, so I wouldn’t have to go to such lengths to make do with overseas ones, right? But I am also deeply aware that it won’t necessarily be them, because they may not want to be in this business. And of course I would very much like the best coding model to be Huawei’s — maybe I could just call the Huawei Cloud service directly. But every company has its own current state, its own culture, its own — the system it has already built. It’s like an immune system: if a cat’s cell suddenly appeared in my body, the immune system would clear it out. That is why the famous book we mentioned earlier, The Innovator’s Dilemma, actually describes exactly this problem. A mature organization usually cannot accommodate something new. But it depends on the organization’s own capacity for self-renewal — or on whether its founder has that vision, right?

The second point is that this vision matters enormously. If you say, I want to monetize this year and rapidly achieve monopoly, that is one approach. If you have a higher vision — say, by next year or in three years, there will be no more hard-to-cure diseases in the world, no more hard-to-debug problems, no more cybersecurity problems — that is a different vision, and it will drive your team and organization to push in different directions. But look at Anthropic — OpenAI created a chatbot, whereas Anthropic’s main revenue comes from coding, right, delivered through API calls. At that point these two businesses have nothing in common, don’t you think? Who they sell to, the different customers, how they charge — from a business-model standpoint there is no similarity at all. So I say: it’s a new business.

And is this model even the best model? Should you try to control the entry point to programming — say, control an IDE like VS Code? Or do you not need to — you sit in the back and just provide an API? You see, different people — even truly great figures — will throw themselves into ferocious competition based on some fantasy. For example, when the internet first took off, some people may still remember the battle between Microsoft’s IE and Netscape. That fight went on for years; in the end even the US Department of Justice intervened, right, and then Bill Gates retired. To us today that’s just a story. But when you look back — was that competition really meaningful? Today Windows doesn’t even ship IE; the default browser on Windows, the Edge browser, is Google’s Chrome browser with a shell around it. So the battlefield you once thought supremely important and must-win — after you won it, you discovered it’s not worth it, it’s nothing; it never brought you that fantastic result, right? Microsoft never became the hegemon of the internet, did it? Instead a mass of new hegemons appeared, and each of them found its own odd track. Nobody expected China to burst onto the scene — Alibaba, Taobao, Alipay, WeChat and all that — it was all new; they didn’t compete with the incumbents on the track the incumbents had defined, did they?

So I think AI’s application layer may be like that too. And for this application layer — in the digital world, the single most important threshold I think everyone needs to cross is this: AI is already extremely capable, very smart, but what it lacks is your personal context. If it does 95% of my tasks better than I do, the one thing for which it has not replaced me is that it has not been brought into my life context or my work context. So I have become a machine that feeds it prompts. Maybe I am inferior to it in every problem-solving capacity, but there is one thing for which it still cannot do without me: it did not attend this conversation between you and me, did not attend my meeting this morning, does not know what KPIs my boss wrote for me, does not know what the most important project of my past six months — perhaps hundreds of meetings’ worth — has been about. When you cross that threshold, you necessarily need an entry point. And that entry point, I personally expect, is the next super, super app.

[4:00:00]

Of course, this entry point will immediately intrude on human boundaries. For example, if it were embedded directly in all my social apps — in WeChat, say — so that the AI could see the context of my every conversation, then it would know everything about all my non-work interactions. It would certainly violate my boundaries, right? On the other hand, it might then no longer need me to describe long-windedly what I want done — it could just go do it for me. Likewise, if it were brought into every piece of software I use for work — able to see my every email, sit in on every meeting with me, see what I use at work, the WeLink that Huawei uses — it would know my entire work context: an always-present companion. It’s entirely possible that — I might be made redundant, because the next meeting might not need me to attend at all; it does everything better than I do.

But at that point it intrudes on another kind of boundary. For one, I would lose my sense of security; for another, my company would say: you have breached information security — these confidential meeting details, right, how could you let a chatbot or agent know them? Might it blurt them out somewhere inappropriate? You see, there is still an enormous amount to be done at the application layer here. I have only described what I think digital AI lacks: the link of bringing it into a person’s context. Whoever solves that first, whoever finds a novel design — and that design may have nothing whatsoever to do with the model itself; it’s a better app — that app may become a super, super, super application. Because if it knows everything you know, and it is more capable than you — then what?

Zhang Xiaojun: You mentioned the innovator’s dilemma just now — but this chip work of yours, within the Huawei system, should also count as a new thing?

Liao Heng: We are not a new thing — Huawei has existed for a long time, since well before I joined. But its dilemma lies in: who is going to define the future? Because this so-called future is a vision — what am I going to do, and how should I do it. In any organization there is a kind of — unless it is your own, unless you are the founder, unless you have every resource, right. Whenever it is not yours — when we are employees, or contributors — we have no choice but to persuade others: OK, here is the future, here is how you’re gonna do it. But often at that moment — for instance, the idea I described earlier of “living in 2030” — to another key decision-maker it may sound like utter delusion. So this requires a great deal of consensus-building. Call it a dilemma if you like — it is unavoidable, so we shouldn’t over-dramatize it either. In other words, in human organizational society, cooperation with one another is inevitable.

Zhang Xiaojun: If today were 2030, what do you see? You said you live in 2030.

Liao Heng: What I see is that we have sufficient manufacturing capacity; I see that perhaps digital AGI has become even better; I see that perhaps we are closer to having a machine that can sweep the floor for me, right. And then I would start to worry about what I myself should do.

Zhang Xiaojun: You have grown more confident over these years, is that right?

Liao Heng: The reason for more confidence is this: when you have faced difficulties ten times as great and found that in the end there was always a way to solve them, you naturally gain more confidence. Having faced those challenges and ultimately crossed those thresholds, it no longer feels so hard.

And there is one topic from earlier that I think deserves one more remark: optical communications. As we discussed, supernodes need to place chips farther apart, and therefore they need to use light. Here, once again, you see that we must respect physical boundaries. We actually built an optical module called Hi1 — it is a 7.2T optical module, whereas the optical modules on the market today are all 800G, that is, 0.8T. So straight away: what is 7.2 divided by 0.8? About 9 times.

Then we immediately face a very interesting question. Once everyone realizes chips have this much bandwidth that needs to be carried outside — with today’s optical modules, picture it: the machine is about the size of this table, your chip sits on a board, and the optical modules sit over on the faceplate side, connector after connector. From the chip to each connector runs a length of electrical cable — cables of maybe tens of centimeters, the longest perhaps 50 centimeters. Imagine it: it’s like an octopus, countless cables flying out of the chip over to a panel lined with optical modules. That is today’s prevailing industry model. The 7.2T optical module we just talked about, by contrast, is placed directly beside the chip. So the connection between chip and optical module becomes a distance of maybe just one to five centimeters — after all, the chip is only about that big. That eliminates that stretch of electrical cabling. This is what we — what the industry — calls NPO: Near Package Optics.

But when we set out to solve this problem, we very firmly chose NPO — we did not take the further step of putting the optics into the package, did not put them inside the chip; we put them beside the chip. And that question arises immediately. If you ever get the chance to interview other people in the optical communications industry, you will find this is a topic countless people have argued over for many years: beside, inside, or outside? What is very interesting is that, for my part, I didn’t need to go through the debate — I directly chose beside. In 2026, this 7.2T optical module absolutely had to go beside the chip — neither outside, nor inside the chip.

Why? Because of the second problem: wherever there is light, there must be a laser — what is the light source? And lasers fail; in a relatively hot operating environment they have a certain probability of failing. Once one fails, you have to solve the problem of how to repair it. And then there is a second question: a great many people, even having chosen to put the optics inside the chip or beside it, then firmly insisted on putting the light source outside. Because once it is outside, if it breaks I just unplug it — pull one out, plug a new one in, swap it — and that solves the failure problem. Whereas we firmly chose: absolutely do not put it outside, it must go inside. And not only inside — I put in two of them: if one fails, I still have another that keeps working, which lowers the failure rate. You see all this — and now you’ll start to doubt me: you, someone from a software background, an algorithms background, who then went and “pretended” to do chips — what qualifies you now to make pronouncements like this, that only this way is best? Well, this topic, once you get into it — it was actually quite a long process.

And it entirely bears out the philosophy I described earlier about how an individual grows within an industry. I mentioned two points: first, open up the funnel — look around, upstairs and downstairs, 360 degrees in every direction, and learn more. So around 2008 or 2009, I had already realized that chips would have to — because their IO, their electrical wires, can only transmit over such short distances — go optical some day. At the time we thought it would have to go optical at 56G, but in fact it only went optical at 224G — off by a factor of four. But even back then I realized we had to bring optics right up to the edge of the chip, or inside it.

With that question in mind, I set out to learn: how lasers work, how to design one, how to modulate the signal, how to receive the signal. And for this — because at the time the optics industry and the chip industry were two completely separate, isolated industries — in the first year I persuaded the company and signed up for a summer school. It happened that a university in the Netherlands was running one — Europe still had something of an edge in optical communications back then — saying anyone could come and learn how to design photonic chips. So I went during the summer holidays, for one week or maybe two. And I found that, OK, I now had a basic understanding of how optics works. That understanding actually became the basis for many important choices I made later.

I immediately realized that every optical device depends on one basic physical quantity, called the refractive index. For any material light passes through, the refractive index represents the speed of light in that material. In a vacuum, light travels at 300,000 kilometers per second, but in glass it is one speed, and in another material — silicon nitride — it is another speed. The second basic fact I learned is that the refractive index is changed by temperature, and no material escapes this: raise the temperature by one degree and the refractive index changes. Third, I learned that lasers fail. Fourth, I learned that for every laser, the light has to be coupled. With electrical wires, “coupling” just means, say, taking a wire and soldering it with an iron — it’s just a discontinuous point, and the wire is connected. But in optics, every single coupling may lose a few tenths of a dB of energy, because when light crosses from one material into another, it loses energy.

You see, these all look like they basically belong to high-school physics — basic facts that students who did particularly well might already know. But you find, surprisingly, that some people who have worked in the industry for decades have forgotten this knowledge — they learned plenty of other things instead. But this knowledge formed my basic view of the chip. First: why shouldn’t you immediately put the optics inside the package (封装)? Because inside that package it is extremely hot — I have a 1,000-watt NPU in there generating enormous amounts of heat. What does 1,000 watts mean? A household electric stove, a fairly high-powered one, might be 2,000 watts, right? So it is very hot. Put anything next to something that hot and it is bound to heat up, and the refractive index will change enormously. So putting it inside creates all sorts of problems caused by refractive-index shifts — best to avoid it.

The second problem: at high temperature, lasers fail. Third — I mentioned another reason earlier — I definitely did not want a single light source placed outside. Because if I have 72 optical channels and I want one light source to feed them, doesn’t that source need 72 times the energy? Seventy-two times the energy produces a hot spot. That is, an extremely high-power beam shining onto one point, which then also has to be split into 72 channels — that one point will inevitably fail. It’s that simple. With that much power — 72 times the power passing through a single point — that point forms a localized hot spot, and the hot spot will burn out your material. Or if glue or dust lands there, it instantly chars black, and it’s no longer an optical communications device — it’s become a cutting machine, right?

So we — perhaps others would do it differently — you see, all of these important choices I’ve just described were intuitive choices. I didn’t run experiments, and I didn’t have some hundred experts verify afterwards whether they were right. I didn’t need verification; I simply struck those options out, because I knew they were bad — so don’t do them. I chose what was easy to do and could be done well. And of course, our current 7.2T module has also turned out very well; in the next generation it may become the main way of interconnecting — it’s cheap, the bandwidth is high, the size is small, and it gets rid of all those extremely troublesome electrical wires. You see, this kind of thinking is purely intuition-driven, purely vision-driven; it doesn’t require discussion with countless people, and yet it may determine whether our system can be spread out across a larger physical space. These ideas all sound very simple when spoken aloud — but someone has to raise the question in the first place.

Zhang Xiaojun: After all these years, do you still feel strong passion for the chip industry?

An Engineer’s Story

Liao Heng: My passion comes more from necessity. Put it this way: if you’re walking down the street and someone — even a stranger — suddenly collapses, you’d feel you ought to go and ask whether something is wrong with them, right? What we see now is that this kind of need is very easy to perceive: we need compute, we need cars that can drive themselves, and so on — these needs are visible everywhere. We need our own smartphones; it can’t be that in all of China there are only iPhones, right? These needs are the proverbial “necessity is the mother of invention” — my passion comes mostly from that. But of course there’s also a negative driving force: when a person isn’t doing anything — or take myself — whenever I’m not doing something particularly good, the odds are I’ll go and do something bad.

Zhang Xiaojun: And what is the mission now? A personal mission, a corporate mission, or something else?

Liao Heng: Huawei’s mission is written out very clearly: to bring digital life to every person, every enterprise, every society. My personal mission is just the feeling that, in a finite life, I should leave behind as many better things as possible — or rather, that what I build should outweigh what I destroy. Because in life you bring nothing when you’re born and take nothing when you die; there’s only the process — and I hope to be of some use in it, of use to others.

You also wrote down two questions for me: one, where exactly is the AI Silicon Valley; and two, how do the ideals of AGI differ from the ideals of Wall Street. I actually already touched on both topics along the way. On Silicon Valley — I mean, where is AI’s Silicon Valley — I’d guess it’s probably in Wudaokou (五道口), (the university district in Beijing near Tsinghua), or else somewhere in Shanghai — Qiantan, or is it called Houhai or something, Binjiang? — over by Binjiang, those places in Xuhui district, the Shanghai Innovation Institute. Why do I make comments like this? Because my own child is over in Wudaokou, still in third year, and we also often go there for exchanges and discussions. I think if you sit down for a meal in any restaurant or café over there, you’ll find at the next table two people in a heated argument about why yesterday’s loss diverged, or how exactly to make reinforcement learning work a bit better. The atmosphere and the people are there.

And why not in Silicon Valley? As I said earlier — it’s not that Silicon Valley no longer has the cohesive pull to attract talent; it’s that these companies’ own desire for monopoly restricts the exchange. An Anthropic employee probably won’t discuss the latest reinforcement-learning techniques with some Stanford undergraduate, because they want to turn that into their competitive advantage — this so-called monopoly is in the process of forming. So I think China has a more abundant supply of talent, with more people actively and energetically working on all kinds of related problems — and for now China is much, much further away from that monopoly. So in this brief window — maybe these three years, maybe five — we are in a period of brilliant stars, a hundred flowers blooming, extreme prosperity, and extraordinarily lively exchange of ideas.

Zhang Xiaojun: So the formation of this open-source culture among Chinese companies — model companies, chip companies, companies of any kind — may have a more far-reaching impact than anything we can imagine right now?

Liao Heng: Yes, that’s how I see it. Or rather, I hope this open-source culture can last a bit longer.

Zhang Xiaojun: Last a bit longer.

Liao Heng: Yes. Because there’s no guarantee at all that the work people are willing to open-source now will still continue next year — so I hope they keep it going a bit longer. If it’s something the founder decides personally, it may last quite a while; if it’s just some department of some company, they may not even have the right to make that choice, right?

Zhang Xiaojun: Right — then they might go closed-source faster.

Liao Heng: Yes.

Zhang Xiaojun: Because it stays inside the company — you don’t know whether it will go closed.

Liao Heng: Right.

Zhang Xiaojun: With more challenges, it might.

Liao Heng: Yes, that’s right.

Zhang Xiaojun: I have a few final quick-fire questions. A favorite food of yours, from anywhere in the world?

Liao Heng: Food... These days I’ve been eating so much coarse grain that I no longer know what I like — or rather, I haven’t thought about what I like in a very long time.

Zhang Xiaojun: A little-known piece of knowledge that everyone ought to know? Actually, you already gave one earlier, so never mind. Based on all the books you’ve read, recommend a few.

Liao Heng: Let me recommend a few books. The first — we’ve recommended it internally at Huawei too — is called The Idea Factory; it should be available on JD.com, though I don’t know if the Chinese edition is titled “Sixiang Gongchang” (Thought Factory). It’s about the history of Bell Labs. The period that resonates most with us, and holds particular lessons, is 1940 to 1945. Because the era it describes was one in which America’s economy, its GDP, was already highly developed, but its science and technology still lagged far behind Europe’s old academic meccas — Britain, Germany. So in those days the top professors all had to have studied abroad — only a Cambridge graduate, or someone from Göttingen, or from some German university, could be a top professor.

