Nvidia and Hugging Face: What the Report Says
Nvidia has reportedly agreed to buy Hugging Face for around 12.9 billion dollars, and neither company has confirmed it. What the reporting actually establishes, what it does not, and three cheap steps for anyone whose production stack depends on the hub.
The Nvidia and Hugging Face acquisition reported this week is not confirmed by either company. The Information reported on 26 August 2026 that Nvidia had agreed to buy the model hub for around 12.9 billion dollars, and CNBC, SiliconANGLE and TechCrunch picked it up the following day. No signed agreement has been reported, and at least one outlet notes the talks could still fall apart. The caveat matters, because if the deal does close, the most widely used distribution point for open-weight models would be owned by the company that sells the hardware those models run on. Here is what is known, what is not, and what a builder who pulls weights from that hub every week should do today.
What has been reported and by whom
The chain starts with The Information, which reported the agreement on the evening of 26 August. CNBC and SiliconANGLE both wrote it up on 27 August, each attributing the figure to that original report rather than to a company statement. TechCrunch framed it as talks closing in rather than a done deal.
What has not happened: no press release from either company, no regulatory filing described in the coverage, no comment from Nvidia or Hugging Face to any of the outlets that asked. Treat every number in this story as one publication's sourcing until that changes.
Why the number is plausible
Hugging Face has never been valuable because of its revenue. It is valuable because it sits at a chokepoint. When a lab releases open weights, the release lands there. When a developer wants those weights, that is where they go. The Transformers library, the datasets hub and the model cards became the default not through a product decision anyone made but through accumulation.
Nvidia sells the compute those models consume. Every developer who successfully runs a 27B model locally is a developer who eventually wants more memory. Owning the place where that decision starts is worth more than the hosting business it comes with. We covered a related move this month when Nvidia opened talks with Rebellions over inference chips, and the pattern is consistent: buy position in the layer above the silicon.
The honest risk, and the honest counterweight
The obvious worry is neutrality. A hub owned by a chip vendor has an incentive to make some hardware paths feel smoother than others, whether that is first-class support for one runtime or which optimised builds get promoted on a model page.
The counterweight is that the assets are hard to capture. Weights are files under permissive licences, and the Transformers library is Apache 2.0 and forkable by anyone with a weekend. The community's willingness to stay is the actual asset, and it is the one thing an acquirer cannot buy outright.
What to actually do this week
Nothing dramatic. Three things that were good practice before this story and are merely more obviously good practice now.
Action | Effort | Why it matters |
|---|---|---|
Pin exact model revisions, not branch names | Minutes | A revision hash is reproducible whoever owns the hub. `main` is not. |
Mirror the weights you depend on in production | An hour, once | Removes a single third party from your critical path. Storage is cheap compared to an outage. |
Record the licence text alongside each mirrored model | Minutes per model | Ownership changes do not alter existing licences, but your evidence that you complied should not live on someone else's server. |
If you are choosing between open weights and a hosted API in the first place, the tradeoffs have not moved: our comparison of open-weight and closed AI models still describes the decision accurately, and running a model locally carries the same risks it did last week.
How to read this story if it moves again
Rumour-stage acquisitions have predictable failure modes. Deals leak before they sign, get repriced, and sometimes evaporate. Separate three questions the coverage blends together: is the deal real, is the price accurate, and does it change anything for you. Today the honest answers are probably, unconfirmed, and not yet. Our guide to telling AI hype from a real shift in your workflow has the longer version of that filter.
FAQ
Has Nvidia confirmed it is buying Hugging Face?
No. As of 28 August 2026 the reporting traces back to The Information, and neither Nvidia nor Hugging Face has commented publicly to the outlets that covered it.
What is the reported price?
Around 12.9 billion dollars, sometimes rounded to nearly 13 billion. That figure comes from the original report, not from either company.
Would my existing model downloads be affected?
Not by the transaction itself. Weights already downloaded are files on your disk, and the licences attached to them do not change because a company changes hands. The risk sits in future availability and future defaults, which is why mirroring matters.
Should I move off Hugging Face?
Not on this news. Reduce your dependency instead of replacing it: pin revisions, mirror what production needs, and keep licence records locally. Those steps are cheap and they pay off regardless of how the deal turns out.
Could the deal still fall through?
Yes. TechCrunch's reporting described talks that had not produced a signed agreement, and unsigned deals at this size are routinely repriced or abandoned.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
