4 SEP 2026 — Nvidia has confirmed it will buy Hugging Face for $12.93 billion, discharging the caution we raised on 1 September when the deal was reported rather than signed. Clément Delangue says he went to Jensen Huang, a year after his company turned down a $500 million offer from the same buyer.
What was confirmed
The price is $12.93 billion. Reports of the announcement add up to $1 billion in employee retention on top of it. Closing is expected in the first half of 2027 and needs regulatory approval, so the better part of a year passes before anything changes hands.
Hugging Face was founded in 2016 and had raised more than $395 million, most recently a $235 million round in 2023 led by Salesforce Ventures in which Nvidia itself participated alongside Google, Amazon and IBM. The platform hosts three million models, half a million datasets and a million applications, used by more than 18 million developers. Reported annualised revenue is about $150 million.
The public commitment from Huang is specific. He says Hugging Face "will remain an open platform for the entire AI ecosystem", and that "Nvidia compute will not be required to build on or deploy through Hugging Face".
The seller made the call
A chip company buying the open-model repository looks like the chip company went shopping. Delangue tells it the other way: Hugging Face wanted a partner who would keep the platform open and neutral, so "we went to talk to Jensen, who offered to do exactly that with us".
If Nvidia initiated, this looks like a chip company seeing strategic value in owning distribution. If Hugging Face initiated, the platform had decided independence was less sustainable than ownership by a partner with an interest in the ecosystem staying broad. Those are different conclusions about the same transaction.
Neither account can be verified from outside, and a founder explaining a sale has obvious reasons to call it a choice rather than a necessity. The $500 million rejection last year still fits Delangue's telling. A company that says no to one price and yes to another has usually changed its mind about the money, not the principle.
Twenty-six times in twelve months
Do the arithmetic on that rejection. Nvidia offered $500 million and was refused; twelve months later it agreed to $12.93 billion, close to twenty-six times as much for the same asset.
The agreed price is around 86 times Hugging Face's reported revenue of roughly $150 million. That multiple pays for strategic position, not cash flow, and it values the repository as infrastructure.
Compare that with what a model lab costs. Nvidia is paying about $13 billion for the place models are distributed from, while the labs producing those models are valued in the tens of billions each. On that scale the distribution layer is cheap. Either Nvidia got a bargain or the labs are expensive.
Where the openness commitment actually sits
The sentence the open-source community will hold on to is "Nvidia compute will not be required to build on or deploy through Hugging Face". Its status is a public commitment from a chief executive, not an enforceable condition of the merger.
The mechanisms that would make it durable are not described. There is nothing on governance after close, no statement about whether the platform keeps a separate board, and no answer on who chooses default configurations or how competing accelerators are treated in documentation, tooling and continuous integration.
None of that requires bad faith to drift. A platform owned by an accelerator vendor will find its defaults tested against that vendor's hardware first, because that is the hardware its new colleagues have. The question for regulators between now and the first half of 2027 is whether a commitment stated in a press release is the right instrument for that risk.
What this says about the moat
Nvidia's advantage has never rested only on silicon. We wrote in September that licensing NVLink to rivals extends the fabric rather than surrendering it, and this purchase is the same manoeuvre one layer up.
Owning the repository does not stop anyone shipping weights for another accelerator. It does put Nvidia at the point where 18 million developers decide what to try first, which is where defaults, examples and the path of least resistance live.
Set against that, the platform's value to Nvidia depends on it staying the neutral place everyone publishes to. A repository visibly tilted towards one vendor would be mirrored elsewhere, degrading the asset in the act of exploiting it. That tension is a better guarantee of openness than the press-release commitment.
What to watch before close
Three public signals will show which way this is going, and none of them requires access to anyone's boardroom.
A competition authority may open a review, since a dominant accelerator vendor is buying the distribution layer for the software that runs on accelerators. Developers may start hedging, which would show up as publishing volume rising at alternative repositories. And the platform's own tooling either keeps accepting contributions that make non-Nvidia hardware a first-class target, or it quietly stops.
Nothing about the first half of 2027 is settled today, and the caution we published on 1 September has been answered on the narrow question it asked. The wider one, whether an open platform stays open under an interested owner, has not been asked yet.
View author profile → · Editorial policy
