Nvidia’s $12.9 Billion Hugging Face Talks Follow a Rebuff

People close to the talks have described a $12.9 billion Nvidia purchase of Hugging Face, a jump from the $4.5 billion value set in 2023.

Last year the hub refused a $500 million Nvidia investment at a $7 billion value, saying it did not want a dominant backer that could steer its choices.

$12.9 Billion With No Public Signature

Separate accounts of the same conversations say a contract has not been signed and the talks can still fall apart. Nvidia and Hugging Face have not commented. Deal talks started after another bidder showed interest, and Hugging Face had been working with a bank to sort through that interest.

A price near $13 billion would be one of Nvidia’s largest purchases, bigger than the $7 billion it paid for Mellanox in 2020. Against about $150 million a year in sales, the multiple is roughly 86 times revenue. Microsoft had held earlier talks of its own, according to people familiar with those conversations.

WHAT WE KNOW

• The figure: People close to the talks have put an agreed price at $12.9 billion, while other accounts of the same process put the talks above $13 billion without a signed contract.
• The silence: Nvidia and Hugging Face have issued no public statement on the reports that landed on August 26 and 27.
• The process: Hugging Face had been working with a bank after a second suitor appeared, and Microsoft had already been in the mix.

WHAT IS UNCONFIRMED

• A closed deal: There is no filing, press release, or named executive on the record saying the purchase is done.
• Terms: Cash versus stock, employee retention, and any promise to keep the hub chip-agnostic have not been published.
• Timing: No close date, no regulator clock, and no description of how AMD, Intel, Google, or Amazon seats on the cap table would be treated.

Until one of those items moves, the number is a reported price, not a completed sale. Stripe’s recent purchase of OpenRouter, a model-routing startup, for more than $7 billion is the closest parallel in the same month: the models can be free, and the layer that routes them is what draws the check.

Hugging Face Already Turned Nvidia Down

Clément Delangue, Hugging Face’s chief executive, spent 2023 selling the company as a shared workshop. Salesforce Ventures led a $235 million Series D in August 2023 that valued the firm at $4.5 billion, with Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, and Sound Ventures in the round. Delangue said at the time the money came with “no strings attached, no commitment,” which he called “really important to us.”

FROM CHATBOT TO SALE TALKS

1. 2016: Clément Delangue, Julien Chaumond, and Thomas Wolf found Hugging Face in New York as a consumer chatbot company.
2. 2018: The team open-sources the Transformers library and the product shifts from a chat app to tools other builders can train on their own data.
3. August 24, 2023: A $235 million Series D led by Salesforce Ventures values the company at $4.5 billion, with rival chip and cloud firms on the same cap table and, Delangue said, no strings attached.
4. Late 2025: Hugging Face turns down a $500 million Nvidia investment that would have valued it at $7 billion, saying it did not want a dominant investor that could sway decisions.
5. June 2026: Delangue says the company has crossed a $100 million annual run rate and is close to profitability, still working through capital raised three years earlier.
6. August 26, 2026: People close to talks describe a $12.9 billion Nvidia purchase; other accounts say nothing has been signed.

A full sale is a different instrument from a large minority check, and that is the kindest reading of the turn. Control still moves. The 2023 round was built so no single chip or cloud vendor could dictate defaults. A 2026 buyout, if it holds, replaces that structure with one owner.

Two Million Models, Thirteen Million Users

Hugging Face’s own spring census said that in 2025 the hub grew to 13 million users and 2 million models, plus more than 500,000 public datasets. More than 30% of the Fortune 500 keep verified accounts. About half the models on the hub have fewer than 200 downloads, and the top 200 models, 0.01% of the catalog, make up 49.6% of all downloads.

THE HUB IN 2025

• People: 13 million users, with more than 30% of the Fortune 500 on verified accounts.
• Catalog: More than 2 million public models and over 500,000 public datasets.
• Concentration: The top 200 models account for 49.6% of downloads, while about half of listed models stay under 200 downloads.
• Geography: Chinese models accounted for 41% of downloads in the past year, and Nvidia is the strongest Big Tech contributor of new repositories.

Alibaba’s Qwen family alone has more than 113,000 derivative models on the hub, more than Google and Meta combined. Independent developers, not industry labs, now take a larger share of downloads than they did before 2022. The workshop is still a commons in the sense that anyone can upload. Traffic, though, already clusters around a few families and a few hardware stacks, which is the map a buyer would inherit.

Why Nvidia Wants the Open-Model Registry

OpenAI, Google, Amazon, and Anthropic have been building their own AI chips so they lean less on Nvidia. A busy open-weight scene gives those labs’ customers somewhere else to train and serve models, which is a way to keep more of the market on Nvidia silicon. Nvidia has already spent tens of billions of dollars on its own open models, and on the hub it is already the loudest Big Tech contributor.

I think Nvidia is very much a community, a platform-based company, and in that respect, I think Hugging Face fits perfectly within that. There is this five-layer cake from Nvidia, and foundational models are one of them.

