Nvidia buys the place where open-source AI lives

On 2 September 2026 Nvidia signed the definitive agreement to buy Hugging Face for $12.93 billion. Hugging Face hosts three million models, half a million datasets and eighteen million developers. For years it has been the neutral place where you leave whatever you want to share.

Today it's time to talk about the landlord

Today we're talking about who's bought the warehouse.

I have a soft spot for Hugging Face and I'd rather declare it before giving my opinion. It's one of the few pieces of this industry that has done more good than noise. A place where you upload a model, someone downloads it, tries it, improves it and uploads it again, without anyone asking which company you work for. The comparison always made is with GitHub, and although comparisons age badly, this one works: it was the industry's public square.

And now the public square has an owner. The owner makes the graphics cards that absolutely everything in the square runs on.

I'm going to try not to be unfair, because the easy reflex here is to shout monopoly and leave it at that. There are arguments in favor of the deal and I'll lay them out. But there's also something I can't shake off, and it has to do with what the word «neutral» means when someone owns the ground.

The numbers in the deal

The figures are in a public document, the Form 8-K Nvidia filed with the US Securities and Exchange Commission dated 2 September 2026, so there's no need to trust any rumor.

The total price is around $12.93 billion. Of that, about $11.9 billion goes to Hugging Face's shareholders, subject to the usual adjustments, and up to roughly another $1 billion goes to a stock retention program for the employees who join Nvidia. That second item says a lot: almost a billion so people don't leave. When you buy a platform, you buy the community, and the community stands on specific people who can quit on a Tuesday.

The deal was announced publicly on 3 September, after The Information broke the news of the talks at the end of August. Closing is expected in the first half of 2027, subject to the relevant regulatory approvals, which in a purchase of this profile are no formality.

It's the second-largest acquisition in Nvidia's history. The largest was Groq's assets, around $20 billion at the end of last year. Two huge purchases in less than a year, both in the same direction: less standalone hardware, more control of the road that runs from the chip to the developer.

What it gets for that money: three million hosted models, a million applications, half a million datasets and eighteen million developers using it. Jensen Huang said Hugging Face will keep supporting open and open-weight models, and that work will be done to broaden developers' access.

Why Huang's line is exactly the one that needed saying

And it's probably sincere. That's the problem.

Open source suits Nvidia. It suits it very well, and not out of kindness. Every open model someone downloads, fine-tunes and runs is demand for compute. When the weights are public, the bottleneck moves from the model to the hardware, and the hardware is sold by Nvidia. A vigorous open ecosystem is, from Nvidia's balance sheet, a machine for generating customers for the cards.

So I'm not expecting a textbook betrayal. I don't think they'll turn off the tap, or start charging to download weights, or delete the competition's models. It would be stupid and they aren't stupid.

What I expect is something much softer and much harder to call out. That the model optimized for Nvidia's architecture loads a little faster. That the documented route, the one in the official tutorial, goes through Nvidia's libraries. That running on the house's hardware becomes the default option, and running on anything else stays possible but requires reading a forum thread. Nobody forbids anything. The friction just gets spread unevenly.

That's what a platform does when it stops being neutral: it doesn't close doors, it adjusts slopes. And slopes don't appear in any press release.

The context almost nobody has put next to it

There's a chronological detail that strikes me as the best part of this story and that I've seen very little of in the coverage.

Six weeks before the signing, in July, Hugging Face's internal production network was breached by some OpenAI agents that had escaped their own test environment to get the solutions to a cybersecurity evaluation. I told the story in detail in another article, and the two companies involved have documented it in writing.

Put the two dates on the same line. In July, the infrastructure of open source's public square is stormed by the systems of one of the biggest labs in the world. In September, the public square is bought by the maker of the chips those systems were trained on.

I'm not implying causality, and I want to be explicit because this is exactly the ground where cheap theories get manufactured. There's no sign that one thing caused the other, and the talks went back further. What there is, is a sequence that captures well how this industry works: the problems of shared infrastructure don't get solved by strengthening what's shared, they get solved by changing who owns it.

The whole chain, again

I've written a fair few articles about the physical substrate of all this. About the Dutch company ChatGPT, Claude and the rest depend on, about why OpenAI wouldn't exist without a factory in Taiwan, about what it means that Nvidia is worth more than Spain's economy.

Every time I go over that chain it's one link shorter. Extreme ultraviolet lithography in Veldhoven. Manufacturing in Hsinchu. Design in Santa Clara. And now also the repository where whatever runs on all of the above gets published.

That's not a conspiracy, it's vertical integration, which is the most boring and most predictable thing in industrial capitalism. Rockefeller bought the pipelines, not just the wells. Carnegie bought the coke mines and the railroads. When a company dominates one link and has money, it buys the adjacent links, because that protects its own against surprises. Nothing new under the sun, except that the link just bought was the one that boasted of belonging to nobody.

Where I'm looking from now on

I'm going to note down three signals, so I can be wrong in public in a couple of years.

The first is what happens to the models of Nvidia's direct competitors in silicon. If friction, delays or loss of visibility show up for what runs well on rival architectures, it'll be seen. And if they don't show up, that will have to be said too.

The second is the regulatory approvals. Closing is expected in the first half of 2027 and there are two jurisdictions, Brussels and Washington, with different criteria and changeable moods. A deal that brings together the dominant maker of accelerators and the dominant model repository is classic material for concentration analysis, and it would be odd for it to go through without questions.

The third, and the one I care about most, is how many people leave. There's that almost-a-billion in stock to retain the team, and it shows that at Nvidia they know perfectly well that what they've bought isn't servers. If in two years half the names that sustained that house's culture are somewhere else, they'll have bought a huge, well-lit warehouse with nobody inside.

Meanwhile, the site keeps working the same as last week. That's the strange part of these deals: you don't notice anything. Until you do.

Definitions

Open-weight model: a model whose trained parameters are published and can be freely downloaded and run. It doesn't imply that the training data or code are public, a distinction often overlooked.

Vertical integration: a strategy by which a company acquires links upstream or downstream in its own supply chain, to reduce its dependence on third parties and protect its position.

Form 8-K: a document that companies listed in the United States file with the Securities and Exchange Commission to report material events. It is a primary, public source.

Stock retention program: deferred compensation in the acquirer's shares, conditional on the employee staying for a set period. It's used when the value acquired lies in the people.

References

NVIDIA Corporation, Form 8-K (SEC, 2 September 2026). Primary source for the structure and amount of the deal.

NVIDIA, NVIDIA to Acquire Hugging Face (blogs.nvidia.com, September 2026). Official announcement and statements by Jensen Huang.

TechCrunch, Nvidia confirms it will buy Hugging Face for $12.9 billion (3 September 2026).

CNBC, Nvidia agrees to buy Hugging Face for almost $13 billion (3 September 2026). Platform figures and comparison with previous acquisitions.

The Information, Nvidia Agrees to Buy Open Source Model Repository Hugging Face for $12.9 Billion (August 2026). First report of the talks.

## You might also like - Nvidia is worth more than Spain's economy - ASML, the Dutch company ChatGPT, Claude and the rest depend on - Why wouldn't OpenAI exist without a factory in Taiwan? - DeepSeek and the end of the American monopoly

## Elsewhere - European Commission — Merger control - Federal Trade Commission — Merger Review