Nvidia and Hugging Face: Why platform access matters
Reuters reports on inquiries by the US Department of Justice into the planned acquisition. At the centre is possible control over the distribution channel for AI models – not an established violation.
According to Reuters, the US Department of Justice is asking about the competitive effects of Nvidia's planned purchase of the AI platform Hugging Face. The agency bases its report of 28 October 2026 on two people familiar with the matter. The department did not comment on specific proceedings, Nvidia declined to comment and Hugging Face did not answer the request. Inquiries are neither a ruling nor proof of an antitrust violation.
The matter nonetheless deserves attention: Hugging Face is not just about a single AI model. The platform serves to share, find and provide models, datasets and applications. Whoever controls such an access point can matter for developers even when their software remains openly licensed.
Signed, but not completed
Nvidia reported to the US securities regulator a purchase agreement concluded on 2 September 2026. The document names a purchase price of around 11.9 billion dollars for shareholders as well as an equity-based retention programme of up to about one billion dollars for employees. Nvidia expects the deal to close in the first half of 2027, subject among other things to the required regulatory approvals.
In a blog post, chief executive Jensen Huang put the transaction at around 12.93 billion dollars. The amounts are therefore not readily identical: the regulatory filing separates purchase price and staff programme, while the blog names a total sum. In any case, the announcement does not mean that Hugging Face is already definitively under new control.
Competition also concerns the route to customers
The publicly accessible US merger guidelines describe a review framework for such constellations. What can be decisive is whether a merger substantially lessens competition because the combined company can restrict access to an important product, service or distribution channel. In doing so, the authorities examine the ability and incentive to restrict, the available alternatives and the importance of the offering in question for competitors.
A complete foreclosure is only one possibility. Guideline 5 also names worsened terms, lower quality, poorer interoperability between systems or delayed access to features. Insight into rivals' sensitive information can be relevant as well. These are general review criteria, not findings about Nvidia's actual conduct in this transaction.
The guideline on multi-sided platforms adds a further question: could an operator favour its own offerings over other participants? A platform connects different user groups; its value also depends on who takes part and which offerings are easily reachable. That is why the question of whether a file may still be downloaded free of charge is not enough for an assessment.
What speaks in favour of the acquisition
Nvidia explicitly promises to keep Hugging Face open. According to Huang's announcement, Nvidia hardware should not be a prerequisite for developing there or for deploying applications. Different cloud providers and accelerators should continue to be supported. The company presents additional infrastructure and development capacity as a benefit of the acquisition.
These commitments are relevant for a fair assessment. They should neither be dismissed as meaningless nor confused with an independent finding. What remains to be examined is how they could be concretely implemented, monitored and, if necessary, enforced. The US guidelines are themselves not a binding decision: the department describes them as a non-binding statement of its review; the law and the facts of the individual case remain decisive.
Why Apertus users may be affected too
Apertus shows how concrete the platform question is for Switzerland. ETH Zurich, EPFL and the Swiss National Supercomputing Centre CSCS published the language model on 2 September 2025. ETH named Hugging Face alongside Swisscom and the Public AI network as an access route. Expert users can download the models and use them for their own projects.
This separates two dependencies: the rights to an open model and the services of a platform for its distribution or operation. A different owner does not turn an existing open licence into an access block. Conversely, the licence does not replace reliable infrastructure. Swiss research groups and companies should therefore document the models, versions and operating routes they need. This includes a test of whether an application with a secured model version also runs outside the previous platform service. Open weights alone replace neither the computing power required nor the documentation for operation. Such tests create a robust fallback option without alleging future abuse. In the end, what will be decisive is not the owner's name alone, but whether choice, switching options and equal access remain in practice.
Sources
- Nvidia: regulatory filing on the purchase agreement, 3 September 2026
- Nvidia: announcement and commitments to openness, 3 September 2026
- US Department of Justice: legal character of the merger guidelines
- US Department of Justice: Guideline 5 on access for competitors
- US Department of Justice: Guideline 9 on multi-sided platforms
- ETH Zurich: publication and access routes for Apertus, 2 September 2025
- Reuters via Yahoo Finance: DOJ inquiries, 28 October 2026
- Cover image: Coolcaesar / Jacek Halicki, Wikimedia Commons
- Cover photo licence: CC BY-SA 4.0
- Article image: Jürg-Peter Hug, Wikimedia Commons
- Article image licence: CC BY-SA 3.0
