Nvidia Buys Hugging Face for $12.9bn and Asks the Open-Source World to Trust It
The world's dominant AI chip supplier now owns the default hub for open models, and its neutrality promises will be tested long before regulators finish their review.
Nvidia has agreed to buy Hugging Face for about $12.9bn, in a deal signed on 2 September and announced the following day. According to the deal terms, roughly $11.9bn goes to Hugging Face shareholders and up to $1bn is set aside as retention equity for staff joining Nvidia. The transaction is expected to close in the first half of 2027, subject to regulatory approval. Chief executive Clem Delangue and his co-founders stay on, and Hugging Face is to run as an independent platform inside Nvidia.
What Nvidia is buying
Hugging Face is where open AI lives. More than 18 million developers, researchers and creators use it, sharing millions of models, around 500,000 datasets and about a million applications. It is the default place a team goes to pull weights, and its libraries sit inside an enormous number of production pipelines. Its revenue, reported at around $150m a year, is tiny next to the price. Nvidia is not paying for the income statement; it is paying for the distribution point of the open-model ecosystem.
Jensen Huang's pitch is that this is an investment in openness. He has committed that users will keep choosing their own models, frameworks, clouds and hardware, and that Nvidia compute will not be required to build on or deploy through the platform. Nvidia already publishes hundreds of its own models and datasets there, so the relationship is not new. What is new is ownership.
Why the promises matter less than the incentives
Nobody should expect Nvidia to block AMD or Google TPUs on day one. That would be crude and would invite exactly the regulatory scrutiny the deal must survive. The more realistic risk is subtler: defaults. Which inference backends are featured first, which quantisation formats get first-class tooling, which hardware gets optimised kernels in the libraries, and which "deploy" button appears on a model card. None of those choices breaks a neutrality pledge. All of them shape where workloads land.
That is why the competition review is worth watching closely. A vertically integrated chip supplier that also controls the most important distribution channel for open models is a textbook case for behavioural commitments, and authorities in the US and Europe have spent the past two years learning to write those, as the Google cases show. Expect requests for binding interoperability and non-discrimination terms rather than an outright block.
What builders should do
- Mirror what you depend on. If your deployment pulls weights from the Hub at build or start-up time, cache them in storage you control. That was good practice before; it is better practice now.
- Watch the libraries, not the website. Changes to defaults in the core Python libraries will tell you more about direction than any blog post.
- Keep alternatives warm. Other registries and self-hosted model stores are less convenient, but having one configured costs little.
The open-model community grew up around a hub with no hardware agenda. It is about to find out whether one can stay neutral when the owner sells the hardware most of it runs on.
Sources