Nvidia buys Hugging Face for $13 billion and promises openness

Thirteen billion, the biggest cheque Nvidia has ever written

On Thursday 3 September 2026, Nvidia announced it had reached a deal to buy Hugging Face, the platform where open artificial intelligence models are traded, for $13 billion. The story comes from Ars Technica, relaying the Financial Times: that is the only source available at this hour, so treat the numbers accordingly.

It is by far the largest acquisition the chipmaker has ever signed, well above the $6.9 billion it paid for Mellanox in 2020 — the deal that opened the datacenter infrastructure door for it. Nvidia, the most valuable company in the world at $5.4 trillion, hopes to close the transaction by 2027. Competition authorities are expected to look closely at the file; Justin Boitano, Nvidia’s vice president for enterprise AI, says he expects them to see it overwhelmingly as good news. Bold forecast. We’ll see how it ages.

Last year, the answer was no

Hugging Face is ten years old, headquartered in New York, named after the hugging face emoji, and lives off premium subscriptions and enterprise services. The platform hosts 3 million models, mostly open, around 500,000 datasets and a million applications; 18 million developers and more than 200,000 companies use it.

And yet last year it turned down a significant Nvidia investment at a $7 billion valuation, precisely to preserve its independence — a $500 million cheque refused late in the year, in part to avoid ending up with a dominant shareholder, according to the Financial Times. You will be relieved to hear that the dominant shareholder problem has now been solved in full.

Open models sell graphics cards

An open-weights model is one whose design is public: you can download it, adapt it and run it on your own machines. Jensen Huang argues that this openness lets start-ups, companies, universities and public institutions build on advanced capabilities without retraining a model from scratch or paying frontier-model rates for every task.

The reasoning holds up, and it happens to point straight at the till. The more open models exist, the more chip buyers exist, and the less Nvidia depends on a handful of very large customers — OpenAI and Anthropic above all, both of which are also building their own processors. The chipmaker had already put money behind open-model players such as Reflection AI, and behind the neoclouds CoreWeave and Nebius. In July, Huang co-signed a letter urging the United States to support open models, against the backdrop of an American debate over possible sanctions on Chinese open-model developers. Openness, here, is not a moral position. It is a distribution channel.

The promise of openness, and what actually backs it

Nvidia says Hugging Face will remain a platform open to the whole ecosystem, that its 18 million developers will keep choosing their models, cloud providers and chips freely, and that the brand will be kept. Nothing in what has leaked describes the mechanism that would make any of that binding: it is a statement of intent, worth exactly as much as you decide to trust it. And the deal hands the manufacturer control of one of the main channels deciding which AI applications spread around the world.

A shop doesn’t have to lock its doors to change what sits on the shelf at eye level.