But in those years, 1940 to ‘45, Bell Labs produced some fantastic work. Bell Labs sits right next to Princeton. Draw a circle with a 100-kilometer radius around it, and you’ll find that at that time, within those 100 — or 50 — kilometers, work of extraordinarily far-reaching influence was produced. First, in computing there was von Neumann at Princeton; then in information theory, Shannon; then for the transistor, Shockley, and also Bardeen, who invented the transistor, right? And the existence of all this was no accident — there was a kind of macro-historical, almost providential set of factors drawing them together. California had not yet risen as a technology center; the technology center of the day really was within that 100-kilometer radius around Princeton. That’s also why I say that when I was at Princeton I didn’t learn all that much — only later, after more experience, in my forties and fifties, did I suddenly come to appreciate, to realize, that the place really did have many of those elements of remarkable people and hallowed ground. I simply missed it, which is a bit of a pity.

And why is this body of work so instructive? Because first, it involves semiconductors. Then Shannon’s work is the foundation of every communication system; von Neumann’s work was precisely on the computing side; and Turing himself was also a Princeton student. So this whole series of works, compressed into a mere five years, actually mirrors and reinforces the big macro factors of our own current five years — maybe our current ten years. Because the Turing machine represents the deepest underlying theory of symbolism, while today’s AI represents connectionism — and connectionism is right in the middle of an explosion, isn’t it? Second, semiconductors — we’ve already talked a great deal about semiconductors; China is precisely in a state of extreme difficulty followed by rebirth, breaking out of the cocoon. And third, communications — China is no longer behind there. But communications and computing are, as I said earlier, one body with two threads — two lines entangled together through the entire industry’s development. So I think these factors — together with China’s real economy, which is already extremely strong, even unmatched in the world — mean that our science and technology must still leap from follower status to a point where, in certain areas, even the world’s most important inventions and creations are born in — perhaps somewhere in Shanghai, perhaps somewhere in Wudaokou, perhaps some combination of the two. I think in this era, this is quietly happening. It’s just separated in time and space — there may be an 80-year gap — but that 80-year gap marks the opening of China’s era.

So I think, regardless of whether there is a so-called trade war or whatever other factors, when these elements combine, we seem to have all the right ingredients — it happens to be the right time, right place, right ingredients — and a chemical reaction is bound to occur. So that’s one book worth recommending. And of course The Innovator’s Dilemma, which I mentioned earlier — that one is probably better suited to business leaders, right: realizing that your success will not keep on succeeding forever, or rather that the very factors behind your success are quite likely to be the chief obstacle behind your next failure.

Of course, there’s another book I’ve recommended and given to our colleagues, called The Rules of Work. It’s probably also sold on JD.com — I don’t know what the Chinese edition is titled. It’s actually extremely simple, a little booklet, about how young people should appropriately view the work environment. Because the vast majority of universities and schools will not teach you these things. The most essential point is this: in a work environment, unless you are purely on your own — maybe you’re a mathematician who can just shut yourself in a room and never interact with anyone — as soon as there is more than one person, there are relatively complex person-to-person interactions. So you need a basic understanding of this, a process of reaching consensus, and a way of operating within the rules — neither going to the extreme of revolution and wrecking the organization, nor letting yourself feel excessively wronged. Anyway, I’ve given this book to a great many colleagues; it helped resolve things for them and brought them a certain relief and comfort. So I think The Rules of Work is perhaps worth a read, especially for younger listeners.

That’s about it. Oh — while preparing, I also remembered the book Edison. First of all, Edison is a super-celebrity. Go back 100 years, and Edison was the Elon Musk of that era, or the Steve Jobs of that era. I think these few people may be heaven’s repeated resurrections — for all we know, the same soul reborn again and again. Or rather, I think they share certain common qualities. As for why I recommend this Edison biography — you can go to gutenberg.org, where there’s a free e-book, because I couldn’t find a print edition, it’s too old — it was written by someone who worked at his side, so it has a strong sense of authenticity. With Edison — people, especially very smart students who did well academically, are often confused: they assume Edison and Einstein are the same kind of figure, the brilliant giants of human history, the great stars in the minds of technically-minded people. In fact they are two completely different types. Edison may never even have attended middle school. He came from a very poor family, dropped out perhaps in primary school, maybe middle school, and then went off to sell newspapers on trains.

But he had an innate engineer’s quality. And this engineer’s quality really comes down to a few basic points. First: do useful things — never make things that are flashy and showy but without substance; you must invent useful things. Second: respect reality — you must first know which problems are worth solving, and not choose to waste time solving meaningless problems. Of course, the book has countless examples of what he invented, and to invent these things he basically used clumsy, brute-force methods — he was not a super-smart person. The so-called clumsy method was: OK, first — perhaps what was extraordinary about him was that he first identified what humanity needed. Hence the lightbulb: there was no electric light, so invent a lightbulb, right? There were no movies, so invent a way to record pictures of moving things — hence film; indeed even the first movie was shot by Edison. There are lots of interesting stories in there. I believe technically-minded readers, students with an engineering bent, will find plenty of meaningful references in it.

There’s one more book, which I read when I had just entered the industry, called The Soul of a New Machine — “Xin Jiqi de Linghun” in Chinese. It’s about — in my story earlier I mentioned a company called Digital, DEC, Digital Equipment — a company of the same era, the minicomputer era, called Data General. The Soul of a New Machine tells how a group of engineers, in perhaps the late 1970s, went about designing a new generation of machine. Why this book? It’s probably of most reference value to those of us who genuinely work on computers. It gives a very real, almost day-to-day account: how each person took part in the project, what difficulties they ran into, how someone was in a bad mood one day and how it got resolved. I even had the good fortune to meet one of the interns described in the book — back then I think he was called Bob. By the time I met him, he was already an EMC fellow, a white-haired old man. So what I want to say is: this book tells the story of engineers, and especially the story of an engineering collective — and that collective’s story is passed down from generation to generation; it keeps repeating. Maybe you were building a minicomputer; now we’re building a chip, or an AI supercomputer, or the next model, or the next promising super-app, right?

It was still enlightening for me, because you see the people who came before you — everything you have been through, the people before you went through as well. And he actually wrote it all down, which is notable especially because engineers are, frankly, a super boring group, right? Mostly introverted, not good at talking, and least of all inclined to write books or memoirs about themselves. So this is an unusual kind of book. What it describes is the real, almost first-person experience of an engineer. So I think for the engineering community, it is well worth a read.

Zhang Xiaojun: When you were talking about engineers passing things down generation to generation, I was thinking: everyone now says software engineers are about to be replaced by AI coding — if that’s the case, maybe people could pivot toward hardware.

Liao Heng: Not necessarily. Since we’re on the topic of coding, I think there are a few things. First, I don’t think you need to worry too much about being replaced. Or put it this way: as long as you have ideas of your own and are unwilling to be easily replaced, you will inevitably think up and invent new needs, and go build the next, more interesting thing — or, whether for your company or for yourself, define something you previously couldn’t do. Now suddenly, with AI behind you, you might have the development capacity of ten people, even a hundred people — so couldn’t you go build something completely new, something that has never existed before? Maybe you won’t earn a higher salary, but I think the thing you just mentioned, the thing so many people worry about — what you should really worry about is whether you’ll just “lie flat” (躺平), or whether you have an inner drive (自驱力) to create more useful things, services, or products. So as I said earlier — at least within our own small organization — I hope everyone has that desire to push themselves forward, rather than only doing something because I told you to do it. Based on our understanding today, what is the key, important bet? Betting on China — and that does not mean I stand for betting against America.

Subscribe now

Image credit is to P. L. on Unsplash

Vietnam shows the distance between the two. In the literature it appears as terrain, a node in the Southeast Asian compute build-out and a question about data centre capacity and the routing of chips, which is a subject I have written about myself. The same state has had a binding, comprehensive AI statute in force since March, and I suspect most people reading this did not know that it had passed, or what it contains. Vietnam is not unusual in this. Governments neither building frontier systems nor trying to are nonetheless deciding what to prohibit outright, whose regulatory model to copy, and how much of a limited administrative capacity to spend on the process. Those decisions are how the technology arrives in most of the world yet, they are often the least legible part of the picture.

What keeps them out of view is not that they are uninteresting but that accessing them is difficult. The texts are in Vietnamese, Spanish, Indonesian. They are issued by ministries whose names and remits differ from one country to the next, so there is rarely an obvious place to begin: Vietnam’s law sits with the Ministry of Science and Technology, Peru’s regime with a secretariat inside the Presidency of the Council of Ministers. Legal status then resists reading from outside, because a strategy document, an ethics code and an enforceable statute all present themselves as things a government has published and that currently apply. The working English-language layer is client alerts from regional law firms which, while careful and often excellent, are written for compliance officers rather than analysts and frequently of uncertain vintage by the time you find them.

None of this is an intellectual barrier. Rather they are just a search cost that can be removed.

That is what I built AI Governance World Map to do, and two of its entries make the case better than a description of the site would.

Vietnam, at drafting

Law No. 134/2025/QH15 passed in December 2025 and took effect on 1 March. It sorts systems by risk and loads the top tier with the obligations one would expect of a European-derived regime: conformity assessment before deployment, registration in a national database, logging, human oversight. Defence and cryptographic uses fall outside it altogether. Penalties run to two billion dong, which is around $76,000, a ceiling low enough that for any firm of scale the compliance cost will exceed the fine for ignoring it. The architecture comes from the EU AI Act by open acknowledgement.

But it departs from that text at three points. Vietnam did not fix its high-risk categories in an annex but delegated the list to the Prime Minister, so the scope of the regime can be revised by executive decision rather than fresh legislation; as of May the list had not been issued. Conformity assessment then defaults to self-assessment for most high-risk systems, with third-party certification held back for a subset, softening considerably the European insistence on external conformity bodies. Foreign providers are asked for a local contact point rather than anything heavier.

What Vietnam loosened or handed to the executive were the parts of the European model that require a standing apparatus of notified bodies to operate, or parliamentary time to maintain. The transfer was selective, and in the direction a state administering the regime on limited capacity would choose. If you want to know what AI regulation will look like across the next thirty jurisdictions to legislate, the pattern of what gets cut in transfer will tell you more than the text being copied.

Peru, after commencement

Ley 31814 passed in 2023 and did very little for two years. Its implementing regulation arrived as Supreme Decree 115-2025-PCM in September 2025 and took effect on 22 January, giving the Secretariat of Government and Digital Transformation authority over technical rules and staggering compliance deadlines by sector across the years that follow.

In May, INDECOPI’s Sala Especializada en Eliminación de Barreras Burocráticas, the tribunal that hears challenges to administrative requirements lacking a basis in law, ruled on which of the regulation’s obligations constitute unlawful barriers and which do not, striking part of it down as an illegitimate restriction on private developers and deployers. The standard dismissal of AI law outside the major powers is that it is declaratory, drafted for the communiqué and never operationalised. Peru’s own administrative machinery had tested this regime and removed pieces of it within four months of commencement.

Nothing exceptional produced that outcome. INDECOPI maintains a standing function to strike out administrative requirements without legal basis, and it applied that function to the AI regulation as it would to any other filing obligation, on application from the businesses carrying the cost. Where a state has working administrative-law institutions and enforcement capacity of the middle range, that machinery does much to determine what an AI framework finally amounts to, and it operates entirely after passage, when outside attention has moved on.

Brazil’s PL 2338 offers the contrast. It cleared the Senate in December 2024 and has sat in the Chamber of Deputies ever since, its biometric surveillance carve-outs unresolved and its floor slot repeatedly deferred. It has drawn more English-language analysis than Vietnam’s statute and Peru’s regulation combined, and it currently binds nobody.

What this is for

Neither statute will shape the direction of AI, and reading either as strategic positioning would be a mistake; borrowing a regulatory architecture is cheap and commits a state to very little. What the two of them document is the life cycle of a transplanted framework, adapted on the way in and eroded once operating, and neither stage is visible from the usual vantage point. Two states with nothing institutionally in common reached for the same template unprompted, and the question worth asking is how many others have. If the answer is most of them, then the substantive convergence in global AI governance has already happened at the level of prohibited categories, risk tiering and conformity assessment, some distance below where anyone is negotiating for it.

The tool treats the major jurisdictions the same way. Each country page reads as a brief rather than a document list: what has been passed, who administers it, what is being argued about domestically, and where the gaps are. Everything is sourced against primary texts, with uncertainty flagged wherever verification fell short. The site itself was built with Claude, with my history degree meaning coding it myself would mean it being finished by 2030. The comparison engine sets any two jurisdictions against each other dimension by dimension, which is where the less familiar entries earn their place: Vietnam beside Brazil is a statute in force against a bill awaiting a floor slot. The timeline runs regulatory action across every tracked jurisdiction in sequence, so movement somewhere you were not watching surfaces without you having to look. I will keep expanding coverage as time allows. I hope it is useful, and do send me your thoughts.

https://aigovmap.netlify.app/

Subscribe now

Importantly it is also a skill you can build. When I first attempted the Astral Codex Ten annual tournament in 2023 I was on par with dartboards and coin flips. Then in 2024 I was able to root out the host of silly mistakes I had made and move up to 214th out of 1,349 entrants. 2025 staying consistent at 309th out of a larger pool of 3,569 forecasters.1

The Astral Codex Ten tournament, which is run by prediction market site Metaculus, is a set of questions covering a whole range but focused on politics, economics and technology. For each question you research and find your own quantified answer, without being able to see those of other forecasters. The scoring then works on how well you predicted an event relative to these peers, rewarding smart probabilistic thinking, knowing when to be a sheep and when to break from the consensus.

The structure for this post is grouping the questions into a number of sections: AI, markets, US politics, global politics and tech & space. In each section I give the question, and then my probability of it being true by the end of 2026 as a %.

AI

Will an AI model reach a 3 hour time horizon with 80% reliability during 2026? - 35%

This question is almost certainly the most important of the whole set. Though it is also I question I myself submitted to the competition so I accept I may be rather biased! Behind the more technical language here is effectively the question of whether AI progress accelerates or remains on its current rapid pace.

The background to this question is the ‘METR graph’ now one of the most prominent measures of AI progress. Rather than measuring AI’s ability to do assorted short Q&A tasks, METR measures AI’s ability to do various tasks (in this case coding ones) and then matches this with how long it takes humans to do the same tasks. This gives you the ‘time horizon’ of the model, how long on average can it work uninterrupted on a task. These time horizons have been growing exponentially as you can see below:

The length of tasks that models can accomplish has been effectively doubling every 7 months. Given that pace of improvement you would see models reaching days and eventually months long time horizons towards the late 2020s and early 2030s. Extrapolating what a month long time horizon means is difficult but probably roughly corresponds with some definition of an artificial general intelligence that can accomplish any cognitive task that a human can.

Eagle-eyed readers will have spied however that the graph above refers to tasks models can complete 50% of the time, when our question is actually about 80% reliability. There is in practice a strong disparity between what models can accomplish inconsistently, and what they can do reliably. Though both have been working on similar trend lines (doubling every seven months), scores on the 80% benchmark are a ways behind in absolute terms:

Looking at the higher reliability is especially interesting as there is a good case that this is a stronger proxy for what we really care about: how useful are these models in the real world. If half the time a model correctly formats my work, and the other half it gives me unusable slop, I am probably not going to bother. Rapid gains specifically in the reliability of AI models could potentially feed through to stronger diffusion and greater real world impact.

Taking the existing trend forward, we would not expect models to reach 3 hours at 80% in 2026. With GPT 5.1 posting the highest score yet at 30 minutes, and roughly two doublings (14 months) between its score and the end of 2026, we would expect it to reach a 2 hour time horizon.

A general rule of thumb for AI forecasting has been to trust in the straight line. Large scale relationships between how much computational power being used and the capabilities of models have held up remarkably well. The UK AI Security Institute has found similar trends in model’s time horizons on cyber tasks to the work done at METR, potentially improving our trust in how reliable the relationship is.

But there is a case for even faster progress. Most obviously you can just look at the graph above and ask what on earth is going on with Claude Opus 4.5. On the 50% graph it is very clearly not where we would expect it to be given the trend line. Yet on the 80% graph it is almost exactly on trend? I have no sense currently of a good explanation of this gap, but it gestures at important uncertainty as to whether we might see AI progress accelerate.

Given the increasing capabilities of models, and their ability to themselves automate larger parts of the AI R&D process, along with the rapid expansion of AI compute as companies race to put up ever larger data centres, we could see the rate of progress shift. Given that we are talking about exponentials, changes in the rate of doubling can have quite dramatic effects in bringing forward the date at which we expect to see even more powerful AI models, and their consequent impacts on the economy, national security and even scarier processes of recursive self-improvement.

Now given all that background, what do I actually think? I place the probability that we hit a 3 hour time horizon in 2026 at 35%. Ultimately hedging towards the existing observed relationship is the best practice, and even the current pace of progress, seeing two doublings through 2026, is a pretty crazy state of affairs.

