Nvidia acquires Hugging Face to control open model distribution

The timing is the buried lede: Nvidia closed this deal two days after Stratechery's September 1st earnings analysis explicitly named distribution control as the company's most viable path to a durable moat against custom silicon. The acquisition reads less like opportunism and more like a thesis being executed in real time.

Stratechery's piece from September 1st (covered here as 'Nvidia Earnings, Dollars Per Gigawatt, Open and Hugging Face') framed Nvidia's core problem as preventing commoditization once custom silicon matures at the major labs. Owning Hugging Face is a direct answer to that threat: if labs defect from Nvidia GPUs for training, Nvidia still sits at the distribution layer where those models reach 18 million developers. Separately, Hugging Face's own September 1st WebGPU kernel release, which pushed inference toward client-side hardware, now reads differently under Nvidia ownership. A platform that was actively decentralizing inference is now controlled by the company most invested in keeping compute centralized and GPU-dependent.

Watch whether Hugging Face's WebGPU and non-Nvidia inference tooling receives the same engineering investment over the next two quarters as it did pre-acquisition. A slowdown or deprioritization there would confirm that neutrality commitments are eroding in practice, regardless of what Jensen Huang says publicly.