Nvidia Buys Hugging Face, the Platform That Turned It Down

Nvidia agreed on September 2 to buy Hugging Face for $12,930,300,000, and Jensen Huang put in writing that builders will not need his chips to use the Hub. Hugging Face had spent the previous year refusing a much smaller Nvidia check so that no single backer could sway it.

The price Huang published opens with 129303, the decimal code for the hugging-face emoji. The close is not due until the first half of 2027, which is when that written promise has to live in the product rather than in a blog post.

Nvidia Wrote a No-Chip Rule Into the Buy

Huang’s post puts the figure to the dollar and then spends more lines on what Nvidia says it will not do. Hugging Face stays an open platform, he wrote, and developers keep the models, frameworks, clouds, inference providers and computing platforms they already pick.

Hugging Face will remain an open platform for the entire AI ecosystem. 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. NVIDIA compute will not be required to build on or deploy through Hugging Face.

Jensen Huang, chief executive, NVIDIA blog

The same note says the Hub will keep backing open-source and open-weight models from every builder, plus multi-cloud and multi-accelerator work, so that NVIDIA compute will not be required to ship on the platform. Nvidia, he added, is already the largest contributor of open models and data on the Hub, with more than 500 models and more than 250 open datasets listed under its own name.

Huang also said Clem Delangue came to him, and that other bidders had tried to buy the company. “It doesn’t matter who the other bidders were; it only matters who wins,” he said. Delangue, Hugging Face’s chief executive, said the two sides signed on Wednesday after he approached Huang over the summer because open-source AI needed more compute, more scale and more visibility.

THE HUB NVIDIA IS BUYING

• Builders: More than 18 million developers, researchers and creators use the platform.
• Catalog: More than 3 million models, 500,000 datasets and 1 million applications sit on the Hub.
• Customers: More than 200,000 companies use it to find, test, customize and deploy models.
• Nvidia’s own listings: More than 500 models and more than 250 open datasets already carry the chipmaker’s name.

Those counts are why the check is large. They are also why a lock to one vendor would shrink the asset Nvidia just agreed to buy.

The $500 Million Check Hugging Face Already Refused

Late last year Nvidia offered Hugging Face a $500 million investment that would have valued the company at $7 billion, more than all the capital it had raised since it started in 2016. Hugging Face turned the money down. The company said it did not want a single dominant investor in a position to sway its decisions.

Clément Delangue, who co-founded the Hub in France with Julien Chaumond and Thomas Wolf and runs it from New York, had a sharper line for why the refusal mattered. Open models, he said, “contribute to democratising AI, to fighting concentration of power which is, in my opinion, the biggest risk in AI.” He also said the company was profitable in 2025 and still had about half of its raised capital on the balance sheet, so it did not need a new round.

Less than a year later he is selling the whole company to the same buyer, at nearly three times the 2023 funding valuation. On Thursday he argued the sale would “allow AI to be more distributed all over the world and avoid too much concentration of power.” He posted the price in full and said Nvidia had committed to keep the platform “open, independent and compute agnostic,” with the founders and the team staying on.

https://x.com/ClementDelangue/status/2095482998674112733

Delangue’s own target in that post is 100 million AI builders who “own their intelligence rather than rent it.” Thomas Wolf, a co-founder, said no company had been a more natural fit, and that Huang had offered Delangue the chance to “double down on building the Hub as an open, independent and compute agnostic platform.” The words are the same ones Hugging Face used when it was still saying no.

Nine Investors Sat on the 2023 Cap Table

In August 2023 Hugging Face raised $235 million in a Series D that valued the company at $4.5 billion. Delangue described that round as an “ecosystem” financing, and said it solidified Hugging Face as “a neutral platform, or the Switzerland” for AI. The new money, he said then, did not come with special access for the corporate backers.

The investor list is the point. Nvidia was one name among chipmakers, clouds and software firms that all wanted a seat on a Hub they did not control.

THE 2023 ECOSYSTEM ROUND

• Salesforce Ventures: Led the $235 million Series D.
• Cloud platforms: Google and Amazon joined the round, after Amazon had already made AWS a preferred cloud for the Hub.
• Chipmakers: Nvidia, AMD, Intel and Qualcomm Ventures all took stakes in the same financing.
• Other strategics: IBM and Sound Ventures completed the group Delangue welcomed by name.

Those partners, Delangue wrote at the time, had already shared more than 1,000 open models and datasets and had more than 10,000 users on the Hub. The design was a cap table that made capture expensive. One of the nine has now agreed to buy the rest.

