Nvidia Is Acquiring Hugging Face for $12.93 Billion: The Open Model Hub Now Has a Hardware Owner

Nvidia announced on September 3, 2026, in a company blog post bylined by CEO Jensen Huang, that it has agreed to acquire Hugging Face for $12,930,300,000. Hugging Face is used by more than 18 million developers, researchers and creators who share more than 3 million models, 500,000 datasets and 1 million applications, and more than 200,000 companies use the platform to discover, evaluate, customize and deploy AI. The announcement states that Hugging Face will remain an open platform for the entire AI ecosystem and that Nvidia compute will not be required to build on or deploy through it, and that founders Clem, Julien and Thomas stay with the team while the brand is retained. ASAP works from Nvidia's official announcement as the primary source to separate what this deal settles from what it leaves open.

The announcement fixes one price and five scale figures

The disclosed purchase price is $12,930,300,000. Nvidia wrote the figure down to the dollar and did not break out cash versus stock. The platform scale presented alongside it comes in five parts: more than 18 million developers, researchers and creators; more than 3 million models; 500,000 datasets; 1 million applications; and more than 200,000 companies.

Nvidia's own footprint on the platform is also quantified. The company states it has released more than 500 models and more than 250 open datasets on Hugging Face and describes itself as the largest contributor of open models and data there. The post adds that Huang recently coauthored an open letter on the importance of open weights to the AI economy.

Financial figures outside the platform do not appear in the announcement. TechCrunch, citing Crunchbase, reported that Hugging Face was founded in 2016 and has raised more than $395 million, with its last round in 2023 at $235 million led by Salesforce Ventures and joined by Google, Amazon, IBM and Nvidia. Citing The Information, TechCrunch put annualized revenue at $150 million. These three figures rest on press reporting rather than Nvidia's announcement, and belong in a different evidentiary tier.

$12.93 billion is roughly 86 times annualized revenue, and that multiple describes the deal

Dividing the two published numbers puts the price at roughly 86 times the reported $150 million in annualized revenue. In software acquisitions, an 86x revenue multiple is not a price paid for cash flow. The asset that can justify it is not revenue but the position that produces it: the distribution path that open models pass through on their way into the world.

Reading the deal as a purchase of that path matches where the announcement puts its emphasis. Huang's stated rationale is not product capability or revenue growth but the size of the community and the importance of open weights to the AI economy. When a hardware company buys a software distribution layer, the return it expects is often not the target's revenue but the durability of demand for its own core business. More widely used open-weight models mean more demand for accelerators to run them.

The same multiple creates pressure in the opposite direction. Justifying 86x requires Hugging Face to keep its neutral position, and the force that sustains that position sits in tension with the ownership structure. Part of why the community trusted this platform was that it was not the asset of any particular accelerator vendor. The higher the price, the stronger the pull to realize the thesis, and the easiest form that realization takes is deeper integration. Deeper integration is also the most common way neutrality erodes.

"Nvidia compute will not be required" is a strong sentence and a narrow one

Nvidia's announcement makes four commitments about Hugging Face, and three of them explicitly guarantee developer choice. Hugging Face remains an open platform for the entire AI ecosystem. Developers choose the models, frameworks, clouds, inference service providers and computing platforms they want. Nvidia compute is not required to build on or deploy through Hugging Face. Multi-cloud and multi-accelerator development and deployment support continues.

All four sentences are prohibitions on explicit coercion. What actually determines a platform's character, though, is not its prohibition list but its defaults. Which model cards surface at the top, what sorting rules search and recommendations use, which infrastructure a new one-click deploy button points at first, which hardware free inference credits run on, which runtime the evaluation and benchmarking tools assume. None of these would violate the sentence about compute not being required, and any of them can move the center of gravity of the ecosystem.

It is worth reading the improvement areas named in the announcement in the same light: platform reliability, safety, model evaluation, inference and deployment capabilities. Inference, deployment and model evaluation are all domains where hardware choice bears directly on results. If Nvidia's infrastructure and engineering go into them, the platform will very likely get better, and the odds also rise that "better" converges on a state achieved best atop one particular stack. Those two outcomes are not mutually exclusive, and they usually arrive together.

