Nvidia (NVDA) Buys Hugging Face for $12.9B, Under the $13B Floor Hugging Face Floated Three Days Earlier
The Information reported Wednesday that Nvidia has agreed to acquire Hugging Face for $12.9 billion, citing a person with knowledge of the agreement. Reuters carried it within the hour. Neither company commented, and no terms were disclosed: no cash-stock mix, no close timing, no retention package, no governance commitments around the platform itself. Everything that determines whether this works is still unreported.
The price is the disclosure that matters. On Saturday, Business Insider reported that Hugging Face had retained a bank to gauge buyer interest at a valuation above $13 billion. On Monday, The Information reported that the company’s annualized revenue had crossed $150 million, up 50% in two months. Those two leaks, in that order, are what a seller’s advisor does when it wants a competitive process. Three days later the asset cleared at $12.9 billion. Below the floated number. No auction premium, no second bidder, no drawn-out process. Nvidia has sat on the cap table since the 2023 Series D, which means it had the relationship, the diligence and the information advantage to preempt before a process could form. It used all three.
The multiple tells the same story from the other direction. At $12.9 billion against $150 million of annualized revenue, Nvidia is paying roughly 86 times. That sounds like a bubble print until you compare it to what Nvidia’s own money underwrote three years ago. The 2023 round valued Hugging Face at $4.5 billion when annualized revenue was reportedly near $30 million, which is about 150 times. Revenue has since compounded roughly fivefold and the multiple has come down by more than 40%, with control included this time. Whatever else is inflating in AI, this particular asset repriced downward on a revenue basis.
That gap between commercial scale and strategic importance is the entire transaction. Hugging Face hosts the better part of a million models and serves tens of thousands of organizations, and it converts something like 3 to 5% of users into paid storage, compute and enterprise tiers. Around 2,000 paying customers. Clément Delangue said in July the company was close to profitability and had barely spent the money it raised in 2023. Being the default place the world fetches open weights from produces $150 million a year. A financial buyer cannot underwrite the difference, because the difference is not a revenue stream. It is a position in someone else’s supply chain. Only the party that owns the rest of that supply chain can pay for it, which is why one bidder was enough.
For the stock, $12.9 billion is noise, and the timing makes that explicit. Nvidia printed $96.2 billion of revenue for the July quarter the same afternoon, up 106% year over year, with Data Center at $89.0 billion and up 117%. It guided the October quarter to $108 billion assuming zero Data Center compute revenue from China. It returned roughly $26 billion to shareholders during the quarter, twice the purchase price. Against $279 billion of supply commitments, the deal is under 5%. Nothing in the model moves. What actually moved the tape was gross margin guided to about 74% from 75.0%, and total debt at $33.4 billion against $8.5 billion at the January year end. Shares closed the regular session at $209.66, down 1.59%, and traded near $220 after hours, still short of the May record above $236 at roughly a $5 trillion market cap. The acquisition was a rounding error announced into a print that had the market’s full attention, which is a reasonable way to bury it.
What Nvidia bought cannot be fabbed. CUDA holds the compute layer through switching costs measured in engineer-years. The hub holds the layer above it, where a model gets discovered, benchmarked, quantized and copied into production. Whoever controls that surface controls which backend is the tested path, which kernel gets the reference config, which quantization format ships in the model card, and which accelerator a developer’s first attempt happens to run well on. Defaults do more competitive work than lock-in because nobody has to be persuaded of anything. Buying the funnel that feeds your own installed base is the highest-leverage $12.9 billion Nvidia could have spent, and it is the one thing a fab cannot produce.
The obvious risk is that the moat is social rather than technical. A git-backed repository of portable weights is one of the easier things in this industry to fork. Its value comes entirely from the fact that everyone agreed to meet there, and that agreement rested on the platform belonging to nobody in particular. AMD, Intel, Qualcomm, Google, Amazon, IBM and Salesforce all participated in the same 2023 round, and all of them ship optimized paths through the hub today. They do not need proof of preferential treatment to react, only a plausible belief in it. Migration away from a hub is cheap compared with migration away from a toolchain.
Then there is the regulatory clock. $12.9 billion clears the reporting thresholds comfortably, and a dominant supplier of AI accelerators acquiring the distribution and discovery layer for the models that run on them is a textbook vertical foreclosure theory in both Washington and Brussels. A second request should be the base case. That is unlikely to block the deal and very likely to slow it, and a long pendency with the platform in limbo does more damage to a community asset than to a hardware business.
The part nobody will price this week is robotics. Hugging Face bought Pollen Robotics, ships hardware from a $100 printable arm up to a $70,000 humanoid, and maintains LeRobot. Robotics datasets on the hub went from roughly 1,100 to roughly 27,000 in a single year. Nvidia has the simulation stack and the foundation models for physical AI and no canonical public repository for the demonstration data those models need. In language, Nvidia is buying the incumbent hub. In embodied AI, it is buying the one that has not been established yet, at a price set entirely by the language business.
The number to watch is not in a filing. It is in the commit history. Every kernel, backend and integration contributed by AMD, Intel, Qualcomm and the independents is public and timestamped, and the rate at which those contributions keep landing after close is the only honest read on whether the developers who make this asset worth $12.9 billion accept the new owner. If that rate holds through the first two quarters of Nvidia control, the price will look cheap. If it thins, Nvidia will have bought a very expensive snapshot of a community that has already begun assembling somewhere else.