Nvidia reportedly buys Hugging Face, the open-model hub, for $12.9B

First reported by The Information, echoed by CNBC, Bloomberg and Fortune; neither company confirms. What it means for open weights, non-Nvidia chips and the Hub.

Nowline AUG 29 3:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The deal: $12.9B for ~$150M in revenue

    The Information reports Nvidia has agreed to buy Hugging Face for about $12.9 billion — roughly an 86x multiple on ~$150M in annual revenue. CNBC, Bloomberg and Fortune corroborate; Business Insider says no agreement is signed yet, and neither company has commented.

  • It already said no once

    Hugging Face turned down a $500M Nvidia investment in 2025 that valued it at $7B, reportedly to avoid one dominant backer. Eighteen months later the whole company is on the table at nearly double that price.

  • Why builders should care: the neutral hub gets an owner

    Hugging Face is where you pull open weights, datasets and Spaces regardless of chip, and its Optimum stack treats Nvidia, AMD, Intel and AWS equally. A GPU vendor owning that layer puts the hardware-agnostic default in question.

  • What stays open — for now

    transformers and diffusers stay Apache/MIT-licensed and forkable, and today's model and dataset downloads keep working. The moat Nvidia is buying is the community, the workflow lock-in, and paid Inference Endpoints that already run on its GPUs.

  • The risk if you're not on Nvidia

    Nvidia's NIM microservices already plug into HF repos; as owner it could make CUDA-optimized containers the default deploy path and relegate AMD, Intel and AWS to second-class — and those vendors may pull back their integration work.

  • What to do this week

    Nothing breaks today. But mirror the models and datasets your pipeline depends on and pin their versions, then watch whether Inference Endpoints and NIM defaults start steering you toward Nvidia-only paths.

  • Elsewhere: OpenAI's Codex CLI hits 0.150.0

    The Codex terminal added @-mentions for task references, response pickers on /copy, auto-titling for unnamed tasks, and Interrupt hooks that fire on command execution — handy if you script the agent.