The Justice Department has started asking Nvidia questions about a takeover the chip maker announced eight weeks ago, according to people familiar with the matter. The deal is Hugging Face, the open model repository, and the price is $12.93 billion.
Reuters reported the inquiries on Wednesday, citing people it did not name. Nvidia has been asked for information. The review is at an early stage, no formal investigation has been opened, and the company has not been accused of anything. Nvidia shares fell 1.2% that day.
The transaction was announced in early September. It carries $11.9 billion for shareholders and a retention package of up to $1 billion for employees, and it was scheduled to close in the first half of 2027. Those dates now depend on how long the government keeps asking.
What the government is examining is not size. $12.93 billion is a rounding error against Nvidia’s market value, and the company buys smaller firms regularly without attracting a second look. Hugging Face is different because of what it sits between.
The company was founded in 2016 by three French entrepreneurs and built the place where developers go to find open models, download their weights and compare results. It hosts hundreds of thousands of models and datasets. Major releases from Meta, Mistral and Stability AI have arrived there before they appeared anywhere else, and the site’s leaderboards became the informal scoreboard for the open model race. Its libraries, including Transformers, are released under the Apache 2.0 license and run inside the training and inference stacks of nearly every AI company, including the ones that compete with Nvidia.
That makes Hugging Face a doorway rather than a product. Whoever controls the doorway can decide what is easy to find, what gets featured, and which hardware a model is tuned to run on first. Nvidia sells the chips that most of those models run on. The concern, according to the people familiar with the inquiry, is whether owning both the hardware and the main distribution point for open software would make it harder for rivals to compete.
The theory is old and the facts are new. Antitrust enforcers have spent two decades on cases about a company owning a platform and a marketplace on top of it. What makes this one unusual is that the asset in question is largely a community. Hugging Face’s value comes from the developers who publish there, and those developers have no contract with Nvidia and no say in who owns the site.
There is a second question, quieter and harder to answer. Open models are popular partly because they can be downloaded, modified and run anywhere. If the repository sits inside a chip company, the incentives around which models get promoted and which get optimized for which silicon change, even if no rule changes. Regulators tend to prefer evidence to incentives, which is why the early phase of a review like this one tends to be long.
Nvidia’s likely defense is straightforward. It does not restrict access to Hugging Face today, it has no reason to restrict it tomorrow, and the open licenses on the software cannot be revoked by an owner. The company can also point out that the models hosted on the site run on competitors’ chips, and that the platform’s growth depends on staying neutral.
The counterargument is that a promise is not a structure. The department’s information requests usually precede one of three outcomes: a quiet close, a settlement with conditions, or a formal investigation that adds months to the timetable. Each of those is now in play for a deal that was supposed to be routine.
The timing matters beyond this transaction. Nvidia has become the largest company in the world by market value on the strength of AI accelerators, and its acquisition strategy has shifted from buying small teams to buying infrastructure that the whole industry uses. Every one of those purchases now arrives with a political cost. A government that wants to be seen supervising AI will find it easier to extract conditions from a company that needs approvals than from one that does not.
For Hugging Face employees, the retention package answers the near-term question and leaves the long one open. Startups acquired by platform owners tend to keep their brand, lose their independence, and discover that the roadmap is being written one floor up. The founders who built the site around community norms will now spend part of next year explaining those norms to lawyers.
For the open model community, the practical effect is a waiting period. Models will keep being published while the review runs. The question that hangs over the repository is what happens to the front page, the leaderboards and the default recommendations once the owner of the site also sells the hardware the site’s users are trying to choose between. Nothing about that changes this quarter. The department has until the closing date to decide whether it wants to test the question in court.
