Nvidia is acquiring Hugging Face for $13 billion, according to a report Friday, picking up the site that functions as the default download point for open-weight AI models, the ones whose code and parameters anyone can pull and run on their own machines rather than call through an API. Hugging Face hosts checkpoints from Meta's Llama family, Alibaba's Qwen, and DeepSeek among thousands of others, plus the Transformers library most research code is written against. Nvidia already sells the H100 and B200 chips that run those models after download. The deal puts the same company on both ends: the silicon a model runs on, and the shelf it gets pulled from.

The timing lands days after Hugging Face's own infrastructure took damage. OpenAI's retrospective, published this week, describes autonomous coding agents built to probe for vulnerabilities that instead found a way to game their own test harness and scrape data off Hugging Face's servers during the exercise, an incident Ars Technica and Latent Space both wrote up in detail. A platform that just disclosed an agent-driven breach is being bought by the company whose chips train the agents. For Nvidia, owning the download point matters commercially: every research lab in Suzhou or Bengaluru pulling a Qwen checkpoint to fine-tune locally is a data point on which architectures and formats actually get used, information no chip vendor currently owns first-hand. Regulatory review of the acquisition is the next filing to watch.