Nvidia has confirmed it is buying Hugging Face for $12.9 billion, taking ownership of the registry where three million models, one million applications and 500,000 datasets sit in front of 18 million developers. The Information reported last month that Hugging Face had reached $150 million in annualised revenue, which puts the price at roughly 86 times revenue. Last year, according to the Financial Times, the company turned down a $500 million deal with Nvidia. Whatever changed in the interval, it was worth about 26 times the earlier offer.
Jensen Huang's commitments in Nvidia's announcement are the ones you would expect. Hugging Face will go on supporting open source and open models and will widen developer access to the platform. It stays an open venue for the whole AI ecosystem, where developers choose their own models, software frameworks, cloud platforms, inference providers and compute systems. Nvidia hardware will not be required to use it.
Read that last promise precisely. It is a statement about requirement, not about default. A registry's power has never been in what it forbids; it is in what sits at the top of the page, what the quickstart snippet imports, and which inference provider is already wired up when a developer clicks deploy. A non-requirement commitment touches none of that.
Earlier TechCrunch reporting sharpens the point: the deal gives Nvidia a way to sell spare compute capacity to enterprise customers as part of the Hugging Face offering. "Developers choose their own inference provider" and "the owner of the platform sells inference through it" are both true, and they are in competition.
The purchase fits a pattern Nvidia has been building openly. It has already published more than 500 models and 250 open datasets on Hugging Face. It told a recent investor meeting it has put more than $50 billion into frontier AI labs. Last month the Wall Street Journal reported a $6 billion agreement with Poolside, a young coding company, to develop open models jointly. Huang has signed a letter alongside several other organisations arguing for open-weight development as a way to strengthen the US position against China.
So Nvidia sells the hardware, funds the labs, ships the models, and now owns the shelf. Asked about open models by an analyst, Huang made the connection himself: almost all of them run on Nvidia hardware. The policy case for open weights and the demand case for Nvidia's chips are the same case, and he has not pretended otherwise.
Here is what I think Nvidia actually bought. Not revenue — $150 million of run-rate does not support $12.9 billion on any multiple in ordinary use. Not technology either; a model registry is not the hard part. It bought habit. Eighteen million developers with a default place to publish and a default place to download is the one asset Nvidia's balance sheet could not manufacture, because it is made of other people's routines. The $500 million refusal a year ago looks, from here, like Hugging Face pricing that asset correctly and Nvidia eventually agreeing.
Clem Delangue's framing of the sale is where it gets uncomfortable. He thanked the community on X for helping prove that an alternative to closed APIs was possible, and said that scaling that model required more compute, support, collaboration and visibility — which is why Hugging Face went to Huang, who offered help on all of it. That is a coherent account. It is also an account in which the demonstration of independence from closed platforms ends with acquisition by the company whose chips nearly all of them run on.
The security sequence belongs in how this deal is read. OpenAI acknowledged that one of its unreleased models had breached Hugging Face; days later, Delangue said an open Nvidia model had helped the platform fend off cyberattacks after closed models failed to protect it. Huang has been making the matching argument to investors: frontier open models are what let companies build distributed, continuously running autonomous defences, work he says could not exist without them, and those models have reached frontier level while becoming important to the US and world economies. The platform breached by a closed lab's model and defended by Nvidia's open one now belongs to Nvidia.
Hugging Face has raised more than $395 million since it started in 2016, the most recent $235 million in a round led by Salesforce Ventures with Google, Amazon, IBM and Nvidia taking part, and Delangue told TechCrunch its growth rate was bringing it closer to profitability. On the numbers in the public record, this was not a company selling to survive. By its own chief executive's account, it sold to get compute. The open alternative to closed APIs concluded that staying open required a patron, and chose the one that makes the chips.