Nvidia announced on September 3, 2026 that it will acquire Hugging Face for $12.93 billion, disclosed in an SEC 8-K filing, with the deal expected to close in 2027. Jensen Huang's announcement was explicit about the terms that matter most to the open-model community: Hugging Face will "remain an open platform for the entire AI ecosystem," developers can pick "the clouds and inference service providers they want and the computing platforms they want," and "Nvidia compute will not be required to build on or deploy through Hugging Face" (Nvidia blog). Hugging Face today hosts more than 18 million developers, 3 million models, 500,000 datasets and 1 million applications, and is used by over 200,000 companies to discover and deploy AI (TechCrunch). That scale is precisely why this deal deserves scrutiny that goes beyond a corporate blog post.
The Steelman: Chokepoint Control Is a Real Antitrust Concern
The strongest case against this deal is straightforward and shouldn't be waved away. Hugging Face is not just a popular website — it is close to the default discovery and distribution layer for open-weight AI, the place where a startup, a university lab, or a government ministry goes to find and deploy a model without training one from scratch. Nvidia is simultaneously the dominant supplier of the GPUs those models run on. Owning both the chip layer and the primary distribution hub for the software that increasingly matters most gives one company visibility into, and potential leverage over, which models get discovered, how they're benchmarked, and which hardware backends get first-class support. Even without an overt act of favoritism, subtle defaults — better documentation for Nvidia's own inference stack, faster support for its formats, a benchmarking layout that quietly advantages CUDA — could tilt outcomes without anyone writing an anticompetitive clause into a contract. Regulators who lived through Nvidia's abandoned 2022 bid for Arm are right to ask whether this rhymes.
It does rhyme, procedurally. The Federal Trade Commission sued in December 2021 to block Nvidia's $40 billion acquisition of Arm, arguing the combined firm would control a critical input — Arm's chip designs — that Nvidia's rivals depend on, and would have "the ability and incentive" to degrade that input for competitive gain; the agency called Arm the industry's neutral "Switzerland" and said the deal would undermine that neutrality (FTC press release). Nvidia walked away from that deal in February 2022 rather than fight the case. Hugging Face is arguably a softer, less concentrated chokepoint than Arm's ISA — model weights aren't proprietary IP that Nvidia could withhold, and switching hubs is technically easier than re-architecting a chip design — but "neutral infrastructure everyone depends on shouldn't be owned by one dependent party" is the same structural logic, and it's worth taking seriously rather than dismissing because the target this time is a website rather than a semiconductor IP license.
Why the Deal Should Still Clear — With Teeth, Not Just Trust
The disanalogy with Arm matters more than the similarity, though. Arm's architecture was a hard, licensed, technical dependency — you cannot build an ARM-compatible chip without Arm's IP, full stop. Hugging Face is a marketplace and hosting layer for weights that are, by definition, downloadable and mirrorable; nothing stops a competitor cloud, a university consortium, or a rival chipmaker from replicating a Hugging Face-like index, and several already exist (Kaggle, GitHub, Ollama's registry, cloud-native model gardens from AWS and Google). The genuine risk isn't that Nvidia can lock the industry out of open models — it's that it could make Hugging Face subtly worse as neutral ground while everyone assumes it's still neutral. That's a real risk, but it's one addressable through enforceable conditions, not one that requires blocking a deal that also brings Hugging Face real capital: the company has never been reported as profitable at anything like this valuation, and Nvidia's balance sheet can fund the infrastructure scaling that keeps a free hub free for millions of users who never pay for it.
This is where regulators, not Nvidia's press release, need to do the actual work. The deal triggers Hart-Scott-Rodino review in the US and a formal merger filing in the EU, giving the FTC, DOJ and European Commission a real window to act before close in 2027 — and Nvidia's own Run:ai acquisition already drew EU competition scrutiny over GPU-market dynamics, so the precedent for asking hard questions here is fresh, not hypothetical. The right posture is proportionate: don't refile the Arm case, but don't accept a blog post as a substitute for a consent decree either. Regulators should extract binding, auditable commitments — model-agnostic search ranking, continued support for non-Nvidia inference backends, published benchmarking methodology, and a firewall between Hugging Face's product roadmap and Nvidia's commercial incentives — with real enforcement teeth if Nvidia backslides after the deal closes and public attention moves on.
The Open-Weight Backdrop
The timing matters because governments are simultaneously deciding how much they want open models to matter. The EU AI Act's Article 53 already gives providers of free and open-source general-purpose models a partial carve-out from technical documentation and downstream-provider disclosure obligations — though that exemption evaporates the moment a model is deemed to carry "systemic risk" (EU AI Act Service Desk). Huang's own framing — that open weights "broaden access to AI and help ensure that AI leadership is distributed across companies, institutions and communities" — lines up with that policy logic. Regulators who agree with the goal of distributed AI leadership have every reason to keep this hub genuinely neutral; that argues for binding conditions on this specific deal, not a blanket presumption against vertical integration in AI infrastructure. Nvidia's commercial interest and the public interest in open-model access happen to point the same direction here — but "happen to" is not "guaranteed to," and that gap is exactly what merger review exists to close.