Nvidia is reportedly paying $12.9 billion for Hugging Face to secure the layer where developers choose which models to run — the point that shapes how the open‑weights ecosystem grows and where the next wave of compute demand lands. The hub now hosts 2.96 million public model repositories, but actively is concentrated in a tiny fraction of them: according to the company’s State of Open Models report, 85.6% have been downloaded fewer than 200 times, and just 1.5% of repositories account for 99.2% of all downloads.
The Information reported the agreement this week, citing one person with knowledge of it, and the figure has since been carried by CNBC, Fortune and Forbes. Neither company has confirmed it, and a Business Insider account of the same talks describes a valuation above $13 billion without a signed agreement. Read the price as reported rather than filed.
Against roughly $150 million of annualized revenue, $12.9 billion works out near 86 times sales and close to three times the $4.5 billion Hugging Face carried after its 2023 Series D. Nvidia joined that round, alongside Google, Amazon, Salesforce, IBM, Intel, AMD and Qualcomm. A multiple like that gets paid for a position in a market, and the position here is the point where developers choose what to run.
What The Catalogue Actually Holds
The platform grew fast over the past year. Public model repositories went from 2.43 million to 2.96 million, datasets from 711,000 to 1 million, and hosted demo applications from 1 million to 1.44 million. Every one of those lines is up by a fifth or more in 12 months.
Underneath the growth, the traffic runs through a very small number of families. Derivatives of Alibaba’s Qwen models account for 151,448 repositories on the hub, roughly 2.6 times Meta’s footprint, and they are being added at 180 to 210 new repositories a day. Google’s derivatives come to 82,506. Only about 3% of 2026 downloads went to models above 70 billion parameters, with the bulk of activity landing on much smaller ones. The busiest part of open-source AI runs on a handful of model families at modest sizes, and the largest of those families comes out of China.
Why The Multiple Is Paying For Position
What gets acquired here is the address developers type when they go looking for a model and the surface where the choice between families gets made. The 2.96 million repositories come along as inventory.
That is a familiar shape for this company. CUDA was given away for free and stayed free for close to two decades, until an entire generation of engineers learned to think in it. The barrier that resulted has always lived in the installed base rather than in the difficulty of writing kernels. Software that everyone already uses is worth more to Nvidia than software it can charge for.
Hugging Face earns $150 million a year and is being valued at 86 times that, because the price is attached to where developers arrive rather than what they pay.
Nvidia’s own quarter gives the number some scale. The company reported $96.2 billion of revenue in the three months to July. At $12.9 billion, the reported price is roughly six weeks of sales for a position at the top of the developer funnel.
Why Open Weights Work In Nvidia’s Favor
An open-weight model is a file. It runs wherever the person who downloaded it already has hardware, and for the overwhelming majority of developers that hardware speaks CUDA. Every fine-tune, every quantization, every local experiment on a small model is compute that gets spent on the accelerator already sitting in the machine.
Closed frontier models concentrate demand inside five or six data center operators who negotiate hard on price and build their own silicon on the side. Open models spread the same demand across hundreds of thousands of developers who buy at retail and never negotiate anything. The wider the open ecosystem gets, the more of the market takes the second shape, which is the friendlier one for anyone selling accelerators.
Washington’s export restrictions accelerated that shift. Cutting off frontier model access taught buyers outside the United States that permission can be withdrawn, and the practical answer available to them was open weights they could hold themselves. The Qwen derivative count on Hugging Face is what that answer looks like eighteen months later.
Whether The Hub Stays Neutral
Everything above depends on Hugging Face remaining the place everyone goes.
Its value comes from neutrality. A developer picks a model there because the hub does not appear to be selling anything, and a repository that starts feeling like a vendor channel loses the property that justified the price. Concerns along exactly those lines surfaced within a day of the report.
The structure of Nvidia’s last large transaction is the reasonable case for this one working. The $20 billion Groq deal in December was written as a nonexclusive technology license with a leadership hire attached, leaving Groq standing as an independent company and carving out its cloud business entirely. Nvidia paid a great deal of money to leave something running the way it was already running.
The test here is narrow and it will be visible. Watch whether AMD, Intel and hyperscaler runtimes still work as well on the hub 12 months from now as the CUDA path does. If they do, Nvidia bought a position on a road it is keeping public. If they degrade, developers will move somewhere else, and $12.9 billion will have purchased an archive that nobody opens.
Where The Value Settles
Distribution keeps turning out to be the durable asset in this cycle. Models are being given away in public by labs in three countries, and the companies capturing the economics are the ones sitting on a layer everything has to pass through. Nvidia owns the compute layer and has now reportedly paid up for the layer where the choice of model gets made. Broadcom holds a comparable position in custom accelerator design, and TSMC holds one in fabrication, for the same structural reason.
Free models make the layers around them more valuable, and this cheque puts a number on how much more.