Why am I at 35% rather than zero? The Opus gap between 50% reliability and 80% seems suspicious. It suggests there is a latent capability for models to do many of these tasks, a sort of ‘reliability overhang’. Which given how important this is to the utility of these models I would expect reliability improvements to be a major focus for the leading firms.

So 2026 most likely sees a similar rate of rapid progress to 2025, but we could see a more general acceleration in model capabilities, or some ‘catch-up’ in terms of reliability that means time horizons could extend further than we would otherwise expect.

Will OpenAI file for an IPO during 2026? - 64%

OpenAI is burning through capital as it looks to grow as rapidly as it possibly can. It has promised hundreds of billions of spending to its various compute partners, across Oracle, Microsoft, Nvidia, AMD as it looks to build a colossal infrastructure ecosystem. On the other side of the ledger it is growing extremely fast and has seen its valuation explode, most recently to $500bn, and reportedly looking to reach over $800bn in Q1.

We have reporting that it is interested in going public, targeting potentially a >$60bn raise at a $1tn valuation. Putting things together it seems likely we see a private funding round in H1 this year, with OpenAI targeting $100bn raised, a massive sum even given the kinds of large sovereign scale players likely involved. Then an IPO most likely in early 2027 or possibly late 2026 to bring a further cash infusion and set it up for future more efficient future equity and debt raising.

The question is therefore just one of timing uncertainty. We can be reasonably confident that OpenAI are looking to go public, they simply might get bogged down in various complexities or difficulties along the way. Or they could look to time it relative to market sentiment, given the kind of punchy valuation involved. A public S-1 filing, the key criteria for this question, can precede an actual IPO by a margin of weeks to months meaning that even a 2027 IPO could see a 2026 filing.

Given that we have clear signals OpenAI is looking to IPO, and is targeting 2026/27, I am at 64% here. Remaining fairly cautious simply because OpenAI is looking to move at some meteoric speed, and while dealing with a huge amount of potential difficulty and uncertainty in relationship to the scale of its infrastructure development and research programme.

Will the U.S. enact an AI safety federal statute or executive order in 2026? - 55%

The short version of this question is somewhat misleading. There is zero chance we see ‘an AI safety bill’ even come close to passing at the federal level. But this question in practice is only looking for an ‘AI safety requirement’ to be included in federal legislation or EOs, which is anything ‘intended to prevent material harm from an AI system’s general capabilities or misuse potential, such as mandated risk assessments, safety evaluations, red-teaming, incident reporting, deployment constraints, or access controls’.

Basically all of these are very reasonable policies, and given how I’d expect AI to only grow in significance this year, including in a number of risky domains such as cyber and bio, the odds that the US federal government does literally nothing seems low. The White House has struck a pretty intense line on AI safety, especially with the push to pre-empt state level AI legislation. But ultimately this hardline polls badly, and clearly flies in the face of a more practical reality.

Is the US Government wants to integrate AI deeply into its national security functions, or in ways that have it interface with essential services, it is not going to want things going horribly wrong with no legal framework to assess accountability for why. These kinds of AI safety ideas will almost certainly just be repackaged and renamed but re-enter in substance. Similar to how the US AI Action Plan calls for interpretability for military AI, we might see ‘compliance audits’ run by PWC rather than ‘safety red-teams’ but as long as these are imposed on some private entity (which to have any teeth they would need to) then the circumstances of this question would be met.

The argument above is effectively a case for why you should expect the Trump administration to do something that it has quite explicitly tried to do the exact opposite of. This is why I’ve netted out at 55% here, because even though there is this wonderful rational case for bringing in these kinds of regulations, this is still the Trump administration so a safe approach is to shrug and integrate a lot of uncertainty.

What will be the highest score achieved on ARC-AGI-2 before 2027? - 85% median

ARC-AGI 2 is a prominent benchmark that tries to assess AI’s capabilities at tricky visual reasoning questions. The goal behind it is to create questions that require ‘true reasoning’ rather than ones which AI models could potentially memorise answers to within their training processes. ARC-AGI 1 saw models initially struggle but then slowly make rapid progress up until they essentially solved the benchmark. So we got ARC-AGI 2! Where models started slow but have begun making rapid progress to solve...

This follows the example of most benchmarks, where once models begin making good initial in-roads, you usually see progress continue rapidly until the benchmark is saturated. Once you have a clear signal that a benchmark is feasible, then firms have a metric by which they can assess their model and potentially find ways of ensuring it can learn to keep progressing its performance via new data sources or post-training techniques such as reinforcement learning.

I expect that we will simply continue to see these trends lines and by the end of 2026 ARC-AGI 2 will be solved. Though don’t fear as ARC-AGI 3 is on its way!

What percent of the top 5 human average score will the best bot score in ACX 2026? - 95% median

A wonderfully meta-level question, just how much better will AI be than me at the exact thing this post is about? Helpfully we have the answer for last year, which was a whole lot.

In 2025 the top bot came 5th on the leaderboard, scoring 95% of the top human score. The next two best bots came in at 30th and 31st scoring 84% of the top human score. For reference I scored 54% of the top human score.

Given that even these now quite old models were able to come very close to the best humans, I would similar or even better results this year. Models forecasting at the level of very good human forecasters, but probably still beaten out by a small set of human superforecasters, though not by very much.

AI models have become incredible compilers and analysers of textual information, so it is fairly unsurprising that they have become very solid forecasters, though the kinds of tacit knowledge and ‘taste’ that elevate true superforecasters might be beyond their reach for the time being.

Will an AI-created song chart in the top 20 of the Billboard Hot 100 before 2027? - 45%

Lots of uncertainty here. AI music has clearly become good enough, such that you can easily not recognise a song was AI-generated. There will be many many AI-generated songs swimming around on the internet seeking out virality. ‘We Are Charlie Kirk’ has blazed the path making it to 21 on Billboard’s US Christian Songs. That we might see another AI-generated song, perhaps less politically and ironically charged than ‘We Are Charlie Kirk’, break through to the mainstream charts seems very possible.

Another pathway is via an established artist using AI to compose and perform either melody or lyrics for them. If you are a slightly washed up pop star and an AI-generated song using your voice goes viral, why not embrace the cash grab being offered up to you.

The main risk here is that the world is already awash with music made by humans, such that very little new can break through at all. There is sufficient Taylor Swift dreck to clog up the top 20 indefinitely, irrespective of anything at all changing in the outside world. With the decks stacked against new music in general, it would take a mega-viral AI song to break through, which while very possible is still just below a 50% probability.

Markets and Economics

Will Nvidia’s stock price close below $100 on any day in 2026? - 17%

One of my strongest predictions last year was putting a lot of confidence behind Nvidia beating the S&P 500. I’m hoping a similar conviction will pay off this year.

For context Nvidia’s share price is currently as of writing $185, after climbing slowly all through the second half of last year, after it hit its lowest point at just below $100 in the aftermath of the Liberation Day market selloff. Nvidia has been posting continued incredible growth numbers, and remains the essential firm to the current global boom in AI infrastructure spending.

It does so while also staying at a pretty reasonable price to earnings ratio, not far away from its historical average even pre-crazy ChatGPT times. Relative to some other riskier players with extremely high valuations on far lower and more uncertain revenue streams, Nvidia has a clear pathway to simply chugging along and spitting off colossal cash flows.

For Nvidia to drop below $100 would not be because of any act of its doing. Rather it would be due to a general large scale downturn in the stock market. It would need to be a ‘bubble pop’ style moment where major players blew up and tanked market confidence across the board. While this scenario is always possible, and the speed at which AI spending has risen, and the leverage that is increasingly involved is pushing us into real boom territory, with indications of a bubble here and there, a major totalising blow up feels unlikely.

Individual players might be challenged, I could see for instance Oracle getting even more over their skis and taking a beating. Some of that loss of confidence would flow back to Nvidia. But ultimately Nvidia has a rock-solid revenue case in the form of hyperscaler spending. The demand for AI inference remains fairly ferocious. I myself have become a convert to Anthropic’s Claude Code, where you use an LLM integrated into the command-line of your computer, to effectively create an incredibly potent AI agent.

The sheer utility these tools can now offer encourages you to simply tear through tokens. 2026 is likely to be the year where we really see the productivity effects of AI become more widespread, and more of the effects of a world where first code-generation, but later more white collar functions, see their effective price drop orders of magnitude. This gives a very solid technological foundation to this AI boom. Absent any of the merits of the tech, leverage and timing mismatches are always a fear, but Nvidia remains one of the best positioned players for this boom. Hence only a 17% chance that Nvidia would see such a major fall from its current lofty $4tn heights.

Will the S&P 500 close above 7,500 at the end of 2026? - 79%

There are two clear bull and bear cases here. First that we are clearly in this huge AI supercycle, with spending exploding even further in 2026 driving a huge volume of physical investment in the US, as well as within cloud and software spending. Then as stated earlier I would expect that the investments and research of the last few years start to feed through more meaningfully in productivity lifts from the actual application of AI itself.

AI is therefore our strong bull case. The bear case is that the Trump administration taking various actions harmful to investors confidence in the US macroeconomy. From continuing with destructive and uncertainty-inducing tariff policies, to various kinds of crony capitalism and putting pressure on the independence of the Federal Reserve. These things are all harmful to investor confidence. But then what is good for investor confidence is continued record earnings receipts.

Ultimately I think that the tailwinds of the AI boom keep pushing up equity markets, and we see the S&P 500 above 7,500. This would be a 9.5% gain from the end of December close, not far from historic average returns of ~10%. My certainty here is lowered by the fact that the world and US economy is rather chaotic currently, but I think the safer bet is that AI has more room to run, and that will be reflected in share prices.

Will the United States experience negative GDP growth during Q1, Q2, or Q3 2026? - 36%

While I see the AI boom supporting US growth generally, the mechanism is more complex with GDP growth than with equity markets. 2025 saw isolated negative quarters of growth for the US, along with some strong ones. With one factor behind this volatility being tariffs, especially big swings in import/export activity that play a large role in changing GDP prints at the margin.

There is a very real chance that even a somewhat underlying strong US economy sees a hiccup quarter, especially as it laps some of these stronger comparables from 2025 that were somewhat artificially boosted by exceptional import/export movements. Putting the odds of a negative GDP print higher for me than that of a significant equities downturn.

US politics

How many of these 15 top US executive branch officials will be out before 2027? - 4 officials

Last year I chose to go out a on a limb by predicting that Trump and Musk would stay firm allies. It is fair to say that one did not pan out as I expected. But I broadly stand by my core analysis:

The recent spat between Sam Altman and Musk over OpenAI’s new Stargate scheme is a good example of how Trump seems to view his coterie of billionaires. In classic court politics style, they are free to snipe at one another, as long as at the end of the day their allegiance lies with him. Musk can exist fairly easily within that framework

The last line being the really unfortunate misjudgement. Though remarkably Trump went through his first year with no turnover at all in his cabinet officials, despite a host of scandals, leaks and conflicts. I do believe this reflects the royal court dynamic, where Susie Wiles as chief of staff in practice has no strong leverage over Trump cabinet members, who rather all individually need to build their own fiefdoms and routes of access to the President.

The relative cabinet serenity of Trump II versus I seems unlikely to last however. Ultimately this year is shaping up to be even messier than the last, with various scandals likely to start dragging down cabinet members with them. Whether it be Pam Bondi over the Epstein Files, Pete Hegseth waging overseas wars, or Tulsi Gabbard trying to secretly get Assad a US green card, we will start to see some consistent turnover.

Will the composition of the US Supreme Court change in 2026? - 20%

Justice Thomas and Justice Alito are the two key figures here, aged 76 and 74 respectively. Both are conservative judges who could face political pressure to step down in favour of a younger figure to even further cement the longevity of the 6-3 republican majority.

Neither however have given any indications of a desire for retirement, and have hired full sets of clerks for future terms. Given republican control of the Senate, and the extent of their majority on the court, the political pressure is unlikely to be overpowering. The main risk here is therefore mostly actuarial with both being well into their seventies.

How will the Supreme Court rule on Trump’s tariffs in 2026? - 55% mixed verdict

The relevant ruling here that on Trump’s use of IEEPA emergency powers to implement tariffs. The most likely outcome here seems to be that some of the tariffs get shot down but others are held up due to political pressures on the court. Most likely to be struck down are the ‘reciprocal’ tariffs based on countries trade deficits of the United States. Buying more of a countries goods than you import is incredibly obviously not a national economic emergency. ‘Something something fentanyl imports’ is not much more convincing but can be dressed up as a reasonable concession to the judgement of the executive branch.

This gives me a probability of 55% for a mixed ruling, 30% for the IEEPA tariffs being entirely shot down, 10% for some other form of ruling, 4% that they are ruled lawful.

Of course the Trump administration has already stated it will just reimpose tariffs through another of the various methods available to them, and given the pusillanimity of the US Congress they will almost certainly succeed.

How many days will the US government be shutdown in 2026? - 30 days median

The October-November 43 day shutdown managed to be the longest in US history, while also flying consistently under the radar in the news cycle, and only being resolved due to the contingent timidity of a group of Democratic senators. None of the dynamics have changed that produced such a long shutdown:

• Democrats don’t trust the Trump Admin, crucially in issues around impounding and whether Trump will simply backtrack on any deal they strike with Congressional Republicans

• Democrats feel a need to stand up to the Trump Admin in some form, and budget deadlock is one of their most effective means of tangibly doing so given they lack control of either House of Congress

• The previous shutdown proved politically positive with Democrats, as it saw Trump’s approval ratings sink and voters largely blaming Republicans for the deadlock, never a given in these kinds of shutdowns

The current deal runs out at the end of January though recent reporting suggests that neither Democrats nor the White House are especially angling for another confrontation. Given the scale of the shutdown in 2025 I don’t expect we see quite that length in 2026, but equally find it very unlikely that we see no shutdown at all. Even if currently Democrats are not minded to use this particular lever, the pressure to take action will inevitably rise as the moment comes, and there is no basis for a lasting bilateral deal that I can see.

Global Politics

Will Keir Starmer cease to be Prime Minister of the UK during 2026? - 45%

Starmer has hit his all time lowest approval rating at -46, and Labour is being hammered in polls. They are behind Reform and the Conservatives, and single digits away from the Lib Dems and Greens. The May elections in Scotland, Wales and local government in England risk an absolute Labour wipeout. We could see Plaid Cymru win in Wales for the first time, Labour are currently polling joint fourth in Wales behind Reform, Plaid and the Greens!

Scotland is equally bad, with Labour trailing both the SNP and Reform by a wide margin, despite the SNP’s extremely long and ineffective tenure in government. Given that Starmer already faces a multitude of poorly concealed knives, with consistent discussion about potential successors and leadership challenges, a wipeout in May could be politically unrecoverable.

Sticking to lines about having real long-term impact, and how actually things are going great - ‘look X metric of depressing British decline has marginally slowed!’ - does not cut it when people do not understand the story of your government, and increasingly have lost faith in the system entirely.

Still, the fact that Starmer does have a huge majority, and has spent years consolidating his control of the party, gives him significant advantages to extend his stay in power. It would take over 80 labour party MPs to trigger a leadership challenge against him, and the act of visibly and publicly backstabbing an electorally successful leader is a high bar for potential challengers. Many will look to bide their time, waiting for someone else to take up the cursed blade and begin the bloodletting.

Will Benjamin Netanyahu cease to be Prime Minister of Israel during 2026? - 35%

Starmer could aim to emulate the ultimate political survivor that is Netanyahu. Despite controlling a bare minimum coalition, facing a verdict of a corruption trial that could give him 10 years in prison, and facing elections in October 2026, the odds are still that Netanyahu rides out the year. He has fashioned himself into the essential piece to putting together fractured and broken Israeli politics, and has become near impossible to jettison. The election could well spell the end of his era, but most likely coalition negotiations drag on all through the end of 2026 while he continues as caretaker, putting off his potential exit into 2027.

Will there be a bilateral ceasefire in the Russo-Ukrainian conflict before 2027? - 50%

Whole lot of uncertainty here. While the US and Ukraine seem to be edging closer along with Europe to some kind of agreed set of terms, there is still a gulf to the expectations of Russia. It is very difficult to see what the acceptable terms of an agreement would be, and the politics within Ukraine are pretty murky to me. Russia has exhausted huge quantities of men and materiel into Ukraine but is capable of going on and on.

Its war economy and state have undergone full metamorphosis to bind almost all aspects of Russia’s future into its overseas conquests. There is no return to normality possible, only temporarily frozen conflict or internal economic implosion. Given how its progress has slowed, a frozen conflict could have its attractions, as a stage in its slow erosion of US and European support for Ukraine. Donbas fort lines playing the same role as Czech ones in the Sudetenland.

