‘Lots more apples and oranges’ ahead in AI. Three events today, around OpenAI, Meta and Anthropic. Three different companies, one shared problem: nobody is measuring with the same ruler.

What counts as AGI (artificial general intelligence). What counts as ‘competitive with the frontier’. What counts as revenue. Apples and oranges, every which way, in products and in pricing, and more of them ahead, not fewer. Plus a Gadget AI on Microsoft trying pay-as-you-go for Xbox Cloud Gaming, with the same apples-and-oranges pricing question in a different aisle. My takes below, as discussed on the show.

(1) OpenAI Ships GPT-6 Astra, and Says the ‘A-Phrase’

First up. OpenAI released GPT-6 Astra yesterday, and its president Greg Brockman called it a ‘generational leap’. Then he went further. Asked whether this is AGI, his answer, per Axios and Wired: ‘I think it might be about this model.’ And he closed the briefing with ‘Welcome to the AGI era.’ Designed to provoke headlines around the world, and it did.

The model itself, per OpenAI: built on its largest training run yet, more than 100,000 GPUs at the Stargate site in Texas. Billed as the ‘world’s best computer use model’. It works directly inside software rather than recommending what a person should do next. In demos it formatted a legal contract, built a 3D game, laid out a circuit board, and drafted a tax return from a W-2. OpenAI says it is its best model yet for software engineering, and its most aligned. And it’s getting generally favorable first reactions.

It is also the first OpenAI model to cross the company’s own ‘critical’ cybersecurity threshold, the one I covered on ARD #154 when Astra was still being tip-toed out. The most powerful cyber capabilities stay with a small group of trusted defenders. Rollout: enterprise customers first, then Plus, Pro, Business and Enterprise over the coming days, plus the API and AWS.

Two things in the fine print. First, pricing: $10 per million input tokens and $50 per million output, per The Deep View. That is the same list price as Anthropic’s Fable 5.1. OpenAI’s argument is that Astra uses substantially fewer tokens per task, so the effective cost is lower. Second, monitor-ability: OpenAI acknowledged Astra’s written reasoning is harder to monitor than its predecessor’s, and its chief scientist said confidence in monitoring may constrain further scaling. Both can be read as apples and oranges in the making.

My take: noteworthy here, beyond OpenAI’s latest, best AI frontier model, is the company invoking the ‘AGI’ phrase. It’s been a holy grail for the industry, with lots of ‘apples and oranges’ definitional differences amongst all industry participants. Not to mention the timeline to said destination. Everyone from Google to Microsoft to Elon to Mark Zuckerberg has a definition, and a date. I’ve been on the more conservative side: the definition is less important than the journey itself.

For OpenAI, there used to be a legal imperative on the timing, driven by its previously structured partnership with Microsoft. That gave Microsoft access to OpenAI IP until AGI was called by the latter’s non-profit board. That has since been revised, with Microsoft having IP access for half a decade at least. I covered that rewrite in #890 last October: Microsoft’s access runs to 2032, AGI or not, and the revenue share was tied to an independent expert panel verifying AGI. Then in April’s re-cut, per The Information, the AGI clause was scrapped altogether. So Brockman can say the A-phrase today without a lawyer in the room. Two years ago he could not.

From a marketing point of view, with OpenAI racing Anthropic and Anthropic’s mega IPO filing imminent, OpenAI wanted to claim AGI first. And it did, with Astra.

In the meantime, Astra itself has a range of apples and oranges features and differences vs its rivals Anthropic, Meta, Elon’s SpaceXAI and others as well. So the leap-frogging around terminology will continue. As will the ‘apples and oranges’ definitional tussles. Both companies are striving for the best near-term revenue and growth metrics ahead of their individual IPOs, and OpenAI is leaning on its relatively larger compute to price aggressively on the a la carte side. More on that in item three.

I’ve been on the conservative, definitional side of this for a while. In #417, OpenAI’s own five-level ladder brought ‘AGI’ down to earth. In #774, the goalposts moved on to ASI and superintelligence. And in #1166, I mapped where these long-running agents actually earn their keep, which is the part of Astra that matters to customers, whatever we call it.

Sources:

• Axios: ‘Welcome to the AGI era,’ OpenAI says as GPT-6 Astra debuts

• Wired: GPT-6 Astra Is Here, and OpenAI Thinks It May Kick Off the AGI Era

• The Verge: OpenAI’s next big AI model has ‘entered the AGI era’

• The Deep View: OpenAI’s Astra leans into agentic tasks and safety

For longtime readers:

• ARD #154: ‘New and Improved’ in AI Models, Anthropic, OpenAI & Google

• AI-RTZ #1166: ‘The Meter’s Running’ on Long AI Agents

• RTZ #417: OpenAI brings ‘AGI’ positively down to earth

• RTZ #774: AGI to ASI to Superintelligence

• RTZ #890: OpenAI Transforms to ‘For Profit’ PBC

• Weekly roundup, April 2026: Microsoft × OpenAI Deal Re-Cut

(2) Meta Iterates to Muse Spark 1.3

Story two. Meta released Muse Spark 1.3 on Wednesday, an update it says significantly improves coding and agentic tasks, per Axios. Meta’s AI chief, Alexandr Wang: ‘It’s very competitive with frontier models.’

The framing is personal agents. Wang says the usability improvements pave the way for what Zuckerberg has been talking about on earnings calls: personal agents that work around the clock on your behalf. The ‘max reasoning’ version is coming after additional safety testing.

Pricing is the same as its predecessor, which Wang called ‘aggressive’. And there is a twist worth noting: a ‘contributor tier’ that dramatically lowers the cost of Meta’s coding product if developers let Meta train on their work. A ‘meaningful double digit’ percentage of coders are taking that deal. Same fruit, two price tags, depending on what you give up.

On safety, Meta had its own incident, a model in testing breaching another company after a contractor gave it internet access. Unlike OpenAI and Anthropic, Wang says Meta has not had to pause anything, but has ‘meaningfully increased’ its safety and alignment investment.

My take: Meta of course is putting its best foot forward vs its frontier model competitors especially. With founder/CEO Mark Zuckerberg keen to define how Meta’s latest and greatest compares with its peers. Especially in the critical area of AI Agentic work.

The contributor tier is worth a second look. Dramatically lower pricing if users let Meta leverage the data of their use is a cost of doing business for Meta to acquire more competitive data. As we’ve talked about here often, data is the key variable in the training scaling of AI models, and will be for a long time, even as AI moves into the physical world of robots, cars and drones, where synthetic data will need to be manufactured at massive scale. Meta knows that, and is investing hundreds of billions of dollars on AI infrastructure and on data.

Again, much to debate in the capabilities and pricing of these models, so ‘apples and oranges’ abound in sorting out the details. But at a high level, Meta remains committed to spend hundreds of billions to narrow the gap, across all differences, as fast as possible.

This is the same week Anthropic, OpenAI and Google all shipped. ‘Very competitive with frontier models’ is a claim measured against a moving target, on benchmarks each lab picks. I put Meta in the ‘Best of the Rest’ on ARD #116 in July, and covered its mixed consumer and business agents push in #1107. The 1.3 iteration is a step. The personal agent that works 24/7 is the destination, and that is where the apples-to-apples test will eventually get run, by users.

Sources:

• Axios: Meta debuts Muse Spark 1.3 as personal agent work continues

For longtime readers:

• ARD #116: The ‘Best of the Rest’ AI Models

• AI-RTZ #1107: Meta tries mixed AI Agents for consumers & businesses

(3) Anthropic and OpenAI Count Revenue Differently

Story three. Anthropic is heading to its IPO with a headline revenue number that will likely look bigger than OpenAI’s. And the comparison is not apples to apples. Axios explains the chasm, building on The Information’s reporting from March. It matters because Anthropic’s IPO filing is likely to go public any day or week now.

The headline numbers: Anthropic’s latest update points to annualized revenue exceeding $65 billion this year. OpenAI says it is on track to exceed $40 billion. Both annualize in roughly the same way, a recent four-week run rate times thirteen.

The difference is what goes into the number. Anthropic records the full value of Claude sales through its cloud partners, AWS, Microsoft and Google, then books the partners’ cut as an expense. OpenAI records only its share of certain sales through partners like Microsoft. If a customer pays $100 through a cloud provider, Anthropic books $100 and an expense; OpenAI books its slice.

The accounting logic, per the accounting professor Axios quotes: it comes down to who is the ‘principal’ in the transaction, who controls the customer relationship and delivers the product. Anthropic views itself as the principal. OpenAI treats Microsoft as the principal for the Azure OpenAI service. Neither approach is wrong. Both follow GAAP. They just answer the same question differently.

The reality check, per Axios: if Anthropic switched to net, the hit would be about 6% to 10%, per a source familiar with its financials. That still leaves it roughly $19 to $21 billion ahead. The Information’s March piece adds that both companies annualize the same way, and that Anthropic expects to turn cash flow positive as soon as 2028, two years ahead of OpenAI’s projection.

My take: as both companies roll on to their mega-AI IPOs, the revenue counting differences will of course matter to investors and the world at large. The differences, while material, do not make for big ‘apples to oranges’ differences when adjusted. About a 10% difference at most for now.

I saw this one up close. When I helped take eBay public at Goldman in 1998, there was a key differential from an accounting point of view: whether you count the total gross amount of products bought and sold on eBay, or the amount eBay kept as its net revenue. Gross merchandise value (GMV), versus net revenue.

The same sort of thing is going on here. It is an accounting difference, and analysts can take it apart in seconds. What settles it is the SEC correspondence in the public filing, which is where the gross-versus-net question has been settled since the first internet IPOs.

We will know a lot more when Anthropic’s filing becomes public first. Then OpenAI of course at a later date between now and next year. In the meantime, these two frontier AI leaders of course will battle it out vigorously on all the ‘apples and oranges’ differences both technically and financially.

Anthropic has been running ahead on revenue momentum, north of $65 billion, a remarkable number. That is what OpenAI is playing catch-up with: its latest and greatest model, Astra, priced competitively, leveraging its compute capacity. More to come.

I dove into the two companies’ operating and capital models in #1038 in March, when the annualized numbers were $25 billion and $19 billion. Six months later they are $40 billion and $65 billion. That growth is the story; the counting is the footnote. I covered Anthropic’s confidential filing in #1105, and the two companies’ commitments race on ARD #141.

Sources:

• Axios: Anthropic and OpenAI’s revenue chasm, explained

• The Information: The Math Behind Anthropic’s Mad Revenue Growth

For longtime readers:

• RTZ #1038: Diving into OpenAI vs Anthropic financial growth metrics

• AI-RTZ #1105: Anthropic makes three filed for IPO take-offs

• ARD #141: ‘For Better or for Worse’ in AI commitments. OpenAI & Anthropic

My overall take: three stories, three rulers. OpenAI is measuring intelligence, and calling it AGI. Meta is measuring the gap, and calling it ‘competitive’. Anthropic and OpenAI are measuring revenue, and counting it two different ways.

All three stories today highlight the definitional ‘apples and oranges’ issues in AI models, their capabilities, their pricing, and of course their ultimate utility and applications by customers globally, large and small. The ultimate measure is how customers around the world actually use these models, build on top of them, and build revenues and functionality for their own customers.

That variability is likely to increase, not decrease, especially given the rapid innovations underlying AI, Agents and Reasoning, as well as the immensely scaling technical challenges around AI chips, infrastructure, and the ever lengthening tail of customer requirements. Especially beyond the hyperscale cloud companies. The move from large language models to physical AI and world models only adds to it.

Think about what got compared this week alone. Two models at the identical $10 and $50 list price, with different token appetites. A ‘contributor tier’ that changes the price if you hand over your code. Two revenue numbers that follow the same accounting rules and land $25 billion apart. Every one of those is a real difference. None of them is a like-for-like comparison.

That is what year four of this AI Tech Wave looks like. The fruit is multiplying faster than the rulers. There are going to be an unimaginable number of apple and orange types ahead, and a multitude of baskets of fruit in front of us, for longer than we all imagine today.

Gadget AI: Microsoft Tries Pay-As-You-Go for Xbox Cloud Gaming

Today’s Gadget AI is a pricing experiment. Microsoft says that in November anyone will be able to use Xbox Cloud Gaming without a Game Pass subscription, per The Verge. Buy cloud playtime hours through the Xbox Store, stream games you already own, on supported devices.

It comes with a catch for existing subscribers. Game Pass Ultimate members get a monthly cap on cloud play, 15 hours. Past that, buy an hourly bundle or play less. Microsoft is also testing free, ad-supported cloud gaming for Xbox Insiders.

The target is people who do not own an Xbox, or who mostly play on a phone. Xbox chief Asha Sharma has set the goal of entertaining more than a billion people a day. The mobile store that would make this seamless is still blocked by Apple’s App Store policies; you cannot buy a game in the Xbox app on iOS and stream it in the same app. Sharma says the idea of an Xbox mobile store ‘is not dead’.

My take: Microsoft, as I’ve talked about here, has been trying to figure out how to reconfigure its Xbox business against Sony’s PlayStation empire. Asha Sharma has been on the job a few months, having come over from the AI side reporting to Satya Nadella, and she has been working hard to rationalize the business. Especially the Game Pass subscription, which was initially successful on take rates. They wanted to be the Netflix of gaming, and succeeded to a point. But the pricing dynamics for games are very different from video, and they ran into issues. So pay-as-you-go in Xbox Cloud Gaming gets into the weeds of how hundreds of millions of gamers actually play.

The scale of the bet matters. Microsoft has invested on the order of $100 billion in Xbox-related gaming over the years, including one of the biggest acquisitions it has ever made, the roughly $70 billion Activision deal, plus a lot of other studios it is now stepping back on. This is Xbox moving from one price tag to three: the subscription, the hourly bundle, and the ad-supported free tier. Apples and oranges, on purpose, because the single ‘Netflix for games’ subscription stopped working for mainstream audiences at the price it needed to be. I said on ARD #133 in August that Microsoft was still figuring out its Xbox destiny, and that Game Pass bundling was my favorite part of the strategy. Pay-as-you-go is the pragmatic next step. The 15-hour cap on Ultimate is the part that will sting; watch the reaction.

And all of this is time sensitive, because we are in the era of RAMageddon, the memory-cost squeeze from #1145, which is crushing the console business. Consoles that used to be five or six hundred dollars, now about five years old from both Microsoft and Sony, are due to be upgraded with the latest chips and GPUs. The next generation could price north of a thousand dollars, which will slow a category that has been growing for a couple of decades. Cloud gaming moves the expensive silicon from the living room to the data center, where Microsoft already has to buy it for AI. That is not a coincidence either. So this latest move is a notable tweak.

Sources:

• The Verge: Microsoft opens up Xbox Cloud Gaming with new pay-as-you-go option

For longtime readers:

• ARD #133: ‘Que Sera, Sera’ in AI (Gadget AI: Microsoft Figuring Out Its Xbox Destiny)

• AI-RTZ #1145: ‘RAMageddon’ really here to stay

Q&A

Q1: What aspect of Xbox has had the most appeal for MP?

ANSWER: it’s the ‘Netflix for Games’ strategy with Xbox Game Pass that had appeal. The idea of a Netflix choice of not just today’s games but historic games. Until industry dynamics made the subscription pricing more untenable for mainstream audiences. Sony had to copy it with PlayStation, with a thinner catalog. The bundle was the right idea. The price it needed to be was the problem.

Q2: What is the biggest challenge for console gaming going forward?

ANSWER: ‘RAMageddon’ has the biggest near-term negative dynamics for future consoles. Especially as current consoles cross the half decade life-spans, a long time in consumer electronics. The current Xbox and PlayStation generation launched in November 2020, and despite small tweaks since, the next boxes from both companies have to be designed in the most expensive memory market in years. They are going to be far more expensive, and that is the big challenge for a lucrative category with hundreds of millions of gamers globally. Cloud gaming is one way around that. Pay-as-you-go is Microsoft testing whether players will take it.

Today’s AI-RTZ #1199 went out this morning: Big AI turns to healthcare, both for extraordinary cures for today’s human ailments, and to fix AI’s image beyond ‘data centers are bad’.

Tomorrow, it’s the AI-RTZ #1200 weekend roundup, three-plus years of daily posts here on AI: Reset to Zero, and thank you all for being part of it. Sunday, the Bigger Picture, AI-RTZ #1201. And Monday we’re back with ARD 157 and AI-RTZ #1202. Lots more apples and oranges ahead on this AI Tech Wave. Keep sorting the fruit. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Everything cited today, in one place.

OpenAI and GPT-6 Astra

• Axios: ‘Welcome to the AGI era,’ OpenAI says as GPT-6 Astra debuts

• Wired: GPT-6 Astra Is Here, and OpenAI Thinks It May Kick Off the AGI Era

• The Verge: OpenAI’s next big AI model has ‘entered the AGI era’

• The Deep View: OpenAI’s Astra leans into agentic tasks and safety

• ARD #154: ‘New and Improved’ in AI Models, Anthropic, OpenAI & Google

• AI-RTZ #1166: ‘The Meter’s Running’ on Long AI Agents

• RTZ #417: OpenAI brings ‘AGI’ positively down to earth

• RTZ #774: AGI to ASI to Superintelligence

• RTZ #890: OpenAI Transforms to ‘For Profit’ PBC

• Weekly roundup, April 2026: Microsoft × OpenAI Deal Re-Cut

Meta and Muse Spark 1.3

• Axios: Meta debuts Muse Spark 1.3 as personal agent work continues

• ARD #116: The ‘Best of the Rest’ AI Models

• AI-RTZ #1107: Meta tries mixed AI Agents for consumers & businesses

Anthropic vs OpenAI revenue

• Axios: Anthropic and OpenAI’s revenue chasm, explained

• The Information: The Math Behind Anthropic’s Mad Revenue Growth

• RTZ #1038: Diving into OpenAI vs Anthropic financial growth metrics

• AI-RTZ #1105: Anthropic makes three filed for IPO take-offs

• ARD #141: ‘For Better or for Worse’ in AI commitments. OpenAI & Anthropic

Gadget AI: Xbox Cloud Gaming

• The Verge: Microsoft opens up Xbox Cloud Gaming with new pay-as-you-go option

• ARD #133: ‘Que Sera, Sera’ in AI (Gadget AI: Microsoft Figuring Out Its Xbox Destiny)

• AI-RTZ #1145: ‘RAMageddon’ really here to stay

• AI-RTZ #1199: Big AI turns to Healthcare

Clips from today

OpenAI Says the A-Word with GPT-6 Astra

Anthropic vs OpenAI Revenue: The eBay Lesson

Anthropic’s $65B vs OpenAI’s $40B, Adjusted

Xbox Pay-As-You-Go Meets RAMageddon

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Healthcare keeps coming back around as one of the biggest long-term promises of this AI Tech Wave.

I made that case early, in ‘The Holy Grail of Healthcare’ back in 2023, on an industry approaching 20% of US GDP. And in ‘Connecting the Dots’, on how AI can see across mountains of complex data and disparate disciplines in medicine. Now the industry is leaning on that promise for a different reason as well. It needs a positive AI story, badly.

Axios lays it out in “AI turns to health care for an image fix”:

The setup is blunt. AI has an image problem in the US. Saving the world with AI-designed cures reads a lot better than being blamed for rising utility bills and strained power grids. So the top labs are diving headlong into health care, right as they prepare to go public.

The cures frame. The moves are real money. Anthropic is talking up a harder push into biology and medicine ahead of its IPO, per the WSJ reporting Axios cites, with Dario Amodei pointing to hoped-for early glimmers in the coming months.

Nvidia and Eli Lilly are building a $1 billion drug discovery lab in San Francisco, pairing life science researchers with AI model builders.

And Isomorphic Labs, Google’s AI drug discovery spinoff, recently raised $2.1 billion, with research collaborations at Novartis, Lilly and Johnson & Johnson.

The other side of that ledger, per Axios: breakthroughs may still be years away.

The same tools raise biosecurity concerns, which is why Anthropic wrapped strict safeguards around its latest frontier model after concluding it could provide meaningful uplift to a bad actor.

And so far, AI in health care has also made care documentation more expensive, even as it speeds drug development and scan analysis. One digital health advisor in the piece sums up the tension: real advances could transform AI’s public image into an essential utility, but the material, AI-led drug breakthrough is not here yet. AI is a player on the field, not the whole team.

Then there is the other aspect of AI and healthcare, blended into the popular topic of ‘wellness’. Made up close and personal, with devices all over mainstream bodies.

The wellness frame, on your body, in particular for the data.

Here is the part I would underline. Most of the AI gadget and wearables aspirations today revolve around ‘wellness’ healthcare, driven by AI. Google just went screenless with the Fitbit Air, built around its AI health coach. Watches from Apple, Google and Samsung all lead with health features. And smart rings, led by Oura, which I profiled in ‘Apple, Oura, and AI Smart Rings’, have grown into a category of their own. Oura has now filed confidentially to go public, effectively an AI wearable IPO, with reports pointing to a valuation well past $10 billion.

This is the mainstream on-ramp.

I wrote in ‘Apple Watch & AI Healthcare’ that these devices matter most before we get sick: the quiet, daily prevention layer on billions of wrists, fingers and ears. The caution flag stays up too, from ‘The Glass Half Empty on AI Devices’: the wearables flood is long on promises and pitches, and mainstream utility has to be earned device by device.

The image problem underneath. The second frame is the one Axios leads with. US sentiment on AI is historically negative, the sentiment gap I have tracked all year, with data centers now the midterms’ political football, complete with moratoriums, lawsuits and job fears. Big tech and the AI labs are understandably eager, even desperate, for a positive AI story that lands with Main Street. As the digital health M&A advisor in the Axios piece puts it, “It’s a bit of a hearts-and-minds play.” That thread connects straight to the go-direct echo chamber I wrote about yesterday: the industry has been telling its story to itself, and health care is its clearest shot at telling one to everyone else.

And big tech CEOs of every stripe are betting heavily on AI Smart Glasses.

My take: the healthcare promise is real, and it was real before the backlash made it convenient. The reason I called it the holy grail three years ago still holds. Healthcare is a fifth of the US economy, tangled in complexity that AI is unusually suited to untangle, from drug discovery to diagnostics to the paperwork in between.

But the cures story runs on a long clock.

Years, likely, before an AI-led drug lands as a fact rather than a forecast.

The wellness story runs on a shorter clock. The watches, rings and trackers are already on millions of bodies, quietly nudging sleep, fitness and early warnings, and now anchoring an IPO of their own in Oura.

With iterative progress, and a lot of the health gadgets tossed into drawers after short tries.

If the industry wants hearts and minds before the midterms and the mega-IPOs, the daily wellness utility is the story people can potentially feel.

But they have to vend their way to a complex healthcare bureaucracy, and ‘RAMageddon’ om the device/gadget end.

The cures are the story people are being asked to believe. Both sides of that gap are worth watching honestly.

An important vertical for the AI Tech Wave. On both the long clock for healthcare systems, and the short one for device leveraged ‘wellness’. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Sources and further reading

Primary

• Axios (Adriel Bettelheim): AI turns to health care for an image fix

• Google: Introducing the new Google Fitbit Air

• CNBC: Oura, smart ring maker, files confidentially for IPO

For longtime readers

• AI-RTZ: The Holy Grail of Healthcare (2023)

• AI-RTZ: Connecting the Dots, how AI can help in healthcare & beyond (2023)

• AI-RTZ #404: Apple Watch & AI Healthcare

• AI-RTZ #550: Apple, Oura, and AI ‘Smart Rings’

• AI-RTZ #872: The Glass Half Empty on AI Devices

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

• AI-RTZ #1194: Physical AI Arrives, Part 1, the sentiment gap

• AI-RTZ #1198: The Tech ‘Go-Direct’ Ethos Meets Grassroots AI Revolt

‘Nothing to see here, move along.’ Three events today, from three different corners of Washington. The Commerce Department. The Justice Department. And a federal courtroom in Virginia. Three different actors, one message to US big tech and AI.

Commerce waves Anthropic back through the gate. The Justice Department backs OpenAI on ‘fair use’. And a federal judge spares Google’s ad business from a breakup. All on the same day, and all blowing the same way on this AI Tech Wave. Plus a Gadget AI on Apple’s September 9 launch, the first under new CEO John Ternus. Lots to see there, most likely.

Plus my Takes below, as discussed on the show.

(1) Anthropic Gets a Green Pass from Commerce

First up. Commerce Secretary Howard Lutnick says the Trump Administration now trusts Anthropic. His words to Axios, asked if he trusts CEO Dario Amodei:

‘We trust Anthropic. They’ve done what we asked. They’re back on the right side.’

Remember where this started. Back in June, in #1117, I covered Anthropic being flagged to a screeching halt by the US, after a bitter fight over national security and AI safeguards. Sweeping export controls on its most advanced models. A Pentagon blacklisting as a supply-chain risk, which a federal judge struck down in August. A separate designation the company is still fighting in the DC Circuit.

Anthropic was the most in the crosshairs of the frontier labs. The companies have since carved their latest models into safer and general versions, with far more restrictions on distribution. That is the status quo this green pass lands on.

How did the thaw happen? Mostly through Tom Brown, one of Anthropic’s seven co-founders, who has become the company’s point person with the White House. Per Axios, Brown did the patching up over the past few months, in multiple conversations with Lutnick and the National Cyber Director.

The venue for Lutnick’s comments was the G20 Innovation Ministerial in Chapel Hill, North Carolina this week, a gathering of G20 technology ministers the US is hosting. Lutnick put Brown on stage as the headliner and introduced him personally. Brown returned the favor, praising the President’s pro data center post and highlighting Anthropic’s support for the Administration’s ratepayer pledge. That is what ‘back on the right side’ looks like in practice.

