NVIDIA buys Hugging Face for ~$13bn, and the open robot stack with it

NVIDIA announced on September 3, 2026 that it will acquire Hugging Face. NVIDIA’s own post puts the consideration at approximately $12.93 billion; press coverage has reported it as almost $13 billion. Hugging Face chief executive Clement Delangue told CNBC his company approached Jensen Huang about an acquisition a few weeks before the announcement — a detail Huang echoed in writing that he was “honored that Clem came to me as he considered the next chapter.”

Read as an AI deal, this is a distribution buy: NVIDIA-stated figures put Hugging Face at more than 3 million models, 500,000 datasets and 18 million users. Read as a physical-AI deal, it is narrower and more consequential. Hugging Face is where the open robotics stack lives.

What robot builders actually get from the hub

A small robotics team without a data-collection budget starts in one of two places: a simulator, or someone else’s checkpoint. Hugging Face is the second. LeRobot bundles the training loops, the dataset format and the pretrained policies that let a lab with two arms and a Jetson board reproduce something in a week rather than a quarter. The community datasets are the closest thing embodied AI has to a shared corpus, and the model cards are how a builder decides which policy to fine-tune before buying hardware around it.

That is a genuine dependency, and it is not symmetric. If the hub changes what is easy to find, what is rate-limited, or what is recommended by default, the effect lands on the small teams first. They are the ones with no fallback. A well-funded humanoid developer can host its own weights, negotiate its own data licences and keep training whatever the hub does; a university lab or a four-person startup cannot.

Two layers, one owner

NVIDIA already sells the accelerators, the simulation stack and Isaac GR00T, its own open VLA line for humanoids. Since July 6, 2026 it has also shipped GR00T and Isaac TeleOp directly into LeRobot workflows — a collaboration announced months before the acquisition. That partnership was uncontroversial while the two sides were separate. Under one owner the same integration reads differently: the company with a model to promote now runs the shelf its competitors sit on.

Nothing in the announcement suggests NVIDIA intends to close the platform, and the commercial logic points the other way. A hub that stops carrying rival checkpoints stops being the place developers go, which destroys the asset that justified the price. Hugging Face’s value to NVIDIA is the breadth of what sits on it, and breadth is the first thing favouritism costs you. NVIDIA also gains little from restricting open robotics models when every one of them is trained on its GPUs regardless of who published it.

A promise is not a structure

That is the honest counter-case, and it is strong. The gap is that open access at Hugging Face is currently a property of the company’s incentives, and after close it becomes a property of NVIDIA’s. Huang’s statement commits to open-weight support and multi-accelerator deployment; it does not commit to a foundation, an independent board seat, a neutrality covenant, or any mechanism a robot builder could point to in three years. Until one is published, developers are relying on a pledge and on the acquirer continuing to find neutrality profitable.

The layer-above-the-robot fight is already the most contested part of this sector — see our coverage of the race to own the layer above the machine. This deal does not resolve it. It moves the referee onto one of the teams.