The criteria for this question is loose, simply the notional agreement to a 30 day ceasefire, not even necessarily that this agreement holds up in practice, which greatly raises the probability of a ceasefire even if the pathway to any kind of longer term settlement seems remote.

Will China attack or blockade Taiwan during 2026? - 9%

The odds of a large-scale Chinese invasion or blockade of Taiwan this year are very small, 1-3% kind of range. Such an operation would be of immense difficulty and complexity, and bear huge political and economic cost. It is not at all clear that Xi is looking to make such a risky gambit anytime soon. Far better to simply continue a process of slow encroachment and coercion, cutting down on Taiwan’s freedom of action, and looking to benefit from the ups and downs of Trump era geopolitics.

The precise criteria of this question leaves open some room for skullduggery. For instance very small scale kinetic attacks on the tiny island of Kinmen just off the Chinese coast would count. Or extended military exercises that deter shipping from a Taiwanese port for 5 days would count as a blockade. Both of these could conceivably occur during some kind of below-war grey zone activities. The odds are unlikely but this is how we climb form 1-3% towards 9%.

Tech and Space

Will TikTok US be banned or sold during 2026? - 93%

The deal for a majority of TikTok US to be sold to a consortium of Oracle, Silver Lake and MGX has already been signed, with it expected to close this month. Given that there seems to be political buy in from both the US and China, I expect that we might finally see the TikTok saga be put to bed, at least in this current form.

Will SpaceX successfully refuel a Starship in orbit during 2026? - 65%

At the start of 2025 Starship seemed to be on a tear. Successfully demonstrating multiple booster catches and getting Starships up into orbit. On that basis I did looked at what the near-term impact of Starship could be on the growth of Starlink and the numbers were pretty crazy. In practice SpaceX’s run of success did not last, with 2025 seeing a whole number of failed Starship tests as the v2 version struggled with a variety of issues in reaching and sustaining orbit.

Such is the risk of doing crazy ambitious projects at the frontiers of space exploration. The question then is whether a 2025 of struggles will produce a more reliable v3 Starship and that same crazy step-change across SpaceX’s business that this kind of next-generation heavy lift can bring.

SpaceX does seem to be an exceptional learning organisation, so I do expect that we will see an increasing number of Starship successes this year. A higher launch cadence, a potential double catch of booster and Starship, and first deployment of the next-gen v3 Starlink satellites. Demonstrating in-orbit refuelling is a very important step for SpaceX’s NASA obligations, and has a decent chance of success, but remains something SpaceX has not yet had substantial experience at, therefore remains fairly uncertain.

Will NASA’s Artemis II complete its mission successfully before 2027? - 75%

Meanwhile NASA is attempting its own major feat. This time not pushing any boundaries but returning us to a milestone that we have not replicated since the 1970s in a crewed orbit around the moon. Given NASA’s lack of risk-tolerance, and the long-lead times to this mission I expect it is more likely they pull off Artemis II than SpaceX cracks in-orbit refuelling. SLS might be a slightly mad boondoggle but throwing a big disposable rocket into orbit should be highly doable.

Subscribe now

Assorted other questions

For a number of the questions I don’t have anything especially interesting to write up, as they travel further outside of my own domains of knowledge, but I give them here along with my predictions for readers who are interested in seeing the full set:

What will be the price of Bitcoin at the end of 2026? - $110k median

Will Donald Trump cease to exercise presidential powers for 48 hours during 2026? - 7%

What will be Donald Trump’s net approval on December 31, 2026? - -14 median

Will there be a ceasefire in the Sudanese Civil War during 2026? - 37%

Will Saudi Arabia and Israel agree to normalise diplomatic relations during 2026? - 42%

How many of the negotiating chapters required to join the EU will Montenegro have closed at the end of 2026? - 18 median

Will the winner of the 2026 FIFA World Cup be a country that has never won before? - 20%

Will the EU require mandatory age verification on social media or AI before 2027? - 40%

Will the FDA approve a psilocybin treatment during 2026? - 25%

Will the US, UK or EU approve a gene editing therapy for a new condition during 2026? - 17%

Will restrictions on the use of the Traditional Latin Mass in the Catholic Church be loosened during 2026? - 25%

Will the WHO declare a Public Health Emergency of International Concern in 2026? - 35%

What will the average global surface air temperature be in 2026 relative to the pre-industrial baseline? - 1.37°C median

Will GTA VI be released during 2026? - 78%

This barrage of investment announcements is undoubtedly a welcome signal for the UK, and a result of extensive work getting the policy machine up and running to facilitate data centre investments via AI growth zones, and negotiate partnerships with leading firms. While this policy work has been progressing, there is still remarkably little clarity in the public discourse around what the UK’s ultimate aims are.

All this investment and various partnerships are often dubbed as ‘Sovereign AI’. Sometimes spoken of as harnessing the economic benefits of AI, building AI sustainably or training BritGPT, there is no clear sense of what AI sovereignty really means. At its worst it risks devolving rapidly into an ‘everythingism’ term, trying to meet every possible policy objective at once. Fusing itself with the very word AI, such that any plan offering merely ‘AI’ standalone would be simply unimportant, nonstrategic, risible.

This would be a critical failure for the UK. The monumental scale of the challenge that the development of increasingly powerful AI systems poses, and the corresponding scale of resources and state capacity the UK has to meet them, demands an unprecedented level of strategic clarity and focus. The UK needs to convert its advantages into economic gain and strategic leverage, risking broad irrelevance if it fails to do so.

I argue that AI sovereignty for the UK is a very narrow metric. It means ensuring our freedom of action to manage and regulate how AI affects us here and how it is governed internationally. The primary barrier to exercising that sovereignty is going to be our one-way dependence on the US AI tech stack. I believe we cannot break or limit that dependence over the next 5-10 years of rapid AI progress, rather what we need to do is change the relationship to one of interdependence.

A UK sovereign AI strategy is one that increases US dependence on UK capabilities. We need to work towards a future where the UK has carved out for itself a dominant position in lucrative and strategically significant parts of the AI value chain. Such that when crunch time comes, and it really matters who is in the room, the UK is understood to be one of those live players.

I have a case for how the UK can best achieve this, though compute is not an essential component. Compute still plays an important role, but largely as a means not an end. Building up a strong stock of public compute is essential to any AI industrial policy, and a reasonable ecosystem of private compute based in the UK, and run by UK neoclouds, is useful for UK national security. The challenge is that going beyond the thresholds of specialised inference and narrow training capabilities presents quickly diminishing marginal returns in terms of the strategic value that this investment is providing. An extra couple of MWs, depending on foreign capital, and being run by US providers gives little real ‘sovereign’ value to the UK. With large-scale compute investments only becoming again valuable once they enable the ability to train models at the current frontier, a level that the UK will not be able to reach.

I am further doubtful of approaches that lean on hardware capabilities such as UK-designed chips or other stages in the semiconductor supply chain, and instead am excited by opportunities for the UK to build a web of software dependencies. Especially in future nodes of the AI value chain that I group together as ‘AI Middleware’ that will become increasingly essential as AI diffuses deeper into enterprises. That will be subject to a follow up post that will explore where UK opportunities are in the value chain, and why AI middleware could be an especially promising approach.

The ultimate means of these approaches, and ultimately any AI sovereignty effort, is shared. The UK needs a focused and aggressive industrial policy. The UK government must pull every lever of capital, compute and talent to give UK firms a leg up in the new era of AI mercantilism. It means making high-risk, high-commitment bets, and accepting the waste, mistakes and failures that will naturally follow. It will require strong political will and focus, prioritising the strategic necessity of this industrial policy over a host of other goals. The follow up to this post will explore in more detail on the exact policy mechanisms the government should be exploiting, but the top three in my current view are:

• Being willing to use the powers of the National Security Investment Act to block potential acquisitions of promising AI startups, including potentially augmenting the Act to block acqui-hires

• Carve off £2-3bn of capital from the National Wealth Fund for a new fund tasked with investing in strategically significant AI firms, aiming to flood the space with capital

• Execute strict prioritisation of public compute resources, with a strong bias towards the Sovereign AI Unit allocating compute towards strategically useful private firms rather than academic users

Subscribe now

What does AI sovereignty mean for the UK?

The UK government has not offered much clarity on this front. The AI Opportunities Action Plan called for a ‘Sovereign AI Unit’ with the mandate of “maximising the UK’s stake in frontier AI”, one of a number of different gestures towards the idea of what AI sovereignty should be:

• The UK should be “an AI maker not just an AI taker”

• The Sovereign AI unit is to “secure the UK’s future as a sovereign AI nation in the context of rapidly advancing AI capabilities”

• Or put another way to “build and harness the UK’s AI capabilities to unlock economic growth and enhance UK national security”

The UK’s compute roadmap offers perhaps the best definition, stating that sovereignty is not “isolation or self-sufficiency” but rather “sovereignty means ensuring we have the ability to act independently and effectively where it matters most”. Sovereignty is about widening the UK’s freedom of action, and giving it maximum leverage over how AI develops and diffuses.

Ultimately there is no single ‘AI sovereignty’, it is not a simple question that can be easily patched with a few thousand more Nvidia chips or a shiny new AWS data centre. Fundamentally states outside of the US and China are set to be dependent on these two major players for the most transformative technology for generations. AI sovereignty is simply this strategic dilemma, and states will respond differently depending on their context. Some will bandwagon, some will balance. States will vie for different nodes of the value chain, and prioritise differently depending on their values. For many developing states the desperate need will be simply to not see the ladder of services-led development kicked out from under them. Smaller states will look to protect languages and cultures from the homogenising power of internet-scale datasets.

The UK can afford to be more ambitious, looking to more fundamentally steer the development and governance of the technology itself. As well as securing strong economic and geopolitical upside from the diffusion of AI systems through its economy and government.

I operationalise AI sovereignty for the UK as therefore primarily being about making the US dependent on unique UK capabilities. The more that the UK can change a dependency on the US into a mutual interdependency, the greater security and leverage it will have.

Each firm, product and service that the US comes to rely on the UK for, chalks up another ‘unit’ of AI sovereignty for the UK. The UK can then cash in those units when it matters. Think of the UK avoiding any compute threshold in a resurrected AI diffusion rule, or being able to place more stringent pre-deployment safety testing requirements on US model developers.

The reality is that the US dominates much of the AI value chain now and for the foreseeable future. The UK is on track to be highly dependent on US capabilities, from chips through to cloud services, foundation models, and applications. Dependence is not in and of itself a bad thing. Simply adopting the US tech stack is the fastest way to AI growth, enhancing public services or building an AI warfighting capability. The challenge is if this relationship becomes too one-sided, rather than being one of mutual interdependence.

AI sovereignty is distinct from other potential goals relevant to AI, whether those be economic growth, military diffusion, or accelerating scientific development. AI sovereignty will require tailored policies, and while it may often mix well with these other goals (e.g. seeding UK AI defence startups) it may also often trade off against them (e.g. coaxing UK advertising firms to use a UK’s startups capabilities over a currently more capable US one). Ultimately AI sovereignty is about building dependency and leverage.

Sovereign AI is not technological independence

The scale and sophistication of the AI value chain rule out strict interpretations of sovereignty as technological independence. China has invested by far the most of any state towards this end, and still faces a number of critical and hard-to-overcome dependencies in lithography equipment or HBM production. The UK quite simply isn’t big enough to come even close to technological independence.

The UK compute roadmap tries to weave between ambition and reality by targeting an eventual ‘full UK-designed compute stack – from chip to system to software’. Given that beyond the chip level much of this software stack is open-source or extremely hard to replicate, this really boils down to a desire to see UK-designed chips. The question is how far does a UK-designed chip give you any real additional ‘marginal sovereignty’ if it remains so strongly dependent on a wider US tech stack.

Given the current US mercantilist AI policy where the emphasis is on full-stack exports, any scenario where the US is cutting the UK off from chips it designs, it is very likely cutting us off from networking equipment, EDA software or TSMC. There is no component of the AI value chain where you can achieve independence in any meaningful sense. A UK chip firm is plenty justifiable on the basis of the potential economic returns, but it has no real strategic logic.

Independence would also be undesirable for the UK. China stands to lose out a great deal from being cut out from global semiconductor and AI supply chains, as specialisation remains an incredibly powerful economic tool. The UK should want to leverage the excellence of TSMC or Nvidia, and ensure it can put the global frontier models to work in its economy.

Sovereign AI is not chasing independence, but rather not being naive about the risks that dependency brings. Building UK capabilities as a hedge against undesirable levels of dependence, and ensuring the UK has as large an option space as possible in steering through the changes that transformative AI can bring to its economy and strategic position. Given the UK’s position firmly within the orbit of the US, the most effective hedge is in building unique UK capabilities. Providing useful tools to the US brings positive influence, while also enabling leverage via the possible threat of their withdrawal.1

What role should compute play in our sovereign AI strategy?

Compute is essential to developing and running AI systems, with the continual rapid scaling of the amount of compute invested into training AI one of the key features of its continual rapid progress. With compute being understood as the ‘lifeblood’ of AI it has naturally come to play a central role in the devising of AI sovereignty strategies. Nvidia et al. have done a great deal to sell ‘sovereign AI’ as primarily being a process of building large amounts of compute.

When frontier AI training clusters were low tens of thousands of chips, costing only some few hundreds of millions in capital expenditures, there was a reasonable case to make. For most developed states this was far from an insurmountable amount of expense for the benefit of catapulting them to the AI frontier. The challenge is that the frontier is accelerating at extreme speed as big tech firms funnel over $350bn into AI infrastructure this year, with this figure growing into 2026 and beyond.

Given that compute investments no longer realistically mean reaching the AI frontier, we need to ask ‘compute for what?’. How does compute actually factor into a sovereign AI strategy for the UK?

There are three relevant categories of compute for the UK:

• Public compute - that owned or allocated by the UK government - this is what the government has itself termed ‘sovereign compute’

• Private compute - this is compute owned and operated by private firms but located within the UK

• Overseas compute - this is compute located overseas and owned/operated by foreign providers but which the UK can rent

Public compute is essential and valuable, giving the strongest strategic returns

Public compute is the most crucial component of the UK’s strategy. It is the resource that the government can use to prioritise high impact academic and commercial research. For seeding promising startups, the ability to provision significant compute quickly is a potent tool. The UK’s stock of public compute is currently fairly small relative to the commercial frontier, with the freshly opened Isambard-AI boasting 5,448 H100 chips. That is around 8MW worth of chips, rendering it invisible on the chart above versus the hundreds of MWs being fielded by US firms.

The stated ambition of the government is to 20x the size of this AI Research Resource by 2030, which would mean roughly matching in 2030 the scale of the frontier in 2024, at 100k H100 equivalent chips. Given the ongoing 30% annual improvement in price/performance of AI hardware, this would mean approximately £1bn of investment out to 2030 in nominal terms.

This scale of investment at least gives the UK something to work with, but could do with dialling up in ambition if we truly believe that we are on a fairly accelerated pathway towards AGI. If the expectation is for truly powerful AI to have appeared by 2030 you could at least double or triple this figure. Most likely via the UK government renting large amounts of compute from private sector operators in the UK or abroad, rather than through self-building more supercomputers in conjunction with universities.

Private domestic compute is less strategically valuable, but depends only on policy execution

Compute built in the UK, but owned and operated by private firms has far less clear utility than public compute. At a high level there is a case to be made that the spillovers from AI adoption are externalities that the private market will not capture, and so it will underinvest in the infrastructure needed for widespread diffusion. This appears to not be the case currently, with very large amounts of capital flowing into this space, and the build out of data centres primarily constrained by available chips and energy rather than financial resources.

Moving down levels of abstraction there are a few potential advantages to domestic private compute:

• You can foster UK-based hyperscale operators

• It provides further ground for a ‘pull-through’ mechanism for UK technologies

• Investment is always welcome

• The UK can fall back on this compute if cut off from external resources

There is a key prerequisite however for any benefits from private compute.

Bypass the UK’s broken planning and energy systems

The UK’s current strategy recognises this, with AI growth zones being created to help circumvent many of the issues with planning crippling most large scale infrastructure projects in the UK. Ensuring that the UK at least has the ability to build data centres if private capital and firms want to do so, is a robustly strong policy approach. The challenge here is almost entirely one of political will and focus. The Department for Science, Innovation and Technology needs to be empowered to carry through on its strategy, and fight off the host of veto holders that usually doom UK projects.2

This requires UK political leaders to do something very alien, and actually face up to difficult trade offs. Climate is by far the biggest sacred cow to slay. Building data centres at any reasonable scale in the UK is going to require some backup gas capacity, in direct contravention of climate goals. Climate is undoubtedly a serious issue, but ultimately the weight of evidence clearly stacks up in favour of AI. Technology capable of profoundly reshaping our economy and strategic outlook over the next five years beats out a rounding error’s worth of global emissions. Data centre energy demand also provides the perfect lever for accelerating the deployment of small modular nuclear reactors as well as restarting new gigascale nuclear efforts, arresting the UK’s energy immiseration and giving us the firm power we need to fully transition off fossil fuels in the future. Recent announcements were encouraging on this front, but there is certainly a long road ahead in bringing these to reality.