Siddy Jobe, fund manager, Eonopolis Exponential Technologies, on Squawk Box Europe

Jobe added that Nvidia wants to be integrated from energy through foundation models and into applications. That is the buyer’s logic in one breath. In July, Nvidia chief executive Jensen Huang used his first post on X to share a letter backing open-weight models that he signed with two dozen other companies, including Hugging Face. OpenAI and Anthropic did not sign. Huang wrote that the world needs frontier closed models and frontier open models, and that open weights help safety, diffusion, and sovereignty.

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.… pic.twitter.com/t02bi51N4C

— Jensen Huang (@JensenHuang) July 24, 2026

Delangue has been on that side of the argument too. After a cyberattack on the hub, he said Hugging Face used an Nvidia-modified version of a Chinese open model to work the incident, and he told viewers that “in this market, probably open models will be kings.” The political fight in Washington over Chinese open weights is the backdrop. Buying the registry would put Nvidia on both sides of that fight: as the loudest defender of open weights, and as the company that would own the place those weights live.

AMD and Intel Paid to Keep This Neutral

The 2023 round was unusual because rival chipmakers wrote checks into the same company. Delangue said those partners had already shared more than 1,000 open models and datasets and had more than 10,000 users on Hugging Face. In 2024 he went further, saying Hugging Face was the only AI startup with commercial work across AWS, Google Cloud, and Azure, and across Nvidia, AMD, Intel, and Qualcomm, which he called the independent and agnostic bet “truer than ever.”

WHO BOUGHT A SEAT IN 2023

AMD still listed Hugging Face as a model-builder partner this summer, next to Meta and DeepSeek, and Hugging Face engineering posts have treated Instinct MI300 as a first-class target in Transformers and text-generation-inference. Microsoft already lets thousands of hub models deploy in Azure AI Foundry. Google Cloud’s use of Hugging Face, per Sacra’s company model, has grown tenfold in three years, into tens of petabytes of downloads a month.

Microsoft kept the GitHub name and let developers run code off Azure after it bought that hub in 2018 for $7.5 billion, which is the hopeful comparison. The difference here is that AMD and Intel paid for a chip-agnostic default, not for a brand that happens to sit on someone else’s cloud. Nvidia can leave every rival kernel in the catalog and still make CUDA the path that works first. Search ranking, featured hardware badges, and which inference backend is one click are enough. A midnight purge of models is the fear that gets posted. The quieter change is what the front page prefers.

A Storefront for Leftover Cloud Time

Nvidia scaled back its DGX Cloud push about a year ago. Hugging Face already sells hosted inference and, with Nvidia, a joint Training Cluster as a Service product that rents large GPU clusters for the length of a training run. People describing the talks have said owning the hub would give Nvidia a way back into that cloud layer without building a new front door, and a way to resell unused capacity the company has pledged to cover in customer compute deals.

Sacra, which models private software firms, puts Hugging Face at 150 million dollars in annual recurring revenue in August 2026, up from an estimated $81 million at the end of 2025 and $100 million in June. Pro seats are $9 a month. Team seats are $20. The firm has been shifting mix away from one-off consulting toward API fees, cloud referral revenue, and a small robotics line after buying Pollen Robotics. OpenAI even chose the hub as the main distribution channel for gpt-oss-120b and gpt-oss-20b, which is the odd picture of a closed lab as a tenant on a registry its chip supplier may soon own.

WHAT THE BUYER WOULD HOLD

• The catalog: The default place to publish, fork, and download open weights, including the Chinese families that now take 41% of downloads.
• The meter: Inference routing, training-cluster hours, and the paid tiers that already turn a free commons into a bill.
• The demand map: Logs of which models, datasets, and hardware targets developers actually touch, which no rival chip vendor can buy at this scale.

That last item is the part of the price that does not show up in a sales multiple. A registry is a sensor. If Nvidia owns it, AMD and Intel still get listings; Nvidia gets to see which listings convert.

The Hub Has Not Said a Word

Builders who want the catalog to stay useful keep circling the same two asks: keep the free tier, and keep odd or uncensored weights findable. Those are product choices, not slogans, and they can change without a press release. Nvidia can say, correctly, that it will not take the library down. It can still decide what is polished, what is mirrored on its own cloud, and which chips get a green check.

Delangue has not explained why a company that would not take Nvidia as a heavy shareholder would now take Nvidia as owner. The simplest answer is the number. Close to profitability on about $150 million of sales is a real business. It is not, on those figures, a $12.9 billion business unless the buyer is paying for distribution and for a veto on who sits in the middle of open-weight traffic.

As of early September, neither Nvidia nor Hugging Face had confirmed the talks, and one account of the process still says they can fall apart.

Disclaimer: This article is news reporting and analysis of reported corporate talks, and it is for information only. It is not investment advice, a recommendation to buy or sell Nvidia stock, or advice on how any uncompleted deal might affect shareholders, employees, or users. Readers who are weighing securities or compensation decisions should consult a licensed financial adviser who can review their own holdings and tax position. Figures, deal status, and company comments reflect the sources available as of September 2, 2026, and can change if a contract is signed, denied, or revised.