The US Department of Justice has reportedly started looking into Nvidia's acquisition of Hugging Face, a deal worth about $13 billion. The reported concerns center on competition in the AI sector and the pressure such an integration could place on customers.

To grasp the significance, one needs to look at the role Hugging Face has assumed in recent years. The platform has become the reference point for distributing, downloading and adapting open-source models: tokenizers, datasets, libraries for fine-tuning and weight repositories. It is not just an archive but a coordination infrastructure for developers and companies that want to keep control of their models, including in self-hosted environments.

That is the crux. Nvidia dominates the market for AI accelerators, and its GPUs are the default choice for training and inference. If the same company also controlled the main hub for open models, the boundary between hardware and software would become much thinner. For anyone using Hugging Face today to download a model and then run it on non-Nvidia hardware, or for those evaluating an on-premise deployment with alternative accelerators, the risk is ending up in an ecosystem where libraries, formats and optimizations push strongly toward the vendor's GPUs.

This is not abstract. Vertical integration changes incentives. Neutral repositories today allow teams to compare implementations, assess how a model performs on different architectures and choose the runtime that best fits their needs. If the repository owner also becomes the hardware seller, documentation, benchmarks and conversion tools could favor its own solutions. For an organization estimating the Total Cost of Ownership of a self-hosted infrastructure, this is not a detail: it can mean higher exit costs, less interoperability and fewer practical alternatives.

Sovereignty is also at stake. Many companies and institutions choose open models and on-premise pipelines precisely to avoid dependence on a single cloud provider or proprietary APIs. An acquisition like this, if confirmed, could weaken that lever because the access point to models would become part of a supply chain controlled by a single entity. The issue is not only classic antitrust: it is the ability of one player to shape the evolution of distribution formats and fine-tuning practices, with consequences for anyone developing or adapting models locally.

The inquiry, still at an early stage, is a structural signal. The real bottleneck in AI is no longer just silicon availability or model quality, but control of the infrastructures that make models accessible. In this shift, the neutrality of distribution platforms matters as much as a GPU's technical specifications. Teams evaluating self-hosted infrastructure today must consider these dependency risks, not just VRAM or tokens per second. AI-RADAR covers these on-premise deployment trade-offs at /llm-onpremise.