Nvidia is reported to be paying $12.9 billion for Hugging Face, the site where developers go to decide which model to run. Neither company has confirmed it. If the figure holds, Nvidia is buying the layer that sits between free model weights and the hardware that executes them, and the hardware belonging to the overwhelming majority of developers speaks CUDA. The choice made on that page is what steers where the next wave of compute demand lands.
The sourcing deserves a sentence of its own. The Information reported the agreement this week, citing a person familiar with the negotiations. CNBC, Fortune and Forbes then carried the same number. Business Insider described the same talks and put a possible value on the platform above $13 billion, while noting that nothing has been signed. So $12.9 billion is a figure that has been published, not a price that has been formalised. Four outlets repeating one number is still one number.
At roughly $150 million in annual revenue, $12.9 billion works out to almost 86 times revenue. It is also nearly three times the $4.5 billion valuation Hugging Face carried after its Series D in 2023 — a round Nvidia joined alongside Google, Amazon, Salesforce, IBM, Intel, AMD and Qualcomm. That cap table reads like an industry consortium paying to keep a neutral layer neutral. The reported price reads like one member of that consortium deciding neutrality was worth owning outright.
Nvidia's own accounts put the sum in proportion. For the three months ended in July the company reported revenue of $96.2 billion. On that run rate, $12.9 billion is about six weeks of sales, spent on a position at the top of the developer funnel.
The catalogue behind that position is thinner than its headline number. Hugging Face hosts 2.96 million public model repositories, but the company's own State of Open Models report says 85.6% of them have been downloaded fewer than 200 times, and that 1.5% of repositories account for 99.2% of all downloads. Nvidia is not buying three million repositories. It is buying the default address that a very small set of files is fetched from, plus everything that never gets opened.
Growth is real on every axis the platform counts. Public model repositories rose from 2.43 million to 2.96 million over the past year, datasets from 711,000 to 1 million, hosted demo apps from 1 million to 1.44 million. Each of those grew by at least a fifth in twelve months.
Where the traffic actually goes is more interesting than how fast the catalogue grows. Derivatives of Alibaba's Qwen models occupy 151,448 repositories on the platform, roughly 2.6 times the number derived from Meta's models, and between 180 and 210 new ones are added every day. Google's models have 82,506 derivatives. Only about 3% of downloads in 2026 went to models above 70 billion parameters. The open ecosystem is not a long tail of frontier-scale weights; it is a handful of moderately sized families, and the largest of them came out of China.
For Nvidia the underlying logic is two decades old. CUDA was given away free until a generation of engineers had learned to work in it. The barrier that resulted has little to do with how hard kernels are to write and everything to do with an installed base. Software that everyone already uses is worth more to Nvidia than software it can invoice for.
An open model is a file. It runs on whatever hardware the person downloading it already owns. Every fine-tune, every quantisation, every local experiment with a small model consumes cycles on an accelerator that is already installed. Closed frontier models concentrate that demand in five or six data centre operators who negotiate hard on price and design their own silicon in parallel. Open models scatter the same demand across hundreds of thousands of developers who buy compute at retail and negotiate nothing. The wider the open ecosystem gets, the more the market takes the second shape — the one that suits a seller of accelerators.
Washington's export controls accelerated the shift. Restricting access to frontier models demonstrated to buyers outside the United States that permission to use a technology can be withdrawn. The practical answer was open weights that can be stored and run without asking anyone. The Qwen derivative count is what that answer looks like eighteen months later.
All of which depends on Hugging Face remaining the place every developer goes. The platform's value is inseparable from its neutrality: a developer picks a model there precisely because the site does not look like it is selling one. The moment the repository starts to feel like a vendor channel, it loses the property that justifies the price. Concerns of exactly that kind surfaced within a day of the report. My read is that Nvidia knows this, and that the shape of the Groq deal is the tell.
That deal, signed in December for $20 billion, was structured as a non-exclusive technology licence with an executive moving to Nvidia. Groq kept its status as an independent company and its cloud business was left entirely outside the transaction. Nvidia paid a very large sum and left the acquired thing operating more or less as it had before. What the current reporting does not say is whether Hugging Face gets the same treatment — whether this is an acquisition at all, or another licence with a door left open. That is the quiet part, and it determines everything about what $12.9 billion actually buys.
The test will be narrow and easy to see. Twelve months from now, compare how well AMD, Intel and the large cloud providers' environments work on the platform against the CUDA path. If they hold up, Nvidia has bought a position on a road that stays open. If they do not, developers move to another site and $12.9 billion has bought an archive nobody opens.
This cycle keeps returning to the same conclusion: distribution is the durable asset. Labs in three countries publish their weights for free, and the economics accrue to whoever owns the layer everyone else passes through. Nvidia owns the compute layer and has now, reportedly, paid for the place where the model gets chosen. Broadcom holds the comparable position in custom accelerators, TSMC in manufacturing. The reason is the same in each case: control of the layer in the middle.
The awkward part is that the layer is worth $12.9 billion only for as long as it belongs to nobody in particular. Nvidia's problem, from the day the deal closes, is that it belongs to someone.