Hugging Face Said No to Nvidia's $500M. It Just Sold for $12.93B.

TL;DR

On September 3, Nvidia confirmed it will acquire Hugging Face for $12,930,300,000. The announcement puts roughly $11.9 billion in shareholders' hands and up to $1 billion into retention equity for Hugging Face staff who join Nvidia. The platform being bought hosts more than 3 million models, 500,000 datasets, and 1 million applications for over 18 million developers and 200,000 companies. Its reported revenue is about $150 million a year, which makes the price around 86 times revenue. Jensen Huang's pledge is one sentence long and it is the whole ballgame: "NVIDIA compute will not be required to build on or deploy through Hugging Face." Nothing changes for your pipeline today. The deal is not expected to close until the first half of 2027, and it has to clear regulators first.

The number that makes this story funny

Late in 2025, Nvidia offered Hugging Face a $500 million investment at a $7 billion valuation. Hugging Face turned it down. TechCrunch reported the stated reason was discomfort at having a dominant investor able to influence company decisions.

Nvidia has now resolved that concern by buying all of the decisions.

CEO Clement Delangue posted the exact figure himself, down to the last hundred dollars, and framed it as open-source AI hitting "an inflection point" ten years in. All three founders and the whole team are going to Nvidia. Delangue told CNBC the plan is "to continue to run independently neutral platform within the Nvidia team," which is a sentence that will be quoted back at somebody in about eighteen months.

Nobody is buying $150 million of revenue for $12.9 billion

Revenue went from roughly $100 million to roughly $150 million in about two months this summer, and the company was described as close to profitable. That is a genuinely good software business. It is not a $12.93 billion software business at any multiple a public-market investor would sign off on.

So price the thing Nvidia is actually buying: the place 18 million developers go before they decide what hardware to rent. It is like owning the recipe site instead of the grocery store. The tolls are small, but you see what everyone plans to cook a week before they shop.

That visibility matters more to Nvidia right now than it did two years ago. Its largest customers are all building their own silicon. A healthy open-model ecosystem, anchored to hardware Nvidia sells, is structural insurance against the closed labs designing their way off the GPU. Buying the index is a cheaper hedge than winning every customer back.

The neutrality problem, stated plainly

Hugging Face's value was never the hosting. It was that the hosting was indifferent. A model card sits there whether the weights run best on an H200, an AMD MI-series card, a TPU, or a laptop CPU. That indifference is why the hub became the default, and it is the one asset an acquisition can quietly damage without anyone filing a bug.

Note who is getting a check here. Hugging Face's $235 million Series D in August 2023, at a $4.5 billion valuation, was backed by Salesforce Ventures, Google, Amazon, IBM, Qualcomm Ventures, Intel, AMD, and Nvidia itself. Two of Nvidia's direct silicon competitors funded the neutral index. Both now cash out of it into Nvidia's hands.

Huang's counter is on the record and it is specific rather than vague, which is worth something: Hugging Face "will remain an open platform for the entire AI ecosystem," with multi-cloud and multi-accelerator support, and Nvidia compute will not be required. Nvidia also arrives with receipts. It is already the single largest contributor of open models and data to the hub, with more than 500 models and 250 open datasets published there.

The precedent everyone reaches for is Microsoft and GitHub, and it cuts both ways. GitHub did not get shut down or wall-gardened. It also did not stay the neutral thing it was, and rivals have spent 2026 running marketing on exactly that complaint. Erosion here would not look like a policy change. It would look like the Nvidia-optimized path being one click shorter than everything else, for four years running.

What actually changes for you, and when

Today: nothing. The transaction is expected to close in the first half of 2027, subject to regulatory approval and customary closing conditions. Between now and then Hugging Face operates as an independent company, because legally it is one.

What you should do is the hygiene you already owed yourself, because a single-vendor default is a risk whoever owns it. Every from_pretrained call in your codebase reaches out to one domain, and most of them do not pin anything.

• Pin the revision. Pass revision= with a commit sha, not a branch name. A tag or main can move under you; a sha cannot.
• Own the cache path. Set HF_HUB_CACHE (or HF_HOME) to a directory you back up, rather than leaving artifacts in a container layer that evaporates.
• Fail loudly in CI. Set HF_HUB_OFFLINE=1 in build jobs so a missing artifact raises instead of silently re-downloading a different one. All of these are in the huggingface_hub environment variable docs.
• Know your second source. If a model matters to production, keep a copy of the weights somewhere that is not a startup being acquired by a chip company.

The parts nobody has answered

Regulators are the real variable. Nvidia already faces active antitrust inquiries in the US and EU over GPU allocation and customer relationships, unrelated to this deal. Bolting on the industry's dominant model-distribution hub is the kind of vertical move that draws a full merger review rather than a quiet clearance, in the US, the EU, and likely the UK. The question examiners will ask is the same one builders are asking: does the company that sells the shovels get to run the map of where the gold is.

There is also a geopolitical wrinkle sitting in Nvidia's own filing. The risk disclosures flag potential government restrictions on Chinese open-weight models hosted on the platform, naming DeepSeek and Moonshot AI. A neutral French startup hosting Chinese weights is a research platform. A US chip giant hosting Chinese weights is an export-control conversation.

And the price only makes sense if the hub stays the default. Recent survey data suggests model-hub procurement has flattened at under a third of respondents while direct sourcing from model providers grew, which means Nvidia may be buying the distribution chokepoint of 2024 rather than 2028. That is a real bet, not a sure thing.

Key Takeaways

• Nvidia will pay $12,930,300,000 for Hugging Face: about $11.9 billion to shareholders and up to $1 billion in retention equity for staff who join.
• That is roughly 86 times reported annualized revenue of about $150 million. Nvidia is buying distribution and demand visibility, not a P and L.
• Hugging Face declined a $500 million Nvidia investment at a $7 billion valuation last year, reportedly over dominant-investor influence, then sold outright for 1.8x that valuation.
• Huang committed on the record that Nvidia compute will not be required to build or deploy through the hub, and Nvidia is already its largest open-model contributor.
• Two of Nvidia's silicon competitors, AMD and Intel, were Series D investors and are being cashed out of the neutral index by Nvidia.
• Nothing changes until the deal closes in the first half of 2027, and it must survive merger review in the US, EU, and likely the UK first.
• Pin your model revisions to commit shas, control HF_HUB_CACHE, and keep a second copy of anything in production. Good practice regardless of who owns the domain.

Sources: NVIDIA to Acquire Hugging Face (NVIDIA blog), TechCrunch: Nvidia confirms it will buy Hugging Face for $12.9 billion, TechCrunch: Nvidia closes in on Hugging Face acquisition, The Register, Fox Business on the deal structure, Engadget, CNBC Squawk Box transcript with Huang and Delangue, Clement Delangue on X, Futurum Group analysis, IBM newsroom on the $235M Series D, NVIDIA to Acquire Mellanox for $6.9 Billion, huggingface_hub environment variables