Nvidia Acquires Hugging Face for $12.93 Billion, Tying Open Source to Its Silicon

Nvidia confirmed on September 3 the acquisition of Hugging Face for $12.93 billion, marking the second largest acquisition in its history, following the $20 billion paid for Groq's assets at the end of last year. In addition to the purchase price, the chipmaker has allocated up to $1 billion in retention bonuses for the team at the French startup. According to Clement Delangue, CEO and co-founder of Hugging Face, he reached out to Jensen Huang weeks before the closing, with other interested parties also in play.

The commercial design weighs more than the absolute value. Hugging Face has become the standard infrastructure through which developers and researchers publish and download models: there are over 3 million hosted models, 18 million developer accounts, and around 200,000 companies using the platform to discover and deploy AI. Bringing it under Nvidia's control means that the distribution counter of the open ecosystem now belongs to the owner of the silicon counter, at a level of vertical integration that AWS, Google, and Microsoft have never achieved.

Why Delangue Sold

Hugging Face had been attempting to monetize free hosting for two years through Inference Endpoints, corporate licensing, and an infrastructure deal with AWS. Private numbers have never leaked, but reports from investors suggested an annualized revenue below $100 million and a cash burn high enough to necessitate a growth round. Huang referred to the platform as a "growth driver" and stated that Nvidia will maintain support for open source and open weight models. The promise is easy to make: Hugging Face's profitability for Nvidia does not come from charging for downloads, but rather from ensuring that every new popular model is optimized for CUDA before it is optimized for any other accelerator.

What Changes for AMD, Intel, and Hyperscalers

AMD reacted to the market opening with an intraday decline, and for good reason. ROCm has struggled to reach parity in support for popular models with CUDA, and Hugging Face's natural integration with Nvidia libraries is likely to pressure this competition. Intel, which based its Gaudi program on open inference benchmarks, loses a neutral interlocutor. For AWS, whose Bedrock was the main enterprise channel integrating models from Hugging Face, the deal introduces contractual uncertainty: Nvidia will have a say on where and how open weights run in the future.

On the hyperscaler side, Google Cloud and Azure need to quickly decide whether to replicate distribution via TensorFlow Hub and Azure Model Catalog or to accept remaining in second place in a discovery channel that is already the benchmark. Historically, they tend to choose replication, but developer inertia is now in favor of Nvidia.

Antitrust in Brussels and New Delhi

The transaction must pass through the European Commission, the U.S. FTC, and likely the Indian CCI. Brussels already signaled in July, during the analysis of the Anthropic-Amazon deal, that it views vertical acquisitions in AI as a distinct risk category, and the base of Hugging Face is in Paris. Nvidia will argue that the market for model hosting platforms is still fragmented, with Replicate, Together AI, and ModelScope in China as alternatives. Brussels is unlikely to accept this argument without demanding commitments for interoperability with competing accelerators.

In India, where Hugging Face is heavily used by academic research teams and language startups for Hindi and other local languages, regulatory concern is likely to shift toward preserving free access. The IT ministry has already called technical meetings for next month. The aggressive timeline Nvidia projected, with a closing expected in the first quarter of 2027, depends on the three regulators agreeing to behavioral remedies rather than structural ones. Historically, when three jurisdictions dispute different remedies, the timeline slips by at least six months.