Nvidia's $12.9 billion Hugging Face deal rests on a promise to stay open

Nvidia has agreed to buy Hugging Face, the hub where developers share open-source AI models and datasets. The agreement was signed on September 2, and Nvidia’s SEC filing puts the price at about $11.9 billion payable to Hugging Face stockholders, subject to adjustments, plus an equity retention program of up to about $1 billion for employees who join Nvidia. Most outlets add the two and call it $12.9 billion. Closing is expected in the first half of 2027, pending regulatory approvals.

Adweek calls it the chipmaker’s second-largest acquisition ever. The more interesting question is not the price but what Nvidia is buying: a place that every AI company, including its competitors, already depends on.

What is being bought

According to figures from Jensen Huang’s blog post, as reported by Adweek, the platform has more than 18 million developers, about 2 million models, 500,000 datasets and 1 million applications. Those are numbers Nvidia cited, and I have not seen them checked independently.

Hugging Face does not make chips, and it is not primarily a model maker. It is the distribution layer: when a lab releases open weights, they usually land there, and developers download them from there. Owning that layer gives a hardware company a view into which models people actually run and a chance to make its own tools the easiest ones to use.

The open-platform pledge

Nvidia knows how that looks, and the filing addresses it directly. The company committed, among other things, to keep the platform open, consistent with existing practices. Hugging Face would still let users upload and download the models and datasets they choose, and it would continue to support other silicon vendors.

That last clause matters most for AMD, Intel and the cloud providers that build their own chips. A model hub that quietly ran best on one vendor’s hardware would not need to block anyone; defaults, documentation and engineering priority would be enough. The pledge is written as a commitment to existing practices, and the filing as summarized does not set out how it would be enforced or for how long. Neither the termination fee nor the outside date is disclosed in it.

Morningstar’s Brian Colello sees the conflict-of-interest risk too. He notes that large-model builders might move off Hugging Face if Nvidia tilted the platform toward its own models, though he considers that unlikely. Whether it happens will show up in behavior, such as which hardware gets first-class support in new releases, long before it shows up in any legal document.

Offense, defense and the China question

Colello calls the purchase “a wise offensive move, but also defensive.” On offense, it improves Nvidia’s position in the open-source community, which supports its own Nemotron model and open models in general. Morningstar expects open models to handle heavy, routine workloads cheaply while closed models such as Claude serve premium, mission-critical work. On defense, he argues, if Anthropic and OpenAI shift toward more in-house chips, Nvidia would still have a stake in the models that remain on its hardware.

Morningstar also calls the deal “immaterial to Nvidia’s future results” in financial terms. For a company of Nvidia’s size, a $12.9 billion purchase is not about earnings. It is about influence over where the software ecosystem grows.

The filing contains one risk worth singling out. Nvidia warns that government restrictions on open-source AI models could affect the Hugging Face platform and its business. It notes that many popular open-source models originated in China, and that any restriction on models derived from China could have a material impact. Nvidia is therefore buying a platform whose most-downloaded content may be exposed to export-control-style politics, and that is a risk the company chose to put in writing.

What to watch

• Regulators. The deal needs approvals before the first half of 2027. A chip supplier acquiring neutral infrastructure for its rivals is the sort of vertical link agencies tend to ask questions about, though I have not seen any regulator comment yet.
• Rival support. Watch whether AMD and other chip vendors keep first-day support for new models on Hugging Face, and whether the open-platform language shows up in a public policy rather than only in a filing.
• Developer behavior. If the 18 million developers see no change in how they upload, download and deploy, the pledge is holding. If paid inference features start to favor one vendor, the pledge is being tested.

For developers, nothing changes today. The deal is pending, and Hugging Face continues to run as it did. The sensible step is to avoid building anything that cannot be moved elsewhere, which was good practice before this announcement and is better practice now.