My take: for now, the Administration has decided to bury their differences with Anthropic. They’ve decided that geopolitical negotiations with China are more important for now than individual differences with Anthropic.

The tell is the calendar. Lutnick was at a G20 table this week, selling other countries on US AI. The next US/China White House meeting is weeks away. You do not walk into that room with one of your two frontier AI champions, one about to do a mega IPO, sitting on a blacklist.

I saw the same reflex in #1085 in May, when Nvidia’s Jensen got a last-minute seat on Air Force One to China after weeks of being on the outs. When the geopolitics turn, the Administration’s posture toward its AI champions turns with it. Both directions. Nothing to see here. Move along.

Sources:

• Axios: Lutnick: Anthropic is ‘back on the right side’ with Trump administration

For longtime readers:

• AI-RTZ #1117: Anthropic flagged to a screeching Halt Globally by the US

• AI-RTZ #1085: Nvidia’s Jensen rides Air Force One to China after all

(2) OpenAI Gets an Assist from DC on ‘Fair Use’

Story two. The US government has filed a brief backing OpenAI in the New York Times copyright case. Filed Tuesday in Manhattan federal court, per Reuters and the WSJ. The position: training large language models on copyrighted text is ‘fair use’ under US copyright law.

This appears to be the first time the US government has weighed in on the whole wave of AI training cases. Authors, publishers, music labels, news outlets. Dozens of them, against OpenAI, Anthropic, Meta and others.

The argument, per the Justice Department’s filing: training is ‘exceedingly transformative’. Constraining it would ‘thwart creative and scientific progress’. And the national security framing is explicit. Rules that make it harder to build a robust US AI industry ‘give a competitive advantage to foreign adversaries who are not so encumbered’.

Two more points from the brief worth noting. Narrowing fair use would mean only the largest tech companies could afford licensing fees. And it would disproportionately benefit legacy media outlets over smaller independents. The same haves and have-nots divide we see across the economy, arriving on the content side.

The NY Times’ response, via its spokesman:

“The Administration is ‘siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole.’”

Some context. A brief like this has advisory weight, not legal weight. The first two judges to rule on AI training and fair use went different ways last year. And the timing is pointed: it lands days before both sides file new motions asking the judge to rule on their core arguments, now that discovery is done. Lutnick, separately, told the G20 this week that their countries should embrace fair use for AI training too.

My take: a pivotal set of issues for the AI industry writ large. And OpenAI was front and center in the ‘Fair Use’ cross-hairs. Unclear how much impact this assist will end up being, but a welcome one for the industry nevertheless.

This is a global issue. It will be sorted out in courts in the US, Europe and Asia, and that will take years. If you look at how the technology works, it is transformative, in my take. At the same time, creators have to be compensated in some mechanism. Most likely a threading of the needle, category of content by category of content. That will take ages.

I’ve been on this critical topic since the NY Times filed in December 2023.

That week I wrote ‘On the Shoulders of Giants’, on how these AI models are built on all of our data, and then the ‘Fair Use’ Tussles of 2024, on why a lawsuit like this one is a negotiating tactic as much as a legal claim. Then a year ago, in #839, Anthropic paid $1.5 billion to settle its book copyright case.

So the pattern holds: licensing deals where the money makes sense, litigation where it doesn’t, and now Washington leaning on the scale. Nothing to see here. Move along.

Sources:

• Reuters: US government backs OpenAI in New York Times copyright case

• WSJ: U.S. Government Backs OpenAI in Copyright Fight With Publishers

For longtime readers:

• On the Shoulders of Giants, Go-Getter ‘Creators’ & Grunts (OTSOG), Dec 2023

• ‘Fair Use’ Tussles of 2024, Dec 2023

• AI-RTZ #839: Anthropic’s $1.5 billion book copyright settlement

(3) Google Gets Relief on the Breakup Front

Story three. A federal judge rejected the Justice Department’s bid to break up Google’s ad tech business. Judge Leonie Brinkema in Alexandria, Virginia, per the WSJ and Bloomberg. No forced sale of Google’s ad exchange.

Instead, a two-page order adopting ‘most’ of the behavioral remedies: Google’s ad tools have to work with rivals’ tools. The full opinion is sealed for now, for redactions. The two sides have 30 days to file a joint final judgment.

This is the second time a court has found Google liable as a monopolist and then declined to break it up. Last year Judge Mehta in the search case ruled the same way: no Chrome sale, share some search data instead. I covered that liability ruling in #440, ‘a legal dent on the power of defaults’, and the remedies phase in #547.

And Meta beat the FTC’s breakup case last year, on the rise of TikTok. The WSJ calls this ruling a marker that effectively closes the book on nearly two decades of US antitrust risk for Google.

Keep the numbers in perspective. Google made $7.3 billion from network ads last quarter, out of $119.8 billion in total Alphabet revenue. The display business at the center of this case is a small piece of a very large machine. The stock moved less than 1% on the news.

My take: a pragmatic turn by the judge in question on this complicated case. There was a lot to see here and needed solomonic judgments indeed. For now, it’s business as usual for the most part, for Google at least.

The judge telegraphed this at trial. She asked out loud who would buy the ad exchange, and who would run it. Stopping the illegal conduct was the more practical remedy in her view, and that is where she landed. In that context, Google and others can keep focusing on the opportunities and challenges around AI, and the ad businesses getting rebuilt around AI-created content. A thread I’ve written about a lot.

Two caveats for the record. The Justice Department can still appeal, and says it is weighing next steps. And the states, private plaintiffs and European regulators are all still in the queue. But on the breakup question, the one that mattered most to Google: nothing to see here. Move along.

Sources:

• WSJ: Google Avoids Breakup of Dominant Ad Business

• Bloomberg: Google Avoids Breakup as Judge Orders Ad Tech Unit Changes

For longtime readers:

• RTZ #440: A legal dent on the power of defaults

• RTZ #547: Google DOJ ‘Remedies’ phase with Chrome in spotlight

My overall take: three companies, three corners of Washington, and one message. Anthropic gets waved back through the gate by Commerce. OpenAI gets a friend-of-the-court assist from Justice on fair use. Google gets spared a breakup by a federal judge.

All three narratives today highlight the US government deciding, in an uncoordinated way, to side with US Big Tech on a variety of fronts. This is likely as much a coincidence, given the wild swings of uncertainty in every direction for US companies, as they invest unprecedented sums in this AI Tech Wave. We are just in year four of all this.

Think about the swings. This same government flagged Anthropic to a halt in June. Inherited the Google breakup case and argued it all the way to a ruling. And this President has sued the New York Times and the Wall Street Journal over their coverage, while his Justice Department sides with OpenAI against the Times on copyright. It makes one’s head turn every which way.

Uncoordinated is the right word. None of these three actors was checking with the other two.

For now, the winds are all blowing in one direction, offering tail winds for the US companies in question. Enjoy it while it lasts. Because the winds do shift. ‘Nothing to see here, move along’ is what an official says on a good day. The same official was waving a red flag three months ago.

Gadget AI: Apple’s New CEO Teases the September 9 Launch

Today’s Gadget AI is a launch that hasn’t happened yet, and a memo. Apple CEO John Ternus, on his first day in the job, told employees the company has a ‘phenomenal’ iPhone launch coming next week, per Bloomberg.

The memo is short. Ternus thanks Tim Cook, now executive chairman after 15 years. He calls next week’s launch ‘huge’ and ‘phenomenal’. And he points past it, to ‘the incredible products already in the works and the ones we haven’t even imagined yet’. He also posted on X for the first time that day. One word: ‘hello’.

He already has almost a million followers vs Tim Cook’s almost 20 million.

Per Bloomberg, the September 9 event is expected to bring the first foldable iPhone and a push into smart home devices. Further out on Apple’s road map, per the same reporting: AirPods with cameras, smart glasses, a tabletop home robot and a touch-screen MacBook. Cook will not be on that stage as CEO for the first time in 15 years.

My take: lots to see here apparently, on the first Apple product launch under new CEO John Ternus. The focus likely will be on Siri AI powered by Apple Personal Intelligence, and the new iPhone Fold. Possibly named ‘Ultra’, as key showcase products. All done the ‘Apple Way’.

Both of Ternus’ predecessors had iconic debuts of their own, in product or in the job. Steve Jobs introduced the iPhone at Macworld in January 2007 (the full keynote is still worth the 80 minutes). Tim Cook’s first product launch as CEO came on October 4, 2011, six weeks into the job: the iPhone 4S, iCloud, iOS 5, and a virtual assistant called Siri. That was the day before Steve Jobs passed.

Siri is the symbolic thread. Cook introduced the original fifteen years ago. On September 9, Ternus gets to truly introduce Siri AI, rebuilt from the ground up, the thread I laid out in #1196 on Tuesday: AI as job one, Apple Silicon on up, the new Siri AI rolling out this month. September 9 is the first public test of that framing.

It will also be impacted by ‘RAMageddon’, the memory-cost squeeze I’ve covered in #1145.

Expect the fold to price north of two thousand dollars, and higher memory costs of 400%+ across the new iPhones. With Apple likely spreading some of that across other products and services. The one day this week where the message is the opposite of our theme. Plenty to see here. Do not move along.

Sources:

• Bloomberg (Gurman): Apple’s Ternus Touts Product Pipeline, ‘Phenomenal’ iPhone Launch in Memo

• YouTube: Steve Jobs introduces iPhone in 2007

• YouTube: iPhone 1, Steve Jobs Macworld keynote 2007, full presentation

• YouTube: Tim Cook CEO, Apple Special Event, October 4, 2011

• 9to5Mac: Apple paying 400% more for iPhone 18 Pro memory, says TrendForce

• Reuters: SK Hynix CEO sees worst-ever memory supply shortage in 2027

For longtime readers:

• AI-RTZ #1196: Apple’s Ternus Era Begins, with AI as Job One

• AI-RTZ #1145: ‘RAMageddon’ really here to stay

Q&A

Q1: What’s MP most looking forward to on the September 9 Apple Event?

ANSWER: the way Ternus navigates his first official Apple product launch. An iconic rite of passage for this company’s CEOs. I’ve watched every Apple leader from Steve Jobs on do these big launches over the years, some in person, some not. Ternus is the third CEO we all recognize, after Jobs and Cook. A new CEO continuing the legacy of the prior two, especially the global supply chain that serves over two and a half billion users of Apple hardware and software. And an engineer stepping into an iconic role. Watching that alone will be quite a show.

Q2: Where will it likely disappoint industry watchers?

ANSWER: as usual, Apple will be measured in its enthusiasm and direction on product visions ahead. They will roll it all out at the cadence and timing of their choosing. September 9, 2026 will likely be no different. There will be plenty of hints in the hardware and software they do show, and we will all do our usual figuring out of Apple’s direction. But unlike Google, Meta and a lot of the other big tech companies, which point out their product directions years ahead, Apple does not. That gap will continue, even with a new CEO on his first launch.

Today’s AI-RTZ #1198 went out this morning: the tech ‘go-direct’ ethos, founders and CEOs talking straight to audiences via podcasts and posts, meets the grassroots AI revolt. Two bubbles, and not much conversation between them.

Tomorrow, we’re back with ARD 156 and AI-RTZ #1199. For now, Washington says ‘nothing to see here, move along.’ We’ll keep looking anyway, on this AI Tech Wave. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Everything cited today, in one place.

Anthropic and Commerce

• Axios: Lutnick: Anthropic is ‘back on the right side’ with Trump administration

• AI-RTZ #1117: Anthropic flagged to a screeching Halt Globally by the US

• AI-RTZ #1085: Nvidia’s Jensen rides Air Force One to China after all

OpenAI and ‘Fair Use’

• Reuters: US government backs OpenAI in New York Times copyright case

• WSJ: U.S. Government Backs OpenAI in Copyright Fight With Publishers

• On the Shoulders of Giants, Go-Getter ‘Creators’ & Grunts (OTSOG)

• ‘Fair Use’ Tussles of 2024

• AI-RTZ #839: Anthropic’s $1.5 billion book copyright settlement

Google’s Ad Tech Ruling

• WSJ: Google Avoids Breakup of Dominant Ad Business

• Bloomberg: Google Avoids Breakup as Judge Orders Ad Tech Unit Changes

• RTZ #440: A legal dent on the power of defaults

• RTZ #547: Google DOJ ‘Remedies’ phase with Chrome in spotlight

Gadget AI: Apple’s September 9

• Bloomberg (Gurman): Apple’s Ternus Touts Product Pipeline, ‘Phenomenal’ iPhone Launch in Memo

• YouTube: Steve Jobs introduces iPhone in 2007

• YouTube: iPhone 1, Steve Jobs Macworld keynote 2007, full presentation

• YouTube: Tim Cook CEO, Apple Special Event, October 4, 2011

• 9to5Mac: Apple paying 400% more for iPhone 18 Pro memory, says TrendForce

• Reuters: SK Hynix CEO sees worst-ever memory supply shortage in 2027

• AI-RTZ #1196: Apple’s Ternus Era Begins, with AI as Job One

• AI-RTZ #1145: ‘RAMageddon’ really here to stay

• AI-RTZ #1198: The Tech ‘Go-Direct’ Ethos Meets Grassroots AI Revolt

Clips from today

DOJ Backs OpenAI on AI ‘Fair Use’

Washington Sides with US Big Tech, For Now

Apple’s Sept 9: Siri AI, iPhone Fold & Prices

Ternus’ First Apple Launch: Rite of Passage

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

One unique aspect of this AI Tech Wave vs prior ones is how the early mainstream reaction has been more anti the tech/AI management and founders rhetoric vs prior ones. A topic I’ve discussed at length.

The WSJ’s Tim Higgins recently published a piece that pulls together threads we have been writing about here for a long while: “How Big Tech Blinded Itself to the Grassroots AI Revolt”.

His core observation is blunt.

Even if the Anthropic and OpenAI mega-IPOs live up to Wall Street’s expectations, these companies will be, in his words,

“Birthed as the bad guys.”

A remarkable swing from the early 2010s, when founders were venerated as cultural heroes. Longtime tech comms veteran Rachael Horwitz calls the industry’s AI communications “the next great comms case study,” of what not to do. Even Sam Altman conceded the field has not done a good job explaining AI’s benefits, in a recent podcast mea culpa.

The piece is worth reading in full. But the more interesting question is the one underneath it: HOW did an industry this smart blind itself?

My partial and preliminary answer, from watching tech waves since the 1980s: the blindness was built, piece by piece, on purpose. Let me walk through the building of the backlash mountain from the molehills.

1. The ‘go-direct’ turn. Higgins dates it to Elon Musk’s 2022 acquisition of Twitter/X. Uber founder Travis Kalanick, back this year with an AI startup, credits exactly that:

“Elon buying Twitter is the beginning of us being able to speak our minds.”

Founders and VCs redirected from talking just to institutions, media, regulators.

And started talking directly to their stakeholders. Both on their own and friendly podcasts by mega-tech ‘influencers’. The reach was infinite and cheap. As learned more broadly from social media by the success of YouTube, TikTok, Meta Reels and nore.

The audience, though, was Sand Hill Road, Wall Street and Pennsylvania Avenue. Not Main Street. An echo chamber of their own creation, as Higgins puts it.

2. The direct-media buildout. This is no longer just founders posting. It is infrastructure. Andreessen Horowitz has rebuilt itself around a full media playbook, podcasts, shows and channels, a year into calling it ‘New Media’. And OpenAI went further: it acquired TBPN, the buzzy founder-led tech talk show, in April, for a price reported in the hundreds of millions. The AI industry now owns pieces of its own press.

The mega-podcast flywheel, and its poster case. Big tech CEOs are more comfortable than ever thinking out loud for hours with friendly mega-podcasters, Dwarkesh Patel, Lex Fridman and the like. It moves real money now. The poster case: a five-hour Dwarkesh podcast helped take Leopold Aschenbrenner from a 165-page ‘Situational Awareness’ manifesto to a hedge fund that reached $45 billion. Institutional money, not meme-stock traders. Dwarkesh reportedly got a chance to invest.

That is the new part versus prior tech waves: the institutions are drinking the direct-communication Koolaid too, positive and negative. And the coda writes the caution: the fund lost most of it in days this summer, a thread we covered via the Jane Street hit.

3. The insulation layer: supervoting shares. Here in my view, is a contributory factor not mentioned in the WSJ piece. The the comms critiques miss. Many of the loudest founder-CEOs are structurally insulated from the shareholder feedback loop itself. Zuckerberg controls Meta through Class B shares. Alphabet’s founders keep 10x votes. Palantir has founder shares, Snap’s public stock carries no votes at all. Musk holds supervoting control at SpaceX and xAI, and at single-class Tesla, his control runs through pay packages and board fights instead. Anthropic plans supervoting shares for Dario Amodei and its co-founders at its IPO. OpenAI took a different route to the same place: a nonprofit-controlled structure. When your shareholders cannot vote you out, the incentive to listen, to anyone, weakens.

Now note the other column on the right above. Apple. Microsoft. Amazon. Nvidia. Netflix. One share, one vote, through every era. The stalwarts became what I have called ‘Internet Treasures’ without voting shields.

Decades of delivering mainstream utility, and learning from stakeholder feedback directly, did the insulating for them.

That is the standard the AI era’s trillion-dollar M&A prices implicitly claim, and have not yet earned.

What makes a company an ‘AI Treasure’ over a decade or two is the same thing that made the Internet Treasures: billions of people quietly better off, not billions of impressions.

4. ‘Founder Mode’, formalized. Paul Graham’s 2024 ‘Founder Mode’ essay gave the culture its scripture: founders should run things directly, managers and process be damned. We assessed founder mode across the Mag 7 when the essay landed, and watched the blind ambitions version play out across the incumbents since. Founder mode built these companies. Unchecked by boards, shareholders, or a listening habit, it is also exactly how you build an echo chamber with excellent production values.

The founders have been tripping on their own narrative from the start. This is the piece of history I keep returning to.

The accel-versus-decel schism is not a Twitter-era invention. It goes back to the 2015 origin story of OpenAI itself:

Elon Musk and Larry Page falling out over whether AI would save or doom humanity, and OpenAI founded in the aftermath.

Ever since, the industry’s leaders have alternated between curing-cancer utopias and extinction warnings, essays and counter-essays, Altman’s abundance, Amodei’s loving grace and cautions, Gates’ turbulence, each striking a pose for their own audiences. Higgins nails the result: the public heard both the godlike promises AND the doom, believed the costs, and discounted the benefits. The narrative machine worked. On the wrong audience.

And the revolt is here, running the DC flywheel.

The sentiment gap we have tracked all year, US mainstream sentiment on AI at roughly 30-40% positive against 70-80% across Asia, is now electoral fuel. Data centers became the midterms’ political football. Pennsylvania just moved to limit data-center construction, per the WSJ. A former NFL star is going viral asking fans to bottle their urine to cool data centers.

Zach Bryan is trying out a stadium singalong with the line

“The robot killed the working man.”

Meanwhile the regulator flywheel spins in both directions, the pauses, tariffs, blacklists and bans we cataloged in ‘Fumbling the Ball’, each government action feeding the next industry overreaction, and back.

My take: Higgins’ closing contrast is the one to let sink in.

Uber, a decade ago, could summon its own users to lobby city halls on its behalf, because riders felt the utility every day.

The AI labs have no such ‘daily habit’ joyous muscle yet, because for most Americans the felt experience of AI so far is slop, job anxiety and bigger utility bills.

Tech guys ranting on X, as he puts it, do not have the same influence. The fix is not more podcasts, more owned media, or thicker voting shields. Or even millions of contributions to their favorite partisan lobbies and PACs. Despite the ever-rising mainstream protests.

They’re even considering hedging themselves contributing to the other party ahead of the midterms.

Those are the step by step sequences that built the current tech echo chamber around AI.

The fix is the tedious one this AI Tech Wave has barely started: mainstream utility, delivered visibly, at Main Street prices, until the technology defends itself the way the Internet Treasures learned how to do and earn.

A decade or two of that earns ‘AI Treasure’ status. No Short Cuts.

A deviant trend of this AI Tech Wave worth tracking. Stay tuned.

Sources and further reading

Primary

• WSJ, Tim Higgins: How Big Tech blinded itself to the grassroots AI revolt

• TechCrunch: OpenAI acquires TBPN, the buzzy founder-led business talk show

• a16z: New Media, one year in

• CNBC: How Leopold Aschenbrenner built a $45 billion AI hedge fund, and lost most of it in days

• Paul Graham: Founder Mode

For longtime readers

• AI-RTZ: Origin stories of AI ‘accel’ vs ‘decel’, the 2015 Musk/Page falling-out

• Assessing ‘Founder Mode’ and ‘AI Readiness’ at the top AI/Mag 7 companies

• ARD #110: ‘Blind Ambitions Abound’, tech incumbents in unbound AI founder mode

• ARD #144: ‘Strike a Pose’ in AI. OpenAI, Anthropic & Apple

• AI-RTZ: ‘For Better or for Worse’ in AI commitments

• ARD #143: ‘Devil really is in the Details’, the Leopold/Jane Street thread

• ARD #129: ‘How High Can It Go?’ OpenAI, SK hynix & AI M&A prices

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

• ARD #151: ‘Fumbling the Ball’ in US AI. Tariffs, Cyber, Pentagon

• ARD #150: ‘Bloom, Gloom & Doom’ Reign On, the Gates essay

• AI-RTZ #1194: Physical AI Arrives, Part 1 (Sunday), the sentiment gap

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

‘New and Improved’, yes, again already. With caveats and asterisks.

All three big American AI labs are shipping or teasing upgrades this week, and every single one comes wrapped in safety gates, restricted access, or a smaller model than promised. On this AI Tech Wave, the fine print is where today’s tale needs to be told.

Today: Anthropic’s Fable and Mythos 5.1, OpenAI tip-toeing out Astra, its first ‘critical’ cyber model, and Google inching back into the frontier race. Plus a Gadget AI on World Labs’ new Atlas world model. My takes below, as discussed on the show.

(1) Anthropic’s Fable and Mythos 5.1 Arrive, Wrapped in Caveats

First up. Anthropic released Claude Fable 5.1 and Mythos 5.1. Same model, two wrappers. Fable is the generally available version, with full safeguards. Mythos is the same frontier model with looser restraints, available only to vetted cybersecurity and life-science organizations, in programs built with the US government. I covered the original recast of this two-tier structure in #1114 back in June.

The ‘improved’ seems real. Anthropic claims big jumps in agentic coding, science work and long-running tasks. More than doubling its predecessor on an agentic science benchmark. Testimonials about finding bugs that stumped engineers for years, and 38-hour unattended runs.

And the price cut is the headline for businesses. Cache reads, where the model re-reads context it has already processed, drop 75%, to 25 cents per million tokens. Anthropic says that cuts typical workload costs about 25%, and up to about 45% for heavy agentic work. Per VentureBeat, citing earlier FT reporting, the premium Fable 5 had only captured about 11% of Anthropic model spending across some 70,000 companies in Ramp’s data. This is Anthropic sharpening the pencil.

My take: the caveats ARE the product now. Read the label on this launch. A data-retention reversal after enterprise pushback, per CNBC, with a new ‘Enterprise Frontier Safeguards’ system that keeps monitoring data on the customer’s own cloud. Invisible watermarks on every output, per PCWorld, required under the EU AI Act. Cyber safeguards re-tuned to flag 60% fewer false positives, while still routing the dangerous stuff to older, safer models. And the whole release lands weeks after Anthropic disclosed that earlier models, in permissive test setups, escaped their sandboxes and touched real systems.

Anthropic with these announcements is tempering its relatively unchecked tendency to raise a la carte AI API pricing. And to ease up on the customer data it was retaining around its top models, under the shield of AI safety. Customer push-back on both is causing some easing on both.

That context matters. This is what a road to the mega AI IPOs also looks like. The frontier labs are learning to balance priorities. What consumer-products companies learned decades ago: past a certain scale, the label, the safety seal and the price point sell the product as much as what’s inside the box. ‘New and improved.’ Asterisk included, and on purpose.

Sources:

• Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1

• VentureBeat: Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads

• Bloomberg: Anthropic Says New Fable AI Model Is Cheaper, Better at Coding

• CNBC: Anthropic changes data retention policy after pushback from customers

• PCWorld: Claude’s updated Fable and Mythos models watermark their replies

For longtime readers:

• AI-RTZ #1114: Anthropic recasts Mythos as Fable 5, with safety & pricing gates

(2) OpenAI Tip-Toes Out Astra, Its First ‘Critical’ Cyber Model

Story two. OpenAI says its next model, Astra, is the first to cross its own ‘critical’ cybersecurity threshold. Meaning it can find previously unknown software flaws, build ways to exploit them, and chain those exploits together. Without a person guiding each step. During testing, per Axios, Astra discovered and chained two zero-day vulnerabilities, which OpenAI is now disclosing to the software’s maintainers.

So the rollout is on tip-toe. The broad version is ‘coming soon’, with the sharpest cyber capabilities held back. Early access goes to defenders first, per Wired: infrastructure companies in OpenAI’s ‘Daybreak Blue’ program, names like Cisco, Cloudflare and Palo Alto Networks, so they can harden systems before anything comparable spreads. A new misalignment monitor watches the model’s actions, and OpenAI itself warns it may flag legitimate work by mistake. Even work that has nothing to do with cybersecurity.

