Nvidia has agreed to acquire Hugging Face, the hub most open AI models and datasets are downloaded from, for $12.9 billion.
Nvidia has agreed to buy Hugging Face for $12,930,300,000. The company that sells you the GPU will now also own the warehouse most open AI models are downloaded from.
Jensen Huang announced the deal on 3 September in a post on Nvidia's blog. The accompanying 8-K filing puts roughly $11.9 billion in cash to Hugging Face stockholders, plus an equity retention programme worth up to about $1 billion for staff moving across. It is expected to close in the first half of 2027, subject to regulatory approval.
Hugging Face is not a household name outside machine learning, but the figures Nvidia quotes explain the price. More than 18 million developers, researchers and creators use it to share over 3 million models, 500,000 datasets and a million applications, and 200,000 companies use it to evaluate and deploy AI. Founded in New York in 2016 by Clement Delangue, Julien Chaumond and Thomas Wolf, it became the default place open weights get published.
Huang's post leads with reassurance about neutrality. Hugging Face stays open to every model builder, every framework and every cloud, and — the line the rest of the hardware industry was waiting for — "NVIDIA compute will not be required to build on or deploy through Hugging Face." Multi-accelerator support continues, which in practice means AMD and Intel silicon keep working. Both companies are Hugging Face investors, as is Amazon, GamesIndustry.biz reports. Nvidia also notes it is already the platform's largest open contributor, with over 500 models and 250 datasets published there.
For a small studio this is not abstract. Hugging Face is plumbing. Speech-to-text for placeholder dialogue, upscalers, a segmentation model behind a texture tool, a local language model driving an editor plugin — a great deal of that resolves to a hub download on first run, often silently, often unpinned. A build machine that fetches whatever is current from a repository someone else now owns is a dependency nobody wrote down.
The useful response is not to rip anything out. It is to pin. Record the exact model revision each tool pulls, mirror the weights you actually ship next to your own content, and read the licence on each one while the ownership is still the ownership you agreed to. Acquisitions rarely break a download URL on day one. They do change hosting policy, rate limits and terms, and this one has eighteen months to run before it even closes.