An investor in the 2023 round became the owner in 2026

A structurally important detail that gets less attention is that Nvidia is not arriving from outside. TechCrunch lists Nvidia among the participants in the 2023 $235 million round, and Nvidia's own announcement describes the company as the platform's largest contributor of open models and data. An investor and largest contributor has become the owner.

That path makes post-deal change unusually hard to measure. A typical acquisition leaves a baseline against which before and after can be compared, but when the acquirer was already the platform's largest content supplier, the baseline itself already leans toward the acquirer. Five hundred Nvidia models and 250 datasets sitting on the platform was the pre-acquisition normal state, and any future increase will not separate cleanly into a consequence of ownership versus a continuation of an existing trend.

Clem Delangue's explanation to TechCrunch suggests the deal came from a judgment about scale rather than pressure to sell: for the mission to happen at larger scale it needed more compute, more support, more collaboration and more visibility, which is why they went to talk to Jensen. Huang wrote that Clem came to him while considering the company's next chapter. The two accounts interlock into a picture of a deal that was not buyer-driven or hostile. A voluntary sale explains the motive, though, and does not guarantee the character of the outcome.

Open weights had a single-point-of-failure problem, and it just became an ownership problem

The structural weakness of the open-weight ecosystem is that 3 million models, 500,000 datasets and the evaluation results around them sit concentrated in one place called Hugging Face. If that place goes down or changes policy, the whole ecosystem feels it. Until now the risk was discussed mainly as availability and policy: outages, content policy changes and the like.

This acquisition moves the same weakness onto a different axis. The single point now sits on a specific chip company's balance sheet. The practical risk comes less from overt discrimination than from quiet reprioritization: which accelerator's kernels get optimized first, whose runtime bugs get fixed first, which weight format becomes the default export target. Decisions like these are made without announcements, look reasonable one at a time, and reveal a direction only in aggregate.

The practical response is not to demand statements but to keep alternate paths alive. The value of open weights lies less in the weights being public than in their being movable through multiple routes. Mirrored repositories, standard weight formats and deployment scripts that do not assume one hub were good habits before this deal, and are now somewhat more concrete insurance.

What organizations can check today, and what they cannot

Most research groups and companies already use Hugging Face for three things: publishing their own models and datasets, sourcing external models, and consulting evaluations and leaderboards. Nothing in this announcement changes any of the three immediately. The post announces an agreement, and it names neither a closing date nor a regulatory timeline.

The items worth checking now sit on the dependency side. Teams that publish there should confirm that copies of weights and datasets exist internally or in another repository, and that download links are not hardcoded into code and documentation, which is sound housekeeping regardless of this deal. Teams that consume models should know how deeply their build pipelines are tied to Hugging Face API tokens and specific endpoints. For organizations running non-Nvidia accelerators, whether the multi-accelerator commitment continues to hold becomes a matter to watch over the next several quarters.

What cannot be checked today deserves equal clarity. The announcement does not say what corporate form Hugging Face will keep, whether any independent governance mechanism will exist, or how existing terms of service and content policies carry over. Handling of billing, paid plans and enterprise contracts is likewise absent. Making a large migration decision before those items are settled is closer to reacting to unease than to reasoning from evidence.

What to watch next

The three gaps in Nvidia's September 3, 2026 announcement map the watchlist for this deal directly. First is the closing timeline and whether competition authorities review the deal. Neither an expected close nor any regulatory language appears in the post, and given Nvidia's position in data center accelerators and the nature of the target, assuming no review would be premature.

Second is whether neutrality moves from sentence to institution. The test is whether the commitment to an open platform turns into something verifiable, such as terms of service language, a governance document or an external oversight mechanism. Third is the behavior of defaults. Deploy buttons on model pages, recommendation ordering and the scope of free inference are the surfaces the community touches daily, and which stack gets proposed first there will reveal the real direction faster than any statement. All three are publicly observable, which makes them more reliable indicators than announcements.

Source: ASAP analysis based on Nvidia's official blog post "NVIDIA to Acquire Hugging Face" bylined by Jensen Huang (September 3, 2026), with funding and revenue figures from TechCrunch reporting of September 3, 2026