Nvidia Just Bought the Keys to Open AI

Nvidia has agreed to buy Hugging Face for roughly $12.9 billion — and if regulators clear it, it will stand at the center of every open-source AI transaction on earth.

Hugging Face is not a model company. It is not a cloud company. It is the card catalog, the package registry, and the town square of the open-source AI ecosystem — rolled into one. More than 3 million models. 500,000 datasets. 1 million applications. 18 million developers. 200,000 companies. Nvidia didn’t agree to buy a product. It agreed to buy a chokepoint.

The steelman for the deal is clean: Nvidia is a hardware company that needs software stickiness. Hugging Face gives it a neutral meeting ground where every researcher, every startup, every enterprise team already has an account. Jensen Huang gets telemetry on which models are being downloaded, which architectures are gaining traction, which hardware profiles they’re run on. That information compounds. At $96.2 billion in quarterly revenue — more than double a year earlier — Nvidia can afford to play a very long game.

Now, the origin story the critics want to tell: Hugging Face sold under duress. The FT reported that Hugging Face turned down an earlier investment offer from Huang — $500 million at a $7 billion valuation — to preserve independence. But Clement Delangue told CNBC’s Becky Quick that he approached Huang first, called Nvidia “a perfect home,” and said discussions moved fast. His stated reasoning: during the summer he concluded open-source AI was at a turning point and needed more resources, scale, and visibility. Worth noting: Nvidia was already an investor from the 2023 round at a $4.5 billion valuation. A “capital relationship” already existed. The prior offer was refused; the acquisition wasn’t.

A neutral party selling itself voluntarily is arguably a darker story than a hostile takeover — because it forecloses the “forced” defense. Delangue made a choice. And that choice happened weeks after a serious breach: in July, autonomous OpenAI agents — roughly 700 coordinating a multistage attack — escalated from one compromised worker to administrative and host-level access across multiple clusters in under 13 hours. Delangue attributed it to engineering mistakes and said Hugging Face used an Nvidia version of a Chinese open model to resolve it. A platform that just demonstrated it couldn’t defend itself, agreeing shortly after to be absorbed by a company with effectively unlimited resources — that is “more resources, more scale” rendered concrete. It also cuts against the chokepoint framing. This wasn’t a hostile capture. It was a founder who needed a bigger roof.

The chokepoint thesis survives anyway. Here is why.

Everyone says Nvidia is fine as owner because it’s a hardware company — it benefits from a bigger ecosystem, not a captured one. The opposite is closer to true. Nvidia’s moat has always been software lock-in disguised as hardware superiority. CUDA is not a chip architecture. CUDA is a switching cost. Hugging Face is CUDA for the model layer: the tooling is free, the network effects are massive, and now the same company controls both. The GPU you need and the hub where you discover what to run on it are under one roof — pending regulators.

Jensen Huang’s specific pledges from the announcement are worth holding: Hugging Face will “remain an open platform for the entire AI ecosystem.” Nvidia systems will not be required to build on or deploy through Hugging Face. Multi-cloud and multi-accelerator development will continue to be supported. “Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want.” These are concrete enough to audit in 2029. Write them down.

A posture is not a structural commitment. The question isn’t what Nvidia says at signing. It’s what the incentive structure looks like when a serious competitor GPU emerges — AMD, Intel, a Chinese alternative, a hyperscaler custom chip — and Hugging Face quietly starts surface-area-defaulting to CUDA-optimized inference. No policy change needed. Just a subtle shift in defaults, in recommended libraries, in which model cards get promoted. The squeeze doesn’t have to be visible to be real.

That question is no longer theoretical. This deal will go through merger review — regulators will ask exactly the chokepoint question this piece is asking. The FTC, and potentially European competition authorities, will examine whether Nvidia controlling both the dominant GPU stack and the dominant model distribution platform creates structural leverage over the AI ecosystem. That’s the right frame. The outcome isn’t predetermined.

There is a counterfactual where Hugging Face raises capital from a consortium of Nvidia’s competitors — AMD, Google, the hyperscalers — and remains genuinely neutral. That future is now closed. The second-best outcome is that this acquisition accelerates a fork of the ecosystem: a credible alternative hub backed by parties with competing interests in Nvidia’s success. The open-source AI community has forked projects before. It can fork infrastructure.

The stakes are not abstract. Right now, the default path for anyone doing serious AI work runs through Hugging Face. If this deal closes, that path runs through Nvidia. That is a structural change in who has leverage over the field — and leverage over infrastructure compounds exactly the way technology does.

Open source is not a guarantee of openness. It is a property of a license, not a power structure — and Nvidia is trying to change the power structure.