NVIDIA agreed to buy Hugging Face, and $1 billion of the price is retention

NVIDIA said on Thursday that it has agreed to acquire Hugging Face, the platform where most of the open-weight model ecosystem is published and downloaded. In a post on NVIDIA's blog, CEO Jensen Huang put the figure at $12,930,300,000. A Form 8-K filed with the SEC the same day breaks that number into two parts worth keeping separate: approximately $11.9 billion payable to Hugging Face stockholders, subject to adjustments, and an equity-based retention program of up to approximately $1.0 billion for Hugging Face employees joining NVIDIA.

That split is not a technicality. Retention equity is compensation that vests over time for people who stay; it is not consideration paid to the company's owners. Nearly every account of the deal published on Thursday used the combined headline number, which is the one Huang chose to lead with.

The timing is more conditional than the coverage suggests, too. The filing states that the definitive agreement was entered into on September 2, and that the transaction is expected to close in the first half of 2027, subject to customary closing conditions including receipt of required regulatory approvals. Nothing has been acquired yet. Headlines saying NVIDIA "has acquired" Hugging Face are running roughly nine months and one antitrust review ahead of the facts.

The scale NVIDIA is paying for

By NVIDIA's own count, Hugging Face is used by more than 18 million developers, researchers and creators, hosts more than 3 million models, 500,000 datasets and 1 million applications, and is used by more than 200,000 companies. Those figures come from the acquirer's announcement and have not been independently audited. The direction, though, is not seriously disputed: for open weights, the Hub is the default distribution point, and a large share of the ecosystem's tooling resolves model names against it at build or run time.

On the financial side, TechCrunch, citing Crunchbase, reports that Hugging Face was founded in 2016 and has raised more than $395 million, most recently a $235 million round in 2023 led by Salesforce Ventures with Google, Amazon, IBM and NVIDIA among its investors. The same report cites The Information putting Hugging Face at roughly $150 million in annualized revenue, and the Financial Times reporting that the company turned down a $500 million approach from NVIDIA last year. Those last two figures are second-hand and we could not open the original reports. If the $150 million number is close, the roughly $11.9 billion purchase price works out to something on the order of 80 times revenue — a multiple that prices position in the ecosystem rather than cash flow.

What NVIDIA committed to, and how firmly

Huang's post is explicit. Hugging Face "will remain an open platform for the entire AI ecosystem," developers will choose their own models, frameworks, clouds and inference providers, and — the sentence that matters most — "NVIDIA compute will not be required to build on or deploy through Hugging Face."

The 8-K repeats the commitment in the more careful register of a securities filing. NVIDIA "has committed to, among other things, keep Hugging Face's platform open, consistent with Hugging Face's existing practices," under which Hugging Face "would continue to permit model makers, developers, and users to upload and download models and datasets of their choosing and to support other silicon vendors."

Both are statements of intent. Neither is, on the public record, an enforceable condition with a named enforcer or a stated duration. The qualifier "consistent with Hugging Face's existing practices" is doing quiet work: it anchors the promise to a baseline that is nowhere defined and that the acquirer will be in a position to interpret. This is worth tracking rather than assuming — and it is the kind of commitment that regulators sometimes convert into binding remedies as a condition of approval.

The risk NVIDIA chose to disclose

The filing adds a risk factor that is unusually candid about what could devalue the purchase. NVIDIA warns that "other parties are actively lobbying" governments to adopt measures that would "restrict or disadvantage open-source models," and that such requirements could "restrict the models or datasets available through Hugging Face, require changes to Hugging Face's platform or practices, delay or restrict offerings, increase compliance costs or result in investigations or enforcement actions."

Many of the world's most popular and successful open-source models originated in China and are then downloaded, revised, fine-tuned, and tested by developers in the United States and worldwide. Any regulatory control or other restriction that limits our ability to provide products and services that support models derived from any region, including China, could have a material impact on Hugging Face's platform.

That is NVIDIA telling its investors, in a document it can be sued over, that the value of the asset depends partly on regulators not restricting Chinese-origin open weights. It is a live policy question, and NVIDIA has just taken a twelve-billion-dollar position on one side of it.

The transaction also has to clear review. The structural question is plain enough to state without guessing at the outcome: the dominant supplier of AI training and inference silicon is buying the accelerator-neutral layer through which open models reach everyone else's silicon. The openness commitments above read as an anticipation of exactly that question.

What changes if you build on the Hub

Today, nothing. There is no API change, no pricing change and no migration to plan. If the transaction closes at all, the filing points to the first half of 2027.

The announcement is still a useful prompt to look at a dependency most teams carry without examining it. If your build or deploy step pulls weights from the Hub by name, you have a single distribution point in your supply chain that is now slated to be owned by a hardware vendor. The mitigations are ordinary engineering hygiene, and they were worth doing before Thursday:

• Pin model revisions to a commit SHA rather than a branch or tag, so an upstream repository change cannot silently alter what you deploy.
• Mirror the weights you depend on in production into storage you control.
• Keep a record of each model's license and its terms as they stood at the version you actually shipped.
• Treat Hub availability as an external dependency in incident planning, not as infrastructure that is simply always there.

For model selection, nothing about which open model is best for a given job changed on Thursday. The thing to watch over the next year is subtler than a policy announcement: whether the Hub's defaults — leaderboard placement, featured inference providers, one-click deployment targets, which quantizations and optimized kernels ship first — begin to correlate with the silicon they run on. That kind of drift shows up in defaults and documentation rather than in press releases, and it is the specific way an openness commitment can be honored in letter while the practical path of least resistance quietly narrows.

What we could not verify

Hugging Face had not published its own announcement on its company blog at the time of writing. Clément Delangue's comments reach this article through TechCrunch's account of his post on X, in which he said open-source AI "needs more compute, more support, more collaboration, and more visibility" and that "that's why we went to talk to Jensen." Interviews he gave to broadcast outlets on Thursday, and several published analyses of the antitrust angle, sit on hosts this newsroom could not reach, so the regulatory discussion above is limited to what the filing itself says. Reporting from August 26 had already put a deal in the range of $13 billion in play, attributed to Business Insider, without a signed agreement at that point.