Nvidia agreed to buy Hugging Face for $12.9 billion

Nvidia said on 3 September 2026 that it has agreed to acquire Hugging Face for $12.93 billion. The deal was announced by Nvidia CEO Jensen Huang on the company's blog. Hugging Face is the platform many software teams use to download open-weight models, datasets, and inference code.

What the announcement says

Nvidia describes Hugging Face at the time of sale as 18 million developers, researchers, and creators, more than 3 million hosted models, more than 500,000 datasets, more than 1 million applications, and more than 200,000 companies using the platform. Those are Nvidia's own figures.

The commitments Nvidia is making in writing are the important part for anyone already building on the platform. Nvidia says Hugging Face will keep its brand, remain an open platform for the whole AI industry, keep supporting open-source and open-weight models, and preserve multi-cloud and multi-accelerator development and deployment. In its own words, "NVIDIA compute will not be required to build or deploy through the platform."

Nvidia also lists what it has already contributed to Hugging Face as a user rather than an owner: 500 open models and 250 open datasets. It calls itself the largest such contributor.

What is still open

The deal close date, the cash and stock split, and the regulatory clearances the transaction still needs are all absent from the announcement. So is any description of how Hugging Face's leadership will report inside Nvidia after close, and any pricing commitment for the paid Inference Endpoints, Spaces, or Enterprise Hub tiers.

The neutrality promise is the sentence to remember

Hugging Face has become the default place teams pull open models and datasets from, and its neutrality was part of why. A single company owning that platform is a bigger change in what teams depend on than a new price on any one model. Two things to watch, starting now.

First, the multi-cloud and multi-accelerator commitment is the most important sentence. Teams running open models on AMD, Intel, Google, or AWS chips are the group that would notice first if inference tooling or default pipelines slowly started to favour CUDA over the next year. Save a link to the announcement, since a promise is easier to check when the exact wording is on hand.

Second, the deal has not closed and has not been cleared. Nothing about how you use Hugging Face today changes because of an intent to acquire. If a team has been building on the assumption that Hugging Face is independent, that assumption will stop being true, but there is no date for the change yet.

For teams pulling weights and datasets, the practical step is the same one that was already good practice: keep model files and dataset artefacts you depend on copied to your own storage or fixed to a specific hash, so a policy change at the platform is a business problem rather than a build problem.

Source

Primary source: Nvidia to acquire Hugging Face. Also covered by Wired and TechCrunch.