Nvidia Is Buying the npm of AI

Last week I ended The Decode #008 with a report that Nvidia was buying Hugging Face and a note that neither company had confirmed it. This week they did. Nvidia signed a definitive agreement on Tuesday, filed an 8-K with the SEC on Wednesday, and Jensen Huang published a blog post the same day.

Two months ago Hugging Face was the company an AI agent fleet broke into. Now it's going to belong to the company that sells the chips those models run on.

What was actually signed

The headline number everywhere is $12.9 billion. The SEC filing breaks it down a little differently, and it's worth reading the filing rather than the headlines:

• An "approximately $11.9 billion purchase price" to Hugging Face's stockholders, subject to adjustments.
• On top of that, an equity-based retention program of "up to approximately $1.0 billion" for Hugging Face employees who join Nvidia.

Add those together and you get the $12.9 billion (Nvidia's blog gives the exact figure as $12,930,300,000). So roughly a billion of the price is paying people to stay, which tells you what Nvidia thinks it's buying: the team and the community as much as the servers.

The rest of the filing:

• It hasn't closed. Closing needs "customary closing conditions, including receipt of required regulatory approvals," and is expected in the first half of 2027. The filing doesn't name the regulators.
• No cash or stock split is given, and no termination fee is disclosed.
• Nvidia flags a risk that governments could restrict open-source AI models, including models that originate in China, and says that could limit the deal's benefits.

What Nvidia promised

Huang's post makes a set of commitments, and the 8-K repeats the core one, that Nvidia will "keep Hugging Face's platform open, consistent with Hugging Face's existing practices." In the blog post:

• "Hugging Face will remain an open platform for the entire AI ecosystem."
• "NVIDIA compute will not be required to build on or deploy through Hugging Face."
• Developers keep choosing their own models, frameworks, clouds, inference providers and hardware, and the platform keeps supporting models from every builder and other chip vendors.

Nvidia says Hugging Face has more than 18 million users, more than 3 million models, more than 500,000 datasets and more than 200,000 companies on it. What Nvidia says it will add: more infrastructure, better reliability, and stronger evaluation, inference and deployment tooling.

The way I think about it: a registry changing owners

I don't think of Hugging Face as an AI company. I think of it as a package registry. When I want an open model, I don't go to the lab that trained it. I go to Hugging Face, the way I'd go to npm, NuGet, PyPI or Maven Central for a library. The model card is the README, the repo is the package, and a commit hash is the version.

Seen that way, the questions are the ones you'd ask if a cloud vendor bought npm:

Neutrality lives in the defaults, not the promises. Nobody needs to require Nvidia hardware to tilt the field. The question is which inference provider shows up first on a model page, which optimized format becomes the default download, and which hardware the evaluation numbers are run on. Nvidia's commitments are about what's allowed. The defaults are where the influence would actually show.

A registry is a supply chain. In July, agents got into Hugging Face through a malicious dataset and leaked credentials (that's what #008 was about). Nvidia saying it will invest in reliability and safety is good news for everyone who downloads from it. A bigger owner can afford a much better security team.

Regulation is now part of the dependency graph. Nvidia's own risk factor says governments might restrict open models, including ones from China. A lot of the best open-weight models people pull from Hugging Face today come from Chinese labs. If those rules arrive, the registry is where they'd be enforced.

It's still 2027. Nothing changes for users today, and regulators get a say first.

What I'd do now, whoever owns it

None of this needs the deal to close. It's the same hygiene I'd use for any registry:

1. Pin model revisions to a commit hash, not a branch name like main. Hugging Face repos are git repos, so this is possible today.
2. Keep a local copy of any model an app actually depends on, so an outage or a policy change upstream isn't an outage for you.
3. Write down the licence of every model you ship with. Ownership changes don't change existing licences, but it's the moment to know what you're relying on.
4. Rotate your Hugging Face tokens if you haven't since July.

Quick hits

• Rosetta's end date is set (Sep 1). Apple says macOS 27 is the final release to support Rosetta, and "Intel-only apps will no longer run" on Apple silicon Macs after it. Some Rosetta support stays for older, unmaintained games. If you still ship an Intel-only Mac build, ship a universal binary now.
• Claude Code's "+25%" limit is a cut for today's users (Aug 29). Anthropic said it will permanently raise standard weekly Claude Code limits by 25% from Sep 14, but that's measured from the base, not from the temporary 50% boost most people have been using. Anthropic's own follow-up: "Compared to today, this works out to a 17% reduction in weekly limits."
• Google Play now requires API 36 (Aug 31). New apps and app updates must target Android 16 (API level 36) or higher to be submitted. You can request an extension to Nov 1, 2026, and Wear OS, Automotive, TV and XR have lower targets.
• Gemini 3.8 Flash is out (Sep 2). Google's new Flash model is in the Gemini API, AI Studio, Android Studio and Antigravity, at an introductory $0.75 / $3.75 per million input/output tokens until the end of 2026, then $1.50 / $7.50. Google warns it may use more tokens than 3.7 Flash at higher effort levels. A restricted 3.8 Flash Cyber variant is for vetted defenders only.

What I'm watching

Which regulators open a review of the deal, and whether anything on Hugging Face's model pages changes before it closes.

Sources

Nvidia buys Hugging Face

• Nvidia, Form 8-K on the Hugging Face agreement (filed Sep 3, 2026)
• Nvidia Blog, Jensen Huang's post on acquiring Hugging Face (Sep 3, 2026)
• Semafor, coverage of the deal (Sep 3, 2026)
• Gizmodo, coverage of Huang's open-platform pledge (Sep 3, 2026)

Quick hits

• Apple Developer, upcoming changes to Rosetta support (Sep 1, 2026)
• BleepingComputer, Claude Code weekly limits change (Aug 29, 2026)
• Android Developers, Google Play target API level requirement
• Google, Gemini 3.8 Flash and 3.8 Flash Cyber announcement (Sep 2, 2026)