Nvidia has agreed to acquire Hugging Face for about $12.9 billion, according to The Information, in a deal that would put the default distribution point for open-source AI models under the company that sells the hardware those models run on. The reporting has been near-unanimous on the strategic logic: Nvidia protects its position by owning the front door to open models, because anyone downloading one still has to rent or buy GPUs to run it.
That reading is correct and incomplete. What almost nobody has put at the center of the story is that Hugging Face already turned Nvidia down, on the record, for exactly this reason. Late last year Nvidia offered a $500 million investment at a $7 billion valuation, and Hugging Face rejected it because, in reporting from the Financial Times since recirculated by TechCrunch, it did not want a dominant investor who could sway its decisions. Nine months later the answer to that concern is selling Nvidia the entire company at roughly 84% more. The independence that was worth declining half a billion dollars to protect turns out to have been a negotiating position.
What $12.9 Billion Buys
Hugging Face is genuinely the center of gravity for open AI development. The platform hosts more than three million models and over a million datasets, with a user base in the millions of developers. The company was recently generating about $150 million a year in revenue, up from roughly $100 million two months earlier, per The Information’s reporting.
Set the price against that: $12.9 billion for $150 million of revenue is about 86 times sales. Nvidia is not paying for the revenue line. It is paying for four things:
- The default place developers go to find, download and publish open models
- Telemetry on what the open-source ecosystem is actually adopting, ahead of the market
- The Transformers library and the tooling layer that sits under a large share of applied AI work
- The ability to make sure the reference path from model to silicon runs through Nvidia
The last item is the one that matters and the one the price reflects. Owning the registry means owning the defaults, and defaults in developer tooling are extraordinarily durable.
Neutrality Was the Product
Hugging Face’s position in the ecosystem was built on not being anybody’s. Meta, Mistral, Alibaba, Google and a long tail of academic labs all publish there, and they publish there because it is the one place that is not owned by a competitor. Developers trust the download counts and the model cards for the same reason.
A registry owned by the dominant supplier of AI compute is a different institution, whatever the operating commitments say at signing. It does not require bad faith to change. It requires only ordinary product decisions accumulating over years: which runtimes get first-class support, which quantization formats are documented, which hardware the benchmarks default to, which integrations ship on day one. None of that is a scandal. All of it moves the ecosystem toward one vendor.
That is the actual risk, and it is worth being direct: the thing Nvidia is buying is partly destroyed by Nvidia buying it. AMD and Intel have spent years trying to make their accelerators a credible second source for open models. Their reference path now runs through a competitor’s property.
Whether It Clears
The deal is not signed. Reporting from TechCrunch and others is explicit that talks had not produced a signed agreement and could still fall apart, and CNBC’s account carries the same caveat. Treat the price as reported rather than final.
Regulatory risk is real but probably overstated by the Arm comparison. Nvidia abandoned its $40 billion bid for Arm in 2022 after the Federal Trade Commission sued, and the reflex is to assume the same fate here. The cases are not alike. Arm was a horizontal problem: Nvidia buying the instruction-set architecture its own competitors licensed. Hugging Face is a software distribution layer with about $150 million of revenue and no direct overlap with Nvidia’s chip business, which makes the traditional market-definition arguments harder to run. The theory that fits is foreclosure of a neutral chokepoint, which is precisely the theory current enforcement has been least willing to press. On this year’s record, including a Justice Department that has been conspicuously accommodating on large deals, we would expect this to clear.
The Pattern Is the Story
This is the second time in two weeks that an AI infrastructure company has bought a neutral layer rather than a model. BTN covered Stripe’s $7 billion acquisition of OpenRouter on the same logic: the routing layer, the meter, the place where the decision gets made, is worth more than any individual model that gets routed to. Nvidia is running the same play one level up the stack, at the discovery layer instead of the inference layer.
Meanwhile the open-source commons that made both targets valuable keeps getting narrower. Meta walked back its open-weight commitment earlier this year, and the neutral hub is now on its way to a chip vendor’s balance sheet.
Our position: this is a good deal for Nvidia shareholders, a very good outcome for Hugging Face’s investors, and a bad one for everybody who assumed the open-model ecosystem had a permanent neutral center. It never did. It had a well-run startup with a strong culture and a cap table that eventually needed an exit, which is a different thing that looked identical for about six years. The lesson worth taking is not that Hugging Face sold out. It is that critical shared infrastructure held inside a venture-backed company is always for sale eventually, and the price was $12.9 billion.
Watch what the large open-model publishers do next. If Mistral, Alibaba and the academic labs keep publishing to a Nvidia-owned registry without hedging, Nvidia bought the whole thing. If a credible neutral mirror appears within a year, it bought an expensive brand and a very good tooling team.