Yesterday the labs argued about cache bills and cyber gates. This morning Jensen wrote a check for the warehouse those models live in.

Nvidia agreed to buy Hugging Face for $12,930,300,000. That is not a round number because it is not a rumor. The 8-K filed before the open says the deal was signed September 2: about $11.9 billion to stockholders, plus an equity retention pool of up to about $1.0 billion for employees who stay. Close is first half of 2027, regulators willing.

The last time Hugging Face had a public valuation it was $4.5 billion, in 2023, in a round Nvidia already sat in. The Verge notes the company turned down a $500 million Nvidia check last year that would have priced it at $7 billion, on the theory that one investor should not own the town square. That theory just got priced.

The filing is the product

Jensen’s blog post is the sales pitch. The 8-K is the operating system.

Hugging Face stays an open platform, Nvidia says, “consistent with Hugging Face’s existing practices.” Model makers keep uploading whatever they want. Users keep downloading it. Other silicon vendors stay supported. “NVIDIA compute will not be required to build on or deploy through Hugging Face.” Clem Delangue, Julien Chaumond, Thomas Wolf, and the rest keep the 🤗 brand.

The census Jensen published is the reason the check cleared:

That is not a model. It is the distribution layer. Closed labs are trying to design their own inference chips. Open-weight builders still pick a hub, then pick a GPU. Nvidia just bought the hub.

The Information has Hugging Face at about $150 million of annualized revenue. Call it eighty-six times sales if you like ugly multiples. Nvidia’s Q2 was $96.2 billion. This is a rounding error with a moat attached.

Nvidia already dumps more open models onto the Hub than anyone else: 500-plus models, 250-plus datasets, per Jensen. Buying the shelves you already stock is either vertical integration or a confession that the next fight is not FLOPs. It is where a coding agent types from_pretrained.

If you still run that agent on a desk instead of a neocloud reservation, a fast NVMe and a quiet UPS remain cheaper than waiting for H1 2027.

Gemini 3.8 Flash works harder. The bill might too.

While Nvidia was signing, Google shipped Gemini 3.8 Flash — third Flash drop in six weeks, three weeks after 3.7. Same introductory price: $0.75 / $3.75 per million input/output tokens through December 31. On January 1 that becomes $1.50 / $7.50. Bookmark it.

The design note is the useful one. Google says 3.8 “works harder”: extra reasoning steps, iterative tool calls, more tokens when the task is ugly. The Verge translated that into English. The list price did not move. The meter might.

Vendor numbers, treat as vendor numbers:

Flash Cyber is not on the public API. It rides Google’s new Fairwind Program: trusted governments, critical-infrastructure operators, software maintainers. CodeMender plus the cyber twin is supposed to write verified patches in minutes inside your cloud, not weeks in a ticket queue. Same polarity as yesterday’s Mythos/Astra split — find-and-fix for the guest list, a cheaper workhorse for everyone else — except this time the workhorse is the Flash SKU and the guest list is named Fairwind.

A used workstation still beats waiting for Fairwind to return your application.

Fair use, now with a national-security header

The other filing this week was not an 8-K. The Justice Department dropped a statement of interest in The New York Times v. OpenAI, telling a Manhattan judge that training on copyrighted text is transformative fair use and that a contrary ruling would “threaten national security.”

The Times called it the first time the department has weighed in on AI training. Associate Attorney General Stanley Woodward billed it as historic. The paper’s line: the administration is “siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole.”

Read the brief as product policy, not literary theory. If training is fair use, the Hub Nvidia just bought keeps filling. If it is not, every from_pretrained becomes a licensing desk.

A hardware security key will not get you Fairwind. It will still beat a reused password the next time a session-stealer walks off with a cookie.

The stack this week is now three objects that look unrelated until you put them on one desk. Nvidia bought the open-weight library and promised, in an SEC exhibit, not to padlock it. Google shipped a Flash model that spends more tokens so it can look like a frontier model at Flash prices, then gated the cyber twin. The DOJ told a copyright court that the training corpus is a national-security input. None of that is a benchmark. It is a purchase order, a meter, and a brief.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.