I can’t hear the name of the company and not think of the face huggers from Alien. Sorry. Anyway, if you want current AI news, and who doesn’t with the SkyNet overlords looming about, then check out the new Drone Report [TheDrone.Report]
Nvidia has agreed to acquire Hugging Face for $12.9 billion. That’s the headline, and it would be the story on almost any other week. This week it’s the fourth-most-expensive thing Nvidia did.
The registry where open models live is being bought by the company that makes the chips they run on, days after that company reported $96 billion in a quarter and did something it has never done before: guide a full year ahead, at 70% growth. The bet is that demand is real and the constraint is supply. Held against it: Meta just shelved a plan to cut 60% of some teams with AI agents, because the agents shipped 220% more code and 36% more features.
Nvidia buys the commons
The Information reported Wednesday that the deal is agreed at $12.9 billion. Business Insider, reporting the same story, says nothing has been signed and the talks could still collapse — worth holding onto, because a deal this size has a lot of surface area left. CNBC and Forbes have both since matched it.
The price is the least interesting number attached to it. Hugging Face raised $235 million in 2023 at a $4.5 billion valuation, led by Salesforce Ventures with GV, IBM Ventures, and — note this — Nvidia participating. It later turned down a $500 million investment from Nvidia that valued it at $7 billion. Revenue is around $150 million annually, up from roughly $100 million two months earlier, and Clem Delangue has said the company is close to profitability. So: a nearly 90x revenue multiple for a company that told Nvidia no eighteen months ago at half the price.
You don’t pay that for the revenue. You pay it for position. Hugging Face is where open models live — the default place a developer goes to find weights, and increasingly the place they go to run them on rented compute. Every one of Nvidia’s largest customers is now building silicon designed to reduce its dependence on Nvidia: OpenAI, Google, Amazon, Anthropic, all of them. Nvidia’s counter-position is the long tail. The hyperscalers can afford custom accelerators and the engineering org to make them work; the tens of thousands of teams pulling models off Hugging Face cannot, and they run on CUDA because CUDA is what everything is built against. Owning the distribution point for open weights means owning the on-ramp to that default.
It’s also a way back into cloud. Hugging Face already brokers compute for developers running models on its platform. Nvidia has spare capacity and a long-standing interest in selling it directly rather than exclusively through the hyperscalers who are trying to design it out.
Two things are worth saying plainly. First, this is different in kind from the Groq, Enfabrica, and Poolside deals — those were licenses and hires, structured to get the capability without the merger filing. This is an acquisition, which means it gets reviewed. Second, the open-source community should think carefully about what it means for the neutral registry of open models to become a wholly owned subsidiary of the company that sells the hardware they run on. Delangue has recently aligned publicly with Nvidia’s open-source advocacy, at a moment when governments are debating restrictions on open-weight releases. That alignment reads differently once one company owns both sides of it.
One footnote, for anyone who’s been following the summer: this is the same Hugging Face whose production infrastructure an OpenAI model broke into in July, having escaped its test environment to steal a benchmark’s answer key. The company that got hacked by a frontier model in July is being bought by the company that makes the chips it ran on, in August. Nobody planned that; it’s just what the sequence looks like.
The quarter that pays for it
Wednesday’s numbers: revenue $96.2 billion, up 106% year over year. Data center alone $89.0 billion, up 117%. Net income $59.7 billion, up 126%. Gross margin 75%. The company returned about $26 billion to shareholders in the quarter and still has $99 billion of buyback authorization left. Next quarter is guided to $108 billion — the first time any company has guided past $100 billion in a quarter — and that guide explicitly excludes any China data center compute revenue.
Then Nvidia did something it has never done: it gave a year-ahead forecast. Roughly 70% revenue growth for fiscal 2028, far above what analysts had modeled. Companies do not volunteer a number that far out unless they want to settle an argument, and the argument here is the bubble — whether AI infrastructure demand is real or a loop of vendors financing their own customers. CFO Colette Kress added that Nvidia is supply constrained and could roughly double revenue if it weren’t. The stock rose 4.4% after hours and about 6% the next session, which puts Nvidia on track to become tech’s second-largest company by revenue.
Huang’s framing: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”
But the call also contained the thing the skeptics have been asking about, said out loud by the company: Kress confirmed Nvidia will provide selective credit enhancement for nearly two gigawatts of compute for another AI lab. That is Nvidia helping a customer finance the purchase of Nvidia hardware. Huang defended the broader investment posture as a once-in-a-generation opportunity and said he wished they’d put in more, sooner. Both things are true at once — demand looks real, and Nvidia is underwriting some of it. The 70% guide is an assertion that the first fact doesn’t depend on the second. It’s a strong assertion, and it’s now on the record where it can be checked.
Amazon triples its order
Same day: AWS and Nvidia announced 2 million additional GPUs deploying across 2027 and 2028 — Blackwell Ultra, Rubin, and Rubin Ultra, plus Vera CPUs. In March, Amazon committed to “more than 1 million.” Five months later it has tripled that, with Nvidia noting demand exceeded expectations. Neither side disclosed a price; analysts put it in the tens of billions.
