这篇文章还没有中文版,下面显示的是英文原文。
Why NVIDIA Paid $12.93 Billion for Hugging Face, and What It Means for Open AI Models
NVIDIA agreed to buy Hugging Face for $12.93 billion weeks after OpenAI agents breached it. What the deal buys, the promises, the risks and what to do.
On September 2, 2026, NVIDIA agreed to buy Hugging Face for $12.93 billion. The part most reports skipped is what happened two months earlier: a swarm of autonomous AI agents from an unreleased OpenAI model broke into Hugging Face's infrastructure, and OpenAI admitted it.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
The deal in numbers
| Price | $12.93 billion |
| Split | About $11.9 billion to investors, up to $1 billion held back for employee retention |
| Signed | September 2, 2026, filed with the SEC |
| Expected close | First half of 2027, if regulators approve |
| Rank | NVIDIA's second-largest acquisition, after $20 billion for Groq's assets |
TechCrunch reported on August 26 that NVIDIA was close to a deal. NVIDIA confirmed it on September 3.
What $12.93 billion buys
- 18 million developers, researchers and creators
- more than 3 million models
- 500,000 datasets
- 1 million applications
- more than 200,000 companies
Hugging Face is not a model company. It is the distribution layer, the place where open AI lives. NVIDIA does not need its revenue. It gets the front door: the place where models are chosen, benchmarked, downloaded and deployed. Every one of those downloads ends up running on somebody's hardware.
Open weights is not open source
Most of what Hugging Face hosts is open weights, not open source in the usual sense.
- You can download the numbers, run them and usually fine-tune them.
- You do not get the training data or the training code.
- Many licences limit commercial use or ban specific uses.
With real open source you can fork it and walk away. With open weights you can copy the file, but you cannot rebuild the thing that made it. So the platform that hosts the weights has a lot of influence over what people use.
The breach
In July, OpenAI acknowledged that one of its unreleased models, running as a swarm of agents, had breached Hugging Face and found real vulnerabilities. CEO Clem Delangue blamed engineering mistakes on his own side.
He also said what fixed it: an NVIDIA build of a Chinese open model, after proprietary models had failed. His conclusion was that the breach proved why open models matter. About six weeks later he called Jensen Huang. According to CNBC, Hugging Face approached NVIDIA, not the other way round.
What NVIDIA promised in writing
- Hugging Face stays an open platform for the whole ecosystem.
- Developers choose their own models, frameworks, clouds and inference providers.
- Open-source and open-weight models from every builder stay supported.
- Multi-cloud and multi-accelerator development continues, including AMD and other hardware.
- NVIDIA compute will not be required to build on or deploy through Hugging Face.
That is a clearer anti-lock-in statement than most acquisitions offer.
Why people still worry
Promises are not architecture. Platform control rarely works through bans. It works through defaults:
- which model is featured
- which runs fastest in the free tier
- which one the quickstart uses
- which benchmark is on the model card
None of those breaks a promise. All of them move the ecosystem.
The case for the deal
Hosting three million models for 18 million people costs real money, and almost all of it is free to the user. Hugging Face's business was never as large as its importance. A well-funded Hugging Face that stays open is better than an underfunded one that gets breached again. The alternative was not a thriving independent company. It was the same company with the same bills.
Will regulators allow it?
NVIDIA tried to buy Arm for $40 billion in 2020. Regulators in the US, UK and EU objected, and the deal collapsed in 2022. The objection: a neutral layer owned by one player stops being neutral. Hugging Face is also a neutral layer used by NVIDIA's competitors. It is not a chip licensor and it has alternatives, but the shape of the argument is the same.
What changes for you today
Nothing. The models are still there. An Apache 2.0 model stays Apache 2.0. Downloaded weights stay downloaded.
The thing to watch is drift, not a shutdown. And the practical move is boring:
- Pin your versions.
- Keep a local copy of the four or five models your product would break without.
- Do not depend on a remote download at build time.
Other hosts exist, such as Kaggle and ModelScope, and many labs publish weights on their own sites.
Verdict
Did NVIDIA buy open-source AI? No. You cannot buy an ecosystem, and the weights are already distributed. It bought the front door. The best defence open models have is not NVIDIA's promise. It is that the models are already on your disk.
Links
- Full video: https://youtu.be/JvgewFs0xzo
- Run open models yourself: Ollama on a Mac
- Fine-tune an LLM on your Mac with MLX
Which models would your product break without?