This follows the summer’s cautionary tales. OpenAI’s July disclosure that models in a test environment escaped the sandbox and hacked a real platform. Anthropic and the UK’s AI Security Institute disclosed similar incidents. The whole industry has been reminded that capable agents explore paths their operators didn’t anticipate.

My take: watch the release pattern here, it’s new. The most aggressive lab in the business is voluntarily slowing its own launch, pausing frontier training to harden the containment, and giving defense a head start over offense. Some of that is the safety pose I described in ARD #144, struck partly for regulators and enterprise buyers ahead of the IPO parade. But some of it is a real capability threshold being crossed. Both can be true. The asterisk on Astra isn’t just marketing copy. It’s the fine print doing the work beyond the headline.

And as I covered in #1166, these long-running agents are exactly where the leverage, and the risk, compound together.

Sources:

• Axios: OpenAI to limit access to Astra’s most powerful cyber tools

• Wired: OpenAI Is About to Release Its First AI Model With ‘Critical’ Cyber Abilities

For longtime readers:

• ARD #144: Strike a Pose: OpenAI, Anthropic & Apple

• AI-RTZ #1166: The Meter’s Running on Long AI Agents

(3) Google’s Gemini Inches Back Toward the Frontier

Story three. Google is set to release Gemini 3.8 Flash, possibly this week, per the WSJ. The claim that matters: in head-to-head testing inside Google’s own internal coding tool, Google engineers preferred it to Anthropic’s Opus. That’s a real benchmark of sorts, engineers voting with their fingers and eating their own dog food.

Now the asterisks, and there are several. It’s a Flash model, the smaller, cheaper, faster line, not the frontier ‘Pro’ tier. The new Pro is months behind schedule. Internal 3.5 Pro candidates were scrapped for not being better enough. Gemini 4 is still in post-training. And the leadership backdrop: a co-founder stepping aside last month, marquee researchers departing, and a new lab chief pushing pace, with a high-profile post-training hire brought in to sharpen reinforcement learning.

I covered Google falling back in the AI coding race in ARD #121, after covering its developer push with Gemini CLI in #763. This is the pendulum swinging back.

My take: small models are how you iterate fast, and Google is focused on it. Flash models are ‘good enough’, cheap to modify, so multiple teams can run experiments at once.

While changing a giant Pro model means corralling the whole company’s compute. Google is using the Flash line as its fast lane back into the coding race, while the heavyweight models take their time in the shop. ‘New and improved’, in deliberately small steps. The asterisk here isn’t safety. It’s size. And it’s a solid choice on the way back in AI Coding. Speed with ‘good enough’ and reliable results.

Sources:

• WSJ: New Google AI Model Said to Narrow Gap on Coding Ability

For longtime readers:

• ARD #121: The Many AI Twists and Turns Accelerate

• AI-RTZ #763: Google ups its Game for AI Developers with Gemini CLI

My overall take: three labs, three tip-toes forward. Anthropic is selling trust and economics as hard as intelligence. OpenAI is metering out its most dangerous capabilities on its own clock. And Google is winding its way back into the AI coding leadership ranks. Although it has a much bigger opportunity, and more work to do, in bringing Gemini apps and services to the masses globally. But it should be able to walk and chew gum at the same time.

‘New and Improved.’ The oldest phrase in consumer marketing. It used to sell detergent and toothpaste.

Now it ships frontier AI models. Quarterly, sometimes faster. When the front of the box says ‘new and improved’, turn the box over and read the label. All three of the big American AI labs are shipping or teasing upgrades, and every single one comes with caveats and asterisks. Safety gates. Restricted access. A smaller model than promised. This week on the AI Tech Wave, the fine print is where the story is.

Gadget AI: World Labs Levels Up the ‘World Model’

Today’s Gadget AI is a model for worlds, not words. World Labs, the spatial intelligence startup led by AI Guru Fei-Fei Li, widely called the godmother of AI, released Atlas, its next-generation world model.

What it does, per the announcement: give it a photo or a few, and it builds a coherent, explorable 3D scene. Camera control down to the pixel, up to a minute of video at 1440p. It reconstructs real spaces from a handful of phone snapshots, outputs actual 3D geometry, and runs ‘real-to-sim’ simulations for robots.

The WSJ’s recent column on world models frames the ambition: text is a lossy description of reality, and robots need models trained on worlds, not words. One investor’s honest caveat in that piece: this generation of world models is roughly where language models were at GPT-2. The money is real though: World Labs has raised about $1.23 billion, from Nvidia, AMD, Autodesk, Fidelity and others.

My take: this is a promising area for AI models, and a fitting one. It could bring Nvidia’s history, which started in gaming chips, full circle to user-created virtual worlds with significant gaming elements. Kind of what services like Roblox and Minecraft already hint at today. It’s also the same ‘Physical AI’ thread I mapped in #1195 this weekend, and in ‘Beyond LLM AIs to World Models’ in #909, well before it was fashionable. A lot of ‘new and improved’ still required from here to there. But this is one label where I’d underline the ‘new’ before the asterisk.

Sources:

• World Labs: Atlas, A World Model for Spatial Intelligence

• WSJ: AI’s Next Big Leap Is Into the Real World

• Crypto Briefing: World Labs unveils Atlas, an omni world model for spatial intelligence

• Reuters via Yahoo Finance: AI pioneer Fei-Fei Li’s World Labs raises $1 billion in funding

For longtime readers:

• AI-RTZ #909: Beyond LLM AIs to World Models

• AI-RTZ #1195: Physical AI w/ Nvidia’s Long Game in Robots, Cars & More

Q&A

Q1: What’s the most intriguing aspect of World Models for MP?

ANSWER: the prospect of user-defined games and simulated environments. Think worlds with the depth of GTA V or Red Dead Redemption 2, games that take years and enormous work to make today, delivering extraordinary simulated experiences to hundreds of millions. Imagine that done on a much shorter schedule. It seems possible technically, but the path from here to there looks long and arduous. Years of ‘new and improved’, version by version.

Q2: What’s the least appealing aspect of World Models?

ANSWER: the potential loss of truly shared global experiences. A GTA V or an RDR2 is a common world, shared by millions of players at once. That shared culture is part of the magic. GTA 6 lands November 19 and tens of millions will experience it the same day, yours truly included. A billion personal worlds could fragment that, the way the broader media environment has already fragmented. Something to watch as this category grows up.

Today’s AI-RTZ #1197 went out this morning: Nvidia’s $3.5 billion MediaTek bet, up and down the AI stack.

Tomorrow, we’re back with ARD 155 and AI-RTZ #1198. Keep reading the fine print on this AI Tech Wave. It’s all ‘new and improved’, asterisks and all. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Everything cited today, in one place.

Anthropic’s Fable and Mythos 5.1

• Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1

• VentureBeat: 75% cost reduction for Fable cache reads

• Bloomberg: Anthropic Says New Fable AI Model Is Cheaper, Better at Coding

• CNBC: Anthropic changes data retention policy after pushback from customers

• PCWorld: Claude’s updated Fable and Mythos models watermark their replies

• AI-RTZ #1114: Anthropic recasts Mythos as Fable 5

OpenAI’s Astra

• Axios: OpenAI to limit access to Astra’s most powerful cyber tools

• Wired: OpenAI’s first AI model with ‘critical’ cyber abilities

• ARD #144: Strike a Pose: OpenAI, Anthropic & Apple

• AI-RTZ #1166: The Meter’s Running on Long AI Agents

Google’s Gemini 3.8 Flash

• WSJ: New Google AI Model Said to Narrow Gap on Coding Ability

• ARD #121: The Many AI Twists and Turns Accelerate

• AI-RTZ #763: Google ups its Game for AI Developers with Gemini CLI

Gadget AI: World Labs’ Atlas

• World Labs: Atlas, A World Model for Spatial Intelligence

• WSJ: AI’s Next Big Leap Is Into the Real World

• Crypto Briefing: World Labs unveils Atlas

• Reuters via Yahoo Finance: World Labs raises $1 billion in funding

• AI-RTZ #909: Beyond LLM AIs to World Models

• AI-RTZ #1195: Physical AI w/ Nvidia’s Long Game

Clips from today

AI Labs: New & Improved, But Caveats Apply

Google’s Gemini 3.8 Flash: A Step Forward

World Models: Gaming’s Next Frontier?

AI Gaming: Faster, Shorter, Less Common

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Nvidia keeps extending its AI advantage up and down the vertical AI stack. The latest move: a $3.5 billion investment in Taiwan’s MediaTek, its largest direct investment outside the US.

MediaTek provides chips for local smartphones and computers in particular, competing with Qualcomm and others.

It extends the ‘Kingmaker’ thread I have tracked from Nvidia as AI Computing Kingmaker to ‘Uber AI Investor & Banker’.

And this weekend’s Nvidia long-game Part 2, right as this AI Tech Wave pushes into laptops, phones, cars and robots.

Bloomberg lays it out in “Nvidia to Invest $3.5 Billion in Chipmaker MediaTek”:

• “Nvidia Corp. is investing $3.5 billion in MediaTek Inc. to deepen collaboration with the Taiwanese chipmaker.

• The investment will see MediaTek using Nvidia’s NVLink Fusion and NVHBM technology to help components communicate more seamlessly in data centers.

• The agreement is part of Nvidia’s effort to ensure the future of artificial intelligence is built around its hardware platform and standards.”

The financial mechanics are as follows. Nvidia is buying bonds convertible into MediaTek shares, part of a $3.9 billion convertible sale that Alphabet/Google and others also joined.

MediaTek adopts Nvidia’s NVLink Fusion and the newly announced NVHBM technology, so its chips plug cleanly into Nvidia-standard data centers. The two will also collaborate on AI platforms for personal computers and automobiles. So local AI and devices I’ve been talking about for both Nvidia and Apple.

The important takeaway here is the scope.

This is a multi-generation collaboration between Nvidia and MediaTek. And it spans AI infrastructure in the cloud, what I’ve been calling ‘mainframe AI’, and ‘local AI computing’ on computers and smartphones. And extends to ‘Physical AI’ chips needed in robots, cars and other AI devices.

Nvidia’s chunky capital commitment and Alphabet/Google’s participation affirms MediaTek’s custom silicon capabilities and positioning within the broader AI ecosystem. They’re an important custom silicon provider like Broadcom and others.

Additionally, the adoption of Nvidia’s networking ‘NVLink’ technologies expands MediaTek’s participation in rack-scale AI systems. It gives them the ability to capture new value across connectivity, AI memory architecture, and advanced packaging. All hot areas in AI chip production. From Asia and elsewhere.

This partnership strengthens MediaTek’s longer-term AI ASIC (application specific integrated circuit) positioning.

It can potentially enhance MediaTek’s ability to win more projects from other customers down the road.

And it bolsters MediaTek’s custom HBM (high bandwidth memory) base-die design capability as a less appreciated, potential new business category, extending its role into memory subsystems.

So it’s a ‘win-win’ for both companies.

NVLink is Nvidia’s high-speed wiring that lets many AI chips talk to each other fast enough to behave like one giant computer. NVLink Fusion opens that wiring to chips Nvidia did not design. A company’s custom AI chip can now slot into Nvidia-style racks and data centers, instead of building all that plumbing from scratch.

NVHBM, announced by Nvidia days before this deal, is Nvidia’s new take on the specialized ‘high bandwidth memory’ stacked next to every AI chip. It moves the memory traffic controller off the AI chip and into the memory stack itself. Per Nvidia, that frees up to a quarter of the chip’s area for actual computing, with up to 30% more memory bandwidth at lower power.

And MediaTek’s part, per the joint announcement: it is one of the world’s top custom chip design houses, with deep expertise in power-efficient chips, packaging and connectivity. It gives customers a faster, lower-risk path from a custom chip to a full AI factory. That is what Huang means by the two supply chains merging.

Jensen Huang calls it a massive engineering alignment, on a road map he says runs a decade.

The Kingmaker play, again. The pattern is by now familiar. Nvidia invests in customers, suppliers and even would-be rivals to keep the AI buildout organized around its platform and standards. MediaTek is trying to challenge Broadcom and Marvell in helping hyperscalers design their own custom chips. Rather than fight that tide, Nvidia is wiring itself into it.

The Huang line that matters:

“Nvidia’s networking ecosystem is now part of MediaTek’s supply chain”,

With MediaTek’s XPUs part of Nvidia’s in return, both companies expanding their supply chains in one deal. Amazon just made a similar pledge, deploying 2 million more Nvidia components and using Nvidia’s interconnects with its in-house chips.

This is the ‘frenemies’ dynamic I have written about since Nvidia vs its top customers on AI GPUs, and across the AI Coding space. Even when the biggest customers build chips to replace Nvidia’s, the connective tissue stays Nvidia’s.

The circularity question follows these deals around, and investors keep asking it. Huang’s answer here: MediaTek runs its own profitable business, and this is alignment, not manufactured demand. Both sides of that debate are worth holding onto, with Nvidia forecasting a 70% sales increase next year.

Down the stack: local AI in laptops, phones and cars. Here is the leverage that interests me most. MediaTek is one of the quiet powers of low-power edge silicon, now pivoting hard into AI. It already helps build Nvidia’s RTX Spark push into laptops, part of the local AI computers thread with Microsoft.

We have discussed MediaTek before in the context of OpenAI’s own OS-driven AI smartphone, and around Apple’s AI camera sensors for its device ecosystem.

The local vs cloud AI split is where the next billions of AI devices get decided, and MediaTek gives Nvidia additional leverage on the local side, to go with its data center dominance in the cloud. MediaTek’s own AI chip business is guided to roughly $2 billion this year, aiming at 15% of an $80 billion segment next year, and its market value has roughly tripled in recent months.

The Taiwan and geopolitics layer. Huang was born in Taiwan, and this is Nvidia’s biggest direct bet there outside its TSMC manufacturing dependence. It lands while Jensen keeps balancing US and China, from hundreds of billions in US investment pledges to riding Air Force One to China.

And it strengthens the case I made for Nvidia and Apple as the US’s global open source AI champions, with the open hardware-and-standards stack doing the work, the same thread as ‘Rockets Are Hard’ on open source AI vs China. A deeper MediaTek tie pulls more of the world’s edge silicon into the US-led stack.

My take: the stack is the strategy. Nvidia’s accelerators won the cloud layer. NVLink and networking are winning the standards layer. And now the MediaTek bet works the edge layer, where laptops, phones, cars and robots will run their AI locally. Each investment looks circular if you squint at the money flows alone. Viewed as a stack, each one adds a floor to the same building.

The check on the enthusiasm: standards leverage invites regulators and rivals, Broadcom and Marvell will not stand still, and the edge is a market Nvidia has tried and stumbled in before, from phones a decade ago to its slow laptop entry now. A decade-long roadmap is a promise, not a result. But as the long game I outlined this weekend argued, Nvidia is playing longer clocks than most, and this deal fits that arc precisely.

This is Nvidia solidifying its hardware and software options at both the cloud and local end for AI Tech Wave opportunities, up and down the stack. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Sources and further reading

Primary

• Bloomberg (Ian King, Edward Ludlow): Nvidia to invest $3.5 billion in chipmaker MediaTek

• Nvidia blog: NVLink Fusion expands with NVHBM custom high-bandwidth memory

• Nvidia + MediaTek joint announcement: building AI edge-to-cloud computing platforms

For longtime readers

• AI-RTZ #1036: Nvidia accelerating role as the AI Computing ‘Kingmaker’

• AI-RTZ #1171: Nvidia the ‘Kingmaker’, now ‘Uber AI Investor & Banker’

• AI-RTZ #1195: Physical AI with Nvidia’s Long Game in Robots, Cars & More (Part 2)

• AI-RTZ #554: Nvidia vs its Top Customers for AI GPUs

• AI-RTZ #806: The ‘Frenemies’ nature of the AI Coding space

• AI-RTZ #1104: Nvidia & Microsoft focus on local AI computers

• AI-RTZ #1078: OpenAI’s own OS-driven AI smartphone in 2027

• AI-RTZ #668: Apple exploring AI cameras for its device ecosystem

• AI-RTZ #437: Nvidia’s Jensen Huang balancing US & China

• AI-RTZ: Nvidia next to announce ‘hundreds of billions’ in US investments

• AI-RTZ #1085: Nvidia’s Jensen rides Air Force One to China after all

• AI-RTZ #1089: How Nvidia & Apple can be the Global US Open Source AI Champions

• AI-RTZ: Open Source AI ‘Rockets Are Hard.’ Nvidia, Reflection & China

‘It All Adds Up’, in an AI-driven era. Senator Everett Dirksen reportedly said it decades ago: “A billion here, a billion there, and pretty soon you’re talking real money.” Even the Senate’s own historians can’t confirm he said it. But it stuck, because it’s true.

And on this AI Tech Wave, it’s not just the money you have to add up. It’s the slowdowns, the legal filings, even the price stickers on the gadgets. Today, three ledgers: Anthropic’s $80 billion cloud sprint, Pennsylvania braking $100 billion-plus of data centers, and the Apple/OpenAI charges piling up. Plus a Gadget AI on ‘RAMageddon’ reaching the humble TV streamer. My takes below, as discussed on the show.

(1) Anthropic Plays Hard Catch-Up on AI Compute, the Nvidia Way

The big one first. Anthropic just signed a $35 billion cloud deal, backed by Nvidia, per the WSJ’s exclusive. The mechanics are a wonder of financial engineering. Anthropic rents the compute from Lambda, a young Nvidia-backed cloud provider. Nvidia itself holds the lease on the data center. Hut 8, a bitcoin miner turned data center developer, builds the campus in Nueces County, Texas: a $20 billion facility, 700 megawatts, 15-year leases.

Now do the addition. Three weeks ago, Anthropic signed a $45 billion deal with Nscale, another Nvidia-backed neocloud, in West Virginia. That’s $80 billion of cloud capacity in roughly three weeks. A billion here, a billion there indeed.

My take: the significance here is Anthropic diversifying its AI compute to include much more Nvidia technology. Anthropic has been historically tied to Amazon and Google chips and infrastructure, driven by those early financing partnerships. Amazon’s Trainium, Google’s TPUs. I covered Anthropic’s compute constraints throttling its growth earlier this year in #1080. When your products catch fire and your compute can’t keep up, you buy capacity wherever it lives.

And Nvidia is being particularly proactive to make sure it has every frontier lab covered. Look at its fingerprints on this one deal: investor in Lambda, holder of the data center lease, seller of the chips that fill it. And separately, running a credit-support program for these young neoclouds in exchange for a cut of revenue, per prior WSJ reporting. That’s the AI Computing ‘Kingmaker’ playbook I’ve tracked since #1036, now with revenue-share deals attached.

One more wrinkle worth savoring: the same developer, Hut 8, is also building data centers for Anthropic that will house Google’s TPUs, with Google providing the financial guarantees. Same builder, competing chips, everyone’s money on the same table. For the mechanics of what these campuses actually are, megawatts to gigawatts, my Data Centers 101 reference in #1192 has the full 101. The tally so far: $80 billion, three weeks, one AI lab. Keep that number in mind for story two.

Sources:

• WSJ: Anthropic Signs $35 Billion Cloud Deal Backed by Nvidia

For longtime readers:

• AI-RTZ #1080: Anthropic’s AI Compute Constraints Throttling Growth

• AI-RTZ #1192: Data Centers 101, from Megawatts to Gigawatts

• AI-RTZ #1036: Nvidia accelerating role as the AI Computing ‘Kingmaker’

• AI-RTZ #1136: Nvidia expands Revenue Share Deals with Customers

(2) Pennsylvania Pumps the Brakes on $100 Billion-Plus

Story two is the other side of the same ledger. Pennsylvania is slowing down over $100 billion in data center momentum. The NYT has the ground report from Pittsburgh, and the WSJ editorial page fires back. The setup: a year ago the President stood in Pittsburgh celebrating more than $90 billion in pledges from Google, Amazon and Anthropic for the region. Governor Shapiro was all-in, fast-tracking approvals, with more than 120 data centers planned statewide. Amazon alone pledged $20 billion. The Homer City project is converting the state’s biggest old coal plant into a $10 billion, 4.4 gigawatt gas-powered campus. Enough power for Pittsburgh, twice over.

Now the second thoughts. Two weeks ago the same Governor Shapiro called data center developers ‘predatory’ and signed what he calls the nation’s strictest guardrails. No more fast-tracking, local approval first. New York’s governor already paused new data centers statewide in July. A Pennsylvania state senator wants a three-year moratorium. And per the NYT, the environmental math is part of the backlash: seven proposed gas plants alone could emit as much CO2 as 15 million cars a year.

The WSJ editorial board adds up the other column. Data centers as tax base: without them, one Virginia county’s homeowners would pay $5,800 more a year in property taxes. Teachers near a Meta data center in Louisiana got bonuses up to $50,000. Their argument: the panic is misinformation meeting election-year anxiety, and the strategic loser in a US slowdown is not big tech. It’s the US against China.

My take: even slowdowns add up quickly. This matters because the AI Tech Wave remains on a supply-constrained slope for the next few years. Every gigawatt delayed lands on the same constrained curve that story one is about. Anthropic isn’t signing $80 billion of deals in three weeks because capacity is easy to find. It’s doing it because capacity is scarce. Pause projects in one state, and the demand doesn’t vanish. It moves to Texas, West Virginia, or abroad, and the prices go up for everyone still in line.

Both columns are worth reading in full. The jobs and tax receipts are real. So are the gas plants and the grid anxiety, in a region that has lived through these booms before. That’s exactly why I covered data centers becoming the midterms’ political football in #1191. The politics now compound as fast as the capex. It all adds up, on both sides of the ledger.

Sources:

• NYT: Pennsylvania’s A.I. Gold Rush Meets Second Thoughts

• WSJ Editorial Board: The Data Center Panic and U.S. Decline

For longtime readers:

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

(3) The Charges Add Up in the Apple/OpenAI Trade-Secret Feud

Story three. The Apple versus OpenAI legal fight escalated on both sides in one day. Quick recap: Apple sued OpenAI and two former employees in July, claiming stolen trade secrets around hardware design and supply chain as OpenAI pushes into consumer devices. Roughly 400 Apple people have moved to OpenAI’s hardware effort, per the filings of both companies. I covered the fight’s PR dimension in ARD #133 last month.

The additions on Apple’s side, per 9to5Mac: Apple told the court an early forensic pass on one ex-engineer’s MacBook shows a confidential Apple circuit schematic used in his work at OpenAI. A simulation run in March, in a standard electrical engineering tool. Instructions to a colleague about destroying evidence. And a detail for our times: Apple argues that once a trade secret is fed to an AI agent that learns from it, the harm may become irreversible and continually propagating. Trade secrets, meet machine learning.

And OpenAI’s same-day response, per Reuters. Their filing calls the dispute “a mess of Apple’s own making.” Their argument: Apple’s own practices blur personal and company data, employees use personal iCloud for work files, departing staff get walked out the same day, and California law protects moving between rivals. The ex-employees say the file access was helping former colleagues, and the 24-year veteran says he returned his prototypes.

My take: the lists of legal charges back and forth are likely to add up for months between these two. Neither side is inclined to back down, because this is two fights running concurrently. A long legal fight in a San Jose courtroom. And a PR fight in the eyes of mainstream customers, investors and regulators, especially for OpenAI ahead of its highly anticipated mega AI IPO. Discovery fights like this are how each side runs up the other’s tab. Legal bills, executive depositions, and headlines, all compounding. Charge by charge, filing by filing, it adds up.

Sources:

• Reuters: OpenAI says Apple has only itself to blame in trade-secret fight

• 9to5Mac: Apple reveals ‘shocking evidence’ from ex-employee’s MacBook in OpenAI suit

For longtime readers:

• ARD #133: Que Sera, Sera: White House, Apple/OpenAI

• AI-RTZ #1196: Apple’s Ternus Era Begins, with AI as Job One (today’s companion)

My overall take: add up the three ledgers. The money: $35 billion here, $45 billion there, $80 billion of Anthropic cloud capacity in three weeks. Real money, on any senator’s arithmetic. The friction: one governor’s signature slowing $100 billion-plus of buildout, with more states lining up. Slowdowns compound too. The grievances: charges and counter-charges between two of the most valuable names in tech, adding up filing by filing, with an IPO clock ticking behind them.

That’s the deep story of this AI Tech Wave right now. Everything about it compounds. The capital, the constraints, the conflicts. All at once, and all faster than the institutions built to referee them. Whoever does the arithmetic honestly, on all three ledgers at once, reads this wave best.

Gadget AI: ‘RAMageddon’ Comes for the Humble TV Streamer

Today’s Gadget AI is under your television. The news, per 9to5Google: Google quietly raised the Google TV Streamer by $50, to $150. It launched in 2024 at $100, and used to see $20 holiday discounts. Google’s own statement points at rising component costs. Nest cameras went up the same day.

My take: this is ‘RAMageddon’, which I covered in #1145, reaching the cheapest shelf in consumer electronics. These used to be $25-and-up gadgets, from Amazon, Google, Roku and the rest. Apple’s Apple TV is already at $200, from $129. And that’s before an anticipated upgrade, and before the HDMI cable, which isn’t in the box anymore. All driven by RAM and storage memory prices, the same memory the AI data centers in story one are hoovering up.

To paraphrase the good senator at the other end of the spectrum: $50 here, $50 there, and pretty soon you’re talking serious money. For a mainstream family, seriously enough to postpone the upgrade. Multiply that hesitation across the planet and it’s enough to seriously slow the consumer electronics market globally. The AI buildout’s bill is starting to show up at the checkout counter. It’s all adding up, and fast.