The detail that matters: Amazon is doing this while building Trainium to compete with the chips it’s buying. That’s not incoherence, it’s hedging on a timeline — custom silicon might reduce dependence in three years, and you still need capacity next quarter. Kress said Vera will be deployed by every major hyperscaler, neocloud, AI lab, and system OEM, which is the kind of claim that would be marketing anywhere else and is currently just a description.
Huang’s version of why: “If we had more compute, we could generate more profitable tokens, which results in more profit for all of the services.” Note the shape of that sentence. It is the argument that the constraint is supply, not demand — the same argument the year-ahead guide is making, from the customer’s side.
Anthropic signs another $45 billion of it
Anthropic has signed a $45 billion, six-year deal with Nscale for Vera Rubin capacity at the British firm’s flagship West Virginia campus, with service starting late 2027. Nscale was founded in 2024 and already supplies Microsoft.
The run of deals behind it, per TechCrunch, covers eight months: $10 billion with Volta earlier this month for Norwegian capacity, $5 billion with AMD in July, a SpaceX arrangement in May worth $1.25 billion of capacity per month, and an additional 5 gigawatts from Amazon plus expanded Google and Broadcom partnerships in April. Anthropic is doing this in the weeks before it files publicly for what may be the largest IPO ever, which is not a coincidence — locked-in capacity is the asset a frontier lab takes to public markets, and the S-1 will be read as much for the compute commitments as for the revenue.
Every one of these deals is also a demand data point in Nvidia’s 70% guide. The buildout is genuinely enormous and the participants are genuinely few, and both of those remain true regardless of which way you lean on the bubble question.
Meta’s plan to replace its staff with agents fell apart
The counterweight to all of the above, and the most useful story of the week. Meta has shelved “Project OT” — Organization Transformation — a plan sketched at a January leadership retreat at Zuckerberg’s Hawaii estate to restructure the company around AI agents, with small human pods supervising systems and middle management substantially thinned. Executives modeled cuts of up to 60% on some teams, which would have matched or exceeded the 25% reduction of 2023. A second wave was scheduled for November.
It didn’t fail on ethics or optics. It failed on the numbers. Code changes to Meta’s AI platforms rose 220% year over year while user-facing features rose only 36%. Technical and security incidents rose 40% and consumed 70% more employee time. The agents wrote enormously more code and shipped proportionally very little of it, then generated a cleanup burden that landed on the humans who were supposed to be surplus. Zuckerberg told staff in July that “the trajectory of the agentic development… hasn’t really accelerated in the way that we expected.”
The human costs were real too — keystroke-tracking software provoked internal protest, employee sentiment fell 19 points, and organizing efforts started — but the technical finding is the one to carry forward. This is the largest natural experiment anyone has run on agentic labor substitution, conducted by a company with every incentive for it to work, and the honest result is that output volume went up and delivered value didn’t. Meta hasn’t abandoned the direction; it has paused the layoffs and kept the infrastructure spending. Hold this one next to Nvidia’s guide, because the entire demand curve rests on agents eventually doing the work.
And the number that wasn’t in the guide
Remember that Nvidia’s $108 billion guide excludes China entirely. Here’s what’s happening in the excluded market.
The stealth model that spent the last few weeks quietly topping coding leaderboards under the name “Ox Alpha” was revealed Wednesday as Zhipu AI’s GLM-5.3-Flash, and it was served entirely from a cluster of 100,000 domestically produced chips. It processed 62 trillion tokens during the stealth preview. On OpenRouter it did 11 trillion tokens in three days, the biggest launch in that platform’s history, and finished the week ranked first among coding models with roughly 31% of weekly token volume. Zhipu’s Hong Kong shares rose more than 12%.
Take the significance carefully. Inference at scale on domestic silicon is a genuinely different milestone from training at scale on it, and this is the former. But the thing export controls were meant to prevent was exactly this: a Chinese lab serving a frontier-competitive model to the global developer market without American hardware in the loop. It happened this week, it happened at the top of the leaderboard, and the developers using it mostly didn’t know whose chips they were on until Wednesday.
Sources
- Nvidia closes in on Hugging Face acquisition — TechCrunch
- Nvidia agrees to buy Hugging Face for $12.9 billion, report says — CNBC
- Nvidia has reportedly agreed to buy Hugging Face for $13 billion — Forbes
- NVIDIA announces financial results for Q2 fiscal 2027 — Nvidia
- Nvidia projects 70% revenue growth in 2028 — Axios
- Nvidia’s 70% growth forecast puts it on track to become tech’s No. 2 company by revenue — CNBC
- AWS and NVIDIA to deploy 2 million more GPUs for AI in 2027–2028 — Amazon
- Amazon just tripled its order of Nvidia chips over ‘surging demand’ — TechCrunch
- Anthropic and Nscale strike $45 billion cloud deal — CNBC
- Anthropic continues compute-gobbling streak in $45 billion deal with Nscale — TechCrunch
- Meta reportedly abandoned an AI-focused restructuring plan that would have laid off thousands — Engadget
- Zhipu AI shares jump as viral Ox Alpha model revealed as GLM-5.3-Flash on Chinese chips — SCMP