Nvidia Buying Hugging Face Is Reported, Not Signed, and Not Confirmed

1 SEP 2026 — Nvidia is reported to have agreed to buy Hugging Face for US$12.9bn, against roughly US$150m of annual revenue and a US$4.5bn valuation in 2023. Neither company has confirmed anything, and no agreement has been signed. What is being repeated as a completed deal is a single sourced report matched by other outlets.

What has been reported

The Information reported on 27 August that Nvidia had agreed to acquire Hugging Face for US$12.9bn, citing a person with direct knowledge of the talks. CNBC, TechCrunch and Fortune matched it. The reports describe an agreement reached in negotiation rather than a signed and announced transaction, and note the deal could still fall through.

Neither company has commented. That silence is itself notable in Nvidia's case, because the company usually responds quickly to reports about it.

Hugging Face was founded in 2016 and operates the hub where developers publish and download open models. Recent annual revenue is around US$150m, up from about US$100m two months earlier. Its last known valuation was US$4.5bn in a 2023 round.

The earlier rejection is the interesting part

In late 2025 Nvidia offered US$500m for a stake at a US$7bn valuation. Hugging Face turned it down, and the reported reason was concern about the influence a dominant investor would hold over a neutral platform.

Nine months later the same company is reported to be selling itself outright to the same buyer. Whatever the concern about influence was, full ownership does not address it — it settles it in the opposite direction.

Price is the likely explanation for the change of heart. US$12.9bn against a US$7bn valuation is a different conversation from US$500m for a minority position, and a board that believed the platform's independence was worth protecting at one number may not have believed it was worth protecting at every number.

What is being bought is not revenue

At US$12.9bn against US$150m, the multiple is around 86 times revenue, which no ordinary software business supports. The purchase is not of an income statement.

Hugging Face is the distribution point for open models. A model that does not come from a hyperscaler's own API is, in practice, downloaded from there, along with the datasets and the tooling around it. That position is close to a public utility for the part of the field that is not vertically integrated.

Nvidia's interest is clear. Every one of those models eventually runs on hardware, and a buyer whose competitors are building proprietary accelerators has a reason to own the place where models are chosen. The stated strategic case adds a route back into cloud services after DGX Cloud was scaled back, and a use for computing capacity already committed.

Revenue that grew 50 per cent in two months

The price is not the most interesting figure in the reporting. Annual revenue is described as roughly US$150m, up from about US$100m two months earlier.

A 50 per cent increase in an annualised figure over two months is not organic growth in a subscription business. It is what happens when a company signs a small number of large enterprise agreements, or when the basis of the figure changes — from recognised revenue to annual recurring revenue, or from one product line to all of them.

Which reading applies matters to the price. If the trajectory is real and continuing, an 86-times multiple on the trailing figure is a much smaller multiple on next year's. If the jump reflects a handful of contracts or a redefinition, it does not compound. Neither company has published financials, so this is not resolvable from outside, and any valuation commentary that treats US$150m as a stable base is guessing.

The neutrality question is real and not simple

An acquisition is not obviously bad for the ecosystem. A well-capitalised owner can fund the bandwidth and storage a repository this size consumes without a business model to match, and Nvidia has an interest in open models remaining viable, since they drive hardware sales.

The objection is structural rather than about intent. A hub owned by a chip vendor has an owner with preferences about which runtimes, quantisation formats and optimisation paths are prominent, and those preferences do not require anyone to act in bad faith to shape defaults over time.

Regulatory scrutiny is the usual answer and it may not apply cleanly here. The two companies do not compete, so a conventional horizontal analysis finds little, and the harm being described is about a chokepoint in an ecosystem rather than a price rise in a defined market.

Why this matters in this region

Open-weight models are the practical route to AI capability for organisations here that will not send data to a foreign API and cannot afford to train anything themselves. That route runs through this hub for teams in Jakarta, Ho Chi Minh City and Kuala Lumpur in a way it does not for a company with a hyperscaler contract.

A change of ownership does not close it, and the sensible response is not alarm but reducing single points of dependency: mirroring the weights you actually rely on, recording the exact revisions, and knowing whether your pipeline would still build if one hostname stopped resolving.

We reported this morning on the breach of Hugging Face by a coordinated set of AI agents. Two stories about the same platform in one day shows how much of the field depends on it, which is the single fact driving both.