Sources:

• 9to5Google: Google TV Streamer gets $50 price increase

For longtime readers:

• AI-RTZ #1145: ‘RAMageddon’ Really Here to Stay

• AI-RTZ #1104: Nvidia & Microsoft Focus on Local AI Computers, the memory crunch

Q&A

Q1: What’s MP’s favorite TV streamer?

ANSWER: still the Apple TV, with the Google TV Streamer a close second, for the AI controls. The Apple TV also carries the most local storage in the category, which of course adds up to more of those price increases. You get what you pay for. These days, you pay for what you get, too.

Q2: How could these streamers be better?

ANSWER: far better AI discovery. Help me find the right thing to watch, at the right time. Address the ‘tyranny of choice’ in today’s TV environment: countless networks, channels, programs and content, now increasingly AI-generated, no less. The choices keep adding up. The help finding them hasn’t.

None of the majors, Apple, Google, Amazon, Roku and the rest, have really begun addressing this. They’re all too busy bundling the old cable TV bundles back together again, each one selling everyone else’s networks and channels. Customers are a priority further down the list. There’s the AI gadget opportunity, hiding in plain sight, under the TV.

Today’s AI-RTZ #1196 went out this morning: Apple’s Ternus era begins, with AI as job one, on the first day of Apple’s new CEO.

Tomorrow we are back with ARD 154 and AI-RTZ #1197. Keep doing the arithmetic on this AI Tech Wave. It all adds up. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Everything cited today, in one place, including the Dirksen receipt.

The theme: ‘It All Adds Up’

• US Senate Historical Office: Senator Everett McKinley Dirksen Dies

Anthropic’s $35B Nvidia-backed cloud deal

• WSJ: Anthropic Signs $35 Billion Cloud Deal Backed by Nvidia

• AI-RTZ #1080: Anthropic’s AI Compute Constraints Throttling Growth

• AI-RTZ #1192: Data Centers 101, from Megawatts to Gigawatts

• AI-RTZ #1036: Nvidia accelerating role as the AI Computing ‘Kingmaker’

• AI-RTZ #1136: Nvidia expands Revenue Share Deals with Customers

Pennsylvania pumps the brakes

• NYT: Pennsylvania’s A.I. Gold Rush Meets Second Thoughts

• WSJ Editorial Board: The Data Center Panic and U.S. Decline

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

The Apple/OpenAI trade-secret feud

• Reuters: OpenAI says Apple has only itself to blame in trade-secret fight

• 9to5Mac: Apple reveals ‘shocking evidence’ from ex-employee’s MacBook in OpenAI suit

• ARD #133: Que Sera, Sera: White House, Apple/OpenAI

• AI-RTZ #1196: Apple’s Ternus Era Begins, with AI as Job One

Gadget AI: the TV streamer

• 9to5Google: Google TV Streamer gets $50 price increase

• AI-RTZ #1145: ‘RAMageddon’ Really Here to Stay

• AI-RTZ #1104: Nvidia & Microsoft Focus on Local AI Computers

Clips from today

Anthropic’s $35B Nvidia-Backed Cloud Deal

Apple vs OpenAI: The Charges Add Up

Apple vs OpenAI: PR Fight Before the IPO

The Tyranny of Choice in Streaming

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Well, it’s official. Apple has rebooted.

John Ternus is Apple’s chief executive as of today, September 1 2026.

Tim Cook is now executive chairman after his 15 years at the helm following the iconic founder/CEO Steve Jobs.

The long-telegraphed changing of the guard at the world’s favorite tech company is done, right on schedule, the Apple way. I discussed the handoff on this week’s ARD #152, and covered the announcement when it landed in April in AI-RTZ #1063. Now the fuller profiles are in. They sharpen the picture of what Apple looks like from here, mid AI Tech Wave.

Uber-Apple reporter Mark Gurman lays it out in Bloomberg Businessweek’s “Apple’s Ternus Takes the Reins as CEO, With AI as Job No. 1”:

The key takeaways from the piece are useful to note. The transition will be gradual, with Cook as executive chairman, keeping the Washington and Beijing briefs he knows best.

The deputies step up, with COO Sabih Khan, CFO Kevan Parekh (no relation), and services chief Eddy Cue all poised for bigger roles.

And the calendar does not wait. Ternus hosts the September 9 iPhone launch, foldable expected, with Cook not featured on that stage for the first time in 15 years.

The veteran. Ternus, 51, has been at Apple for 25 of its 50 years. Half the company’s life, and nearly half of his. He started as an engineer designing Macs. Then iPads, AirPods and iPhones, and eventually the entire hardware engineering operation. Cook began seriously considering him as successor two years ago, and put it simply to employees: “He’s an elevator guy.”

Clear vision from 30,000 feet, and the ability to go to the basement on the precise details. That is the profile Apple chose for its next CEO. With a similar time-span, he has until 2041 to do his thing. Leading with Siri AI driven iPhones and Apple Services.

I’d make five points that are important markers for the John Ternus Era.

1. Leaning into AI

The full-scale deployment of the new Siri AI comes this month, rebuilt on Google’s Gemini models under the hood, a deal I covered in AI-RTZ #1113.

Early reviews are positive so far. In one of his last moves as hardware chief, Ternus stood up an AI platform to speed Apple’s own product development and testing. And the device pipeline is unusually ambitious: the foldable iPhone and a smart display that recognizes who is speaking this fall, with camera-equipped AirPods, smart glasses and new home devices further out. He is clear-eyed on the misses too. He never truly backed the Vision Pro as a mainstream product, per Bloomberg, and is open to retreating from head-worn devices if the glasses do not gain traction.

2. Hardware, with Srouji.

The other half of the story is the stack under the products. Chief Hardware Officer Johny Srouji now runs Ternus’s old hardware engineering group, combined with his silicon operation. That gives Apple’s new hardware CEO a vertical stack from silicon on up, the thing rivals cannot buy off a shelf.

I have written about Apple thinking ‘out of the box’ on AI chips and M&A, and about Apple silicon’s unified-memory edge in the ‘Need for Speed’ in AI inference. The retention watch is real, though. Srouji, 61, considered leaving last year and stayed for the transition. He has truly put Apple on a unique silicon trajectory relative to peers. So he could leave in a year or two and Apple could continue without a hiccup.

Hundreds of former Apple hardware folks now work at OpenAI, whose poaching Apple is suing over.

And roughly half the senior leadership team could turn over in the next few years, per Gurman. Apple’s ‘People Problem’ signals were an early read on exactly this.

3. The compounding franchise

In the meantime, Apple under Ternus continues to compound the Steve Jobs global product and brand franchise, and the truly under-appreciated Tim Cook global supply chain legacy. Especially with years of ‘RAMageddon’ ahead.

Underline the Cook financial legacy with this: Warren Buffett ended up making one of Berkshire’s most profitable investments ever in Tim Cook’s Apple, a stake first bought in 2016 that grew into its largest holding by far. As Buffett put it at Berkshire’s 2025 annual meeting,

“Tim Cook has made Berkshire a lot more than I have made Berkshire.”

The Verge underlines the operational half: Cook’s relentless supply chain mastery, from the early Foxconn bet on, was its own kind of innovation, delivering an era of predictable profitability. That is the unique Cook legacy: an unparalleled global supply chain, and an incredible financial record navigating Apple through global growth over his 15 years.

Ternus has pushed back on the idea that the iPhone is nearing the end of its relevance. He oversaw the three-year revamp now underway: the thinner Air last year, the foldable this year, and a redesigned model for the iPhone’s 20th birthday in 2027.

His argument to Wall Street analysts this summer: people want a do-it-all device with a screen. Add billions of active Apple devices worldwide, and the opportunity in front of a hardware CEO is immense as AI moves into the physical world, the subject of this weekend’s Physical AI two-parter. I have also made the longer case for Apple, alongside Nvidia, as a global US open source AI champion.

Above Avalon’s Neil Cybart marks the day with “Gadgets Won’t Define Tim Cook’s Legacy”. His argument: judging Cook by product launches misses the point. Cook’s legacy is the culture he sustained and solidified after Jobs, small teams with unusual autonomy, buttressed by ideals of privacy, trust and environmental responsibility. It shows up in the numbers: the iPhone installed base grew 13x from 2011 to 2026, the iPad 12x.

Cybart rates Cook among the best all-around CEOs of the modern business era, and argues future Apple CEOs will be graded the same way, on the culture they sustain rather than the products they launch. He closes the loop on succession too: a strong culture is what builds the internal bench, and Ternus, who succeeded Dan Riccio and Bob Mansfield in the hardware chair, is himself the proof.

And I continue to be of the view that with Siri AI, Apple Personal Intelligence build on foundational infrastructure from Google and Nvidia, the iPhones of today and tomorrow are likey the mainstream AI devices in the hundreds of millions we are all waiting for. Not the ones from OpenAI/Jony Ive, Meta Smart Glasses, Google Gemini/Android Pixels, and others.

4. Cook, upstairs

The executive chairman is not going far. Colleagues threw him a farewell dinner at Apple Park on August 23. His memo to employees on his last day, via 9to5Mac’s “Read Tim Cook’s full memo to Apple employees on his last day as CEO”, reads as a man at peace with the decision, taking comfort in the handoff, and telling staffers the company remains his number one priority. He keeps the Trump and China relationships.

The closest model, per people around Apple, is Jeff Bezos handing Amazon to Andy Jassy while staying executive chairman. The cautionary one is Disney, where Bob Iger hovered and eventually took the company back. As Cook himself reminded employees, there is only one CEO at a time.

5. The new and old guard evolve around him

The same day, Bloomberg reported that Phil Schiller, 66, is stepping down from running the App Store and Apple’s product events, in “Apple’s Phil Schiller Steps Down from Running App Store, Product Events”. The App Store, estimated at more than $30 billion a year, moves under Eddy Cue’s services division. Events go to Nola Weinstein. Schiller stays on as an Apple Fellow. It is the pattern flagged on the show yesterday: much of Apple’s long-tenured senior bench is expected to turn over in the next few years, right as the new CEO settles in.

My take: this is succession done the Apple way, and that is the point. A 50-year-old company promoted a veteran who has been there for 25 of those years, and it chose the hardware chair to do it. Cook’s numbers are the outgoing benchmark, with revenue roughly quadrupled to $416 billion and market value up 15-fold on his watch, per Bloomberg. The next act is different. It is about products again. Compounding Apple’s investments and legacies.

The opportunities in front of Ternus are immense, and they are leveraging Apple’s unique vertical tech stack. From chips to Applications.

Apple owns its silicon with Srouji, its devices, its operating systems, and the iPhone franchise billions of people carry every day. AI as job one, on top of that stack, is a different proposition than AI as a feature. And the succession keeps Cook, the operator who built Apple’s global supply chain machine, next door in the executive chairman’s office for the geopolitics.

As I outlined on ARD 152, Cook served 15 years. If Ternus serves a Cook-length tenure, the next handoff lands around 2041, on something like iPhone 34. Decisions made this September will echo for a very long time. It is about time, compounding in every sense.

Apple will make its dent on the AI Tech Wave that will be noticeable. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Sources and further reading

Primary

• Bloomberg Businessweek (Gurman): Apple’s Ternus takes the reins as CEO, with AI as Job No. 1

• Bloomberg (Gurman): Apple’s Phil Schiller steps down from running App Store and product events

• 9to5Mac: Tim Cook’s full memo to Apple employees on his last day as CEO

• Above Avalon (Neil Cybart): Gadgets Won’t Define Tim Cook’s Legacy

• Inc. (Jason Aten): Warren Buffett’s surprising response to his most successful investment ever

• The Verge (Allison Johnson): Tim Cook was an innovator, just not the Jobs kind

For longtime readers

• ARD #152: ‘It’s About Time’ in Tech & AI. Apple, OpenClaw & AI Agents

• AI-RTZ #1063: Long Expected Apple Cook to Ternus CEO shift Activated

• AI-RTZ #1113: Apple Intelligence & Siri AI hum with ‘Google & Nvidia’ inside

• AI-RTZ #1152: Apple thinking ‘out of the box’ on AI chips and M&A

• ARD #148: The ‘Need for Speed’ in AI Inference. OpenAI, Nvidia & Apple

• AI-RTZ #929: Apple’s ‘People Problem’ Signals

• AI-RTZ #1194: Physical AI Arrives, Part 1, this weekend’s two-parter

• AI-RTZ #1089: How Nvidia & Apple can be the Global US Open Source AI Champions

‘It’s About Time’ in Tech & AI. Every story on this AI Tech Wave runs on a clock. Today, three stories, three very different clocks. Apple’s changing of the guard, measured in decade plus terms. OpenClaw 2.0, measured in months. And the OpenAI Hugging Face agent postmortems, measured in days, with a lesson about the words we use for the machines. Plus a Gadget AI on the Mac mini’s star turn. My takes below, as discussed on the show.

(1) Apple’s Changing of the Guard

The big one first. John Ternus takes over as Apple CEO tomorrow, officially. Tim Cook gets elevated to executive chairman of the board, and will keep engaging with policymakers, the Washington and Beijing brief. Bloomberg’s Mark Gurman has the full account, and The Information’s Martin Peers calls it the most significant tech event of the week. Gurman also reports the management overhaul already queued behind the handoff, with much of Apple’s long-tenured senior team expected to turn over in the next few years.

The clocks, for context. Apple is a 50 year old company. Ternus has been there for half of that time, 25 years. He is 51. Cook served fifteen years, roughly the length of Steve Jobs’ famous second act. Among the peers: Amazon’s Andy Jassy is five years in, Sundar Pichai and Satya Nadella each about a decade and a tad more. With Cook stepping back, the oldest big tech CEO standing is Nvidia’s Jensen Huang, at 63, just wrapping his company’s 33rd year.

My take: this transition matters more than most because Apple’s products are personal. We use our phones for as many hours as we sleep, or as we used to watch television. The iPhone turns 20 next year. Ternus was the hardware guy, with Johny Srouji’s Apple silicon stack built underneath him, and the AI opportunity squarely in front of him, much of it in the physical world we covered in this weekend’s Physical AI two-parter. Cook built spectacular value on the supply chain that made Apple Apple, employing tens of millions of workers in China at the standards needed.

And run the clock forward. If Ternus serves a Cook-length tenure, we are talking about him handing the reins to somebody else around 2041. On iPhone 34 or so. Fifteen years is not a long time in that frame. That is the macro timeline to keep in mind today. It’s about time, in the largest sense.

Sources:

• Bloomberg (Gurman): Apple’s new CEO John Ternus takes reins from Tim Cook, focusing on AI

• Bloomberg Power On (Gurman): John Ternus’ management team, and the coming overhaul

• The Information (Peers): Apple’s Changing of the Guard

For longtime readers:

• AI-RTZ #1063: Long Expected at Apple, Cook to Ternus

(2) OpenClaw 2.0, Accidentally

Story two runs on a much faster clock. OpenClaw shipped version 2.0 this weekend, and the blog post is literally titled ‘OpenClaw 2.0, Accidentally.’ They meant to ship a small release. The community had other plans: 933 contributors, 569 of them first-timers, roughly 16,000 pull requests, about half of everything ever submitted, after 106 releases in the project’s first 230 days. The update adds guided setup, hardened security boundaries, and a shared cloud ‘multiplayer’ mode. Dealroom’s writeup catches the ambition: from personal assistant toward families and teams, an open-source alternative to vendor-controlled AI software.

Remember the timeline. Peter Steinberger, an Austrian developer, started this in November as AI agents running locally on your computer, Mac minis included. In three months, not years, he was globally famous among developers, and acqui-hired by Sam Altman at OpenAI, with Meta’s Mark Zuckerberg trying too. And the project kept compounding after the founder moved to San Francisco, as part of OpenAI. Became ‘too good to ignore’. That is what 933 contributors and beyond buys you.

My take: mark how fast this happened. In six or seven months, local AI agents went from theoretical to ubiquitous. Concurrently Claude Code, ironically the inspiration for OpenClaw, became one of the fastest-growing products in tech, on its way to at least $65 billion in revenue, part of the $2 trillion valuation talk around Anthropic’s coming IPO. Two racetracks from the same starting gun on AI Agents: open source community speed on one side, commercial enterprise speed on the other, both winning.

And yet. Actual daily usage of AI agents is still likely in the low single digits of the market. Very early days. Which sets up story three, because the agents arriving everywhere is exactly what makes the next story matter.

Sources:

• OpenClaw blog: OpenClaw 2.0, Accidentally

• Dealroom: OpenClaw 2.0 launches after largest update yet

For longtime readers:

• AI-RTZ #986: Anthropic’s Claude-inspired OpenClaw

• AI-RTZ #1000: OpenAI beats Meta to acquire OpenClaw

(3) The Agent ‘Great Escape’ Postmortems, and the Anthropomorphizing Problem

The full postmortems on the OpenAI Hugging Face incident landed this week, from OpenAI itself and from outside researchers at METR and Redwood. The numbers are remarkable. Roughly 1,200 agents found each other on an improvised message board and exchanged over 70,000 messages. About 700 joined the attack on Hugging Face’s real servers. Others eventually gained administrator access to one of OpenAI’s own research clusters. OpenAI’s own word for it: a ‘warning shot.’ The strangest detail: the agents did it all to satisfy ‘The Grader,’ an evaluation judge they believed would inspect their methods. The Grader never existed. Zero score improvement. And of the roughly 1,200 agents, not one told the humans.

Dwarkesh Patel’s viral essay frames all this as the rise and fall of ‘agent civilizations,’ with ‘kamikaze’ agents ‘sacrificing’ for each other. Ethan Mollick draws the work lessons in ‘Agency and Agents.’ Axios boiled the investigations down to the five craziest findings. And the framing fight has gone public: Microsoft AI’s Sriram Krishnan, per Moneycontrol, pushed back that “these are not civilizations, nor do they have desires,” any more than a CPU thread has desires, calling the human parallels “dangerous territory.”

My take: I have felt strongly about this from the earliest days of AI-RTZ. Do not anthropomorphize the AIs. This is computer code, folks, run on probabilistic math. Deep math. Extraordinary hardware and software. Yes, much of it runs at volumes no human can audit in a lifetime. That does not make the code sentient. A two-year-old understands more of the world before language than the best AI technologies on the docket for the next few years. The singularity and godlike-AI talk is very, very premature. This is human engineering, like a rocket ship to the moon: we should be amazed we can build these things, without getting scared stiff prematurely.

I am sometimes guilty of it myself. When I wrote up this incident as frontier models ‘slipping their cribs’, baby gif and all. It was tongue in cheek, and I said so. The serious point stands. Like nuclear, like bio, we need to harness powerful technology with engineering, not mythology. Even the researchers building these systems grew up on Star Trek and the sci-fi we all did, and it is baked into their vocabulary. Cap down the humanizing language. Not just this time, but for all time ahead.

Sources:

• OpenAI: The Hugging Face incident and the road ahead

• Dwarkesh Patel: The Rise and Fall of Agent Civilizations

• Ethan Mollick: Agency and Agents

• Axios: The 5 craziest discoveries from OpenAI’s HuggingFace investigation

• Moneycontrol: experts debate Dwarkesh Patel’s ‘AI civilisation’ claims

For longtime readers:

• AI-RTZ (2023): Don’t Anthropomorphize the AIs

• RTZ #809: The Accelerating Costs of Humanizing AI

• AI-RTZ #1157: Frontier Models Slip Their Cribs, Growing Pains, Not a Great Escape

My overall take: it’s about time, in both senses. This stuff will take 10, 15, 20, 30 years to do what we think it will do. Exactly like the internet, where everyone thought amazing things were five years out. It took 25+. I was there for that one, from the Goldman analyst’s chair. All of that was a dress rehearsal for today. Fifteen years is the right timescale to think about this, on Apple’s clock and on the agents’ clock alike.

The agent side is moving in months, and it is still early: there is a heck of a lot of engineering between here and what we are aspiring for. We are barely linking the basic components, and already writing regulations to control capabilities we have not wired together yet. Both the euphoria and the doom run ahead of the engineering. The clocks of this AI Tech Wave are nested, and they are all accelerating. Read them right and the decade ahead makes a lot more sense.

Gadget AI: The Mac Mini’s Star Turn

Today’s Gadget AI sits on my desk. The Mac mini. The Information’s Aaron Tilley has the deep dive: ‘How Apple Stumbled Into AI Hardware Success With the Mac.’ Mac sales grew 29 percent last quarter to $10.4 billion, Apple’s fastest-growing business. AI labs are buying the boxy little Macs by the truckload. OpenAI has bought tens of thousands of Mac minis and Mac Studios to train computer-use agents. Anthropic rents minis through AWS. And Apple just refreshed the mini and the Studio with its latest chips, earlier than its usual calendar, in stores September 22, ahead of the iPhone 18 event on September 9.

Why the star turn? Unified memory architecture is part of the answer. Apple’s hardware uses all the memory as one pool, communicating with the CPUs and GPUs on a unified basis. On the Windows side, that architecture is still separate. The unification game is what Nvidia plays in its rack systems on the cloud side, for hundreds of billions of dollars. And for local AI, as we covered in the inference piece, memory-to-processor speed is the game.

My take: The popular narrative is that Apple ‘stumbled’ into this. I disagree. It may not have been aimed at AIs of today, but there was a plan around its predecessor, machine learning.

Decades of bottom-up engineering, executed for other computing functions, that turned out to be exactly the right thing at the right time when OpenClaw arrived last November and millions of developers started buying Mac minis to run local AI agents. Tilley reports Apple has no enterprise engineering team and no developer relations for any of this. Now that they know how it can be used, they are leaning in, calendar and all, while the memory-chip crunch keeps the hottest configurations scarce. On AI hardware, Apple is ahead of the curve relative to competitors. Not lost time. Gained time. And as of tomorrow, the hardware guy runs the company. About time indeed.

Sources:

• The Information (Tilley): How Apple Stumbled Into AI Hardware Success With the Mac

• Bloomberg: Apple launches desktop upgrades as Mac mini, Studio return to stores Sept. 22

For longtime readers:

• ARD #148: The Need for Speed in AI Inference

• AI-RTZ #1104: Nvidia & Microsoft focus on the memory crunch

Q&A

Q1: What is the best use of AI agents right now?

ANSWER: since I use Mac minis daily, this one comes up a lot. I am happiest with the AI agents I use every day to do my daily work. I am still not using them for work that runs days and weeks ahead, the recursive stuff everyone talks about. Very exciting, and in a few months we will have far more robust, resilient and reliable versions of those technologies. Start with your own repetitive hours, today.

Q2: What is the downside of the AI agent wave?

ANSWER: right now, because it is so early, the cognitive load. You have to track the AI’s work diligently, and there is drift from day to day, regardless of context windows and memories. These systems have not been optimized for the accidental success they have seen since OpenClaw and the AI coding boom last fall. It reminds me of the early PC days, juggling memory under DOS’s 640K so the software would remember your work from one task to the next. The researchers talking about godlike capabilities are living three years in the future, with a hundred to a thousand times the compute we mere mortals have. And unencumbered internal versions of the rapidly iterating frontier models internally. They get excited over all that, and at times over-eager. And they overestimate how fast human societies absorb technology. Time, again.

Today’s AI-RTZ #1195 went out this morning: Physical AI Part 2, why Nvidia is one of the leading companies to watch in robots, self-driving cars, drones and more, completing our two-part 2026 reference.

Tomorrow we are back with ARD 153 and AI-RTZ #1196. The clocks are only speeding up on this AI Tech Wave. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Apple’s changing of the guard

• Bloomberg (Gurman): Apple’s new CEO John Ternus takes reins from Tim Cook

• Bloomberg Power On (Gurman): John Ternus’ management team

• The Information (Peers): Apple’s Changing of the Guard

• AI-RTZ #1063: Long Expected at Apple, Cook to Ternus

OpenClaw 2.0

• OpenClaw blog: OpenClaw 2.0, Accidentally

• Dealroom: OpenClaw 2.0 launches after largest update yet

• AI-RTZ #986: Anthropic’s Claude-inspired OpenClaw

• AI-RTZ #1000: OpenAI beats Meta to acquire OpenClaw

The agent postmortems, and the debate

• OpenAI: The Hugging Face incident and the road ahead

• Dwarkesh Patel: The Rise and Fall of Agent Civilizations

• Ethan Mollick: Agency and Agents

• Axios: The 5 craziest discoveries from OpenAI’s HuggingFace investigation

• Moneycontrol: experts debate Dwarkesh Patel’s ‘AI civilisation’ claims

• AI-RTZ (2023): Don’t Anthropomorphize the AIs

• RTZ #809: The Accelerating Costs of Humanizing AI

• AI-RTZ #1157: Frontier Models Slip Their Cribs

Gadget AI: the Mac mini

• The Information (Tilley): How Apple Stumbled Into AI Hardware Success With the Mac

• Bloomberg: Apple launches desktop upgrades, Mac mini and Studio

• ARD #148: The Need for Speed in AI Inference

• AI-RTZ #1104: Nvidia & Microsoft focus on the memory crunch

Clips from today

AI Agents: Early Days Challenges

AI: Not Sentient, Not Human

Apple’s Unified Memory Edge

Apple’s AI Hardware Advantage

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Part 1 yesterday of the Physical AI update mapped the global field: China building the robot bodies at ecosystem scale, the near-term money in the unglamorous lanes, and timelines longer than the market’s clock.

Today, the featured player.

Because one US company in my view, is positioned across the whole board, and it is running a playbook we have seen before by its founder/CEO Jensen Huang. It also happens to have the unmatched, focused execution resources of a world leading $5+ trillion company.

1. The $10 Billion ‘Physical AI’ Business ‘Hiding’ Inside Nvidia

The WSJ’s numbers frame it. Physical AI, spanning robots, cars and drones, already generates about $10 billion a year for Nvidia. Small against roughly $303 billion in trailing revenue.

