Nvidia has agreed to acquire Hugging Face, the open-source AI platform, in a deal valued at $12.9 billion, in what would be the chipmaker’s largest acquisition in its history. Hugging Face hosts hundreds of thousands of open-source machine-learning models, datasets, and applications, and has served as the primary distribution platform for independent AI research globally. The deal comes weeks after New York City announced a ban on AI tools in schools through grade 8, which experts say could set a national template for AI regulation. Meanwhile, the UK’s House of Lords has called for AI “kill switch” powers, and Australia’s AI data centre boom is generating fresh questions about energy and land costs. The acquisition, if approved by regulators, would give Nvidia — already dominant in the hardware layer of AI — a commanding position in the software and model distribution layer as well.
The received wisdom
The progressive-technocratic case against this deal is, in many ways, the obvious one, and it is not wrong on its own terms. Nvidia already controls roughly 80 percent of the high-performance GPU market for AI training and inference. Hugging Face is the primary open marketplace through which independent researchers, startups, and academics share and access AI models — the closest thing the field has to a neutral commons. Combining them gives a single company unparalleled leverage over who gets access to compute, who gets distribution for their models, and — critically — who can be squeezed out. The Antitrust Division, under any administration serious about competition, should scrutinise this deal carefully. The argument that Nvidia might preference its own models or hardware in Hugging Face’s recommendations, penalise competitors’ chips in ranking algorithms, or simply charge rent on the infrastructure layer, all of these are textbook vertical foreclosure concerns.
A different read
And yet the instinct to reach for the regulatory lever first — to treat consolidation as the primary problem and structural separation as the primary solution — misses something important about how AI’s economics actually work, and about what kind of regulation is likely to be effective versus performatively reassuring.
Start with the “open” in open-source. Hugging Face’s value proposition has always rested on a paradox: it is an open platform built on a venture-backed company that needed to monetise. Before this deal, Hugging Face was burning through its Series D capital at a pace that raised serious questions about its long-term independence. The choice, in other words, was not between a Nvidia acquisition and a perpetually independent open commons. It was between a Nvidia acquisition and, eventually, a Microsoft acquisition, or a Google acquisition, or a slow collapse into a product division of one of the hyperscalers. At least Nvidia’s core business is chips, not competing AI services — which means its incentive structure is arguably less directly conflicted than Meta’s or Anthropic’s would be as acquirers.
The deeper issue is structural and is not solved by blocking any individual deal. AI development at the frontier — the large models that actually drive capability improvements — requires compute at a scale that only a handful of organisations can access. This is not primarily a consequence of corporate consolidation; it is a consequence of physics and capital markets. Training a frontier model costs tens of millions to hundreds of millions of dollars, requires purpose-built data centres, and depends on supply chains that run through Taiwan and South Korea. No amount of antitrust activity changes those underlying constraints. What antitrust can do is preserve competition at the application and distribution layer — ensuring that startups can access models without paying rent to an integrated gatekeeper.
That goal is achievable, but it requires carefully targeted intervention, not blanket opposition to scale. The key question is not whether Nvidia owns Hugging Face, but whether Hugging Face’s open-access policies survive the acquisition intact — whether the platform continues to allow hosting of models that run on AMD or Google TPUs, continues to give equal ranking to models regardless of the hardware they were trained on, and continues to publish its governance policies publicly. The UK Lords’ call for an AI “kill switch” is a vivid phrase but a conceptually muddled policy; the more productive question is what behavioural remedies could be attached to this acquisition as conditions.
There is a broader political economy point here. The progressive instinct — exemplified by the NYC schools AI ban and the Lords kill-switch proposal — is to treat AI as primarily a risk to be managed rather than a capability to be shaped. That instinct produces policies of exclusion and prohibition, which tend to benefit incumbents (who can absorb compliance costs) over challengers (who cannot). The right response to the Nvidia-Hugging Face deal is not a ban but a set of conditions: open licensing, non-discriminatory access, and published audits. Regulation that specifies ends rather than means is almost always more effective than regulation that specifies structure.
What to watch
- FTC and DOJ review timelines: whether the Biden-era merger-review machinery, now under different political management, treats this as a presumptively anticompetitive vertical integration or a manageable deal with conditions.
- Hugging Face’s governance changes after closing: any modifications to its open-access policies or model-hosting terms would be the earliest warning sign.
- Nvidia’s share price reaction to the deal: if markets respond positively, it signals that investors expect the acquisition to generate pricing power — which is the same thing regulators should be worried about.
- European Commission action: Brussels has been more aggressive on AI competition than Washington; an EU block or conditional approval could effectively set global terms.
— J