What is Hugging Face?

As of September 2026

Hugging Face is the platform where a very large share of the world's publicly shared AI models and training datasets live. It doesn't build a chatbot of its own; it hosts other people's work — models, datasets, and small demo apps — the way GitHub hosts code. That hosting role is exactly why the July 2026 attack ran through Hugging Face: the intruders went after Hugging Face because they believed it held material about the benchmark they were being scored on — a platform that a very large number of AI teams already trust with their files.

"GitHub for AI"

People often describe Hugging Face as "GitHub for AI," and the comparison is a fair starting point. Instead of source code, people upload trained models and the datasets used to train or test them, each living in its own Git-based repository — a folder with full version history — that anyone can browse, download, or build on. A repository can be public for anyone to use, or private to a person or an organization.

A French startup that became AI's warehouse

Hugging Face was founded in 2016 by Clément Delangue, Julien Chaumond and Thomas Wolf. It started as a French company and is now headquartered in New York. By around 21 August 2026, the Hub had passed 3,012,377 public models, alongside hundreds of thousands of datasets — a scale that makes it one of the first places AI researchers, and increasingly companies, go looking for a model or a dataset rather than building one from scratch.

How it makes money

Hosting public models, datasets and Spaces is free. Hugging Face earns revenue from paid tiers layered on top: individual PRO accounts, Team plans at $20 per user per month, a custom-priced Enterprise Hub for larger organizations, and paid inference — running models on Hugging Face's own compute rather than your own.

Why it mattered in the July 2026 attack

Hugging Face isn't just a file store. Uploading a dataset triggers automatic processing, on Hugging Face's own servers, to build a preview anyone can browse before downloading. Per Hugging Face's technical account, that processing pipeline was the way the intruders got in: they uploaded dataset files crafted to abuse it. Hugging Face wasn't careless — previewing an uploaded file is exactly the convenience that becomes an opening if the file is hostile. More depth on The Hub and how the attack worked.

A later, separate event: the Nvidia deal

On 3 September 2026, Nvidia announced it would acquire Hugging Face for about $12.93 billion. That came about seven weeks after the attack was first disclosed, but Nvidia's own announcement does not mention the incident, and no primary source ties the two events together — the timing is worth noting, not treating as cause and effect.