Choose UK operators over involving UK capital

A key question for private compute build outs is who do we want ultimately owning and operating these compute resources? Relying almost entirely on US hyperscalers or Middle Eastern capital feels intuitively opposed to a ‘UK sovereign’ strategy. Though ultimately this question mirrors the broader one of how much compute the UK really needs. One or two UK or European based operators are an effective hedge against dependence on the US, and useful for the most sensitive national security workloads. But beyond this necessary minimum, policy should be driven by the economics of AI data centres, which appears currently somewhat unappealing.

One potential view is that UK capital needs to be prioritised such that we don’t lose out on this component of the AI boom. In reality, the case seems closer to we should really want others holding the bag for what are highly risky and speculative investment plans.3 The bulk of data centre investment is coming from the US hyperscalers, who can still relatively easily fund this expense from their huge cash flows. Here there is very little reason or ability for UK capital to play a role.

Outside of the hyperscalers, investment is coming primarily from ‘neoclouds’, startup wannabe hyperscalers who are plunging billions into renting or acquiring data centre space and power, and splurging even more to fill these with Nvidia chips. The UK is home to one promising neocloud candidate in Nscale, which recently signed a major Stargate deal with OpenAI in Norway and now one for the UK. It also hosts Fluidstack, a GPU aggregator platform drawing on these many various smaller neoclouds. Europe also has a number of neocloud contenders looking to take advantage of its compute build out.

Ultimately the neocloud business is likely to prove a scale game, with a few contenders winning big (current favourites being Coreweave in the US, Nebius in the EU), and a long tail of smaller providers struggling to stay afloat with paltry margins. The dream neocloud business case is getting nice early allocations from Nvidia, spinning up large clusters of chips, and selling these to major AI firms at stable, preferable prices in multi-year deals.4 The reality for most neoclouds will be aggressive commoditisation in the on-demand pricing market, getting new chips too late to benefit from initial scarcity prices, and begging your accountants to not change their useful life assumptions.5

If major controllers of UK capital want to take on this risk, they are welcome to do so, there aren’t exactly many other good sources of return in the current UK economy. But there is no strong reason to think that the UK government should try to ensure UK capital doesn’t get ‘left out’ of this process. Lower-risk areas of the value chain e.g. data centre construction and hosting are perhaps more attractive in terms of risk-adjusted returns for infrastructure capital. Generally the more lucrative portions of the value chain are currently upstream (Nvidia/TSMC) and will increasingly be downstream (high-margin cloud services, AI applications and their subsequent productivity benefits).

The capital provider is almost certainly less important for security and strategic purposes than the nationality of the operating company. While capital can provide control over corporate governance and strategy, and indirect influence through the lure of further capital access, its role is much more arms length than that of the operating firm. The real security and data privacy risks lie with the operating company. The UAE cannot ‘take back’ capital it has already sunk into a UK-based data centre, the US could potentially significantly disrupt the UK economy by putting pressure on US cloud providers.

For critical UK government and national security workloads there is a very good case to make that some provisioning (or at least the option of provisioning) from a non-US cloud operator is wise. Most promising are interventions to reduce the capital costs of UK-based neoclouds. Since 70-80% of the total cost of a GPU cluster is in the capital rather than operational costs (where the UK suffers from expensive electricity and planning), effective interventions to reduce this capital cost could make UK providers more price competitive. These could be direct financial support, loan guarantees (potentially via the National Wealth Fund) or anchor commitments from public compute investments or other public sector compute/cloud consumers.

Data centres are investment and that is always welcome

Data centres are large capital investments that generate a reasonable number of construction jobs and spill over effects into UK industrial supply chains. Ideally local governments would be able to capture the business rates they generate, to compensate for the underwhelming local economic benefits from operation, rather than having these flow to the treasury, but even absent this mechanism data centres are not heavy costs on local communities. Land, water and energy are real costs, but they don’t generate significant pollution for instance, and can be integrated into local heating systems to better utilise their waste energy.

In almost every case receiving this investment in the UK will be better than not receiving it, and small growth boosts are still welcome given the UK’s prevailing macro conditions. Though in strategic terms this effect is largely irrelevant.

Compute is a useful but incomplete hedge against geopolitical uncertainty

One potential advantage of fielding large amounts of private compute in the UK would be that in cases where the UK is cut off from access to foreign compute, it has a stock under its jurisdictional control on which it can rely. This again is most importantly an advantage for the case of ensuring some UK-operated and domestic based compute, sufficient to serve critical national security or public sector needs.

Once we pass this threshold we return to the challenge of how US-dominated the AI value chain is. In cases where the UK is cut off from European or Chinese compute, this is likely a small and short-term effect, with the UK unlikely to be buying compute at large scale from either. The greater concern is a fall out with the US, or a wider US-dominated ecosystem (e.g. if we were purchasing AWS compute located in the UAE). Being able to switch to using private compute in the UK instead seemingly gives a hedge, but in reality much of this private compute would still be operated by US firms, and dependent on access to US AI models. If the US can remotely hinder our ability to effectively use this compute, and potentially cut us off from the models that make this compute actually useful. Then our hedge has brought us very little benefit.

In any scenario where the UK government is seizing control of Azure data centres, the US government would be taking commensurate actions against our ability to effectively use this compute. Rendering simply geographic possession of the chips fairly useless. Having geographic diversity in our sources of external compute seems a stronger approach to mitigating this risk.

Ultimately how effective a stock of UK compute is depends on some underlying assumptions around how rapidly AI is progressing, and how intensive the US limiting action is. Compute is a more useful hedge where AI progress is slower, and where the US cut off is not complete and the UK can still access some important compute inputs even if not chips.

What about being at the frontier?

One way that private domestic compute could be highly useful is if it enables a UK-based frontier AI developer. If the US were to cut us off from model access, we could simply train our own. The challenge is that this requires a cluster at sufficient scale to match those being built in the US. Given lead times, and the existing uncertainty around how well the UK can execute on planning reforms, it seems essentially inconceivable currently that the UK could build such a frontier cluster by 2030. Anthropic has assessed that the frontier will be 2GW in 2027 and 5GW in 2028, with likely further multi-site integration the norm into 2029/30. This is simply a colossal scale of infrastructure, which the UK cannot race with from its current bewildered and belated standing start.

The UK also lacks a frontier AI developer capable of using this infrastructure. While previously Deepmind might have been able to be understood as a ‘UK-based’ frontier developer, this is no longer. Post its integration with Google Brain, despite being headquartered in the UK, it is far more tightly integrated into the wider Google superstructure, with a more internationally distributed staff. It is also tightly bound into Google’s infrastructure strategy, building a >GW scale multi-campus training cluster in Iowa. The scale of this investment and how tightly integrated available compute is to frontier training capabilities, mean that dislodging Deepmind from a wider Google structure is likely no longer theoretically possible.

The UK should learn from the experiences of France and Canada in their support of developers vying to be at the frontier. Both Mistral and Cohere are severely lagging behind the leading US developers both in financial terms with poor revenue performance, and increasingly in model performance. They lack the symbiotic relationship with tech giants that benefit OpenAI and Anthropic, and have struggled to build useful or lucrative products.6 Putting significant resources behind an unbelievably challenging run of the red queen would be folly.

Overseas compute is a useful final piece of the portfolio

Much of the UK’s compute needs will be met with supply based outside its borders. Given that the UK is likely to face a persistent cost disadvantage with its domestic compute build outs, and the relative unimportance of latency considerations, this is not a major strategic issue for the UK. As long as the UK can meet the minimum threshold it needs for running sensitive and secure inference needs, then there is no real strategic advantage to whether additional compute is based within the UK or coming from outside.

The UK will remain expensive, but not insurmountably so

As explored earlier, the bulk of the total cost of building and running a GPU cluster is in the capital costs of acquiring the chips. In most cases this will mean purchasing from Nvidia or AMD, where the cost is not meaningfully higher if the cluster is in the UK or UAE. The exceptions to this being hyperscalers, notably Google and Amazon, that deploy their own internal AI accelerators, cutting out Nvidia’s very healthy margins.

The key driver of cost differentials is instead energy. This means both the industrial price of electricity and the cost of building backup gas or battery storage. The UK does especially poorly in terms of the industrial cost of electricity.

Norway & UAE are rates for energy-intensive industry, 2025 data for US, 2024 for the rest7

This represents a meaningful harm to the UK’s cost competitiveness, though it can be somewhat compensated for via other means of support, for instance the capital support discussed above. The UK is relatively more competitive in the cost of building new gas plants, though it still lands at a significant premium to the UAE, and a smaller one to the US.

EIA estimates for the US, Al Layyah plant for the UAE, Watt Direction estimate for the UK

Given the cost differential between UK compute and that located elsewhere, the UK is likely to want to import a decent proportion of its compute from overseas, but is by no means barred from substantial amounts being based in the UK.

Latency is unlikely to be an important constraint on compute imports

Latency is primarily going to be a consideration for consumer internet platforms integrating AI, where tens of ms differences can translate into meaningful costs to engagement and conversion. While currently most AI tokens are likely consumed in consumer usage in ChatGPT and Google search we should expect this to shift over the next few years to become more strongly enterprise weighted.8 For any use case involving larger models, or reasoning models the geographic latency will rapidly become a rounding error to the second of delay before models produce output tokens. This will be especially true for agents and systems working asynchronously. Co-workers are high latency tools, and the same will be true for most AI enterprise solutions.

For these consumer platform use cases it is also entirely adequate to import this compute from the EU, where latencies remain low, though the US or UAE starts to become more problematic. For instance, the UK probably does not need OpenAI compute more closely located than Stargate Norway, which will be entirely adequate in latency terms.

Latency from London, Gemini TTFT

The other candidates for low-latency requirements are industrial or robotics use cases. Generally I view these as heading towards specialised on-premises or on-device solutions to achieve maximally low latency, which might be necessary e.g. for robotics. The inbetween of relying on a large data centre in close but still tens of ms distance, seems an unlikely solution to win out.

Given the relative lack of latency-sensitive workloads in the UK’s future demand profile for AI, we should be more comfortable simply importing this compute from areas where it is cheaper. It is likely wise to avoid too many single points of failure e.g. heavy dependence on the US or UAE, as disruption to these data centres or subsea cable infrastructure could be problematic. Rather the UK should be comfortable importing compute, from all three of the US, Europe and UAE.

Compute is a means not an end

Compute is an essential component of the UK’s AI sovereignty but largely as an instrument for enabling the UK to build strengths in other areas of the AI value chain. Building out AI compute is not in and of itself a recipe for sustained or meaningful UK leverage over other states, and does not alter the dynamics of the UK’s heavy one-way dependence on the US AI tech stack.

Altering this dynamic will require building UK strengths in areas where we can generate interdependency, by building successful national champion firms that bring UK tech abroad. Though this is an immensely challenging task, as we need to make good bets under significant technical and economic uncertainty.

The UK’s strengths sit outside compute

Ultimately compute is a means not an end for the UK to build AI sovereignty. Instead we need to look to our real advantages in an abundance of talent, strong institutions and a potent digital services economy. Given a sufficiently focused and aggressive industrial policy we can build powerful positions in key future nodes of the AI value chain and ensure that Britain stays in the room when that crisis comes. Interdependence can be much more powerful for the UK than compute infrastructure. Do subscribe to read all about what the UK’s sovereign AI portfolio should be when that follow-on piece drops!

Subscribe now

I worked on this piece as a Summer Fellow at the Centre for the Governance of AI, which was the perfect environment to do so. My greatest thanks go to all those at GovAI, staff and fellows, who kindly took the time to read various drafts and debate these ideas with me. As well as to all those outside GovAI I was able to meet and think through these arguments with!

Yet Singapore's experience also reveals the fundamental limits of digital sovereignty in the AI era. Even this exceptionally wealthy and efficient city-state, blessed with geopolitical positioning that most nations cannot replicate, remains forced to rely on foreign infrastructure for cloud computing, semiconductors, and foundational AI technologies controlled by US and Chinese ecosystems.

Singapore's case thus illustrates both the possibilities and constraints facing smaller nations in technological bipolarity. While countries should indeed seek specific niches within the AI value chain based on their national characteristics to maintain relevance and influence in global policy direction, even the most successful middle powers will face technological dependencies on US and Chinese infrastructure. Singapore's experience shows that maximising agency within these constraints rather than pursuing impossible independence offers the most viable path forward, though success requires the rare combination of economic prosperity, institutional efficiency, and strategic positioning that few nations possess.

(Memorial to Admiral Sir Clowdisley Shovell., Sebastiano Ricci; Marco Ricci (1725))

Singapore’s AI policy

Ostensibly, Singapore appears poorly positioned for AI leadership. Its tropical climate and limited available land make compute-intensive operations expensive, with a population of only 5.4 million leading to a lack of scale in terms of data generation. However, rather than attempting impossible technological sovereignty, Singapore recognises the nation's inherent constraints and focuses instead on gaining a strategic advantage over aspects of the AI value chain. In doing so, Singapore's outsized relevance on the global AI stage originates from positioning itself as a research-driven melting pot where applied innovation flourishes, enabling Singapore to lead regional AI development while building critical capabilities in R&D and talent cultivation.

Importantly, Singapore’s early recognition and swift pivot to AI was not a flash in the pan. Rather, it originates from decades of state-led digital planning. From the 1980 National Computerisation Plan through IT2000 and Intelligent Nation 2015, methodical institutional development enabled decisive action when AI emerged as critical technology, launching its National AI Strategy, the first country to do so, in 2019 and committing a further $743 million through NAIS 2.0 in 2023.

Building on efficient regulatory policy, the Singaporean government’s stewardship extends to direct deployment, running over 200 AI implementations across healthcare, transport, and urban planning. These government entities serve as production testbeds that attract international investment while building practical expertise, with public sector demand de-risking enterprise adoption for commercial markets. This has created an exceptionally efficient regulatory environment where AI companies can be established within 24 hours and ethical guidelines like the Personal Data Protection Act are quickly implemented, a result of seamless integration between AI R&D and government deployment.

The results of Singapore's effective government policy have been to increase the levels of invested capital in AI to the point where, according to CSET, government supported R&D spending as a percentage of GDP now exceeds the United States eighteen-fold. However, it has also allowed for Singapore's continued position as a neutral territory where both the technological superpowers can operate. Major American firms like Google, Meta, and Salesforce have established AI research centres alongside Chinese giants including Alibaba and Huawei, creating a unique competitive advantage where Singapore benefits from knowledge transfer and long-term investment from both technological superpowers simultaneously. In doing so, Singapore positions itself at the coordination layer of the AI value chain, becoming an intermediary investment hub where both superpowers conduct research, test applications, and access regional markets.

Ultimately, Singapore's achievements demonstrate how smaller nations can secure outsized influence by leveraging institutional efficiency and financial capital rather than competing simply on raw compute scale. The city-state has transformed existing strengths into a unique value proposition as a staging post for global AI development, occupying the critical applied research and governance layers of the AI value chain. Rather than building costly hyperscale infrastructure or training frontier models, Singapore has become the essential platform where AI technologies transition from research to real-world application, where global talent congregates, and where both superpowers can operate without triggering bilateral tensions. However, success illuminates fundamental constraints: Singapore's influence extends only to the application and policy layers of the AI stack, whilst remaining entirely dependent on American and Chinese infrastructure for cloud compute, semiconductors, and foundational models that ultimately determine technological sovereignty.

(Lawrence Wong | Prime Minister of Singapore)

So how much do the United States and China dominate the AI value chain?

Put simply, the answer is overwhelmingly. Far from democratising innovation, artificial intelligence has reinforced a global digital hierarchy with the United States and China now controlling key layers of the AI value chain, from semiconductor design and cloud infrastructure to training data and deployment platforms. This dominance is not incidental but rather the result of scale, capital, and system-wide integration that few other nations can realistically match. As AI systems grow more complex and capital-intensive, the barriers to entry continue to rise, leaving most countries structurally excluded from full participation.

At the hardware level, semiconductors reveal the depth of this dominance. The United States and its allies control over 90 percent of global semiconductor manufacturing equipment, and American firms such as Nvidia and AMD lead in chip design, with Nvidia alone capturing more than 80 percent of the data-centre GPU market in 2023. Yet the fabrication of these chips takes place primarily in Taiwan , through TSMC, who manufacture the majority of the world’s most advanced semiconductors. While the United States consolidates its advantage through design leadership and control of chip design software, it relies on a network of aligned states for fabrication. China, by contrast, is investing heavily to localise its entire semiconductor supply chain, though it remains several generations behind. For smaller nations, the capital required to participate meaningfully in this layer—often exceeding ten billion dollars for a single fabrication facility—renders entry effectively impossible.