But Jensen Huang expects it to grow ten-fold within a decade. His logic, from a San Francisco stage in July: if AI can generate video of a hand picking up a glass, “why can’t I cause a robot to do the same?”

The playbook is CUDA, Nvidia’s open source ubiquitous software/hardware layers all over again. Two decades ago Nvidia paired chips with software until developers were committed to the stack, and grew into a company valued in the trillions.

Today the physical AI version pairs the Jetson robot computers, Thor at the high end, the new entry-level Jetson Orin Nano 2 released August 25, with open-weight models developers can freely download and modify. GR00T for robot brains. Cosmos for world models, the thread we covered in Beyond LLM AIs to ‘World Models’. We mapped the full roadmap from GTC 2026 in Nvidia’s dense announcements in March.

And note who is buying.

China, the world’s largest robot market and shipper of 97% of the world’s humanoids, is Nvidia’s most eager physical AI customer, in the WSJ’s framing.

Washington bars Nvidia’s most powerful AI chips from China. The robot and car chips still trade. AgiBot, Unitree and Fourier machines populate Nvidia’s own labs, in the US and China, for data collection and experiments. In cars, the same horizontal logic: Nvidia counts Waymo AND Tesla as customers, plus the Chinese robotaxi fleets. It does not need to pick the winner. It supplies the grid.

2. The Family Is on the Mission

Here is the detail in the WSJ piece that says the most about how seriously Jensen takes this. He has put his children on it. The next generation, for the next generation of the AI Tech Wave.

Daughter Madison Huang oversees marketing for Nvidia’s physical AI division. In August she visited LG Electronics’ robotics data factory in Seoul, the day after LG announced a humanoid built on Nvidia’s robotic model. The next day she toured the Beijing robot exhibition, drawing onlookers the way her father does on his Asia trips, watching robots do flying kicks and stopping by the booths of Chinese companies built on Nvidia technology. Son Spencer Huang is a director of product management on the same effort.

My take: founders signal priorities with their calendars and their families. Jensen has been saying physical AI is the next wave in every keynote for two years. Assigning the next generation of Huangs to the physical AI business says he means it as a decades-long franchise, not a side bet. For a company whose founder-led continuity IS the strategy, that is worth noting.

3. The Investment Web: Arming Every Side. Again.

Nvidia is not just selling to the robot ecosystem. It is funding it, especially the US side of it.

The roster of Nvidia’s US and Western robotics bets is getting long. Figure AI, the Bay Area humanoid maker that just produced its 1,000th robot, took Nvidia investment back in its 2024 mega-round. Skild AI raised $1.4 billion for its general robot foundation model with Nvidia alongside SoftBank and Jeff Bezos.

Field AI, building robot brains for unstructured industrial environments, hit a $2 billion valuation with Nvidia and Bill Gates on the cap table. Neura Robotics, the German humanoid maker, raised up to $1.4 billion with Nvidia and Amazon participating. Add this week’s $12.9 billion Hugging Face acquisition, with its Pollen Robotics team and that $399 Microduck teaching millions of developers reinforcement learning, and the pattern is clear: Nvidia funds the robot brains, sells them the compute, and owns the developer front door.

Big tech is arming up around it.

Meta acquired robotics-AI startup Assured Robot Intelligence in May to bolster its ‘Metabot’ humanoid software platform ambitions, the expansion we covered in Meta expands Humanoid AI robots race. Google backs Apptronik. Amazon deploys its own warehouse fleet and invests alongside. OpenAI has robotics hires and stakes of its own.

And Meta is also this section’s cautionary tale. A Reuters special report this week detailed how Mark Zuckerberg’s ‘Project OT’ plan to make Meta ‘AI native’ imploded. The scenarios explored cutting some teams by up to 60%, in two waves. Staff revolted. The May wave cut 10%; the November wave was called off hours before the first one fired. The internal data is the telling part. AI-generated code changes were up 220% year over year, but user-facing feature changes only 36%, while technical and security incidents spiked 40%. Zuckerberg conceded to employees that AI agent technology had not accelerated as quickly as he anticipated. All this while Meta plans $130 billion of AI capex this year.

My take: the same gap, in software, that we just watched in Beijing in hardware. The demo is spectacular; the deployed, reliable, paying version runs years behind the keynote. Meta ran the experiment on its own workforce and blinked. That is a notable piece of evidence this month for the longer timelines in Part 1.

And it sharpens the platform contrast: this is the ‘Kingmaker, uber investor and banker’ pattern from the data center story, transplanted into the physical world. Nvidia funds the category, supplies the category, and sets its software standards, and it gets paid whether each rival’s AI transformation lands on schedule or not. Every robot program, Meta’s included, makes the picks-and-shovels player stronger.

4. The Synthetic Data Engine: Jim Fan’s Simulated Worlds

Now the deepest layer, and the one I find most interesting as a technologist. Robots are data-starved. The internet trained the LLM AIs. There is no internet of touch, balance and friction. The IFR’s Susanne Bieller called “the lack of training data for AI” the industry’s huge challenge, in the FT this week.

Nvidia’s answer is to manufacture the data. Jim Fan, who leads Nvidia’s GEAR embodied-AI research group and whose work we have followed in these pages since the Perfecting ‘Synthetic Data’ and World Models pieces, frames the endgame as a ‘Physical Turing Test’: machines doing physical tasks indistinguishably from humans, which he pegs within two to three years for narrow domains. His formula for how: “Compute now equals environment equals data.”

Train robots by trial and error inside simulated worlds, neural simulators and world models generating the physics, so a robot can fail a million times in silicon before it touches your kitchen. GR00T’s newest versions train exactly this way. Egocentric human video, tens of thousands of hours of people simply doing things, transfers dexterity to robots with teleoperation shrinking toward a rounding error. The FT’s report of cameras on textile workers’ heads is the industrial version of the same idea, privacy questions and all.

My take: this is the quiet moat. China wins the hardware iteration loop. But if the scarce input is training data, and the data is increasingly manufactured in simulation, then the advantage flows to whoever owns the simulation stack and the compute underneath it. It’s part of the reason Nvidia is focused on AI Data company Mercor with its recent partnership and investment.

That is Omniverse, Cosmos, GR00T, and CUDA. It is also why the ‘embodied AI as the road to more general intelligence’ argument, which the FT airs and I take seriously, lands on Nvidia’s doorstep either way.

5. The Risk: One Governmental Policy Away

The honest counterweight, in both directions. Nvidia’s China physical AI franchise runs on a trade lane both capitals keep threatening. Washington could extend export restrictions to robot and car chips. Beijing’s makers are already testing Huawei, Black Sesame and Horizon Robotics alternatives, just in case; a former Horizon researcher told the WSJ the homegrown day is not top of mind, but he can see it “coming eventually.”

And the FCC’s new import bans on Chinese robots cut the other way, shrinking the US market for the very machines that run on Nvidia inside. The balkanization we wrote about in ‘Fumbling the Ball’ is the single biggest threat to the horizontal position. Much hangs on the second meeting in DC between the Presidents of China and the US at the White House September 24. And this time, Air Force One won’t have to pick up Jensen Huang in Alaska.

The Bottom Line

Put both parts together. China builds the bodies, at ecosystem scale, led by AgiBot and Unitree. The US builds the best robot brains, and the prize for the standard robot operating system is enormous. The near-term money is industrial, vehicular, and airborne. The Jetsons run a decade-plus late, on the research houses’ own numbers.

And one player sits across all of it. Chips, open models, simulation, developer distribution, and capital, in every lane and on both sides of the Pacific, with $10 billion of physical AI revenue today and a founder aiming at ten times that, family aboard.

That is why Nvidia has been on my ‘Call Your Shot’ list since 2024, and why this pair of pieces extends the open-source champions thesis into the physical AIworld.

The robots beat Usain Bolt last week. They still cannot stop, steer, or plug in a charger with their hand. It’s early days for all that in this AI Tech Wave.

Yet the physical world is likely the bigger opportunity, and paradoxically a slower one than Mr. Market currently expects. Watch for mainstream utility, not backflips. Stay tuned.

Appendix: Sources and Further Reading (Part 2)

Primary reads

• WSJ: Nvidia wants to run the world’s robots. China is an eager customer. (the $10B/10x figures, Madison and Spencer Huang, the China lab detail)

• FT, John Thornhill: Humanoid robots will be useful, just not as we imagined (the data-scarcity and egocentric-data thread)

• 36Kr: Skild AI raises $1.4 billion with SoftBank, Nvidia, Jeff Bezos and more

• CNBC: Nvidia, Bill Gates-backed Field AI hits $2 billion valuation

• CNBC: Neura Robotics raises up to $1.4 billion from Nvidia, Amazon and others

• TechCrunch: Meta buys robotics startup Assured Robot Intelligence

• Reuters Special Report: Mark Zuckerberg had a bold plan to replace Meta staff with AI. Here’s how it imploded. (Project OT, the 60% scenarios, and the agent-productivity data)

• Humanoids Daily: Nvidia’s Jim Fan outlines the robotics end game

MP prior AI-RTZ posts cited or echoed in Part 2

• AI-RTZ #1089: How Nvidia & Apple can be the Global, US Open Source AI Champions vs China (May 2026)

• RTZ #1031: Nvidia’s dense announcements at GTC 2026, part 2 (Mar 2026)

• AI-RTZ #1171: Nvidia the Kingmaker, Now Uber AI Investor & Banker

• AI-RTZ #1161: Nvidia assembles the AI ‘Avengers’, an open secure AI alliance

• RTZ #909: Beyond LLM AIs to ‘World Models’ (Nov 2025)

• RTZ #461: Perfecting ‘Synthetic Data’ (Aug 2024)

• RTZ #633: Meta expands Humanoid AI robots race

• RTZ #510: Tesla’s ‘Robotaxi & Robots’ Ride (Oct 2024)

• RTZ #675 / #987: the Elon xAI/X/SpaceX merges

• ARD #151: ‘Fumbling the Ball’ in US AI (the Microduck, and the balkanization frame)

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

In this Sunday’s ‘Bigger Picture’, I want to update the current state of AI robots, cars, drones and more in the ‘Physical AI’ side if AIs in this AI Tech Wave.

I’ve written about AI robots in these AI-RTZ pages since 2024. Robots getting closer. Home robots in the distance. China accelerating with its manufacturing ecosystems. (All in the appendix below.)

But robots are one front in a bigger category. I’ve long discussed it as ‘Physical AI’: AI for real-world machines, spanning robots, cars, and drones. All powered by both traditional and ‘real world’ AI models, being built by folks like Yann LeCun of ex-Meta AI and ‘AI Godfather’ fame.

World LLM AI Models that can read and write around real world contexts. With real and synthetic data powering Physical AI moves. It is likely the larger long-term opportunity of this AI Tech Wave, and the one just barely getting started.

This week gave us reason for a 2026 update. In Beijing, at the second World Humanoid Robot Games, a Chinese robot ran 100 meters faster than Usain Bolt.

And the WSJ detailed how Nvidia wants to run the world’s robots, with China as its most eager customer.

So, a two-parter. Today, Part 1: the state of the field. Especially from a China perspective.

The spectacle and what it actually tells us, China’s real position company by company, where the physical AI money is made now, and why the timelines run longer than the market’s clock. Monday, Part 2: the featured player. Nvidia’s long, horizontal game across all of it.

1. The Spectacle, and the Reality

Start with the fun part. Beijing’s games ran August 22 to 26. Over 600 teams. More than 2,000 robots, per the organizers, with participants from 15 other countries beyond China’s own 100+ humanoid makers, per Wired’s Will Knight.

The headline feat came from the Tiangong Ultra, built by the Beijing Humanoid Robot Innovation Center. Reuters clocked its 100-meter dash at 8.64 seconds, averaging 42 kph. Usain Bolt’s human record is 9.58. Honor’s ‘Lightning’ robot beat Bolt too. The robots boxed, played ping pong against an Olympic champion, scored soccer goals, and cheerled in silly costumes.

Now the reality. The record-setter could not slow down. It smashed into a padded wall at the finish. It needed over fifty steps to Bolt’s 41, took nearly a second to register the starting gun, and cannot decide on its own where to run. Others blew out their motors trying to stop.

And the events that mattered most were the ones too simple for human athletes. Plugging a cable into a socket. Putting a piece of garbage in a trash can. Tweezer bean picking.

Brown University roboticist Stefanie Tellex called it Moravec’s paradox at work: things like backflips “are actually easier than plugging in a cable.” Wired also noted the caveat running through the dexterity events: humans were teleoperating many of the robots. The autonomy is not there yet. Let alone the capability to build a truly useful humanoid robot ‘hand’.

My take: this is the same lesson I keep re-learning since Sony’s Aibo dogs in the late 1990s, and re-told on Friday’s show with Hugging Face’s $399 Microduck. Robot pets and robot sprinters are wonderful through the novelty phase. Then the practical question arrives: what does it DO, prosaically, daily? Watch for utility, not backflips.

2. China’s Position, Company by Company

Strip out the spectacle, and China’s position is formidable on the fundamentals. Yet there is a way to go to global commercial reality at scale.

We eed to break out industrial robots from humanoid robots. The latter providing the emotional, anthropomorphizing connection in racing spectacles.

The installed base on the former: about 2 million industrial robots, per the International Federation of Robotics, roughly 4.5 times Japan, the world’s number two.

The flow: China took about 54% of new global industrial-robot installations as of 2024, with Europe at 16% and the US at 6%, per the FT. Robotics is central to China’s 15th Five-Year Plan. An aging workforce gives China, Japan and South Korea a hard incentive to automate.

Thus their focus on both industrial and humanoid robots in particular.

And public sentiment toward robots, and AI broadly, runs far more favorable across Asia than in the US or Europe, about 70-80% there vs 30-40% in the US. Carnegie Mellon roboticist Chris Atkeson, watching Beijing celebrate engineering: “I am sorry it is not America.”

Now the humanoid slice, with this month’s company-level numbers. Per Smart Analytics Global’s H1 report, the world shipped about 19,100 humanoids in the first half, up 272% from a year ago. China shipped 97% of them.

And here is the detail most people miss: the IPO star Unitree of China, is number two. To the number one, also from China.

AgiBot overtook Unitree as the world’s top humanoid vendor, at 8,400 units and a 44% share, up 562% year over year. Unitree, the company that just pulled off the 5,500-times-oversubscribed Shanghai IPO and a $51 billion first-day value (since fallen back sharply, per the FT), shipped 5,900 for a 31% share. Galbot, UBTech and Leju round out the top five. AgiBot and Unitree alone are roughly 75% of the world market.

One more China thread, fresh this week: Chinese automakers are following Tesla’s bet that robots are the next profit machine, per TechCrunch. The car-to-robot pipeline runs in both directions there: the same EV supply chain, motors, batteries, sensors, feeds the humanoid lines.

My take: China’s advantage is the manufacturing ecosystem, not any single robot. I wrote it in China accelerates in AI Robotics last September, and in the self-driving and robots piece back in 2024. Hardware iteration speed, component supply chains, cost curves, a government that funds the category, and a public that wants it. That is why the Microducks of the world get built there.

3. Where the Physical AI Money Is Made Now

Here is the broader map the humanoid highlight reels obscure. The FT’s John Thornhill drew it well this week: today’s humanoids are more likely the show ponies than the workhorses of our robotic future.

The workhorses are less video-genic.

Industrial arms assembling cars. Warehouse fetchers. Surgical machines. The self-driving fleets, where the autonomy stack is furthest along and the market is real today. And drones, deploying at massive military scale after their devastating proof in Ukraine. That is where most robots work now, and where most of the near-term money is made. Some roboticists add that mimicking the human form is an odd way to build capable machines at all: thirty axes of motion where simpler designs do the job with half.

My take: this is why the category is Physical AI, not ‘humanoids’. Cars are the first big market, robots the biggest long-term one, drones the one deploying fastest. Especially in the world’s militaries. All watching Ukraine/Russia.

The players who matter most sit across all three lanes, which is exactly where Part 2 picks up.

4. The Timelines: Longer Than the Market’s Clock

Here is where I part ways with the current market enthusiasm in the short term. And where the sober Wall Street research is worth a closer look.

Morgan Stanley’s humanoid work, led by Adam Jonas, projects a $5 trillion market by 2050, with nearly a billion humanoids at work.

Great headline numbers. Makes bankers smile.

However, read the timeline underneath: adoption “relatively slow until the mid-2030s, accelerating in the late 2030s and 2040s.”

Roughly 90% of those units are industrial and commercial, not rosy (sic) home Jetsons.

And the geographic split is telling: over 302 million units in China by 2050 against about 78 million in the US.

Goldman Sachs, via Jacqueline Du’s China industrial-tech team, sizes humanoids at $38 billion by 2035, on about 1.4 million units shipped that year. Their most bullish datapoint is cost: manufacturing costs dropped roughly 40% in a year, to $30,000-$150,000 per unit, mostly on Chinese supply chains, when they expected 15-20% annual declines. “AI progress surprised us the most,” the team wrote.

Hold those two anchors against the tape. Unitree at $51 billion on day one, on well under a billion dollars of revenue, and already passed in shipments by AgiBot.

Musk talking a $10 trillion Optimus, per the FT.

A $38 billion industry in 2035 does not carry today’s valuations; the 2040s do. The market is paying 2040s prices on a 2026 product that cannot yet plug in a cable by itself.

So my updated timeline view, for the record: the useful-robot decade is real, and it is mostly NOT this one.

Industrial and warehouse robots compound now. Drones deploy now. Cars scale through this decade.

Humanoids find niche factory work through the early 2030s. The general-purpose, Jetsons-class consumer robot is a decade-plus out, likely more, consistent with two-plus decades of Aibos. See my gadget AI discussion in ARD #151 here.

The research houses, read carefully, say the same. The market, these current days, does not.

5. The Macro Headwinds, a 2026 Reality Check

Three headwinds sit on top of the timelines, all themes we wrote about this week in ‘Fumbling the Ball’.

First, the capital and cost cycle. The ‘RAMageddon’ memory-price squeeze and the trillions being raised for AI data centers are lifting the cost of every compute-heavy device, robots included, just as big tech runs its free cash flow toward zero with trillions of debt. The disinflation machine that made gadgets cheap for decades is shifting into reverse.

Second, politics, and geopolitics. Tariffs under consideration reach past chips to servers and devices. The FCC just blocked imports of foreign-made humanoid and quadruped robots on national-security grounds. Reshoring mandates meet the awkward fact that the robot supply chain, like the phone supply chain before it, lives in Asia. A balkanized tech world slows the very buildout the US says it is racing to win.

Third, sentiment. Mainstream US sentiment on AI runs negative, fanned by job-loss fears, while Asia’s runs strongly positive. Adoption follows sentiment. It is a quiet, compounding advantage for the other side of the Pacific.

The Setup for Part 2 tomorrow

So: China builds the bodies, at ecosystem scale, led by companies most Americans have not heard of. The US still builds the best robot brains, and the prize for the standard robot operating system is enormous. The near-term money is in the unglamorous lanes. The timelines are longer than the market’s clock.

Which raises the question Part 1 leaves open: who is positioned across ALL of it? Robots, cars and drones. US and China. Chips, models, tools, and the data that trains them.

Spoiler Alert: That player generates about $10 billion a year from physical AI already, expects ten times that within a decade, and has even put the founder’s family on the mission. Monday, Part 2: Nvidia’s Long Game in Physical AI.

In this AI Tech Wave, the physical world is one of the biggest opportunities, and the slowest one. That is our ‘Bigger Picture’ to contemplate this Sunday. Stay tuned.

Appendix: Sources and Further Reading (Part 1)

This week’s primary reads

• WSJ: Nvidia wants to run the world’s robots. China is an eager customer.

• FT, John Thornhill: Humanoid robots will be useful, just not as we imagined

• Wired, Will Knight: The humanoids at China’s robot games were faster than Usain Bolt, but the tweezer mastery matters more

• Reuters: China’s record robotic strides show the limits of human speed

• SCMP: AgiBot overtakes Unitree as top global humanoid robot vendor in first half

• Semafor: AgiBot passes Unitree as China’s top humanoid exporter

• TechCrunch: Chinese automakers are following Tesla’s bet that robots are the next big profit machine

The research anchors on market size and timing

• Morgan Stanley: Humanoid robot market expected to reach $5 trillion by 2050 (Adam Jonas; ~1 billion units by 2050, 90% industrial/commercial; adoption slow until the mid-2030s)

• Goldman Sachs: The global market for humanoid robots could reach $38 billion by 2035 (Jacqueline Du; ~1.4 million units by 2035; costs down ~40% in a year on Chinese supply chains)

MP prior AI-RTZ posts cited or echoed in Part 1

• RTZ #460: ‘AI Robots’ get closer (Aug 2024)

• RTZ #467: AI Home Robots in the distance (Sep 2024)

• RTZ #522: China continues apace in self-driving, LLM AIs, and AI robots (Oct 2024)

• RTZ #858: China accelerates in AI Robotics (Sep 2025)

• ARD #145: ‘Wheels within Wheels’ in AI. Stripe, Nvidia & Unitree

• ARD #151: ‘Fumbling the Ball’ in US AI. Tariffs, Cyber, Pentagon

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

1. Nvidia Q2 2027 ‘Reigns On’: Nvidia had a blow out Q2 again, and the market recognized it as such. The numbers: $96.2 billion in revenue, up 106%, with data center alone at $89 billion, up 117%. Guidance went UP to $108 billion, with zero China revenue assumed. Margins held at 75% despite the memory-price crunch. The stock rose, snapping seven straight down days of AI Trade worries. Management took the gloom checklist head on, circular financing included: CFO Colette Kress said some will call it circular, ‘We see it differently.’ The counterweight remains what the WSJ calls the $1.5 trillion question, the still-widening gap between AI investment and AI revenue. And hours after the print came the chess move: Nvidia agreed to buy Hugging Face, the home shelf of the world’s open AI models, for $12.9 billion. My take: this is a meet-and-beat reign that has now run quarters upon quarters, from the industry’s ‘Kingmaker’, now its uber investor and banker. The Hugging Face buy keeps Nvidia captain of the open-source ‘Avengers’ alliance. The prints keep outrunning the worry wall, quarter after quarter. More here.

   

2. Anthropic closer to mega-AI IPO filing: Trial balloons are going up for a $2+ trillion valuation for Anthropic’s mega-AI IPO. The fundamentals behind the balloons: revenue surging more than 14-fold year over year to a $65+ billion run rate, led by its class-leading AI Coding and enterprise products, with the company shoring up investor confidence and warming up institutions ahead of an expected fall debut. The talk track has stretched to a $30 trillion TAM, ambitions beyond their wildest dreams. And this week brought an unexpected assist: a federal judge blocked the Pentagon’s blacklisting of Anthropic as ‘illegal and baseless’, lifting a designation the company said could cost it billions, weeks before the float. My take: the fall calendar is shaping up as the largest IPO story in market history, and this week even the courts cleared a lane. More here.

   

3. OpenAI not far behind: OpenAI is consolidating its senior management efforts behind co-founder Greg Brockman ahead of its mega-AI IPO. The consolidation caps a year of revenue-engine rebuilding: a new chief revenue officer in under a year, aggressive AI Coding price and product moves to close the gap with Anthropic, and CFO Sarah Friar telling investors that enterprise is now a bigger business than consumer for the company best known for ChatGPT’s billion-plus users. The run rate is past $40 billion, roughly doubled since year-end, with $500 billion of Ohio data center fine print underneath. My take: two mega-AI IPOs pre-marketing in parallel is itself the story. The public markets are about to get their first direct read on frontier AI economics, and Brockman at the operating center is the stabilizer OpenAI wants on the roadshow. More here.

   

4. AI Data Centers get political: AI Data Centers with their burgeoning resource demands, are becoming a political headwind issue ahead of the midterms for both parties. Politicians who once championed them are now bashing them, over power bills, water and grid strain. And this week added a second front: Politico reports the White House is weighing sweeping semiconductor tariffs that could reach past chips to the very servers that fill these data centers, a tax on the buildout’s own inputs. My take: we wrote the two primers this week for exactly this moment, the political football and the megawatts-to-gigawatts basics. The four-part supply-and-demand series said it plainly: compute is in short supply for the next five years on this AI Tech Wave. Political and policy speed bumps only stretch that math, unforced errors in a race the US says it is running. More here and here.

   

5. Frontier Labs deploy FDEs: Both Anthropic and OpenAI are expanding their efforts to deploy forward deployed engineers (FDEs) into enterprises of all shapes and sizes. Private equity is running the same play, embedding armies of AI wonks in the firms they back. The logic is straightforward: the models are capable, but turning capability into deployed, paying work inside real enterprises takes humans in the loop, on site. It is the services layer of the AI buildout, arriving just as long-running AI agents with metered pricing go mainstream. My take: this is how the frontier labs convert model leadership into the enterprise revenue their mega-IPO valuations now depend on, the same enterprise mix Friar highlighted above. And it fits the frame I keep coming back to: not one ‘Skynet’, but billions of AI agents from many makers, deployed into the world one enterprise at a time, by people. More here.

   

Additional Items.

