Nvidia bought the library

Here is the CEO of Hugging Face, explaining why open AI models matter. Open weight models, he said, "contribute to democratising AI, to fighting concentration of power — which is, in my opinion, the biggest risk in AI."

Concentration of power. The biggest risk in AI. His words.

Four days ago he agreed to sell his company to the largest chip manufacturer on earth for $12.93 billion.

This is not a story about a hypocrite. It's a story about a trap that open source walked into with its eyes open, and it's more interesting than the gotcha version. It has a villain that might not be a villain, a betrayal that might have been the right call, and a structural problem that nobody has solved.

Start with what was promised.

The neutrality was the cap table

In 2023, Hugging Face raised $235M at a $4.5B valuation, and the interesting part is who from: Google, Amazon, Intel, AMD, IBM, Qualcomm, Salesforce — and Nvidia. Direct competitors, in a single round, deliberately.

Clément Delangue told Axios it "solidifies our position as a neutral platform, or the Switzerland ... for AI."

He wasn't being poetic. He was describing an ownership structure. If everyone holds a slice, nobody holds you. Neutrality wasn't a value statement, it was a defensive arrangement of the equity.

Then in 2025, Nvidia reportedly offered $500M for a larger stake at a $7B valuation, and Delangue turned it down. The reported reason: one investor shouldn't steer the company.

That's not a rumor about his principles. That's a man declining half a billion dollars to protect a structure.

Nine months later, he sold that investor everything.

What was actually sold

There are over three million AI models you can download right now, for free, and most people have never heard of the site they sit on.

Hugging Face doesn't make a single one of them. Meta trains Llama. Google trains Gemma. Alibaba trains Qwen. OpenAI released Whisper. They all put the finished files in the same public place.

It's a library. It doesn't write the books.

You download one and it runs on your own machine — free, offline, unobserved. That's the entire product.

And it started as a chatbot app for teenagers, in New York, in 2016. That flopped. Then in late 2018 the team rebuilt one of Google's models in about a week and gave their version away, and that week got them more attention than two years of the actual company. In 2019 they killed the chatbot and became infrastructure.

Nobody planned this. Which makes the price look absurd.

Why anyone pays $12.93 billion for file storage

Hugging Face earns an estimated $150M a year. Nvidia paid roughly eighty times that, and nobody pays eighty times revenue for an income statement.

The technology doesn't explain it either. It's object storage with version control on top. A competent team builds it in six months.

Here's the proof: Nvidia already had one. Their own model registry, running for years. It never caught on outside their own ecosystem. They had the product and paid $12.93 billion anyway, because the product was never the hard part.

What can't be built is three million existing files and the companies that upload them. Millions of scripts with the address hardcoded inside. Every tutorial and README pointing at the same door. A decade of accumulated defaults. "Where do I get a model" has exactly one answer, and changing that answer is a marketing problem — one that costs more than $12 billion.

Microsoft learned this the expensive way. They built their own GitHub competitor, killed it, then paid $7.5B for GitHub. You can build the software. You can't build the habit.

The neutral platform helped you not buy Nvidia

Hugging Face doesn't only store files. Its libraries are a large part of what lets you run these models on hardware that isn't Nvidia's — AMD chips, Intel chips, Google's TPUs.

That's what made the neutrality load-bearing rather than decorative. It is now being bought by the company it helped you avoid.

The outrage reading is the lazy one, though, so here's the other side properly. Hugging Face went to Nvidia — Delangue approached Jensen Huang weeks before the deal. Neutral infrastructure still has to eat, and $150M against petabytes of free bandwidth is thinner than it sounds. Switching costs here are unusually low, because weights are just files. And everyone forecast an exodus when Microsoft bought GitHub in 2018; it never came.

The strongest argument is the incentive one. If AI consolidates into three closed labs, Nvidia sells to three customers who hold all the leverage. If thousands of companies self-host, Nvidia sells to thousands. Open models sell graphics cards — Nvidia may have more reason to keep this open than a cloud provider would.

Signed, not closed

Most coverage is getting one thing wrong: Nvidia has not bought Hugging Face. The agreement was filed with the SEC on September 2 — $11.9B to shareholders plus about $1B in retention equity — but it isn't expected to close until the first half of 2027, pending regulatory approval. In 2020 Nvidia signed a $40B deal for Arm. It collapsed in 2022 and never closed.

So how this goes is still open, and there's a cheap public test: watch where the next major open model lands first. If Meta, Google and DeepSeek keep releasing on Hugging Face on day one, the neutrality was structural and survived the sale. If they start hedging, it was always just a promise — and promises have owners now.

Open weights were supposed to stop AI being controlled by a handful of companies, and they worked. Anyone can download a serious model tonight and nobody can stop them. But almost everybody downloads it from the same address.

The models decentralized. The distribution didn't. That's the trap: you can give the product away completely and still end up with a chokepoint, because convenience concentrates even when ownership doesn't.

Delangue was right that concentration of power is the biggest risk in AI. He was watching the wrong layer.

If you build on open weights — does any of this change where you'd pull from? Or is a file just a file?