Cloud infrastructure further reinforces this asymmetry. As of 2024, only 32 countries host AI-ready data centres, with American firms such as Amazon, Microsoft, and Google controlling over 65 percent of global cloud services. Collectively, they plan to spend more than 300 billion dollars on AI infrastructure in 2025. In parallel, China is scaling its domestic compute capacity through state-backed firms such as Alibaba Cloud and Huawei, supported by industrial policy that prioritises technological self-reliance.

Data and model development add another layer of structural advantage. American firms benefit from the global dominance of English and expansive user bases, enabling the collection of vast, high-quality datasets that improve model performance over time. Chinese companies, drawing on a domestic population of 1.4 billion, generate unmatched volumes of Chinese-language data through the firms such as Baidu, Tencent, or Alibaba. Even under export controls, firms such as DeepSeek and Moonshot AI have released models approaching international benchmarks. Both superpowers also control the tools and frameworks on which global AI development depends such as TensorFlow, PyTorch, and cloud-based APIs, further embedding their dominance throughout the software layer. For most countries, the absence of linguistic reach, platform ownership, and domestic scale limits any realistic pathway to sovereign model development or deployment.

To be clear, other nations play important roles within specific layers. Taiwan is essential to chip fabrication, Japan leads in robotics, and the European Union has set global standards in AI regulation. Yet without direct control over compute, data, or foundational models, these contributions remain partial. The AI value chain is tightly coupled, and each layer depends on upstream integration controlled by the superpowers. This consolidation reflects more than just temporary leadership. It is a structural outcome shaped by economies of scale, population size, network effects, and sustained capital investment. For every other country, the costs of competing across the full AI stack now exceed the bounds of national feasibility. Even targeted efforts at localisation tend to remain dependent on US or Chinese infrastructure at some point in the chain.

The result is a deeply uneven AI economy. The gap between the two superpowers and the rest of the world is no longer merely one of capability but of possibility. For middle powers, the question is no longer how to catch up, but how to remain relevant within a value chain that has already been consolidated.

So what can ‘middle powers’ do?

Singapore’s experience delivers a crucial lesson for middle powers: achieving full technological sovereignty across the entire AI stack is not only unrealistic for most countries, but also unnecessary. Instead, the best way to maintain agency and influence is to identify strategic opportunities within the AI value chain that match national strengths and circumstances.

The routes taken by the United States and China rely on vast internal markets, tremendous financial resources, and highly integrated ecosystems. Fundamentally, middle powers cannot hope to match these advantages in the foreseeable future. However, not all hope is lost, with Singapore’s example illustrating how it is still possible to shape the global AI landscape by efficiently and effectively focusing on targeted areas. Even without direct control over foundational infrastructure or frontier models, smaller nations can secure roles of real significance.

At the same time, it is essential to recognise that Singapore’s approach is shaped by factors that are not replicable for most countries, including its history, scale, institutional efficiency, and geopolitical position. Rather than attempting to copy Singapore, other middle powers must carefully evaluate their own demographic, economic, and institutional strengths to design national AI strategies that suit their specific contexts. This might involve focusing on specialised R&D hubs, fostering startups, or playing a convening role in regional or international AI collaboration. By identifying and investing in these tailored strategic niches, middle powers can maximise their agency within the AI value chain and avoid being left behind completely by the increasingly dominant US and China.

Ultimately, the core lesson is that middle powers do not need to compete across every layer of the AI stack to remain relevant. Instead, they should invest in complementary capabilities and seek partnerships where their contributions add unique value to the currently dominant ecosystems. Singapore’s pragmatic focus on leveraging what it does best, rather than pursuing unattainable self-sufficiency, provides an important example. For other middle powers, long-term influence in the AI era will depend not on matching the superpowers in every dimension, but on wisely selecting and developing areas of true differentiation and strategic advantage.

Subscribe now

While the rest of big tech is plunging hundreds of billions into AI infrastructure, and reorienting their businesses towards what is clearly the most important technology of this century. Apple is spending single digit billions a quarter, and is twiddling around with generated emojis, restructuring apps and exploring the technological cul-de-sac of the Vision Pro.

Apple’s dithering has left it two valid options. Accept its backward AI position and cede control over the fundamental software of its ecosystem to an external partner. Or do what Apple has always done best, and find what it needs in the talent and industry of China.

It might already be too late

While Apple has fallen woefully behind on AI, its core business is solid, it is not going to collapse and can live a relatively long profitable existence, but its stance as one of the titans of big tech is absolutely under threat. On hardware it still excels, with a world-class chip design team. It is notable that if you are someone with too much time and too much money, that for some reason really wants a local LLM, you are probably going to buy a flavour of mac. Unfortunately Apple has completely missed the boat on AI software.

Apple Intelligence sort of made sense, you design the local LLM and offload when you need to, trying to get the best of both worlds. Though in practice you get pretty much nothing, as the local model is deeply unintelligent, the server model never worth querying, and ChatGPT integration highly awkward.

Getting AI right really matters! While Apple Intelligence is just a weird dud for now, and broadly has zero effect on the wider utility of my phone, this is going to change. AI is becoming ever more powerful, and is now truly coming into its own in the ability to autonomously handle tasks. Delegating to an asynchronous AI agent is going to become increasingly useful, and AI will edge closer and closer to being a primary interface for many digital tasks.

Phones are really useful, they aren’t going away anytime soon, but ecosystem lock-in only goes so far. The more that Apple clearly lacks useful AI features that its competitors possess, the weaker its position will be. Apple is already behind on useful features such as AI image editing, and is likely to fall further behind in the realm of intelligence assistants and AI search. Then when paradigms do increasingly shift - the top bet right now appearing to be minimally intrusive AI-glasses - Apple will find that maybe it can make the hardware happen, but it will lack the fundamental AI capabilities that lets this new hardware paradigm flourish.

To regain its AI relevance it needs both a world-class research team and access to solid frontier scale AI infrastructure.

So they just buy Anthropic?

One result of Apple’s delays is that its options are increasingly withering away. OpenAI and Anthropic are rapidly growing out of where Apple would be an attractive partner let alone buyer, while other startups increasingly fall further behind the frontier.

Partnering with OpenAI for Apple Intelligence seemed to be a first step towards greater things. Apple taking OpenAI under its wing, showing the up and comers how it is done. In practice OpenAI is increasingly unable to work with its closest and far more obviously synergistic partner Microsoft. Buying out Jony Ive’s firm and moving into hardware shows OpenAI to be on a path that almost certainly takes them off the cards for Apple.

So it's Anthropic right? Anthropic is much more vibes-aligned with Apple. Its next reported funding round only takes it to a $150bn valuation. Not cheap, but given the circumstances a pretty good deal. You get some great value for that $150bn, a world-class team, rapidly growing business, and a potent model to plug into the Apple ecosystem. Claude for the masses, they’re gonna love it.

Unfortunately it is far from clear that Anthropic would sell. Anthropic desperately needs capital to fuel its ambitions to be at the frontier, this is reportedly why it is now considering Middle Eastern capital, but it is far from clear that it needs Apple to access a sufficient scale of capital. OpenAI’s most recent round is apparently five-times oversubscribed, Anthropic’s revenue is growing at a bonkers rate, and it has a commitment to an AI safety mission, and an ambition that goes far beyond iPhone assistants.

Amazon (and Google) also got there first. Amazon has put $8 billion of investment in, and pushed Anthropic to tightly integrate into its infrastructure stack. Hundreds of thousands of Trainium chips, and eventually GWs of power are Amazon’s offer to Anthropic. The reality is that money is pretty easy to find, credible frontier AI training clusters are not. If Apple were to buy Anthropic it would itself be ending up dependent on billions of dollars worth of AWS compute capacity. More importantly it would find its future existentially dependent on AWS building out tens of billions of dollars more compute for Apple’s exclusive use!

So if it's a no on Anthropic, who have we got left? Meta has poached anyone worthwhile who was willing to be poached, so a mad recruiting drive is out the window. xAI seems rather unlikely to be selling. Mistral is both not worth the money, and would have the whole awkward situation of the French government and European sovereign AI ambitions. Cohere is perhaps worth even less, frankly Canada should probably be glad to sell it off, but they don’t quite seem to recognise this reality yet. This takes you to the end of the list for true model developers, only one option remains.

Apple goes (back) to China

Truly how else could it have gone? What else would so evocatively match the long-curve of Apple’s history. If you can’t make it in America, you’ve gotta go to China.

China has a wealth of good options. Incredible talent and a moribund venture capital ecosystem leaves a number of excellent potential bargains. Deepseek, as the newly minted national champion, might be a little ambitious. Moonshot could be more on the cards, and is very credibly close to the frontier with its Kimi models. MiniMax or Zhipu would both be excellent options.

If going the acqui-hire route then poaching wholesale teams from Alibaba, ByteDance or Tencent would all be excellent approaches. Luring this talent to the US is likely impractical, but Apple is already dropping $250 million to expand its Singapore office, and could relentlessly work to bring in as much Chinese talent as possible. Worst case scenario, an office in Shenzhen/Shanghai, and simply triple whatever it is you are currently spending on lobbying and PR.

Given the right framing they could even get Trump’s support.

“These shady TikTok guys are building huge datacentres in Malaysia with their Oracle cronies, now we don’t want to lose those great American tech exports, so how about we just take over all those Nvidia chip orders?”

“Why would we want all these Chinese maths nerds building AI for China when they could be building AI for America?”

With GWs of AI capacity in South East Asia, and a cracked research team in Singapore, Apple could bring itself back as a live player. It would be able to design competent local models, and reasonably capable larger models for its cloud servers, and could still purchase some additional model innovations from OpenAI or Anthropic if necessary.

Crucially it would have preserved its strategic autonomy, it would have a viable outside option from any partnership it entered into, and could pursue technical pathways that leveraged its specific hardware advantages, rather than being caught in a race of simply chasing Google innovations.

The alternative is that Apple increasingly loses all control over the key technology that will define the future. Large scale general models continue to crush smaller ones, with Apple having neither the technical talent or compute infrastructure to remain anywhere close to the frontier. It retains some formidable hardware in the near-term, but increasingly finds its control over the software layer eroded.

Google keeps improving Gemini’s ability to handle personal context and act as a truly useful personal assistant. This lets it jump between form factors, and create ecosystem effects much more powerful than copy-pasting between your iPhone and Macbook. It potentially puts its frontier AI to work in accelerating its hardware R&D, reducing the gap with Apple here.

Unless Apple acts soon, Jony Ive and Sam Altman shall one day soon waymo into Cupertino as conquerors, ready to turn the ruins of this fallen tech empire into simply an especially aesthetic new cluster of GPUs.

Thanks to Miles Kodama for triggering the idea for this post and helpful feedback!

Subscribe now

• China’s state-led approach has rapidly built AI infrastructure and talent, enabling both private and some state-backed firms to emerge as global competitors. However, this comes with persistent inefficiencies, as political incentives often prioritise visible short-term achievements over deeper innovation

• Political incentives and short-term targets have produced impressive scaling, but also inefficiencies, resource fragmentation, and limited progress in foundational innovation.

• While China’s model has created the conditions for ecosystem-wide growth, its long-term sustainability is uncertain, especially as increased state intervention could constrain the autonomy that has enabled recent private sector successes.

Introduction

China’s rapid advances in artificial intelligence (AI) are often attributed to the power of state intervention. Since the late 2010s, the Chinese government has set ambitious national targets, poured billions into AI research and infrastructure, and signalled to local officials that AI is a pillar of national rejuvenation. Yet, as with other strategic sectors, the reality of how state intervention plays out on the ground is far more complex than a simple tale of top-down success. Through exploring how China's massive state investment has built the infrastructure that enables both direct successes and unexpected private innovations like DeepSeek, we examine why this approach produces such mixed results, and consider whether China can maintain this balance as political pressures for greater control over strategic AI companies intensify. Finally, we'll see what this means for the future of AI development compared to Silicon Valley's market-driven model.

(Source: Tjeerd Royaards)

How State Intervention in AI Works in China

The Chinese state's approach to AI reflects its broader political economy, where central planning meets decentralised implementation. Since declaring AI a strategic technology, Beijing has crafted a comprehensive framework centred on the 2017 "Next Generation Artificial Intelligence Development Plan," which established ambitious targets including global AI leadership by 2030. Multiple ministries coordinate this effort, including the Ministry of Science and Technology, the Ministry of Industry and Information Technology, and the National Development and Reform Commission—a structure that provides comprehensive coverage but sometimes leads to overlapping responsibilities.

Provincial and municipal governments translate these broad national directives into localised implementation plans with varying degrees of adaptation to local conditions. When Beijing announced its national AI strategy, over 20 provinces released their own development plans within 18 months. This rapid response demonstrates the system's capacity for mobilisation, though many plans share striking similarities rather than leveraging regional comparative advantages. Shenzhen focuses on hardware integration, Shanghai on financial applications, and Beijing on fundamental research—yet all three simultaneously claim aspirations to become comprehensive "world-leading AI innovation centres."

Career incentives fundamentally shape how officials engage with AI development. Local leaders advance their prospects by delivering visible achievements in priority sectors within their 3-5 year appointments. The cadre evaluation system rewards technological progress alongside economic growth and social stability, creating motivation to support AI initiatives. This incentive structure excels at rapidly directing resources toward nationally prioritised sectors, though it can sometimes prioritise short-term visible outcomes over longer-term capability building.

Local governments have developed several mechanisms to support AI firms, with particular attention to small and medium-sized enterprises (SMEs). The "Little Giants" programme identifies promising SMEs for targeted support based on technological capabilities and growth potential. By 2023, over 12,000 companies across sectors had received this designation, including more than 300 AI firms. Companies selected for the programme receive coordinated backing that includes preferential access to government contracts and research partnerships, support that has proven valuable for firms developing specialised AI applications in manufacturing and healthcare, though selection sometimes favours politically connected enterprises over the most innovative. This targeting approach extends to broader industrial organisation through initiatives like Suzhou's "economics-driven consortia" and the "Supply Chain Architect" (链主) programme, which designate leading enterprises to coordinate development across value chains, successfully addressing certain innovation gaps while occasionally creating artificial collaborations that lack market-driven spontaneity.

Collectively, these interlocking mechanisms have generated substantial outputs. China now publishes numerous AI research papers, has built extensive networks of AI research facilities, and is widely viewed as a secondary, but growing, power in relation to the USA for AI. These achievements demonstrate the system's capacity for mobilisation and resource allocation, though questions remain about efficiency and innovation quality. These mechanisms have successfully constructed an elaborate AI support infrastructure; whether this machinery effectively produces genuine innovation and competitive advantage requires examining actual outcomes rather than administrative inputs.

Does State Intervention Work? Successes, Inefficiencies, and Structural Limitations

China's state intervention in AI reveals a complex pattern where impressive scale and deployment speed coexist with persistent inefficiencies that constrain innovation quality. Understanding how state support functions requires examining both its direct effects on individual companies and its broader role in creating ecosystem conditions that enable seemingly private innovation to flourish.

SenseTime's trajectory illuminates tensions within China's approach. Founded in 2014, the company received substantial early support through government procurement contracts for facial recognition systems, providing stable revenue streams whilst developing core technology. Local government subsidies reduced R&D costs, and facilities like the Shanghai AI Tower offered below-market computing resources that accelerated algorithm training. By 2020, SenseTime had deployed technology across 150 cities and achieved a valuation exceeding $7.5 billion.

Yet SenseTime's stock price has since fallen 75% from its IPO, revealing how political considerations often override market signals in resource allocation. The company's valuation reflected state-sponsored ambitions rather than sustainable commercial fundamentals. Political incentives favour showcase projects that demonstrate visible technological progress within officials' 3-5 year promotion cycles, often sustaining firms regardless of their actual market performance. Government VC funds have invested $184 billion in AI firms between 2000-2023, yet such investment largely targets visible applications promising immediate results rather than foundational research capabilities. Companies like CloudWalk, which received substantial subsidies despite persistent losses and declining market share, illustrate how political priorities can maintain underperforming firms without improving their technology or business models.