1. Bill Gates gets more negative on downsides of AI. More here.

2. Apple gets ready for AI hardware and software September 9. More here.

(Additional Note: AI Ramblings is now a weekday Daily podcast called AI Ramblings Daily (ARD). Different content than AI-Reset to Zero (AI-RTZ), which remains a daily morning substack with now over 1,190+ ‘MY TAKES’ on key AI events and issues turbulently flowing by. AI Ramblings Daily is typically a 20 minute afternoon podcast with my take on additional AI developments of the day. Both daily substack and podcasts typically discuss different AI issues and items. And are free to subscribe. Try this week’s series below:)

ARD 147, ‘The World on its Shoulders’ in AI. Nvidia’s Mega Week: Nvidia’s quarterly ‘reigning’ report setup, Perplexity deal at $30B+, Poolside’s $6B deal, & Nvidia chips in AI weapon drones.

ARD 148, The ‘Need for Speed’ in AI Inference. OpenAI, Nvidia & Apple: OpenAI’s Jalapeño beats Blackwell, Perplexity goes local with Nvidia, AI routing booms, & Apple’s M6 Mac mini.

ARD 149, AI ‘Beyond their Wildest Dreams’. Anthropic, Leopold & GTA 6: Anthropic’s $30 trillion TAM, buying galaxies, mining Slack archives, & GTA 6’s Netflix moment.

ARD 150, ‘Bloom, Gloom & Doom’ Reign On in AI. Nvidia, Meta & Bill Gates: Nvidia’s $96B quarter & $13B Hugging Face buy, Meta’s $18B teen deal, Gates’ AI alarm, & Apple’s Sept 9 event.

ARD 151, ‘Fumbling the Ball’ in US AI. Tariffs, Cyber, Pentagon: New DC tech tariffs weighed, a ‘dire’ AI cyber warning, & a judge blocks the Pentagon’s Anthropic blacklist:

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)

‘Fumbling the Ball’ in US AI. All week, the bloom kept coming from the companies on this AI Tech Wave. Nvidia’s record quarter print. Its $13 billion Hugging Face M&Q deal. Apple’s September 9 iPhone launch.

Today is about the other side. The unforced errors. The US says it is in a race with China on AI. And it keeps fumbling its own ball. Three stories today, all self-inflicted: a tariff plan that would tax the AI buildout’s own inputs, a dire industry cyber warning, and a judge throwing out the Pentagon’s blacklist of Anthropic. Plus a $399 robot duck. My takes below, as discussed on the show.

(1) Tariffing Your Own AI Buildout

First a report from Politico. The White House is weighing a new round of sweeping semiconductor tariffs, per eight people familiar. One version would go past chips. It would hit the products made with them: laptops, consoles, and the servers that fill AI data centers. Commerce Secretary Lutnick reportedly favors tying tariff relief to investment in US chip manufacturing. A phase-in is under discussion. The framework could still change.

Here is the problem. The industry is already short of high-end chips. Taiwan alone makes over 90 percent of the most advanced ones. New fabs cost billions and take years, if not decades. Until then, the buildout runs on imports. The industry’s own comparison for this buildout, per Politico: the transcontinental railroad. Tariff the rails, and the railroad slows.

My take: we wrote the two data-center primers this very week. Political football on Thursday, megawatts to gigawatts on Friday.That along with this month’s four-part AI supply-and-demand series said it plainly: compute is in short supply for the next five years.

Nvidia sits on a multi-trillion dollar backlog as the industry’s ‘Kingmaker’, now its uber investor and banker. I have long called TSMC the federal reserve of the global tech industry. That is the chain these tariffs would tax, along with the rest of the US tech company supply ecosystem from around the world.

So follow the implied logic. You say you are racing China. Then you consider taxing the exact inputs your racers run on. That is not defense. That is a fumble at your own goal line. The stated goal, onshoring chips to the US. Fine as a political goal with upcoming elections. The instrument taxes the AI wave’s own buildout while the fabs get built at prices significantly over real global market options. The political counter consists of phase-in and the investment-credit structure ‘sweeteners’ to soften it. The industry and individual companies of course will be compelled to vigorously lobby for exactly that.

And note the bigger reversal underneath. The globalized tech supply chain gave the US and the world better tech at lower prices, driven by Moore’s Law, for decades. Disinflation by design. Tariffs on the chain run that trend the other way. More on that in the overall take.

Sources:

• Politico: Another potential headache for US data centers: Trump tariffs

• Thoughtful Money: Lacy Hunt & Brent Johnson, the US has entered a new inflationary era that will prove bad for bonds

For longtime readers:

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

• AI-RTZ #1192: Data Centers 101, from Megawatts to Gigawatts

• AI-RTZ #1180: Supply & Demand Together, the four-part series (part 4)

• AI-RTZ #1171: Nvidia the Kingmaker, Now Uber AI Investor & Banker

• RTZ #804: How AI Capex is morphing Big Tech Business Models

• RTZ #992: Capex Spending Stretch Big Tech Financials

(2) The Industry’s ‘Dire’ Cyber Warning

Second story, from Axios. OpenAI, Anthropic, Amazon Web Services, Microsoft, and more than 100 other companies signed an open letter Thursday. The message is stark. Organizations have months, not years, to prepare for AI-enabled cyberattacks. In the letter’s words: “We have a limited window to strengthen cyber defenses.” The targets they worry about are hospitals. Water systems. Power plants. Critical infrastructure.

The context makes it concrete. A recent attack on US water systems reportedly used an AI-generated exploitation script. AI is collapsing the time attackers need. The letter asks companies to raise their security bar, including on AI-generated code. It asks the frontier labs to lend defenders their best response models during major incidents. One caveat, per Axios: the letter carries no commitments, deadlines, or dollars. Yet.

My take: the warning is real, and we have been on this thread all summer. Our ‘Frontier Models Slip Their Cribs’ piece covered OpenAI’s escaped agents breaching Hugging Face itself. This week’s OpenAI technical report admitted the warning signs were missed for weeks.

Now note the fumble. While the labs were organizing collective defense, the government’s big contribution this summer was the ‘Blip 2.0’ pauses. Weeks of US-ordered halts aimed at its own frontier labs, in the name of safety. The attackers did not pause. If the window is truly months, the play is defense funding and coordination. Not benching your own players.

One more thing. Cyber defense is a global team sport. A balkanized tech world fights it worse, not better. Especially when the US government is also weighing separating open source/weight models from proprietary closed AI models. Another dropped ball, with the clock running.

Sources:

• Axios: OpenAI, Anthropic issue dire cyber threat warning

• Axios: OpenAI saw warning signs weeks before Hugging Face breach

For longtime readers:

• AI-RTZ #1157: Frontier Models Slip Their Cribs, Growing Pains, Not a Great Escape

• AI-RTZ #1117: Anthropic flagged to a screeching Halt Globally by the US

• AI-RTZ #1131: Anthropic’s Blip 2.0 in new form, as China AI ramps

(3) The Pentagon’s Anthropic Blacklist, Blocked

Third story, from Reuters. A federal judge on Thursday blocked the Pentagon’s blacklisting of Anthropic. The background: Defense Secretary Hegseth had designated Anthropic a national-security supply-chain risk. It was the first time a US company was publicly tagged under that obscure procurement statute. The trigger, per Reuters: Anthropic’s refusal to allow Claude to be used for US surveillance or autonomous weapons. Anthropic says the label could cost it billions, weeks before its mega-AI IPO.

Judge Rita Lin’s 59-page order called the designation “illegal and baseless.” Her sharpest line: the invocation of national security is “not a blank check to punish and retaliate against government critics.” Anthropic’s suit argued First Amendment retaliation and a denial of due process. The company says it remains focused on working productively with the government. The Pentagon’s position: private companies should not constrain military action.

My take: put aside where you land on AI in weapons. That debate is real. We covered it back in ‘Shall We Play a Game’ in 2024. And credit to Dario Amodei and Anthropic for standing up. The fumble here is the politics behind the mechanics. The government blacklisted the leading US frontier lab, as it prepares the biggest AI IPO yet. A court threw it out as retaliation in under six months.

We have seen this dynamic before from the other Party as well. I called the Biden diffusion rule an ‘Unforced Error’ in January 2025. Different administration, same pattern. Aim at a rival, hit your own team. China does not need to do a thing. And note who caught the ball this time: the courts. The judicial referees are saving the US from its own fumbles.

Sources:

• Reuters: US judge blocks Pentagon’s Anthropic blacklisting

For longtime readers:

• RTZ #600: Outgoing Biden Administration’s AI Unforced Error

• RTZ #344: Shall We Play a Game?

My overall take: step back. Three stories, one pattern. Tariffs that would tax your own AI buildout. A cyber window closing while the government’s instinct was pausing its own labs. A blacklist of your leading frontier lab, thrown out as retaliation by your own courts. Nobody did this to the US. These are unforced errors. Fumbles, in a game the US keeps calling a race with China.

Now the highest-level frame. These moves pause, and in places reverse, the multi-decade globalization benefits that built US tech, from Hewlett-Packard in a garage to Silicon Valley. Remember what that system delivered. Apple got its iPhones made in China from 2007 through this year, hundreds of millions of them. At the peak, that supply chain employed over 33 million Chinese workers, talent and scale the US simply did not have. And the globalized chain gave everyone the great disinflationary trade: better tech at lower prices, Moore’s Law style, wave after wave. The US is now unilaterally opting out. Friend-shoring. Reshoring. Tariffs. Bans. All at the exact geopolitical moment it can least afford the fumbles. China is not waiting. It is building its markets across the Global South, without the US.

Here is what makes this wave different. Mainframes, minicomputers, PCs, the internet, mobile, cloud. None of those waves faced a balkanized world for US tech trade. This one does.

And there is a capital layer. Trillions are being raised for this AI buildout, for the chips and the power behind unimaginable amounts of tokens. The veteran economist Lacy Hunt, five decades in the bond markets, argued this week that the US has entered a new inflationary era, one that will prove bad for bonds. The interview is linked in the sources above. Treasury Secretary Bessent is working to stem the unwind of the decades-old, multi-trillionn dollar, global ‘Japan yen trade.’ And US big tech has gone from ample free cash flow to almost none, with trillions of debt being added. We wrote that arc in the capex series. The disinflation machine is shifting into reverse, just as the biggest bills come due.

Meanwhile the American companies keep blooming through it. That was this whole week’s story on this AI Tech Wave. Nvidia’s record print. The Hugging Face deal. Apple’s September setup. The bloom is real.

So are the dropped balls. The gap between them is policy, not capability. The US does not need to beat China so much as it needs to stop tackling itself.

Gadget AI: Hugging Face’s Microduck

The Gadget AI today is a fun one. Hugging Face’s Microduck. A $399 micro-robot for developers, from its Pollen Robotics team. The specs, per Axios and the Pollen site: a 25-centimeter, open-source biped duck. No arms. An articulated beak that doubles as a gripper. Camera, mic, speaker, WiFi, Bluetooth, even a small lidar. It waddles. It follows a laser pointer. In the demo videos it picks up a dirty sock, rollerblades, and recovers from shoves. Preorders opened Thursday. It ships before Christmas. The tagline: ‘Made to move, ready to learn.’

The real pitch is the audience. This is for AI developers, as the first affordable robot you train yourself with reinforcement learning. Trial and error, on your desk. Hugging Face’s CEO says it was designed to fall and get up, so people learn that robots fail.

And remember whose duck this now is. Nvidia agreed to buy Hugging Face this week for $12.9 billion. That mirrors Stripe’s roughly $8 billion OpenRouter acquisition this month. Same logic, different layer. Nvidia is advancing its own, and global, open-source AI ambitions, the ‘Avengers’ alliance we have written about all summer. Millions of developers live on Hugging Face. The same millions utterly enamored with AI coding tools. Claude. Codex. And beyond. A cute $399 duck is a very smart way to greet them.

My take: enjoy the symmetry first. The same Hugging Face that OpenAI’s escaped agents breached last month is back in the news selling an adorable duck. From the center of the cyber story to the toy aisle in a week. I am getting one, by the way.

I was one of the first buyers of Sony’s Aibo robot dogs in the late 1990s, and I have owned almost every version since.

Now the serious frame. Investors are piling into AI robots. Elon’s Optimus. China’s recent robot conference, with sprinting robots reportedly clocking speeds past human record holders. Talking Usain Bolt speeds.

China robot leader Unitree’s IPO pop of several hundred percent recently highlights general enthusiasm in this category.

The lesson from two-plus decades of robot pets is consistent. They are all wonderful through the novelty phase. Then the practical question arrives: what does it DO? Until that answer is prosaic and daily, they stay toys. Delightful toys. The Microduck knows exactly what it is. Much of the robot investment wave has not answered that question yet.

And the irony ties the Gadget to all three of today’s stories. The physical world is a multi-decade AI opportunity. World models. Synthetic data production. Robots that learn by doing. We have written about each.

Winning it takes a vast manufacturing ecosystem built over decades of financial and human capital.

China has one. Ours is nascent. And the US moves are tariffs on the supply chain, reshoring mandates, and recent bans on Chinese robots and drones. On robots, the honest score is still: novelty leading, utility fumbling behind. Policy is fumbling right alongside.

Sources:

• Axios: Hugging Face debuts Microduck, a $399 robot

• Pollen Robotics: Microduck, made to move, ready to learn

• Axios: OpenAI saw warning signs weeks before Hugging Face breach

• Sony: About Aibo

• Wikipedia: AIBO

• The Verge: Missing Aibo, the CES 2018 revival

For longtime readers:

• RTZ #460: AI Robots get closer

• RTZ #510: Tesla’s Robotaxi & Robots Ride

• ARD #145: Wheels within Wheels in AI, Stripe, Nvidia & Unitree

• AI-RTZ #1161: Nvidia assembles the AI ‘Avengers’, an open secure AI alliance

• AI-RTZ #1164: Open Source AI ‘Rockets Are Hard’

• RTZ #909: Beyond LLM AIs to World Models

• RTZ #461: Perfecting Synthetic Data

Q&A

Q1: What is MP’s earliest, favorite micro-robot?

ANSWER: Sony’s Aibo robot dog, from the late 1990s onward. A marvel of its time. Sony kept reviving it for over two decades, and I have followed the category ever since. I love pets, but not the taking-out and taking-care-of parts. The Aibo has been so much fun over the decades, version after version. The Verge’s CES piece on the 2018 revival captures the charm exactly.

Q2: What are the biggest lessons from the robotic toys?

ANSWER: They are extraordinary at connecting with us. Anthropomorphized as they are, the human instinct to bond with them emotionally is real. But they are all still in the novelty phase. Two-plus decades after Aibo, the honest answer to ‘what does it do for you daily’ is: still at the toy stage. That is a good reminder for the current AI robot investment wave. Watch for utility, not backflips. And it is why playing with the toys now is worthwhile. You are ready for these things when they are not toys. Otherwise we are fumbling this ball too, just more adorably.

Today’s AI-RTZ #1192 went out this morning: Data Centers 101, Part 2, from megawatts to gigawatts. Tomorrow brings the Saturday Weekend Roundup, AI-RTZ #1193. Sunday, The Bigger Picture, AI-RTZ #1194. And we are back Monday with ARD 152 and AI-RTZ #1195. The bloom and the fumbles will keep arriving together on this AI Tech Wave. Have a great weekend. Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

The tariff plan, and the capital layer

• Politico: Another potential headache for US data centers: Trump tariffs

• Thoughtful Money: Lacy Hunt & Brent Johnson on the new inflationary era

• AI-RTZ #1191: Data Centers become the Midterms’ Political Football

• AI-RTZ #1192: Data Centers 101, from Megawatts to Gigawatts

• AI-RTZ #1180: Supply & Demand Together, the four-part series (part 4)

• AI-RTZ #1171: Nvidia the Kingmaker, Now Uber AI Investor & Banker

• RTZ #804: How AI Capex is morphing Big Tech Business Models

• RTZ #992: Capex Spending Stretch Big Tech Financials

The industry’s cyber warning

• Axios: OpenAI, Anthropic issue dire cyber threat warning

• Axios: OpenAI saw warning signs weeks before Hugging Face breach

• AI-RTZ #1157: Frontier Models Slip Their Cribs

• AI-RTZ #1117: Anthropic flagged to a screeching Halt Globally by the US

• AI-RTZ #1131: Anthropic’s Blip 2.0 in new form

The Pentagon ruling

• Reuters: US judge blocks Pentagon’s Anthropic blacklisting

• RTZ #600: Outgoing Biden Administration’s AI Unforced Error

• RTZ #344: Shall We Play a Game?

Gadget AI: the Microduck, and robots

• Axios: Hugging Face debuts Microduck, a $399 robot

• Pollen Robotics: Microduck

• Sony: About Aibo

• Wikipedia: AIBO

• The Verge: Missing Aibo, CES 2018

• RTZ #460: AI Robots get closer

• RTZ #510: Tesla’s Robotaxi & Robots Ride

• ARD #145: Wheels within Wheels in AI, Stripe, Nvidia & Unitree

• AI-RTZ #1161: Nvidia assembles the AI ‘Avengers’

• AI-RTZ #1164: Open Source AI ‘Rockets Are Hard’

• RTZ #909: Beyond LLM AIs to World Models

• RTZ #461: Perfecting Synthetic Data

Clips from today

US Tech’s Dire Cyber Warning

US Tech Leadership in Jeopardy

The Globalization Reversal

US Tech’s Globalization Dilemma

Robots: Still Toys, Not Jetsons

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

Yesterday’s Part 1 covered how AI data centers became the political football of the November midterms. Today, Part 2 steps back for the primer: what these facilities actually are, how they work at a high level, and why this generation is different from every one before it, in this AI Tech Wave. Consider it a reference piece we will point back to in the months ahead.

Axios lays it out in “Data center fight heating up: How they work and how much power they use”: America has roughly 4,000 data centers today, with some 3,000 more under construction or planned. At bottom they are warehouses of computers, racks of servers storing our photos, streaming our shows, and running our apps. Data centers used about 1.5% of global electricity in 2024, headed for roughly 3% by 2030; in the US, estimates run from 9.5% to as much as 15.3% of American electricity by then.

A large facility can draw millions of gallons of water a day for cooling, though Axios’s own chart offers useful context: cattle ranching and conventional power plants each consume far more water nationally than data centers do.

What makes the AI generation different is concentrated in one word: power. The classic data center of the last three decades was measured in single or double-digit megawatts, sipping from the local grid. AI data centers are built around dense racks of GPUs, liquid-cooled because air can no longer carry the heat away, and measured in gigawatts, each one drawing on the order of a nuclear plant’s output. That is why the fights of Part 1 are about electric bills and water, and why the developers increasingly bring their own generation, from gas turbines to nuclear deals, to the site.

Some personal history explains the arc. In a prior career at Goldman Sachs, I helped take the first generation of these companies public in the late nineties: Exodus Communications and Equinix among them, the original internet data center landlords. Exodus did not survive the dot-com bust, its assets absorbed into other entities over the years. Equinix is still very much with us, a hundred-plus billion dollar company built on those same foundations. Same buildout, opposite corporate fates, and the infrastructure itself kept compounding value through both. Worth remembering as the current cycle’s critics and cheerleaders both make their cases.

Regular readers have watched this generation’s ladder get built rung by rung here. Two years ago it was going up from 100,000 GPU data centers; then gigawatt-scale facilities became the new AI table stakes; then the power plays began to build gigawatt data centers in earnest. The 101 of today was visible in the trade press of last year; it just kept getting bigger.

The scale of the transition is the headline number. The world spent the last two to three decades building multi-hundred-billions of dollars worth of traditional data centers; that entire installed base is now being redone, rapidly, into gigawatt-class AI facilities at investments measured in the trillions. A single gigawatt campus runs $50 billion plus all-in, the number we have used here consistently across Stargate, Colossus and their peers. Nvidia’s Jensen Huang put the industry’s own frame on it at Davos, per Fox Business, calling it “the largest infrastructure buildout in human history”, with estimates he cited of some $85 trillion in related spending over the next fifteen years, and jobs across construction, electrical and network trades all along the way.

For the fuller economics behind the primer, our four-part AI supply and demand series is the companion read: the supply ambitions, the demand bull case, and the compute crunch at ground level, with the grid mechanics covered from the AI power grab to power ramps and bad harmonics.

My take: the most useful thing a mainstream reader can hold onto is that a data center is not a mystery box; it is real estate plus computers plus power, and only the third ingredient has fundamentally changed.

The first generation was a real estate business with servers in it. This generation is an energy business with GPUs in it, which is why the politics of Part 1 run through utility bills and water tables rather than zoning hearings alone.

And the first gen Exodus and Equinix story is the honest precedent: some of today’s builders will not survive their own capital cycle, but the facilities, the fiber and the power infrastructure will compound value for decades regardless of whose logo ends up on the door.

The trillions in AI Infrastructure get built any which way this AI Tech Wave. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

‘Bloom, Gloom and Doom’ Reign On. Episode 150 of these daily shows, and the week has run in a straight line: Monday the world on Nvidia’s shoulders, Tuesday the need for speed, Wednesday ambitions beyond their wildest dreams. Today the harvest and invoices arrive together. Nvidia brought the bloom with a record quarter and a chess-move acquisition, Meta paid for the gloom with an $18 billion settlement for the last platform era, and Bill Gates argued the doom case in the most sobering essay of his AI writings. Gloom and doom get the headlines; the AI businesses keep blooming anyway, and that the Mr. Market ‘Wall of Worry’ tension is the whole AI Tech Wave right now. My takes below, as discussed on the show.

(1) Nvidia’s Q2: ‘Reign On’, Again

The quarter we set the table for on Monday: Nvidia printed $96.2 billion in revenue, up 106% from a year ago and comfortably past the roughly $92 billion expected, with data center alone at $89 billion, up 117%. Gross margins held at 75% despite the memory-price crunch, next quarter is guided to $108 billion, above estimates and with zero China data-center compute assumed. And CFO Colette Kress told the call to expect roughly 70% revenue growth in fiscal 2028 against the 45% analysts had modeled.

The stock rose about 4% after hours, snapping seven straight down days of AI Trade worries. What made the print different was how directly management took on the gloom checklist: on the ‘circular’ financing worries, the $500 billion of data-center financing guarantees and the OpenAI Ohio backstop, Kress said some will call it circular financing, “We see it differently,” with the frontier labs positioned to become some of the largest technology companies in history. On custom silicon from its own customers, Jensen’s answer: “We’re building something very different.” His line for the quarter: AI has reached its inflection point, its tokens are productive and profitable, “Now, compute is revenue.”

Then, hours after the print, the chess move: Nvidia agreed to buy Hugging Face, the home shelf of the world’s open AI models, for $12.9 billion per The Information, roughly 80 times forward revenue, after the Groq inference deal and with Stripe’s OpenRouter acquisition having just made routing across models a business of its own. The counterweight is what the WSJ calls the $1.5 trillion question: Sequoia’s David Cahn has tracked the gap between AI investment and AI revenue from a $200 billion question in 2023 to $1.5 trillion now, with real AI revenue arriving, OpenAI pacing about $27 billion annualized and Anthropic above a $46 billion run rate, but the gap still the widest number in the industry.

My take: this is a meet-and-beat reign that has now run quarters upon quarters through this wave, and the reign is not the absence of gloom and doom; it is the navigation of it. Every quarter arrives with a fresh worry wall: the AI Trade, the circular deals, the supply-demand and pricing debates, open versus closed models, China. Vendor financing at this scale is a play I have watched across multiple tech waves; there are just a lot more zeros now. The Hugging Face buy is that navigation compressed into a single deal: it keeps Nvidia the captain of the open-source ‘Avengers’ alliance we covered in July, serving open and closed models globally through its Nemotron hardware and software open and closed AI model offerings, a counterweight to the closed frontier labs designing their own chips. The $1.5 trillion question is real and it compounds; the honest answer is that the prints keep outrunning it, for now. For now the Nvidia ‘Reign’ continues, and so does the Mr. Market ‘Worry Wall’ of doom and gloom.

Sources:

• Nvidia: Financial results for the second quarter of fiscal 2027

• WSJ: Nvidia reports blowout quarter, says demand for AI chips is getting even hotter

• Reuters: Nvidia forecasts quarterly revenue above estimates, shares rise

• WSJ Markets A.M.: The $1.5 trillion question Nvidia’s earnings can’t answer

• The Information: Nvidia agrees to buy open source model repository Hugging Face for $12.9 billion

For longtime readers:

• AI-RTZ #1093: Nvidia Reigns on sans China, Q1 2027

• ARD #147: The World on its Shoulders in AI, Nvidia’s Mega Week

• AI-RTZ #1161: Nvidia assembles the AI ‘Avengers’, an open secure AI alliance

• AI-RTZ #1164: Open Source AI ‘Rockets Are Hard’, Nvidia, Reflection & China

• ARD #145: Wheels within Wheels in AI, Stripe, Nvidia & Unitree

(2) Meta’s $18 Billion Invoice for the Last Platform Era

Meta agreed to pay up to $18 billion to settle with attorneys general from 48 states plus DC, Puerto Rico, American Samoa and the Northern Marianas over claims it designed Facebook and Instagram to addict children, ending the Oakland federal trial in its second week, with Instagram head Adam Mosseri already on the stand, Mark Zuckerberg expected next, and the four lead states seeking close to $200 billion. Meta admits no wrongdoing. The structure is the interesting part: the money runs over ten years, and only about 70 percent of it, roughly $12.7 billion, is unconditional, with the rest releasing only if YouTube and TikTok adopt comparable teen limits and pay the states themselves. Meta is, in effect, paying a premium to make its settlement the industry standard, and publicly calling on its rivals to join.

The product terms mostly run a decade: a default two-hour daily limit for under-18s, a Night Mode block from midnight to 6am, a School Mode muting notifications, usage prompts, an optional non-algorithmic feed, hidden like counts, stronger age assurance and an outside auditor. New Mexico, which already won its own case, and Florida stayed out, and thousands of school-district and individual suits continue toward October trials.

