SANTA CLARA, 29 AUG 2026 — Reports that Nvidia has agreed to buy Hugging Face for about US$12.9bn have been run as settled news by several outlets. The deal is not settled.
Neither company has confirmed the acquisition of the largest hub for open-weight AI models, and the distinction matters for anyone deciding what to do next.
What has actually been reported
The original reporting came from The Information and was picked up by CNBC, Forbes, Fortune and Bloomberg. CNBC cited a source familiar with the matter confirming that acquisition has featured in ongoing and recent talks, which is a narrower claim than an agreed deal. Neither Nvidia nor Hugging Face responded to requests for comment.
The figure varies by outlet. Most reporting puts the purchase at US$12.9bn, with the company valued above US$13bn overall, and headlines have rounded to whichever number reads better.
That is the entire evidentiary basis for the story. The Information has a strong record on deals of this kind, so the report is credible. It remains a report rather than an announced transaction.
The scale, if it is real
This would be by a wide margin the largest acquisition Nvidia has ever completed. Mellanox, bought in 2020, cost US$6.9bn. The only larger attempt was the US$40bn bid for Arm, which Nvidia abandoned in 2022 after regulators in three jurisdictions made clear it would not clear.
The relevant precedent is not about price but about position. The Arm bid collapsed because Nvidia proposed to buy a supplier its own competitors depended on. With Hugging Face, the same question arises one layer up the stack.
Nvidia is already a shareholder
These talks did not begin with a cold approach, a fact largely absent from the coverage. Nvidia was an investor in Hugging Face's August 2023 Series D, a US$235m round led by Salesforce Ventures that also included Google, Amazon and IBM and valued the company at US$4.5bn.
An existing investor negotiating to buy the whole company is a different situation from a cold approach. Nvidia has had board-level visibility into the business for three years, and the reported US$12.9bn price represents roughly a tripling of the US$4.5bn valuation it paid into.
It also complicates the neutrality question in a direction worth noticing. Google and Amazon are on the same cap table, and both compete with Nvidia in AI silicon. Whatever they think of the reported deal, they are not disinterested observers of it.
What US$12.9bn buys
Hugging Face's last publicly reported revenue figure was around US$50m annualised, in 2023. The company has not disclosed a current number and it has certainly grown, but no plausible growth rate makes US$12.9bn a conventional revenue multiple.
The price is not based on revenue. It is for Hugging Face's position as the place developers go first: the default namespace for model weights, the source of the download counts that function as a popularity ranking, and the home of libraries that a large share of AI code imports by habit. Distribution of that kind is not something a competitor builds by spending the same money.
That strategic value is exactly what a competition authority would examine. The justification for the price is also the basis for the regulatory concern.
Why the neutrality question is not rhetorical
Hugging Face is where open-weight models are published, benchmarked and downloaded, and a great many of them are trained or served on silicon that is not Nvidia's. AMD, Intel, Google's TPUs, Chinese accelerators and a growing set of inference startups all depend on that hub being indifferent to what the model runs on.
Nothing about ownership forces that indifference to end. The models are hosted under their own licences, the libraries are open source, and a hub that visibly favoured one vendor would lose the community that makes it valuable. Nvidia has every commercial reason to keep it neutral.
What changes under new ownership is who decides the defaults. The suggested runtime, the quantisation formats with first-class support, the kernels in the reference path, the construction of the leaderboards, the top of a model card — none of these requires a deliberate decision to disadvantage anyone. They accumulate.
This is the same structural point we made when AWS acquired DuckDB's maintainers. The licence protects the code. It does not govern the roadmap, and it has nothing to say about which path is the easy one.
Nvidia has recently avoided doing exactly this
The most interesting context is a deal Nvidia structured six days earlier to be something other than an acquisition.
Nvidia paid around US$7bn for Poolside's technology and 109 staff without buying Poolside. That structure is not accidental. A licence-and-hire avoids merger review, avoids assuming liabilities and avoids the optics of a dominant supplier absorbing a company in its own ecosystem.
If Nvidia has agreed to buy Hugging Face outright, it has chosen the opposite path for a much larger sum, in a market where the antitrust argument is more obvious rather than less. Either the company has concluded the review risk is manageable, or the reported deal is less settled than the headlines suggest. Both readings are available, and the second is the one the absence of confirmation supports.
What to watch
The first signal to watch for is confirmation. A deal this size requires a filing from a US-listed acquirer, and until one appears it exists only in reporting.
The second is the regulatory posture. Nvidia's position in AI accelerators is stronger now than its position in GPUs was when Arm was blocked, and buying the distribution layer for models that run on rival hardware is a legible theory of harm rather than a speculative one.
The third is what Hugging Face's largest non-Nvidia users do. If AMD, Intel or the Chinese labs begin mirroring weights elsewhere or investing in an alternative hub, that will say more about the perceived consequences than any commitment made at announcement.
Until a filing appears, what is known is that reputable outlets report a deal, and the two companies involved are silent.
View author profile → · Editorial policy
