NVIDIA buys Hugging Face for $12.93 billion and promises not to tie open models to its chips

After weeks of rumors, NVIDIA has confirmed its purchase of Hugging Face — the hub through which the bulk of open machine learning models is distributed. The sum is stated down to hundreds of dollars: $12,930,300,000, that is, about $12.93 billion. The announcement was made on September 3, 2026.

What is being bought is the model distribution layer. According to NVIDIA, Hugging Face hosts more than 3 million models, 500,000 datasets and 1 million applications; the platform is used by over 18 million developers, researchers and creators, as well as more than 200,000 companies. For those who download the weights of open generative models for images, video and audio from it, the deal raises one main question: will the hub remain neutral with respect to the hardware these models run on?

Huang promised that NVIDIA chips will not become mandatory

NVIDIA CEO Jensen Huang listed the commitments right in the deal announcement. Hugging Face “will remain an open platform for the entire AI ecosystem”; developers themselves choose models, frameworks, clouds, inference providers and compute platforms. As a separate point: “NVIDIA compute will not be required to build on or deploy through Hugging Face.” The platform will continue to support open models and open-weight models from any developer, as well as multi-cloud and multi-accelerator development and deployment. The signature brand with the hugging emoji is being kept.

NVIDIA points out that it has already published more than 500 models and over 250 open datasets on the platform, and calls itself the largest supplier of open models and data for Hugging Face. The closing timeline, the payment terms and the founders' future roles are not disclosed in the announcement.

A year earlier, Hugging Face turned NVIDIA down

Last year Hugging Face rejected a $500 million offer from NVIDIA — the Financial Times wrote about it. This time the initiative came from the other side: in the blog post Huang writes that Clément Delangue approached him himself while thinking over the next stage for the company. In a post on X, Delangue thanked the community for the platform becoming an alternative to closed APIs and explained the logic of the deal: for this to work at greater scale you need “more compute, more support, more collaboration and more visibility,” and that is exactly what Huang offered.

The economics of the transition look like this. The company was founded in 2016 and has raised more than $395 million in total, according to Crunchbase. The last round was in 2023: $235 million led by Salesforce Ventures, with Google, Amazon, IBM and NVIDIA itself among the investors. The platform's annual revenue is estimated at roughly $150 million (data from The Information), and in a July interview with TechCrunch Delangue said the growth rate was bringing the company close to break-even. At that revenue, the deal price corresponds to about 86 annual revenues — the payment was for the hub's position, not for its cash flow.

Why NVIDIA needs the model distribution layer

Judging by the sequence of deals, NVIDIA is building out the stack from the top down: the Mellanox purchase gave it networking, Run:ai — compute orchestration, the agreement with Groq — inference hardware. Hugging Face covers the level at which developers find and fetch models. Openness, meanwhile, is not new rhetoric for NVIDIA: for years the company has been publishing its own models, libraries and datasets openly and distributing them, among other channels, through Hugging Face, and Huang co-authored an open letter on the importance of open weights for the AI economy. In parallel, the company is investing in model development on the other side: according to The Wall Street Journal, it struck a $6 billion deal with Poolside, and on its latest earnings call it spoke of more than $50 billion invested in frontier labs.

The direct benefit is obvious even without promises of exclusivity. As TechCrunch notes, a dominant hardware vendor benefits from an open ecosystem that it controls: it is convenient to tune the platform for it, and spare capacity can be sold to corporate customers bundled with Hugging Face services. The owner of the hub also gets to see which models are actually downloaded and deployed.

Hence the main objection. The value of Hugging Face rests largely on hardware neutrality: it holds weights that run the same way on different hardware. The parallel with Microsoft's purchase of GitHub suggests itself, but with a caveat: Microsoft maintained the platform's neutrality under competitive pressure, while for NVIDIA that pressure is noticeably weaker — it already dominates compute for AI. Immediate changes probably should not be expected, but companies would do well to assess the risks of a deeper integration of the platform with NVIDIA's tools in advance.

The regulatory part is non-trivial as well: NVIDIA has been under antitrust attention in several jurisdictions at once for years, and its attempt to buy Arm fell apart precisely on arguments about the neutrality of an infrastructure company. Promises that now sound like lines from a press release may well turn into formal conditions for approving the deal — and in that sense the phrase about NVIDIA compute not being required, placed in the announcement, appears to be addressed not only to the community.

For CG specialists the practical takeaway is restrained for now: until the deal closes nothing changes, the repositories and weights stay where they are. After that, much depends on the details. If open models for generating images, video and 3D keep coming out on Hugging Face without being tied to a specific hardware vendor, the deal will remain unnoticeable for studio pipelines. But if integration with NVIDIA's software stack turns out to be noticeably deeper than support for alternatives, the habit of keeping the entire open ecosystem in one place will start to cost more than it does now.

Community reaction

The official announcements from Jensen Huang and Clément Delangue gathered thousands of reactions: congratulations from Elon Musk, Mira Murati and Emad Mostaque sit alongside developers' distrust. Part of the community sees the purchase as a powerful boost for open models and a blow to closed services like OpenAI. Another part believes NVIDIA is buying not so much a company as the ecosystem's front door — the place from which millions of people download models. The main argument is about neutrality. Many accept the promise not to tie Hugging Face to NVIDIA chips as a necessary condition for trust, but skeptics demand proof in the product itself: don't squeeze out other vendors' models, don't cut downloads and don't strangle support for AMD and other accelerators. There are also harsh assessments along the lines of “if you can't win the race, buy the whole track” and accusations that the founders' goal is simply to cash out the company.

“If you can't win the race, buy the whole track.”

— @Rohit_wish (X)

“NVIDIA says Hugging Face will remain open and won't require their compute. Good: it is precisely Hugging Face's genuine neutrality that is the reason developers trust it. The deal will become very interesting if that ever changes.”

— @nexgencoded (X)

“‘NVIDIA will be a great home for Hugging Face.’ Ha, no. Well, have fun destroying what you had.”

— @dr3dn0t (X)

What drove the deal: fear of losing the developer ecosystem, control over open models, overheated AI market valuations?