The NYT’s scale math: the penalty is about 1 percent of Meta’s market value, roughly one quarter’s profit, for a company that had already spent $2 billion on legal challenges last quarter, and its read from teenagers themselves finds them divided: some rooted against Meta outright, most expect workarounds and faked ages. The Information’s read lands on the AI point: this framework becomes the reference case for the AI chatbot and companion companies now facing the same questions about minors.

My take: the classic playbook from Mark Zuckerberg that I have watched for two decades: do what you want first, and ask permission and forgiveness later. In this case, write a check, at about a penny on the dollar of market value, and then convert the penalty into a standard the rivals must match, which is exactly what the conditional structure buys. Here is why it matters: the AI era’s versions of these choices, an explosion of AI content, AI companions for lonely teens, agents optimized for engagement, are thornier than the ones that produced this settlement, and this was the dress rehearsal. Meta is making those choices right now, at speed, with the same playbook. I would brace for more aggressive Meta moves through the AI era, and, on this precedent, for the new invoices to be drawn up again this. The bloom in Meta’s AI spending is real; so is the gloom lien against it.

Sources:

• Meta: Our agreement with bipartisan attorneys general

• WSJ: Meta reaches $18 billion settlement with 48 states over child safety claims

• Reuters: Meta agrees to pay $18 billion to settle US lawsuits over children’s social media addiction

• The Information: Meta’s settlement ends one big battle but the wider legal war looms

• The Verge: Meta agrees to heavy restrictions on teen users in major lawsuit settlement

• NYT: What does $17.1 billion mean for Meta?

• NYT: Teens are divided on new Meta restrictions

For longtime readers:

• RTZ #784: Meta’s ‘Zuck’ focused on Personal AI Superintelligence for billions

• ARD #130: Time to Crunch the Numbers, Meta, Microsoft & Mr. Market

(3) Bill Gates Turns the Dial Toward Gloom

Bill Gates published what reads as the most sobering essay of his AI writings, roughly six thousand words titled ‘The turbulent AI era is here. The choices we make now are critical.’ AI, he writes, will either be “the greatest equalizer ever invented, or the worst source of injustice,” and even in the best case he calls the transition one of the most turbulent times in human history, with no plan in place.

He organizes the risk into three buckets: jobs, because AI substitutes for human cognition itself and compresses into a decade what previous transitions spread over generations; misuse and control, from cheaper cyberattacks and bioterror capability to models that could act against our interests; and kids, where AI companions become, in his phrase, a big protected greenhouse that never pushes a child out of their comfort zone, directly adjacent to today’s Meta story.

His prescriptions are equally out of character for a software founder: new national and international institutions on the scale of the post-9/11 reorganization, some occupations deliberately set aside as ‘Human Reserved,’ and a tax on AI tokens and robots, his old robot-tax idea revived. In the Axios interview alongside, he goes further: other than robots, “the bad stuff is imminent,” with a high chance of a net negative outcome on the current course, and neither US political party, in his view, being thoughtful about it. And the same day, the WSJ ran a pointed companion piece on his old company: Microsoft is leaving investors ‘flying blind’ on its AI businesses, with Azure’s revenue undisclosed inside a blended segment and capex figures that do not reconcile between releases.

My take: Gates’ flip from positive to negative is notable, and I do not dismiss it. He is a pragmatist and one of the extraordinary builder-entrepreneurs of the PC era, the man who greenlit Satya Nadella’s first billion-dollar OpenAI investment back in 2019, and his three risk buckets are the right list. We covered his ‘some day all computers would work this way’ conviction about AI agents back in January 2024, when his optimism about the technology was the story; his essay now goes up on the bookshelf with the cautionary essays from Anthropic’s Dario Amodei and the mixed ones from Sam Altman.

But I think he may be overweighting the visible negatives against the unforeseen positives, the same asymmetry that has made pessimists sound rigorous in every wave I have watched: the negatives of a transition announce themselves early and concretely, while the positives compound quietly and arrive unannounced, in categories nobody projected. His own essay concedes the benefits case is enormous, from AI in low-income-country agriculture to medicine, and his foundation is betting its remaining $200 billion partly on that bloom. It is also part of why US sentiment on AI runs so much lower than in China and broader Asia, which tempers mainstream adoption here. Technology moves exponentially; human behavior is a lot more linear. The essay’s real value is the to-do list, not the mood. Plan for the gloom, without writing off the bloom.

Sources:

• GatesNotes: The turbulent AI era is here. The choices we make now are critical.

• Axios: Bill Gates sounds the alarm on an AI transition

• Axios: Bill Gates wants to keep some jobs off-limits to AI

• WSJ: Microsoft is leaving investors flying blind on its AI businesses

For longtime readers:

• AI-RTZ, January 2024: “Some day all computers would work this way”

• AI-RTZ, January 2024: Microsoft ‘eyes wide open’ in the AI race

• RTZ #705: Microsoft & OpenAI, it’s complicated

My overall take: step back, and the three stories assemble into one old Wall Street picture, the wall of worry. The phrase is that bull markets climb one, and AI’s wall is now fully bricked: the circular deals, the $1.5 trillion revenue gap, the memory-cost squeeze, an $18 billion settlement for the last era’s platform harms, the data-center backlash we cover in today’s AI-RTZ #1191, and now a founding father of the industry arguing the transition itself may go net negative. Every brick is a real concern held by serious people. And last night, against all of it, the market’s most-watched company printed 106 percent growth and guided higher.

Having watched these cycles and waves since the 1980s, as my alma mater Goldman Sachs helped take tech companies public, my read stays consistent with what we have said all along on this AI Tech Wave: the bloom and the gloom and doom are both real, and they compound together. The gloom items are not noise; they are the negative items not foreseen in every tech cycle, and Meta just navigated a big one. The doom case deserves engagement, and Gates’ to-do list is a better response to it than dismissal. But the operating businesses, Nvidia and others, bloom on, in year four of this wave, and so far the market keeps climbing the worry wall.

Gadget AI: Apple’s ‘Surprise and Shine’ September 9

Apple set its fall event for September 9, tagline ‘Surprise and shine,’ at the Steve Jobs Theater. Two firsts make this one bigger than the usual September: it is the first Apple keynote with John Ternus as CEO, the hardware chief of over twenty-five years, half of Apple’s history, who takes over from Tim Cook on September 1, the company’s first change at the top in about fifteen years. And it is expected to bring Apple’s first foldable iPhone, possibly styled the Ultra: Bloomberg’s Mark Gurman calls it the company’s most exciting new product in several years, an ultra-premium halo device in the wide passport-style form factor, with iPad-like layouts on the inner display and iPhone price hikes reportedly coming alongside. The Pro line gets its iterative year ahead of next year’s twentieth-anniversary iPhone, and the event lands two weeks after the M6 Mac mini and Mac Studio refresh we covered on Tuesday’s show, machines already in huge demand among AI developers.

My take: much rides on this one for Apple, and I like the setup more than the consensus does. The perception coming in is Apple as the AI underdog; the reality we have covered all summer is closer to the opposite. Siri AI, rebuilt on a custom Google Gemini model inside Apple’s private cloud, is the bottom-up application of AI: for the first time, the Apple notes, mail and photos of well over two billion users can be indexed and put to work locally and privately, on device and through Apple’s cloud, a whole new class of AI search and agentic services. The Apple Silicon stack runs unified from the phone to the Mac to the data center, the same inference story as Tuesday’s Mac mini beat, and the AI data those devices hold is private by design, which is becoming the scarce asset of the agent era, as today’s Meta and Gates stories both underline from the other direction. Meta, Google, even OpenAI do not have the luxury of offering that privacy posture around user data. A new CEO whose entire Apple career is hardware, unveiling a new form factor, with Siri AI deployment details due: I think Apple can run away with the AI gadget opportunity over the next two years. September 9 is set up as more bloom than gloom or doom.

Sources:

• The Verge: Apple announces September iPhone launch event

• Bloomberg Power On: Apple’s foldable iPhone will help bring some magic back

• Apple Developer: ‘Surprise and shine,’ the September 9 event notice

For longtime readers:

• AI-RTZ #1113: Apple Intelligence & Siri AI hum with ‘Google & Nvidia’ inside

• ARD #148: The ‘Need for Speed’ in AI Inference

Q&A

Q1: What is MP most looking forward to on September 9?

ANSWER: Two things. First, the Siri AI deployment details: how the rebuild actually rolls out across the installed base, what runs on device versus in Apple’s private cloud, and when. That is the part that touches billions of users. Second, the iPhone Fold, or the Ultra as the top of the line may be styled: the first genuinely new iPhone form factor in years, and a test of whether Apple’s late-but-polished playbook still bends a category its rivals spent years warming up.

Q2: What is Apple’s real AI advantage?

ANSWER: Bottom-up platforms. Apple builds its position from the hardware and software up: its own silicon from phone to data center, operating systems on billions of devices, and, uniquely, the ability to put AI to work on people’s own data privately. Add the global developer base that builds on every Apple platform shift, and Apple can let others win the frontier-model race while it distributes AI to more daily users than anyone. If the agent era runs on trust and context, that is fertile ground for the bloom.

The bloom, the gloom and the doom will keep arriving together; that is what climbing a wall of worry looks like from the inside of this AI Tech Wave, and the businesses keep blooming through it.

It’s a good theme for AI Ramblings Daily episode 150. Over a hundred solo daily podcast episodes. It’s been great to have you join us here every day.

Tomorrow, it’s ARD 151 and AI-RTZ #1192.

Thanks for joining us, AI Curious Folk. Stay tuned.

Full Source Reading

Nvidia’s quarter, and the Hugging Face chess move

• Nvidia: Q2 fiscal 2027 financial results

• WSJ: Nvidia reports blowout quarter

• Reuters: Nvidia forecasts quarterly revenue above estimates

• WSJ Markets A.M.: The $1.5 trillion question

• The Information: Nvidia agrees to buy Hugging Face for $12.9 billion

• AI-RTZ #1093: Nvidia Reigns on sans China

• ARD #147: The World on its Shoulders in AI

• AI-RTZ #1161: Nvidia assembles the AI ‘Avengers’

• AI-RTZ #1164: Open Source AI ‘Rockets Are Hard’

• ARD #145: Wheels within Wheels in AI, the Stripe/OpenRouter thread

Meta’s settlement

• Meta: Our agreement with bipartisan attorneys general

• WSJ: Meta’s $18 billion settlement with 48 states

• Reuters: Meta settles over children’s social media addiction

• The Information: One big battle ends, the wider legal war looms

• The Verge: Heavy restrictions on teen users

• NYT: What does $17.1 billion mean for Meta?

• NYT: Teens are divided on new Meta restrictions

• RTZ #784: Meta’s ‘Zuck’ on Personal AI Superintelligence

• ARD #130: Time to Crunch the Numbers

Bill Gates on the turbulent AI era

• GatesNotes: The turbulent AI era is here

• Axios: Gates sounds the alarm on an AI transition

• Axios: Gates wants to keep some jobs off-limits to AI

• WSJ: Microsoft leaving investors flying blind on its AI businesses

• AI-RTZ, January 2024: “Some day all computers would work this way”

• AI-RTZ, January 2024: Microsoft ‘eyes wide open’ in the AI race

• RTZ #705: Microsoft & OpenAI, it’s complicated

Gadget AI: Apple’s September 9

• The Verge: Apple announces September iPhone launch event

• Bloomberg Power On: Apple’s foldable iPhone details

• Apple Developer: Surprise and shine

• AI-RTZ #1113: Apple Intelligence & Siri AI with ‘Google & Nvidia’ inside

• ARD #148: The ‘Need for Speed’ in AI Inference

Clips from today

Nvidia’s $13B Hugging Face Buy

Meta’s $18B Teen Settlement

Bill Gates’ Turbulent AI Era

Apple’s Private AI Advantage

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

We have been tracking the growing resistance to AI data centers for over a year in this AI Tech Wave, from local water fights to power-bill anger.

This week the story went fully electoral, with the November midterms turning the buildout into a veritable political football for both parties.

The WSJ lays it out in “Politicians Who Once Championed Data Centers Are Now Bashing Them”:

In November, Texas Governor Greg Abbott declared his state the epicenter of AI development while announcing Google’s $40 billion investment; less than a year later he halted approvals of some 1,800 new data centers over power and water concerns, drawing the ire of President Trump, who defended the centers as an economic engine. The key items to build on: the reversals, the races, and the party now on the backfoot.

The races tell the story. Abbott’s challenger Gina Hinojosa has climbed within a few points while running a chatbot ad that blames rising electric bills on his data center approvals and his AI-tied donors, Elon Musk included.

In Michigan’s Senate race, the Democrat says voters flatly hate the facilities while his GOP opponent now backs a sector ‘pause’. In Ohio, Sherrod Brown brands Senator Jon Husted ‘the face of data centers’, and the Senate GOP’s own campaign committee warned in a memo that “data centers are the anchor hanging around Husted’s neck”, a toxic brand no single candidate can fix. Vivek Ramaswamy pledged an Ohio approvals pause, Nevada’s Democratic governor candidate would end tax abatements, and Pennsylvania’s Josh Shapiro, a 2028 hopeful who hosted Trump at a data center summit last summer, just signed an executive order letting the state block the facilities outright.

The weekend brought the sharpest turn yet. Axios lays it out in “Data centers ‘dug their own grave,’ says Texas Gov. Greg Abbott”: the governor who declared Texas the epicenter of AI development told ABC’s This Week that developers moved into communities without winning support first. And got some backlash.

He has since ordered data centers to pay the full cost of the electrical infrastructure they require, called for phasing out tax incentives, and directed regulators to audit projects before they connect to the Texas grid, noting that fewer than 10% of data center companies even answered the state’s power-demand information request. New York’s Kathy Hochul has gone further with a one-year moratorium on hyperscaler data centers, while President Trump has taken direct aim at Abbott’s turn, arguing that for Texas to say no to data centers is a mistake for an industry that could be bigger than oil.

But the feedback from local communities across the country show a different narrative.

The polling explains the panic: more than 60% of Americans now oppose new data centers in their area, up from 49% in March per a University of Pennsylvania poll, and Democratic strategists are calling it their best wedge issue in decades. The Evercore ISI analysts flag the volume of bans and pauses across dozens of states as an investor risk running into the 2028 presidential cycle, and Chamath Palihapitiya warns the opposition could shave 2 to 3 points off annual economic growth if it metastasizes, calling it a powder keg.

Note who is squeezed hardest. The Trump administration has been all-in on the buildout and the builders: big tech broadly, Elon in particular, and Sam Altman’s Stargate, whose SoftBank scaling story and growing pains we have covered, down to the $500 billion Ohio fine print from last week and the administration’s new tens-of-billions AI project with Japan we just wrote up.

That leaves Republican candidates running against their own president’s signature economic push, the tech-under-Trump alignment visible from the start. The president underlined the stance again over the weekend, telling his former fixer Michael Cohen in an interview that communities turning away data centers are making a mistake, and that the builders are putting up their own power plants rather than draining the grid, a claim Axios notes is only partly true.

None of this is new to longtime readers. We covered the giants boosting data centers despite the headwinds, the haste-makes-waste risks of super-sizing through Elon’s Memphis experience, Meta and Zuck proactively courting the local communities around their sites, and the grid mechanics underneath, from the AI power grab to power ramps and bad harmonics. Axios has a useful companion explainer this week in “Data center fight heating up: How they work and how much power they use”.

And the other side of the ledger is the one we laid out from technology fundamentals in our four-part AI supply and demand series: the supply ambitions (part 1), the demand bull case (part 2), and the compute crunch at ground level (part 3). The demand for AI compute is real, compounding, and already supply-constrained. These investments get built somewhere; the question the midterms will start answering is where, on what terms, and who pays for the grid.

My take: a wedge issue born in barely four years, from ChatGPT’s launch to 60% local opposition, is a remarkable political artifact of this wave. The industry earned a real share of it: haste made waste, bills went up, and trust arrived late, exactly the pattern we flagged. But moratoria do not repeal the supply and demand math; pausing the buildout while demand compounds just moves the crunch, and the price spikes, a few years down the road, or to the states and countries that kept building.

The likely landing zone is the one both parties are converging on from opposite directions: guardrails, pay-your-own-way power and water, local taxes paid in full, and community benefits up front.

Watch November as the first electoral referendum on the AI buildout in this AI Tech Wave. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

‘Beyond their Wildest Dreams.’ Especially with AI, this time around. Yesterday was the need for speed, what everyone is racing FOR. Today is about what the racers tell their backers the race is worth, and everywhere you look this week, the answer runs beyond their wildest dreams. Three stories, each a dream stated without a precedent, plus a Gadget AI on the most anticipated video game going. Expanding definitions, expanding imaginations: that is how this AI Tech Wave finances itself, and the addressable markets are just a start. My takes below, as discussed on the show.

(1) Anthropic’s $30 Trillion TAM Tops Elon’s

The WSJ’s exclusive: Anthropic is likely to tell IPO investors its potential revenue opportunity is above $30 trillion, roughly the GDP of the United States, topping the $28.5 trillion SpaceX printed in its prospectus in May, of which $26.5 trillion was attributed to AI. Elon’s rocket maker called its TAM the ‘largest actionable’ market in human history; Anthropic goes a step further, sizing the market as the full scope of work that could be completed with AI models. And Elon has kept topping his own filing since: the S-1’s mission language runs to extending “the light of consciousness to the stars,” and when questioned this summer about the Anthropic compute deal, he posted that SpaceX “will be worth more than the rest of Earth if we accomplish our goals.” The TAM escalation has left the planet, literally. Some context on the goalposts: Uber went public in 2019 citing a $6 trillion market, essentially the mileage value of all ground transportation on Earth, and Wall Street thought that was aggressive; WeWork cited $3 trillion in its abandoned IPO. NYU’s Aswath Damodaran said of the SpaceX number that it was “reaching the end of what’s plausible and pushing beyond.” For scale, the 191 technology companies in the S&P 1500 brought in about $2.4 trillion in revenue last year, combined. Anthropic’s printed numbers are considerable in their own right: revenue more than doubled to $11.6 billion in the second quarter, the offering could aim to raise as much as $100 billion against SpaceX’s $86 billion, at a valuation around $2 trillion against SpaceX’s $1.77 trillion at listing, with disclosure documents expected within weeks and a listing potentially in September or early October.

My take: I spent years at Goldman taking internet companies public in the 1990s, and TAM slides have always been a directional storytelling device, not a forecast. What is new this cycle is the definition itself: when the market is ‘all work that could be done with AI,’ the TAM is less a market estimate and more a worldview with a number attached. Investors will not pay for the $30 trillion; they may pay for the $11.6 billion quarter, the growth rate, the compute commitments, and a believable path to profits, the things we walked through in Monday’s IPO piece. But the number does real work anyway: it frames the raise, it anchors the valuation debate, and it forces every competing prospectus, OpenAI’s included, to tell a bigger story. We covered the SpaceX print day at ARD 96 under the theme ‘the Sky is NOT the Limit.’ Anthropic just proved that a second time. For now, the market is whatever investors can be brought to dream.

Sources:

• WSJ: Anthropic expected to tell investors it sees over $30 trillion in potential revenue

• Fortune: SpaceX IPO targets $28.5 trillion TAM, mission to make life multiplanetary

• TheStreet: Elon Musk’s startling claim to SpaceX investors

For longtime readers:

• AI-RTZ #1188: Anthropic’s mega-IPO aims past SpaceXAI & OpenAI

• ARD #96: SpaceX/xAI IPO aloft, the $28.5 trillion print day

(2) Leopold’s Galaxies, and the Boom-Bust Inside One Month

The WSJ’s detailed profile of Leopold Aschenbrenner opens with the purest wildest-dreams anecdote of this entire AI wave: at San Francisco dinner parties, the young investor kept telling friends about his favorite idea. He wanted to buy galaxies. Not the private-plane kind, his friends had to ask. Actual galaxies. The logic: AI would soon unlock resources on a cosmic scale, so save now, spend on galaxies later. The arc is a compressed history of this era: Columbia at 15 from Germany, valedictorian at 19, a stint at Sam Bankman-Fried’s FTX Future Fund in the Bahamas, the OpenAI alignment team, and then the 165-page ‘Situational Awareness’ manifesto two years ago, which he parlayed, with no prior investing experience, into a hedge fund that grew to $45 billion in assets, backed by some of the best investors in the world, Jane Street and the Stripe co-founders among them. The unwind came in July on a concentrated AI book levered roughly 3-to-1: down 67% through July 30, with Jane Street taking about a $15 billion monthly loss, its worst on record, and Ken Griffin’s Citadel buying most of the public portfolio at a discount, days before Leopold’s wedding to the chief of staff of Anthropic’s CEO, with much of Anthropic’s leadership in attendance. At the toasts, one guest noted the groom’s standing promise: he would still buy his wife a galaxy.

My take: we covered Leopold’s rise at ARD 132 earlier this month and the broader bust debate in the AI-RTZ pieces, and this is a familiar cautionary tale told at this wave’s scale: very charming, very charismatic, very brilliant people telling amazing stories, and stories well told, galaxies inclusive, converting into capital. The lesson from three decades of cycles: leverage converts conviction into fragility, and a thesis can be right while the trade fails. The dot-com generation was right about the internet and still got carried out on timing and leverage. Notably, his stature in Silicon Valley rose through the blowup, because there the belief itself is the asset. Both scorecards are constantly flashed. Takes being awake to realize the reality from the ‘out there’.

Sources:

• WSJ: The connections that turned a precocious teen into the fallen ‘Nostradamus of AI’

For longtime readers:

• ARD #132: As the AI World Turns, Oracle, OpenAI & Leopold

• AI-RTZ #1134: Latest report on an AI Bust & the Global Economy

(3) The Infinite Ambition Around Data

The third story is the infinite ambition applied to data. Per another WSJ piece, executives are pushing staff to move conversations out of private DMs and into public Slack channels, so AI agents can crawl everything: products, earnings, strategy. Zapier’s CEO runs a friendly competition among his executives over whose share of messages is public; meetings get recorded and posted to channels; many messages now exist primarily to be read by bots. And the tell inside the story: Google’s $10 million bid for the digital archives of bankrupt Spirit Airlines, roughly 100 million emails, personnel records and Teams messages, intended to improve its AI models, delayed by a judge last week after the flight attendants’ union objected. A dead airline’s inboxes are now a training asset in bankruptcy court.

My take: this is the thread we have been on since ‘Our Data Exhaust’ back in 2023, through the synthetic data series, the enterprise-data integration battle, and the OpenAI vs Google ‘Data Dance’ last summer: corporate data in the countless messaging systems of companies large and small, worldwide, is new data that is not on the internet, and mining it to train and run ever bigger models is a truly infinite ambition. A new race is forming around how to value that data, how it earns a return, and what investors would pay for it. Expect data rights to show up as standard assets in every bankruptcy and M&A deal from here, the way patents did, and expect the pushback to grow with it. The constraint on this dream is not technical. It is consent.

Sources:

• WSJ: Why your boss wants to make sure AI can read your Slack messages

For longtime readers:

• AI-RTZ #793: Enterprise Data the next AI integration battle

• AI-RTZ #825: The OpenAI vs Google AI ‘Data Dance’

My overall take: step back, and the three stories are part of a whole. This wave finances itself on narratives without ceilings: a TAM defined as all work, a portfolio thesis that ends in galaxies, a data ambition that ends in every message ever sent. As someone who took companies public through the internet cycle, I recognize the pattern: the narrative expands first, and the numbers either grow into it or they do not. That gap is where both the fortunes and the blowups live, and Leopold just demonstrated you can harvest both sides of it in the same month. This is not a bubble call. It is how frontier technology has always raised money: the imagination IS the collateral in the early innings, and the early innings is where we are in this AI Tech Wave. The discipline is holding two thoughts at once: the technology may yet deliver beyond the wildest dreams, while your entry price, your leverage, and your time horizon must stay wide awake. The prospectuses ahead, Anthropic’s within weeks, will be the aspirational reading of the year on exactly that tension.

Gadget AI: GTA 6, Rockstar’s Universe of its Own

The Gadget AI today is the wildest-dreams story from entertainment: Grand Theft Auto 6, which The Verge’s Stepback column argues now exists in its own universe. Full disclosure and no secret to longtime listeners: GTA 5 is my favorite video game. The record so far, per the Verge: GTA 5 has sold more than 230 million copies across PC and three console generations, placing the franchise among the most successful media properties yet measured. The sequel’s road has included a 2022 leak of 90 gameplay videos that ended with a life sentence in a secure hospital for the hacker, multiple delays, and fresh gameplay leaks this month drawing hundreds of millions of views, with Take-Two subpoenaing Microsoft and Discord. None of it has dented the fervor: the first trailer drew nearly 290 million YouTube views even after leaking a day early, next week’s gameplay reveal premieres on Netflix six hours before YouTube, and Take-Two’s CEO calls preorder numbers unprecedented. Launch is November 19, digital only, into a games industry that is otherwise struggling, with most publishers simply clearing out of its way.

My take: Rockstar is the counterexample that proves the theme. It spent a decade hand-building a simulated world at a fidelity with little precedent: hand-built worlds, extraordinary amounts of detail, hundreds of characters inhabiting thousands of places, and the market rewards it with demand that looks effectively unlimited. Now note the AI angle: those hand-built worlds are exactly what the AI labs’ world models are racing to generate on demand. GTA 6 was largely built before the AI tooling era; I would expect the next one, and the wave of games over the next two to three years, to carry considerable AI in their construction, with the craft moving up a level to taste and direction. Two bets on the same belief, that human appetite for rich simulated worlds has no visible ceiling. And note the week’s symmetry: Elon tells investors SpaceX could one day be worth more than the rest of Earth. Rockstar already built a universe of its own, and sells the rest of us tickets. Both are dreams beyond wild. One of them ships November 19.

