Nvidia buys the front door to open AI as closed labs

What happened

Hugging Face is the go-to hub for open AI models. Nvidia says more than 18 million developers use the platform to share over 3 million models, 500,000 datasets, and 1 million applications. More than 200,000 companies rely on the service. Huang says Hugging Face will stay open to everyone. Nvidia hardware won't be required, and other clouds and chips will still be supported.

In the Nemotron Coalition, it's working with Mistral AI, Thinking Machines Lab, Perplexity, Black Forest Labs, Cursor, LangChain, Reflection AI, and Sarvam on another open model. That doesn't put Nvidia in the lead. It sits behind the strong Chinese models. In return, the Nvidia model runs much faster in certain scenarios. 1 tuned for its chips. The purchase would now add influence over the most important showcase for this competition.

Why it matters

By its own account, Nvidia is the biggest contributor of open models on the platform, with over 500 models and more than 250 datasets. Both sides tell different stories about who approached whom. Huang says co-founder Clem Delangue came to him while thinking about the company's next chapter.

Co-founder Thomas Wolf, on the other hand, writes on LinkedIn that Huang made Delangue the offer to build the hub into an open, independent, and hardware-neutral platform. Wolf calls Nvidia the best-fitting partner for the company's mission, spanning open weights, robotics, and science. Nothing changes for users today, he says. Open AI stands at a turning point where scale and compute matter more and more.

Nvidia makes its money on compute infrastructure, and anyone using the API from OpenAI or another AI provider has mostly ended up on Nvidia chips. But the handful of big providers are now building their own accelerators. Google, Amazon, and OpenAI are doing it, and so, more recently, is Anthropic.

Open models, by contrast, run across many clouds, inside companies, at universities, and in government agencies, a customer base that doesn't build its own chips. Huang boiled the logic down for Axios in a single line. " Nvidia has run its own open models for a while now. With Nemotron, the company releases weights, large parts of the training data, and the recipes for training and post-training.

What to watch

Then there's the hosting side, since Hugging Face distributes models and also rents out the compute time to run them. Nvidia had gone down this road once before and backed out. It scaled back its own DGX Cloud because it didn't want to poach rental customers from its big buyers.

Since then, its Lepton marketplace routes jobs to partners like CoreWeave, Lambda, and Nebius instead, and Hugging Face plugs into it through a cluster service. In late July, Nvidia reported $36 billion in commitments from agreements with AI cloud partners. Under these deals, Nvidia's commitments shrink when the partners sell their capacity to third parties. Nvidia also partly insures cloud partners against unused capacity. A large developer hub as an extra sales channel would come in handy here.