While a high degree of waste is inherent to early-stage innovation everywhere, the Chinese system is distinctive not just in the scale of its misallocated capital but in how political incentives often prolong the lifespan of underperforming firms. In the US, failed startups typically exit swiftly, freeing up resources and encouraging rapid experimentation. By contrast, some state-supported Chinese companies can persist for years, with misallocated capital accounting for around 12% of sector output in 2023, compared to 8% in the US. While this extended support does help expand the overall talent pool, it also creates a secondary effect: by insulating firms from market discipline, it can dampen competitive pressure, slow the reallocation of both capital and expertise to more promising ventures, and risk entrenching suboptimal business models. Thus, while China’s approach accelerates talent development, it also amplifies the structural costs and opportunity losses associated with prolonged inefficiency.

Despite these limitations, state intervention does excel at addressing specific market failures in AI development. Early government backing reduced capital constraints for compute-intensive research, provided early markets for technology validation, and coordinated complementary resources like data and infrastructure. When aligned with genuine commercial potential, such support significantly accelerated SenseTime's initial commercialisation, enabling rapid deployment across Chinese cities and establishing market presence that would have taken longer through purely private funding.

More importantly, the massive investment in flagship AI companies like SenseTime has functioned as an ecosystem catalyst, expanding the talent pool, accelerating infrastructure development, and increasing the overall scope of AI activity in China far more rapidly than market forces alone would have achieved. Former employees from these state-supported ventures likely disperse throughout the sector, carrying expertise to new companies and contributing to broader technological diffusion.

‘Private Innovation’ and Deepseek

The emergence of companies like DeepSeek reveals how state intervention's true impact extends far beyond direct subsidies to individual firms. DeepSeek operates as a privately-owned company with no direct government funding, yet its success demonstrates the ecosystem effects of China's broader AI industrial policy.

DeepSeek's technical capabilities rest on ecosystem effects from years of public investment. While the company bought its own Nvidia chips and operates independently, it benefits from talent pools expanded by state subsidies to AI firms and research networks established through university partnerships. When SenseTime and similar companies received massive government backing, they created spillover effects, training researchers, accumulating knowledge, and sector-wide expertise that now benefit private firms like DeepSeek.

DeepSeek's achievements in large language model development illustrate how ecosystem-level investment can enable genuine innovation even when flagship companies like SenseTime struggle with market performance. Rather than indicating a failure of industrial policy, such cases show that success should be measured by the growth of overall technological capabilities across the sector, not solely by the performance of individual companies. Although political pressures, regional competition, and a short-term focus create challenges for many firms, private companies operating within this environment can often navigate these obstacles more effectively, benefiting from the infrastructure and talent cultivated through state support. In this way, China’s industrial policy has built an innovation ecosystem that delivers real advances, though it does so amid significant inefficiencies and resource waste.

China's AI development model continues producing impressive quantitative outputs and application-level deployment whilst facing persistent challenges in fundamental innovation quality and resource efficiency. The approach excels at scaling proven technologies and building physical infrastructure, but the pathway to these capabilities involves sustaining numerous underperforming firms and fragmented resource allocation. Companies operating within China's state-built AI infrastructure can achieve remarkable efficiency and technical capability, yet the ecosystem enabling such success required massive public expenditure on projects that failed to deliver commensurate returns.

Industrial policy's ultimate measure may be its ability to expand overall technological capacity rather than the market performance of any single recipient. The question remains whether such ecosystem-level benefits justify the substantial inefficiencies required to achieve them, particularly when comparing the costs of state-led development against alternative approaches to building AI capabilities. Yet even accepting that China's wasteful but effective approach has successfully created conditions for innovation, a more fundamental question emerges: whether this model can sustain itself as AI becomes increasingly strategic. Can privately-owned firms like DeepSeek continue to flourish outside direct state control, or will political imperatives eventually override the market dynamics that enable their success?

China and Silicon: Divergent Models

DeepSeek's emergence within China's state-built AI ecosystem reveals a fundamental tension that distinguishes Chinese development from Silicon Valley's approach. Whilst private firms can currently exploit the infrastructure and talent pools created by massive public investment, their operational independence remains contingent on political tolerance rather than market protection. Silicon Valley firms operate within legal frameworks that provide predictable boundaries for state intervention, whereas Chinese companies like DeepSeek exist in a space where political imperatives could override market dynamics at any moment.

Both systems allocate resources through competitive mechanisms, yet the nature of accountability differs fundamentally. Silicon Valley's venture capital ecosystem subjects firms to continuous market validation, creating selection pressures that favour technologies with clear monetisation paths and defensible intellectual property. China's approach subjects firms to dual pressures: market performance and political alignment. Companies must simultaneously satisfy commercial viability and contribute to national technological objectives, often creating conflicting priorities that become more acute as AI gains strategic importance.

The key structural difference lies in the trade-off between ecosystem benefits and direct control. China's massive public investment has successfully created conditions for innovation, apparently enabling firms like DeepSeek to achieve remarkable technical capabilities whilst operating with constrained budgets. Yet maintaining genuine innovation requires allowing market forces sufficient space to operate, creating a tension between the state's desire for strategic control and the need for autonomous technological development. Silicon Valley avoids such tensions through clearer institutional boundaries, though at the cost of coordination failures that prevent the rapid national-scale deployment China can achieve.

As AI becomes increasingly central to national competitiveness, China faces an unresolved dilemma. The ecosystem approach that enabled DeepSeek's success depends on preserving space for market-driven innovation, yet strategic imperatives create pressure for more direct state control. Huawei's trajectory from relatively autonomous private enterprise to a company operating under state direction illustrates how rapid success can alter the relationship between government and private firms.

(DeepSeek founder Liang Wenfeng. Source: The Times)

More direct state influence over companies like DeepSeek appears increasingly likely as AI's strategic importance grows. While DeepSeek remains genuinely private today, the Chinese state operates under a political logic predicated on comprehensive control over strategic sectors. As AI capabilities approach levels that directly affect national security and economic competitiveness, allowing flagship technology companies to operate with full autonomy sends a message of limited control for which contradicts with CCP governance principles. The intensifying technological competition with the United States compounds these pressures, creating imperatives for coordinated national strategy that fragmentary private development cannot satisfy. The question is not whether DeepSeek will face increased oversight, but how extensive that control becomes, and whether it remains private in name while losing autonomy in practice.

Yet predicting the implications of increased state control for Chinese AI development remains difficult. Greater coordination might accelerate deployment in priority areas whilst potentially constraining the experimental innovation that enabled DeepSeek's technical achievements. The challenge lies in determining whether China can maintain the ecosystem benefits that fostered genuine innovation whilst satisfying the political imperatives that demand more centralised control over strategic technologies.

Conclusion

Through sustained state investment and intervention, China has rapidly established the infrastructure and developed the talent base essential for future advances in AI. This foundation has enabled a new generation of private firms, exemplified by DeepSeek, to emerge as genuine competitors on the global stage. While a number of state-affiliated enterprises have also found success, many others have struggled to adapt, highlighting the uneven outcomes of China’s approach. Nonetheless, the state’s commitment to building capacity has created the conditions necessary for continued growth and innovation, allowing China to close the gap with international leaders and assert itself as a major player in artificial intelligence.

Yet, this success is accompanied by persistent inefficiencies and structural tensions. The same mechanisms that enable swift deployment and widespread adoption can also sustain less competitive firms and fragment resources, sometimes at the expense of deeper innovation. While the current model has delivered results, its stability is not guaranteed. As the state’s likely role in guiding the direction of leading firms like DeepSeek grows, the delicate balance between fostering genuine innovation and maintaining political oversight may become harder to sustain. Whether China can preserve the dynamism that has propelled its rise in AI, while accommodating greater state intervention, remains an open question as the sector’s strategic importance continues to increase.

Subscribe now

• A leading AI lab would still be accountable to its investors and compute partners

  • Lab CEOs may be very powerful agents but they are still bound to their principals

• Monetisation of more powerful AI models will require more decentralised ecosystems and cannot be effectively accomplished within a single firm, reducing the concentration of economic power in a single lab

  • Trickier pathways to monetisation will limit labs ability to increase their internal compute consumption and slow overall progress

The AI 2027 scenario is the best and most fleshed out case for an intelligence explosion dynamic, providing concrete forecasts and arguments that this piece will look to question. AI 2027 is very much worth reading, but at the most basic it envisions a world where the leading AI firms develop systems increasingly capable of accelerating the pace of AI progress. Rapid progress in automating AI R&D leads to an arms race between the US and China as the leading firm in each looks to progress as quickly as possible to an artificial general intelligence (AGI) and on from there to a superintelligence (ASI).

This rapid development faces few economic constraints, with the leading lab understood to have effectively unlimited access to capital and rapid revenue growth. We argue that economics will likely act as an important brake on the pace of AI progress. Developing an AGI will require an ecosystem of actors providing capital, supply chains and end-demand. This ecosystem will affect the decision making capabilities of the leading lab, likely pushing towards greater focus on monetisation and competition in the market, requiring more compute dedicated to external needs, reducing the potential for internal recursive development.

Labs don’t exist in isolation

The AI 2027 largely assumes that all the key decisions by the leading AI lab are undertaken by its executive team, with little external input in this process. Once models progress to significantly greater capabilities through 2027 there is an increasing role played by the US executive and national security state, but otherwise the key decision makers remain the lab executives. The board of OpenAI is understood to have entirely lost control of the firm, with little visibility over internal developments.

To an extent this makes sense considering the very short timelines within the scenario. The nature of bureaucratic corporate structures means that the lab CEO, if they desired to, could act as a fairly free agent versus its various principles. Though we’d argue that the AI 2027 scenario still takes this too far.

At no point in the AI 2027 scenario is the leading firm profitable, rather its revenue is being quickly sunk into further compute expenditures, to justify the many hundreds of billions being spent on capex to fulfill these compute needs. Through the period of the scenario OpenAI would need to be continually raising more capital to fund this expansion. This would require at the least engagement with its existing investors, but likely also bringing in new investors and further tapping debt markets. Given the rate of capabilities advance it would be unlikely to struggle in accessing this capital, but the process would remain an important constraint on the time and focus of the executive team, and an important conduit for external insight into the firm’s activities and decision making.

The scale of the compute infrastructure build out would also require engagement with other partners within this ecosystem. This access to compute is likely to be both more significant and more difficult than the access to capital. Both OpenAI or Anthropic would remain strongly bound to their hyperscale compute providers, even in worlds where they increasingly self-build, as OpenAI is trying to achieve with Stargate. This self-build means cutting out one partner with significant leverage over you (Microsoft) but requires a new larger equity partner (Softbank) and new suppliers further down the supply chain (Oracle for servers, TSMC for in-house chips etc.).

In AI 2027 the only actor capable of slowing or stopping the leading lab is understood to be the US executive. In reality the leading lab would either need to be leaning on a hyperscale compute partner to take on the capex cost, or bring in a far larger amount of capital to fund this build themselves. In either case these players would be additional stakeholders that could influence the leading lab’s development and deployment decisions. Or it could be the case that Google is the leading lab, where its public listing would become a further condition limiting its ability to unilaterally decide and execute its chosen strategy.

While the OpenAI boardroom drama demonstrated how boards can be fairly impotent principals versus a more cunning agent, it also showed the importance of external players. Sam Altman’s actions would not have been possible without the support of Microsoft. Even given large capabilities advancement and a clear lead a firm like OpenAI would not be able to shed its various economic dependencies on its compute and capital providers.

Monetisation is far more of a challenge than assumed

AI 2027 relies on work done by FutureSearch on how feasible it would be for OpenAI to scale to a $100bn ARR by mid-2027. FutureSearch forecast the scale and composition of OpenAI’s revenue, and compare OpenAI’s time to $100bn against that of other previous tech companies, and find that it is in line with the trend. In doing so FutureSearch largely assumes that OpenAI functions like these previous tech platform companies and in doing so misses likely some important features that make AI monetisation different.

ByteDance was able to scale from $1bn to $100bn in only six years as it was scaling a viral consumer platform with very little marginal cost, and a lucrative advertising model. The same is not true for OpenAI. Its consumer business has much more meaningful marginal costs and a far less lucrative model. While OpenAI is planning to work towards monetising free users, almost certainly through some variety of advertising, for the time being these are simply a cost, with OpenAI needing users to upgrade to their $20 subscription.

FutureSearch compare OpenAI’s subscription business to Netflix to sense check the scale it could have achieved by mid-2027. This is reasonable but raises another example of how poor OpenAI’s economics are, as where Netflix upon reaching a large enough scale was able to keep its content spend fixed and reach much higher levels of profitability, the compute needs of supplying a ChatGPT sub mean this will never be the case to such an extent for OpenAI.

The FutureSearch model also opts to disregard API revenue, and instead assume the bulk of OpenAI’s revenue will be coming from products that directly automate work. They argue that API revenue as it depends on the growth of firms building on OpenAI’s models is less certain, and that it is unclear how far API revenue has a competitive moat against open-source models or other closed-source developers.

This competition does not occur however with the automating agents that OpenAI creates, with their being able to charge $20,000 a month(!) for vaguely defined R&D research agents, $10,000 for software engineers, and $2,000 for knowledge workers. In the scenario OpenAI is leading but other labs are <1 year behind and OpenAI is often prioritising its best models for internal use rather than actively looking to commercialise them. OpenAI seems highly unlikely to be able to charge the average salary of a US software engineer as its pricing point, with competition pushing this down significantly.

They also model OpenAI’s Enterprise revenue as separate from these agents that are replacing workers, though it is not clear quite where the boundaries of the two would be for categories such as knowledge workers. Why would there be such a gulf between a $50 a month ChatGPT Enterprise sub and a $2,000 a month replacement knowledge worker? Presumably the growth in replacement knowledge workers would partly cannibalise the spending on ChatGPT Enterprise as the keen adopters that had been driving that revenue were the first to switch up to more powerful agents.

Economics is an important brake on AI 2027 style scenarios

AI 2027 largely assumes that future markets for AI are largely going to look like B2B SaaS for cognitive labour. OpenAI could rapidly scale its profitability by selling more advanced chatbots and standardised automation agents. The reality is likely to be much more challenging, with model-making companies competing fiercely and margin being a product of either traditional tech moats (large scale consumer/enterprise platforms) or an ability to effectively monetise the frontier of intelligence. The latter is very possible, but is far harder to do. It requires an ecosystem of firms above the model layer that can mould and shape that intelligence to fit the shapes of existing organisations and processes and extract real value in the economy.

It is that ecosystem that can transform the economy. Jack Wiseman in his review of AI 2027 explains very well the significance of understanding these market dynamics:

“Overall research output is most sensitive to growth in R&D compute, because of its effects on experimental throughput. But the authors’ expectations for R&D compute budgets are downstream of ungrounded expectations for automation and revenue. With more grounded expectations for automation, R&D budgets would be lower, and so research output would be less, so capabilities progress more slowly, so automation happens at a more reasonable pace”

AI systems will keep getting more powerful, and will be capable of accelerating AI R&D, but this process will be limited by the need to fund and monetise these models. Some of Sam Altman’s time will be dedicated to efforts to speed up internal R&D, but much will go to convincing a jumpy Softbank to pony up another $10bn or TSMC to build out a new advanced packaging plant. The CEO of Accenture will want a call before committing to some huge new sum to roll out OpenAI’s newest agent, the California AG will have a new question about new appointments to the non-profit board.

The companies developing these powerful AI systems are still ‘normal’ companies. In meaningful ways this limits their ability to continuously concentrate more and more power within themselves.

• They are dependent on sources of capital, supply chains, and customer ecosystems.

• They rely upon the legal and political governance that facilitates these markets.

• They need to rapidly grow to justify explosive valuations and voracious consumption of cash.

After the development of Artificial Super Intelligence this equation increasingly breaks down. But prior to that point these economic realities will provide an important break on the rate of AI progress and crucial mechanisms for ensuring that the world envisioned by AI 2027 does not come to pass.

Subscribe now

This blog was written as a project for the Bluedot Economics of Transformative AI course and my thanks go to my discussion group for plenty of interesting ideas and discussion, and to BlueDot for putting the course together!

Key Takeaways:

• The unpredictability of US trade policies may paradoxically undermine America's technological containment strategy in Southeast Asia, creating strategic openings for Chinese cloud providers despite export controls.

• While American hyperscalers maintain technological superiority, Chinese providers are leveraging price advantages and wider regional presence to target cost-sensitive markets where "good enough" performance suffices.

• Southeast Asia's cloud market appears headed toward strategic fragmentation rather than winner-takes-all: US dominance in high-value sectors, Chinese growth in price-sensitive segments, and local champions securing nationally-sensitive workloads.

The global race for technological leadership in Artificial Intelligence is increasingly tied to the control of foundational cloud infrastructure, with Southeast Asia emerging as a critical battleground. While American cloud providers currently enjoy a leading market share, the region's future AI landscape may be less certain than it would have been only a short few years ago. The Trump administration's simultaneous pursuit of reshoring manufacturing through broad tariffs while containing China's technological rise has created an unintended contradiction, leading to America's competitive vulnerability in Southeast Asia stemming not from technological disadvantages but from perceived inconsistency in US policy.