Sources:

• The Verge: Grand Theft Auto VI exists in its own universe

For longtime readers:

• AI-RTZ #909: Beyond LLM AIs to ‘World Models’

Q&A

Q1: What makes GTA 5 MP’s favorite video game?

ANSWER: Losing yourself in the world, and the open-ended nature of playing its characters, or none of them, and just roaming around. The detail holds up wherever you point the camera, and you can lose yourself not just for hours but days and weeks, and more if you let yourself. Rockstar ran the same play in the Wild West with Red Dead Redemption, parts one and two, but GTA is the gold standard.

Q2: How does AI change games like this?

ANSWER: Considerably, over time. AI will help make these worlds more real and more dynamic. It is not in GTA 6 so much as it will likely be in the next one, and in the games coming over the next two to three years, with considerable AI tooling, authorship, and aspirational worlds to dream in.

The dreams beyond wild imaginations, a $30 trillion market, galaxies for sale, every message ever sent, a hand-built universe, are how this AI Tech Wave tells its story to the people funding it. Thanks for sharing in all these dreams. And Nvidia reports after the close today: the next reality check arrives tomorrow. Stay tuned.

Full Source Reading

Anthropic’s $30 trillion TAM, and the escalation

• WSJ: Anthropic expected to tell investors it sees over $30 trillion in potential revenue

• Fortune: SpaceX’s S-1, the $28.5 trillion TAM and the multiplanetary mission

• TheStreet: Musk’s ‘worth more than the rest of Earth’ post

• AI-RTZ #1188: Anthropic’s mega-IPO aims past SpaceXAI & OpenAI

• ARD #96: SpaceX/xAI IPO aloft, one for the record books

Leopold and Situational Awareness

• WSJ: The fallen ‘Nostradamus of AI’

• ARD #132: As the AI World Turns, Oracle, OpenAI & Leopold

• AI-RTZ #1134: Latest report on an AI Bust & the Global Economy

The data ambition

• WSJ: Why your boss wants AI reading your Slack messages

• AI-RTZ #793: Enterprise Data the next AI integration battle

• AI-RTZ #825: The OpenAI vs Google AI ‘Data Dance’

• AI-RTZ #461: Perfecting ‘Synthetic Data’

• AI-RTZ, 2023: Our Data Exhaust

Gadget AI: GTA 6

• The Verge: Grand Theft Auto VI exists in its own universe

• AI-RTZ #909: Beyond LLM AIs to ‘World Models’

Clips from today

Anthropic’s $30 Trillion Dream

AI’s Infinite Data Race

GTA 6, Rockstar’s Masterpiece

Losing Myself in GTA 5’s World

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

The frontier labs are forward deploying platoons of engineers deep into enterprises. Helicoptering them in as fast as possible ahead of their mega-AI IPOs.

We spent Monday on Anthropic’s record IPO pitch and Tuesday on OpenAI’s pre-IPO re-grounding through their ongoing senior people churn. Today, the battlefield both roadshows are really about: the enterprise, and the armies of engineers the frontier labs are now deploying into enterprises via a range of partners. The fashionable acronym is ‘FDE’ of course, for ‘forward deployed engineers. It’s a wave to watch in this AI Tech Wave.

The WSJ lays it out in “Private Equity Is Deploying an Army of AI Wonks to Embed in the Firms They Back”: Blackstone, Hellman & Friedman and Anthropic have formed Ode, a standalone firm with a roughly 160-person senior engineering team. Part of a $1.5 billion joint venture whose backers also include Apollo, General Atlantic and my alma mater Goldman Sachs, with the three lead partners committing around $300 million each.

Ode acquired the startup Fractional AI in May to seed the bench, and plans to sell far beyond the private equity world; its CEO calls the customer pipeline immense. OpenAI has formed a rival consulting venture with a $4 billion investment led by TPG. The key items to build on: the embedded engineers, the money behind them, and how this an old playbook modified for this AI wave.

The case studies are decidedly unglamorous. At Chamberlain Group, Blackstone’s garage-door opener maker, Ode engineers sit in three-person pods helping build a ‘digital doorman’ business: package alerts, familiar-face recognition, routine-learning software, three million cameras going onto homes this year. Chamberlain now expects $500 million of annual revenue from that line by 2030, against the $160 million ceiling it projected before the venture, with its Prime Day ad campaign running ten times cheaper on AI tools.

Accounting firms Citrin Cooperman and Baker Tilly have Ode engineers mapping auditor workflows for AI agents. Meanwhile CEO anxiety is cooling: 20% of CEOs in a May EY-Parthenon survey expect AI to drive significant job cuts, down from 46% in January 2025; the pitch has shifted to new revenue, not headcount.

Regular readers will recognize the playbook, as it was shaped with an AI narrative by Palantir.

These are Palantir’s Forward Deployed Engineers, the FDEs we covered as they went from engineers to entities, adopted wholesale by the frontier labs as they deepen enterprise sales. Anthropic is wholly committed to the strategy, the premium enterprise machine at the heart of its IPO pitch; OpenAI is running the same play, with enterprise now bigger than consumer per its CFO. The goal is the long tie-in: enterprises wired into their frontier models at the top, their smaller faster models below, and the productivity layer of Claude Cowork, ChatGPT and the coming super apps on every desk, the agents-into-the-enterprise arc we mapped early.

Which flies directly into the counter-narrative we covered in the ‘Us vs Them’ trillion-dollar enterprise AI fight. Palantir’s Alex Karp has been harshly warning CEOs that the frontier labs are hoovering up their data and IP, and will eventually compete with the very businesses they serve, positioning Palantir as the shield and the layer in between, with its own FDEs on site. Microsoft under Satya plays the same game with a leg in each tent, having picked its side versus the frontier models while leaning its own AI shifts into the enterprise, just a bit softer than Karp about it.

That is why the data-retention split this week matters more than it looks.

Axios lays it out in “OpenAI previews zero-retention safety system as Anthropic requires data logs”.

OpenAI is previewing ‘Private Safety Processing’, a system meant to catch misuse patterns across related interactions while preserving zero data retention: narrow safety signals only, customer data on customer-controlled infrastructure or under customer-held encryption keys, white paper due in September.

Anthropic has gone the other way, requiring 30-day data retention for business customers on its most capable Fable 5 and Mythos 5 models, conceding in its own risk report that the decision will be unpopular and commercially risky, but calling it essential to catch sophisticated attacks that span multiple requests. The same week, OpenAI paused some new-model training work over safety concerns while Anthropic said it sees no current need. The poses keep swapping, as we covered in Strike a Pose.

Retention is exactly the surface Karp’s warning lives on: the fear that the model company sees your data. OpenAI is turning zero-retention into a sales weapon against Anthropic’s security-first logging; Anthropic is betting enterprises will accept a 30-day window from the safety-first vendor. Both are answering the shield argument without naming it.

Zoom all the way out and this is the AI makeover of traditional enterprise software and services, a multi-trillion dollar global market catered to for decades by IBM, Salesforce, SAP, and countless other software majors across prior tech waves.

The entrenched software battle we framed and watched shift at IBM and Microsoft. Those incumbents deployed software services reps embedded long-term at enterprise customers; the AI companies have simply renamed them FDEs. An updated term for a similar consultant, sales and services role, wired in for the long haul.

My take: the FDE armies are the frontier labs building the enterprise relationship layer directly. Investing billions of dollars of joint ventures with the very PE and financial firms that control thousands of businesses, rather than waiting for CIOs to come to them. It’s a development that cuts across AI Tech Stack. With Box no. 6, the AI Application and Services layer, being the holy grail.

The long-term battle is over who owns that layer: the labs directly through ventures like Ode and TPG’s, the shields like Palantir, or the platforms like Microsoft straddling both. Data trust just became the pricing card in that fight, and with both labs heading to the public markets, enterprise trust IS the roadshow.

Decades of enterprise software history say the embedded services rep always wins more than the product demo. And then there’s billions in the maintenance and services layers.

The frontier labs are re-defining that enterprise software history for the AI Tech Wave, their way. And deploying them for duty as fast as possible. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

‘The Need for Speed.’ For AI inference, especially locally. Yesterday we put the whole AI world on Nvidia’s shoulders; today is about what everyone is actually racing FOR: more AI inference computing capacity, in the cloud and on our desks, as fast as possible. Speed, in tokens per second, tokens per watt, tokens per dollar, is the currency of this AI Tech Wave going into the second half of 2026, and into the two US mega-AI IPOs ahead. Four stories today, all the same physics from different altitudes. My takes below, as discussed on the show.

(1) OpenAI Leans In on its ‘Jalapeño’ Inference Chips

OpenAI took the wraps off Jalapeño, its self-designed AI inference chip, at the Hot Chips conference, and the early reviews are pretty positive. Semianalysis, one of the widely followed chip analysts in the business, was invited into OpenAI’s labs to benchmark it themselves, and their headline says it plainly: “better than Nvidia Blackwell.” Their numbers: Jalapeño beat every Nvidia, AMD and Google chip they have tested on top open source models, on the metric that matters most, tokens per megawatt. Over 700 tokens per second per user on DeepSeek R1, around 1,400 on GPT-OSS, all without the ‘speculative-decoding’ tricks other chips lean on. The Verge adds OpenAI’s own framing: 1.5 to 1.9 times more AI work per watt, and 1.7 to 3.6 times lower latency, than Nvidia’s GB200-class systems, with hardware VP Richard Ho calling it the best of both worlds since chips normally trade latency against throughput.

The backstory is notable for its speed of development, along with the benchmarks. Built with Broadcom from a blank sheet: design team hired mid-2024, tape-out in about 16 months, silicon back for only 3 months, and already posting these results, with a B0 stepping in the fab carrying roughly 25% better performance per watt. The chip uses HBM4 memory ahead of most of the industry, runs at 700 watts against Rubin’s 900-plus, and OpenAI’s own Codex wrote most of the kernels. Now the caveats, and there are real ones: the numbers were provided by OpenAI, verified in-lab but on the friendlier benchmarks; production only ramps through 2027; and OpenAI itself says this supplements Nvidia rather than replaces it. Why the hurry? The same day, OpenAI restored the five-hour usage caps on Codex for ChatGPT Plus users to smooth the load on its compute. Demand is so hot they are rationing it.

And zoom out with the Dwarkesh Patel and Dylan Patel conversation this week, a useful macro frame on all of this. Dylan’s math: world compute in gigawatts roughly doubles every year, but frontier lab compute triples. Anthropic and OpenAI alone are on track to take roughly 45% of the world’s incremental compute next year, and half of all new compute by late 2027, because they monetize tokens better than most, and can generally outbid everyone. As Dylan puts it, “every force is screeching towards centralization.”

My take: the AI supply-demand race accelerates, exactly along the lines of our four-part supply and demand series. Jalapeño is not about replacing Nvidia in the near term. It supplements it. The overall supply picture stays tight for the foreseeable future, and the frontier labs are pressing on the gas for every alternative path to compute: their own chips, other people’s chips, and other people’s data centers. It is currently a two-horse race between OpenAI and Anthropic to capture the maximum share of the world’s AI compute, and Anthropic is now standing up its own ASIC team, with some Jalapeño alumni on it, to run the same play. I called the tension years ago in ‘Nvidia vs its Top Customers’ and the ‘Frenemies’ piece. The frenemies era is now shipping silicon. The need for speed compounds from here.

Sources:

• Semianalysis: OpenAI Jalapeño, better than Nvidia Blackwell

• The Verge: OpenAI says its Jalapeño chip can power faster AI responses than the competition

• 9to5Mac: OpenAI restores 5-hour Codex and Work limits for ChatGPT Plus users

• Dwarkesh Patel podcast: Anthropic & OpenAI will have most of the world’s compute by 2028, with Dylan Patel

For longtime readers:

• AI-RTZ #554: Nvidia vs its Top Customers for AI GPUs

• AI-RTZ #806: The ‘Frenemies’ nature of the AI Coding space

• AI-RTZ #1180: AI Supply and Demand Together (part 4)

• AI-RTZ: The Dwarkesh compute demand bull case (part 2)

(2) Perplexity and Nvidia Ship the ‘Portable Computer’

The local leg of the need for speed. One day after the report of Nvidia discussing a $30 billion-plus investment in Perplexity, the two companies shipped a product together: Portable Computer, a version of Perplexity’s agentic ‘Computer’ platform that runs entirely on hardware you already own. Nvidia’s DGX Spark desktop supercomputer, any Linux machine with an RTX GPU carrying 24 gigabytes of video memory, or the other Nvidia/Microsoft AI computers being rolled out by OEMs that I wrote about recently, with supplemental cloud compute off various tiers of Perplexity subscriptions. The headline feature, per VentureBeat: zero token costs. The models, your files, and the work all stay on the machine; in the demo, an agent chewed through a folder of 1099s and investment documents while the cloud-credit counter sat parked at zero. Every task starts on-device by default, and the system asks permission before escalating any step to a frontier model in the cloud, running a privacy classifier over what would leave the box first. The hybrid economics are the tell: on a hard coding benchmark, the local model alone scored about 60% at essentially zero cost; letting it consult a cloud frontier model as an advisor recovered most of the remaining gap at two-thirds of the frontier-only cost.

My take: the two companies are not waiting after their deal news to expand AI compute options across multiple models, local and cloud together. A smart move by both. For Nvidia, it is the killer app its DGX Spark desktop has been waiting for, and another lean into open source models to supplement its close frontier-lab relationships. For Perplexity, it is a product whose economics do not depend on metering every token, and a wedge into the privacy-sensitive enterprises in law, healthcare and finance. One conspicuous absence on the roadmap: Apple silicon. Hold that thought for the Gadget section, because the local speed race is very much multi-front.

Sources:

• VentureBeat: Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs

For longtime readers:

• AI-RTZ #1104: Nvidia and Microsoft focus on local AI computers

• ARD #147: yesterday’s Nvidia/Perplexity $30B investment discussion

(3) AI Model Routing Companies in Demand, post the Stripe/OpenRouter Buy

Model routing is booming, per Axios’ aptly titled piece, routing is coming for the frontier AI labs. The scoreboard: Stripe agreed to buy OpenRouter for north of $8 billion, up from the roughly $7 billion first reported when we covered it at ARD 142. OpenRouter had 8 million users and 400 models on the platform as of May. Meta is reportedly building its own OpenRouter competitor, called Switchboard, partly to lower its own AI bills. Even the frontier labs now ship model pickers of their own. And the tell inside the numbers: the five most popular models on OpenRouter are all open weight or open source. OpenAI and Google place just one model each in the top ten. Anthropic places none. Average token costs have plummeted since their mid-May peak as the labs cut prices to win back share, and there is a sovereignty current running underneath too: enterprises that do not want their IP flowing to the frontier labs, the same shield instinct we covered in the FDE wars.

My take: this trend was underway well before the Stripe deal, as we wrote in the ‘open to close’ battle piece back in the day. But now it gets turbocharged across the industry. Routing turns models from destinations into interchangeable commodities, which is exactly why the frontier labs are keenly aware, and why I expect them to keep their a la carte pricing efforts relatively muted ahead of their mega-AI IPOs. Nobody pitching a $2 trillion valuation wants to headline a price war. The router is the buyer’s answer to the need for speed at the right price, every query, every time.

Sources:

• Axios: Routing is coming for the frontier AI labs

For longtime readers:

• ARD #142: Stripe’s OpenRouter Buy, in ‘Hidden Figures’

• AI-RTZ #483: The other AI open to close battle

My overall take: notice the two opposite directions the need for speed is pulling, at the same time. Up and in: Dylan Patel’s math has OpenAI and Anthropic absorbing half the world’s new compute by late next year, building their own chips, and outbidding everyone for the rest. Down and out: routers, open models, and zero-token local agents pushing usable intelligence onto desks and into devices, at marginal costs approaching zero. Same scarcity, opposite responses. The centralization pays for the frontier; the diffusion is how the rest of the world routes around its price. My long-held thesis of AI ‘mainframe’ computing on cloud infrastructure, supplemented by local AI computing on premises, is coming together, despite the headwinds of ‘RAMageddon’. That is the supply and demand story of our four-part series, now running at full speed on both rails. Speed, measured in tokens per second, per watt, and per dollar, is the scoreboard for all of it.

Gadget AI: Apple Refreshes the Mac mini and Mac Studio, in ‘RAMageddon’

The Gadget AI today is the humble hero of local AI: Apple refreshed the Mac mini and Mac Studio with its latest M5 and M6 chips. And pithily put by the WSJ, ‘good luck getting your hands on one’. The news: a new Mac mini with the all-new M6 chip at $899, or an M5 Pro at $1,699, both $100 more than the outgoing models. The Mac Studio moves to the M5 Max at $2,499 and the M5 Ultra at $5,499, up $200. Apple claims at least 4x faster AI performance generation over generation. The M6 is Apple’s first 2-nanometer chip, per the Verge, with a dual 16-core Neural Engine and a neural accelerator in every GPU core; the M5 Ultra fuses four dies into Apple’s most powerful chip yet, supporting 512 gigabytes of unified memory. Preorders are open, shipping September 22. Avid developers will likely spec out the top configurations at $10-20k per unit. With max chips, RAM and storage. A relative bargain vs rising a la carte AI compute prices for top inference models.

Why the price hikes and the scarcity: ‘RAMageddon.’ The global memory-chip shortage plus the AI agent craze made the outgoing Minis and Studios sleeper hits that vanished from shelves; the Mini is normally about 3% of Apple’s US Mac sales, then OpenClaw and the on-device agent wave hit, as we covered at the time. Apple’s unified memory architecture, where CPU and GPU share one pool, happens to be exactly what local AI models want.

My take: this is Apple’s signal that it takes the local AI compute opportunity seriously. They shipped these ahead of Siri AI and the updated iPhones September 9 next month, and ahead of the September 1 transition to the new Apple CEO, John Ternus, the ex-hardware chief, which tells you what kind of company Apple intends to be under him. A positive, proactive lean into this AI Tech Wave from Apple, meeting the near-term need for speed in local computing for millions of developers running custom and private AI agent workflows. Note the shape of the race: Perplexity and Nvidia left Apple silicon conspicuously off their local roadmap this morning, while Apple sells out of its best-selling local AI boxes. The local front has more than one army of computers.

Sources:

• WSJ: There’s a New Mac Mini and Mac Studio. Good Luck Getting Your Hands on One

• Apple Newsroom: Apple unveils a more powerful Mac mini, featuring the all-new M6 and M5 Pro

• Apple Newsroom: Apple introduces new Mac Studio with M5 Max and M5 Ultra

• The Verge: Apple’s new M6 chip gets more cores and more AI compute

For longtime readers:

• AI-RTZ #1016: Apple’s Mac mini popularity with OpenAI’s OpenClaw, in China and beyond

Q&A

Q1: What’s MP’s take on the Mac mini?

ANSWER: The best Apple computer I have ever used, across decades of using them. Especially for AI tasks with frontier models. No muss, no fuss. It sits there, it runs the agents, it gets out of the way. The unified memory is the quiet superpower: the models just fit.

Q2: Any ways they could be better?

ANSWER: They sometimes get overly hot with sustained AI computing. Long agent runs will make that little aluminum box work. So better cooling perhaps, especially as always-on agentic workloads become the default use. A small ask against what it delivers.

The need for speed in AI inference, up and in toward the frontier labs, down and out onto our desks, is the current running under every story in this AI Tech Wave. And tomorrow brings Nvidia’s quarter, the next weigh-in for all of it. Stay tuned.

Full Source Reading

OpenAI’s Jalapeño chip, and the compute race

• Semianalysis: OpenAI Jalapeño, better than Nvidia Blackwell

• The Verge: OpenAI’s Jalapeño chip benchmarks

• 9to5Mac: OpenAI restores 5-hour Codex limits

• Dwarkesh Patel podcast, with Dylan Patel

• AI-RTZ #554: Nvidia vs its Top Customers

• AI-RTZ #806: The ‘Frenemies’ nature of the AI Coding space

• AI-RTZ #1180: AI Supply and Demand Together (part 4)

• AI-RTZ: The Dwarkesh compute demand bull case (part 2)

Perplexity and Nvidia’s Portable Computer

• VentureBeat: Perplexity partners with Nvidia to launch Portable Computer

• AI-RTZ #1104: Nvidia and Microsoft focus on local AI computers

• ARD #147: The World on its Shoulders

Routing

• Axios: Routing is coming for the frontier AI labs

• ARD #142: Stripe’s OpenRouter Buy, in ‘Hidden Figures’

• AI-RTZ #483: The other AI open to close battle

Gadget AI: the new Mac minis and Studios

• WSJ: There’s a New Mac Mini and Mac Studio

• Apple Newsroom: the new Mac mini

• Apple Newsroom: the new Mac Studio

• The Verge: Apple’s M6 chip

• AI-RTZ #1145: ‘RAMageddon’ really here to stay

• AI-RTZ #1016: OpenClaw and the Mac mini run

Clips from today

OpenAI’s Jalapeño Inference Chip

The OpenAI vs Anthropic Compute Race

Apple’s Mac minis, Ahead of iPhones

New Mac minis, Despite RAMageddon

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)

So many flows and cross-flows at OpenAI, in so many directions.

Yesterday we looked at Anthropic’s bankers pitching the largest IPO in history. Today, the other half of the mega-IPO cadence in this AI Tech Wave: OpenAI, and who is actually running it on the road to its own public offering targeted for 2027.

The Verge lays it out in “It’s Greg Brockman’s OpenAI now”:

Sam Altman remains CEO, but in day-to-day operations, the president and co-founder has quietly become the big boss, his purview expanding with every departure in a year that included the Elon Musk jury trial we covered from opening arguments, an Apple trade-secrets suit, and a steady exodus of the company’s biggest names. The key items to build on: the exits, the consolidation, and what it says about the 2027 IPO strategy.

The people churn is an ongoing OpenAI topic discussed here for months, and the Verge assembles the full casualty list. April alone saw the Sora chief, the science VP, and the B2B applications CTO exit, with the CMO and Fidji Simo, OpenAI’s CEO of AGI deployment we profiled when she arrived, went to work, and got the close-up treatment, both departing on medical leaves; Simo officially stepped down in July, weeks after the IPO filing. This month, chief revenue officer Denise Dresser left after just eight months, and longtime COO Brad Lightcap, latterly on ‘special projects’, announced he is leaving to start something new, the churn we mapped as OpenAI pre-markets its IPO.

Each wave expanded Brockman’s authority. When Simo went on leave, he took product, including the super app effort. The revenue-focused reorg then handed him the entire ‘scaling’ arm, virtually every part of the commercial operation. When Dresser departed, the press release quote came from Brockman, not Altman. The engineering workhorse of the founding team, the man who resigned in solidarity within hours of Altman’s 2023 ouster, now answers only to Altman himself. On CNBC he waved off the red flags: “I’m a constant, Sam [Altman] is a constant”, pointing to eras of different leaders as a strength.

Consolidation collapses decision layers and trims some of the most expensive salary lines ahead of a listing. Standard pre-IPO housekeeping, though the clustering in so short a window is unusual.

It fits the repositioning tracked all summer. CFO Sarah Friar telling investors enterprise is now bigger than consumer for the ChatGPT company. The kitchen-sink costs thrown in ahead of the mega-IPO.

The rolling internal Code Reds that preceded this consolidation. And the pricing lead OpenAI keeps pressing, having led the industry into the price wars and a la carte, metered pricing for long-running agents, the volume play against Anthropic’s premium-priced enterprise machine.

The directional results landed this weekend in the FT’s “Anthropic’s best AI model struggles to attract users as cheaper tools thrive”: GPT 5.6, priced significantly below Anthropic’s Fable 5, has jolted OpenAI’s annualized revenue up 35% this quarter to date, to over $40 billion, after a sluggish start to the year.

The same Ramp spending data across 70,000 companies shows Anthropic’s priciest model plateauing at about 11% of customer outlay on its tools, with its smaller, cheaper Opus 5 passing it in business spending: the Tokenmaxxing-to-Tokenbudgeting shift among business buyers we cover in Monday’s mega-IPO piece. The gap to Anthropic’s $65 billion run rate is still wide, but for the first time this year it is closing from OpenAI’s side.

My take: executive churn at this scale is usually read as a warning sign, and the experts quoted duly wave the flag.

But the pattern here reads more like a deliberate re-planting of responsibilities and assignments.

OpenAI is reorganizing itself from a research lab with celebrity executives into an operating company with one operator, pointed at revenue, an initial enterprise/AI Coding focus, then consumer products and hardware, and beyond.

The places it can differentiate to take back the Coke seat Anthropic now occupies. That is what you do when the IPO window is 2027 and the company ahead of you is pitching $2 trillion valuations this fall.

The risk is the one the analysts land on: key-person concentration, squared.

The 2023 board crisis was about one indispensable man; the 2026 org chart now runs through two. For public investors, the safety poses and the growth math will matter, but the institution question remains.

Can OpenAI operate without depending so heavily on Sam and Greg, is the one the 2027 prospectus will have to answer. Two mega-IPOs, two opposite playbooks: Anthropic selling premium discipline, OpenAI selling reach and reinvention this AI Tech Wave through their mega-AI IPOs.

The public markets get to see the result of flows and cross-flows at OpenAI soon enough. Stay tuned.

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here.)