What the $12.93 Billion Filing Buys

Nvidia’s Form 8-K, filed on September 3 for an agreement dated September 2, splits the headline figure. Stockholders are due an approximately $11.9 billion purchase price, subject to adjustments, and Hugging Face staff who join Nvidia are in line for an equity retention program of up to about $1.0 billion. The filing is the legal home of Huang’s blog line: Hugging Face would keep letting model makers, developers and users upload and download the models and datasets they choose, and would “support other silicon vendors.”

THE DEAL IN THE 8-K

The 8-K also flags a risk that has nothing to do with AMD. Many popular open-source models, Nvidia wrote, originated in China and are then downloaded, revised and tested worldwide. New limits on those models “could have a material impact on Hugging Face’s platform” and on Nvidia’s own results. Other parties, the filing says, are already lobbying governments to restrict or disadvantage open-source models.

This is Nvidia’s second-biggest purchase after it paid $20 billion for Groq assets in December 2025. Huang told an interview the Hub is “a large growth driver of our company.” Justin Boitano, Nvidia’s vice president and general manager of enterprise computing, told a briefing the Hub is structurally a “deconcentration platform” and that regulators should read the deal as a way to spread AI rather than concentrate it. That is the argument Nvidia will carry into merger review. It is also an argument that only holds if the Hub still ships other people’s silicon.

The Hub Still Ships Code for Rival Chips

Hugging Face already publishes an interface with AMD ROCm and AMD Ryzen AI, so Transformers and Diffusers models can run on AMD stacks rather than on Nvidia GPUs. A separate package, Optimum Habana, is the Optimum tools for Intel Gaudi accelerators, covering Gaudi 1, 2 and 3. Optimum as a whole also points builders at AWS Trainium, Google TPUs and Intel OpenVINO.

That code is the practical form of Huang’s sentence about other computing platforms. If those libraries stay current, the Hub can still be the place a lab lists a model that never touches CUDA. If they rot, the pledge becomes a blog archive. The 8-K’s phrase is “support other silicon vendors.” Support, in software, is a shipping schedule.

Owning the front door is how Nvidia gets the Hub’s model choice layer without forcing a CUDA login on day one. Default buttons, recommended inference paths and the order of hardware docs can move traffic even while rival runtimes remain a click away. The worry that keeps coming back from people who actually fine-tune on the Hub is not a midnight ban on AMD. It is a quieter tilt, in which the easiest path still leads to Nvidia gear because that is the path the new owner knows how to pave.

Mira Murati, who is building Thinking Machines Lab after a stretch as OpenAI’s chief technology officer, replied to Huang with a clean congratulations and called Nvidia “a great home for Hugging Face to continue expanding access to open models.” Elon Musk’s reply was one word, “Congrats!” The named-exec thread was applause. The product argument is still the libraries, the defaults and the ranking of deploy buttons.

The Pledge Has to Last Until 2027

Delangue said he and his two co-founders will join Nvidia after the deal. Huang said the Hugging Face brand stays. Between those two sentences sits a close that the 8-K will not even attempt before the first half of 2027, because the filing needs regulator sign-off in more than one capital. Nvidia has told staff and investors the Hub is a deconcentration story. Merger offices will ask whether a chip vendor can own the main listing desk for open models and still leave rival accelerators a first-class path.

THE PATH FROM REFUSAL TO SALE

1. 2016: Clément Delangue, Julien Chaumond and Thomas Wolf found Hugging Face in France; the company later bases itself in New York.
2. August 2023: A $235 million Series D values the Hub at $4.5 billion, with Nvidia sitting beside AMD, Intel, Google and Amazon on the cap table.
3. Late 2025: Hugging Face turns down a $500 million Nvidia investment at a $7 billion valuation, saying it does not want one dominant backer.
4. Summer 2026: Delangue approaches Huang; talks move in a few weeks.
5. September 2, 2026: The companies sign. Nvidia files the 8-K the next morning.
6. September 3, 2026: Huang publishes the $12,930,300,000 figure and the no-chip rule; Delangue posts the compute-agnostic pledge.
7. First half of 2027: Expected close, after customary conditions and required regulatory approvals.

Until that close, the Hub is still an independent company with AMD and Intel code in its docs and a blog post that says Nvidia chips are optional. After it, the test is whether those libraries, those listings and those deploy paths still look optional when the owner also sells the GPUs. Huang bought a hug priced into the dollar figure. He also bought a promise that the hug does not squeeze anyone else off the Hub.