While export controls on advanced chips directly target China's AI capabilities, the broader pattern of policy volatility—marked by sudden tariff implementations, unexpected pauses, and shifting regulatory frameworks—could paradoxically undermine these containment efforts. This unpredictability might create new opportunities for Chinese cloud providers to gain ground in Southeast Asia, where long-term infrastructure decisions require stable partnerships. Through focusing on rapid deployment and competitive pricing, Chinese firms may be able to leverage erratic US policy to grow market share while building foundational infrastructure for regional AI adoption, with their ability to present themselves as potentially more reliable partners, despite technological constraints, proving increasingly attractive to Southeast Asian markets seeking predictability.

Though Southeast Asian nations will likely continue their longstanding strategy of geopolitical hedging, increasingly by fostering their own domestic cloud champions alongside foreign partnerships, the uncertain nature of US trade and technology policy might, on balance, open a door for China to deepen its influence in the region's cloud ecosystem. Any such shift could begin to reshape Southeast Asia's AI infrastructure landscape, potentially altering the balance of power in the digital economy over the coming years. However, before examining how policy unpredictability creates opportunities for Chinese providers, we must first understand the current competitive landscape in Southeast Asia's cloud market.

(Patrick Chappette | Copyright 2018 Cagle Cartoons)

The Current Southeast Asian Cloud Landscape

American cloud giants have firmly established their dominance in Southeast Asia's cloud market. Early movers, such as Amazon Web Services (AWS), gained a significant head start and set the standard for cloud services in the region. While precise regional figures are scarce, AWS continues to command a substantial share globally, holding 33% of the worldwide cloud infrastructure services market as of Q4 2024, which represents the most recent verified data available. This dominance has faced increasing challenges in recent years, not only from local providers better able to navigate the regulatory landscape but also from the emergence of Chinese cloud providers. These Chinese firms, while often perceived as not yet possessing the technological sophistication of their American counterparts, have competitively undercut prices while expanding their geographic coverage across Southeast Asia through strategic deployments of availability zones. Alibaba Cloud operates 10 availability zones and companies like Huawei also aim to match, compared to AWS with 7 availability zones. While a quantitative advantage in availability zones does not definitively equate to overall compute power, it signals a commitment to building a wider regional footprint, with McKinsey data indicating that China's public cloud market is expected to reach $90 billion by 2025 supporting this commitment. American providers still enjoy the lion's share of the Southeast Asian cloud market, a position underpinned by technological strengths and long-standing customer relationships; however, they now face a credible and growing challenge from Chinese firms who are beginning to dilute their market share.

(Alibaba Cloud Data Center Location in Zhangbei, Hebei province, China)

The competition between American and Chinese providers is further shaped by their different service emphases. Drawing parallels to the global market, American providers have been progressively shifting from basic Infrastructure-as-a-Service (IaaS) offerings toward higher-margin Platform-as-a-Service (PaaS) capabilities. In contrast, Chinese providers have developed expertise in lower-margin but widely accessible infrastructure deployments. This difference is particularly significant as Southeast Asia's market remains predominantly IaaS-focused, creating natural opportunities for Chinese providers whose domestic experience aligns with the region's current needs. However, this landscape is not static. Chinese cloud firms are increasingly motivated to shift toward more profitable software and platform services, mirroring the evolution seen in Western markets. Yet this transition faces varied adoption rates across Southeast Asia's heterogeneous economies. Financial services and technology sectors in Singapore and Malaysia may rapidly embrace PaaS offerings, while manufacturing-heavy economies like Vietnam and Indonesia could maintain IaaS dependence for longer periods. This uneven transition creates a complex competitive environment where Chinese providers can leverage their IaaS expertise in some markets while simultaneously developing PaaS capabilities for more technologically advanced sectors—a flexibility that may prove advantageous amid shifting policy conditions. Additionally, while the current dominance of American cloud providers remains largely CPU-centric, with Central Processing Units serving as the foundation of general-purpose computing tasks, GPUs are now widely viewed as the future of cloud computing. These specialised processors, essential for powering artificial intelligence and machine learning workloads, represent the next critical phase of cloud infrastructure development. As competition intensifies across Southeast Asia, cloud providers' ability to effectively deploy GPU-accelerated computing resources will increasingly determine their market position and long-term influence in the region.

In the current landscape, American cloud providers maintain a significant advantage in GPU capabilities through preferential access to cutting-edge hardware. AWS and Microsoft Azure deploy NVIDIA's A100 and H100 processors across their infrastructure, establishing a performance edge for AI workloads that Chinese competitors cannot directly match. The advantage stems largely from U.S. export controls instituted in 2022 and expanded in 2023, which specifically restrict the sale of advanced AI chips to China and affiliated entities. These regulations prevent chips with performance above specific thresholds from reaching Chinese companies, effectively blocking access to NVIDIA's flagship processors. After initially permitting NVIDIA's China specific H20 chips, lower performance alternatives to the flagship H100 processors, the U.S. government reversed course on April 9, 2025, requiring export licences for these processors. This latest policy shift further complicates strategic planning for cloud providers across the region, with enforcement varying significantly by country. Singapore's stringent implementation places fundamental limitations on Chinese providers' high performance AI offerings, while more flexible approaches in Thailand and Indonesia allow for broader deployment, creating an uneven competitive landscape that mirrors the broader volatility in the U.S. China technology relations.

In terms of the deployment itself, Chinese providers have responded pragmatically by deploying previous-generation GPUs not covered by restrictions and integrating domestic alternatives like Huawei's Ascend processors. Though these deliver only 50-70% of NVIDIA's performance, they enable Chinese firms to establish "good enough" AI computing capabilities across their more extensive regional footprint. Furthermore, the introduction of Huawei’s Ascend 910C processor, which promises improved performance compared to earlier models, reflects ongoing efforts to mitigate reliance on imported technology. Additionally, prior to the April 2025 licensing requirements, Nvidia's H20 chip had been gaining adoption by Chinese firms such as ByteDance and Tencent due to its affordability and strong inference capabilities. These new restrictions will likely accelerate Chinese providers' existing strategy of developing alternative solutions alongside deploying previous-generation GPUs not covered by restrictions. By combining these technical solutions with significant cost advantages (typically 20–30% lower than American equivalents) and greater geographic distribution, Chinese providers are creating competitive AI infrastructure despite technological constraints. Ultimately, this strategy may allow them to effectively serve the majority of regional AI workloads where absolute performance is less critical than accessibility, cost-efficiency, and local availability.

Policy Unpredictability as Strategic Advantage: Potential Reshaping of Southeast Asian Cloud Competition

(The Economist Weekly Cartoon)

"We must as a nation be more unpredictable. We are totally predictable. We tell everything... We have to be unpredictable. And we have to be unpredictable starting now." - Donald Trump, Foreign Policy Speech, April 27, 2016

Unpredictability, whether you view it as strategic or not, has without doubt been especially relevant in recent US policy. Manifesting itself in US trade policy most clearly are three dimensions of volatility: shifting tariff frameworks, inconsistent export controls, and unpredictable regulatory stances. While those more bearish on China within the administration will argue that the, as it stands at the time of writing, 145% tariffs on China will place pressure on their economy, the broader pattern of policy inconsistency may create unexpected opportunities for Chinese cloud providers in Southeast Asia, potentially undermining the very containment strategy these policies were designed to advance. To understand these dynamics, we must examine how policy unpredictability might create asymmetric impacts on US and Chinese cloud providers due to their contrasting corporate structures, supply chain integration, and investment timeframes.

​​Tariff volatility creates unequal impacts on US and Chinese cloud infrastructure deployment strategies in Southeast Asia. What began as Trump's calibrated tariff system—49% on Cambodia, 46% on Vietnam, 10% for Singapore—quickly shifted to blanket 10% rates for most countries while China's tariffs soared to 145%. For US providers, these fluctuations introduce significant challenges. Their globally distributed supply chains, often involving components from multiple countries, become more difficult to manage when tariff regimes change unexpectedly. Microsoft's decision to delay a $1 billion Indonesian data centre expansion in 2024 illustrates this problem. Initially announced in 2022, the project faced multiple financial recalculations as tariff policies shifted, ultimately leading to postponement when component import costs rose unpredictably. Chinese providers face a different set of circumstances. While they confront higher direct tariffs targeting Chinese exports, their more vertically integrated domestic supply chains offer greater insulation from the cascading effects of global tariff volatility. Companies like Alibaba and Huawei have developed supply ecosystems predominantly within China, allowing them to maintain more predictable cost structures despite targeted trade barriers. Huawei’s Ascend 910B processors exemplify this advantage: manufactured domestically by SMIC using a 7nm process, they deliver up to 400 FP16 TFLOPS in theoretical performance, comparable to NVIDIA’s A100 in specific tasks. The forthcoming Ascend 910C, expected later this year, promises even greater performance. While these chips face production challenges due to fabrication limitations, they highlight how Chinese firms are leveraging homegrown technologies to mitigate reliance on foreign supply chains.This difference in vulnerability potentially creates a strategic advantage for Chinese cloud providers in Southeast Asian markets. While US firms must recalculate expansion plans with each policy shift, Chinese providers can leverage their supply chain stability to commit to more consistent regional deployment timelines, potentially securing partnerships that value reliability. This dynamic highlights the fundamental mismatch between cloud infrastructure's 3-5 year planning horizons and policy environments that now change quarterly or monthly, creating particular uncertainty for GPU-accelerated infrastructure essential for AI applications.

Export control inconsistency might further amplify these challenges, especially as the market moves from CPU-centric workloads toward the GPU-intensive computing necessary for AI applications. US firms operating with quarterly performance pressures and shareholder expectations could struggle to commit to long-term infrastructure investments when the underlying technology access rules remain in flux. The administration's approach to Nvidia's H20 chips illustrates this pattern of uncertainty. Following periods of apparent permission and conflicting signals about potential restrictions, the government ultimately imposed license requirements in April 2025. However, this pattern of rumours, conflicting signals, and eventual decisions may create planning challenges for both US providers (uncertain which technologies they can deploy) and their Southeast Asian customers (unclear which platforms offer long-term stability). Chinese providers, by contrast, might adapt to technology restrictions through organisational structures more aligned with long-term planning, where state ownership of key cloud players like China Telecom and more general alignment between corporate and national strategic objectives enables larger, longer-term commitments. Alibaba's recent $53 billion investment in AI and cloud infrastructure, surpassing its total spending in this area over the past decade, exemplifies this capacity for substantial long-term commitments despite external uncertainties.While this figure still falls well short of the capital expenditures by American hyperscalers like AWS, Microsoft and Google, the dramatic increase is indicative of Chinese providers' growing confidence in navigating policy volatility. Once hampered by regulatory crackdowns and slowing profits, Alibaba is now making a major strategic bet on cloud and AI infrastructure that suggests confidence in navigating policy volatility. Despite these potential advantages, Chinese providers must overcome significant challenges that persist regardless of US policy volatility. Lingering concerns about data security and privacy among potential Southeast Asian clients, the political sensitivity of Chinese technology in sectors like finance and government, and their historical difficulty in developing robust software ecosystems atop their infrastructure offerings all present ongoing obstacles. This software ecosystem weakness—evident in China's domestic market where PaaS offerings have struggled to gain traction—remains a vulnerability that American hyperscalers continue to exploit through their more mature developer platforms and enterprise software integrations. Nevertheless, these structural characteristics of Chinese cloud providers might still allow them to make more confident infrastructure commitments despite market uncertainties, though their GPU offerings would likely remain at a technological disadvantage.

(Source: Financial Times)

Regulatory unpredictability across Southeast Asian markets could create potential opportunities for Chinese-Southeast Asian partnerships, particularly in the IaaS sector where Chinese providers have developed significant expertise. While US firms have focused investments in regulatory-stable Singapore (exemplified by AWS's recent $6 billion expansion), Chinese providers might pursue more aggressive deployments across diverse regulatory environments. Alibaba Cloud's expansion into Thailand and Huawei's rollout of data centres in Indonesia demonstrate how Chinese providers may leverage their experience navigating complex regulatory landscapes to build regional presence despite policy uncertainties. Their approach emphasising flexibility by deploying standardised infrastructure designed to accommodate evolving data sovereignty requirements rather than waiting for regulatory clarity might prove advantageous in certain markets. The strategy could facilitate partnerships with local operators, such as Indonesia's Telkom collaborating with Alibaba Cloud while simultaneously maintaining AWS relationships. These partnerships may allow Southeast Asian entities to maintain balanced relationships with both American and Chinese providers, hedging against unpredictability from either direction.

Ultimately, policy unpredictability is likely to reshape Southeast Asia's cloud landscape through targeted market share shifts that favor Chinese providers in cost-sensitive sectors while American hyperscalers maintain their leadership in high-performance computing and advanced services. American providers will likely maintain their leadership in Singapore and Malaysia, particularly in financial services and multinational enterprise segments that require high-performance GPU deployments and advanced PaaS offerings. However, Chinese providers are positioned to gain meaningful market share in Vietnam, Indonesia, and Thailand—especially in public sector contracts, small and medium enterprises, and manufacturing industries where IaaS capabilities and cost considerations often outweigh cutting-edge performance requirements.

Over a 3-5 year horizon, American cloud providers will likely see their regional market share erode from current dominant positions, particularly in markets where data sovereignty concerns and cost sensitivity drive decision-making. This erosion represents a strategic setback for US technological influence in Southeast Asia, undermining America's position in one of the world's fastest-growing digital economies. While not catastrophic, this dilution of market position could undermine America's broader technological containment strategy and allow Chinese providers to establish footholds in sectors that, once secured, may be difficult to reclaim. The irony is that policy measures designed specifically to constrain Chinese technological expansion may inadvertently accelerate it in certain Southeast Asian markets by creating opportunities Chinese firms are structurally better positioned to exploit. Nevertheless, it must be noted that alternative scenarios remain. For instance, a greater privileging by ASEAN nations for security concerns could prompt a minimal-change scenario where American providers maintain their dominant positions despite policy inconsistencies, as Southeast Asian nations prioritize security considerations over cost advantages. However, this security-first approach appears unlikely given the region's diversity of strategic alignments and the economic imperatives driving their digital transformation agendas. ASEAN countries are unlikely to reach consensus on security-based technology restrictions, especially when their rapidly growing digital economies have much to gain from the foreign investment and transformative potential these cloud infrastructures provide.

The most likely outcome will involve a more diversified cloud ecosystem across the region. Local cloud champions like Telkomsel's NeutraDC in Indonesia and VIETTEL IDC in Vietnam may capture growing portions of government workloads and regulated industries in their respective countries. These local providers may strategically position themselves as neutral alternatives, leveraging nationalistic procurement policies while selectively partnering with both US and Chinese hyperscalers for technical capabilities they cannot develop independently. However, their impact will likely remain limited to specific regulated sectors as they lack both the scale and technological capability to challenge either Chinese or American cloud providers more broadly. For US strategic interests, this represents a concerning development, not because Chinese providers will displace American hyperscalers outright, but because they may secure sufficient market position to influence technical standards, data governance practices, and digital infrastructure decisions across a region that represents a crucial technological battleground. Overall, for American cloud providers, current regional market leaders, any movement toward a more equitable market share represents a net loss of influence and revenue, with Chinese providers standing to gain the most significant portions of redistribution.

Conclusion

The unexpected consequences of US policy unpredictability may reshape Southeast Asia's cloud landscape in ways that undermine rather than advance American strategic interests. While tariffs and export controls were designed to contain China's technological rise, their perceived inconsistency has created opportunities for Chinese cloud providers to position themselves as more reliable partners in an industry where stable, long-term infrastructure planning is essential. Over the next several years, Southeast Asia's cloud ecosystem will likely become more diversified, with market share redistributed along geographical and sectoral lines. American providers will retain their dominance in Singapore and Malaysia's financial sectors and in advanced computing applications, whilst Chinese alternatives steadily capture market share, particularly in price-sensitive segments and basic infrastructure services. Meanwhile, local champions will establish defensible positions in industries where data sovereignty requirements create natural protection from foreign competition.

Ultimately, market fragmentation challenges US technological influence in one of the world's fastest-growing digital economies. As Southeast Asian nations pragmatically balance relationships between competing powers, America's leadership position increasingly hinges on policy consistency rather than technological superiority alone. Even advanced American cloud offerings may struggle against Chinese alternatives that offer greater stability, superior pricing, and vertically integrated supply chains strengthened by China's vast domestic market. These advantages create competitive edges in markets where cost-efficiency and rapid deployment often outweigh absolute performance, potentially allowing Chinese cloud providers to secure footholds that will prove difficult to dislodge.