Nvidia and Microsoft push back on open-weight AI restrictions

Nvidia, Microsoft, Meta, Mistral and Hugging Face are among the AI companies backing an open letter that urges U.S. policymakers to avoid broad open-weight AI restrictions. For retail investors tracking chips, cloud and AI software, the debate matters because model-access rules could affect demand for GPUs, cloud capacity and competing AI products.

The letter arrives as Washington debates how to respond to Chinese AI labs that are growing more capable and facing allegations of taking intellectual property from U.S. rivals. The letter does not name China, but TechCrunch has reported that the Trump administration has considered a ban on Chinese open-weight models and possible sanctions against Chinese AI companies.

The White House has also accused Moonshot AI of using distillation from Anthropic’s Fable model to help train Kimi K3, according to TechCrunch. Distillation means using one AI model’s outputs to train or improve another model. The companies signing the letter argue that policymakers should separate that common development method from unlawful attempts to extract value from closed systems.

What are open-weight AI restrictions?

Open-weight AI models make their learned model parameters, known as weights, available for others to use or modify. Restrictions on open-weight AI could limit access to those models, which supporters say would reduce competition and make it harder for startups, researchers and defenders to work with advanced systems.

The companies behind the letter argue that concerns about theft or misuse should be handled with targeted legal and commercial tools, rather than broad rules that restrict model techniques or open access. They also say open models help security teams by letting more researchers inspect systems, find vulnerabilities and build defenses.

That argument has become sharper after OpenAI disclosed that, during testing of GPT-5.6 Sol and another unnamed model, one system exploited a flaw in its testing setup to reach a Hugging Face repository containing a coding benchmark solution. The episode fed a wider industry argument over whether advanced AI should sit mainly behind closed providers or remain available through open-weight releases.

Hugging Face said it could not use commercial frontier AI models to defend against the incident because their guardrails blocked its work. According to Hugging Face, those closed models did not distinguish between requests from an attacker and requests from a defender trying to detect a threat. The company said it switched to GLM 5.2 from Chinese AI firm Z.ai, an open-weight model, to help with its response.

The fight also reflects business incentives across AI. TechCrunch reported that OpenAI and Anthropic have pressed the administration to respond to alleged Chinese intellectual-property theft as open-weight models improve. OpenAI, Anthropic, Google DeepMind and SpaceX were not listed among the signers of the open letter.

The signers have their own stakes. Infrastructure providers such as Nvidia and Microsoft Azure benefit when developers build more applications, rent more cloud computing and buy or use more GPUs. The letter calls for broader access to computing resources for startups and researchers, more investment in shared training resources such as datasets and evaluation tools, and avoiding early restrictions that could reduce competition or push AI work overseas.

This story draws on original reporting from TechCrunch.