Nvidia, Meta, Microsoft, and 22 other prominent organizations have issued a joint warning to U.S. policymakers about the risks of imposing extensive controls on open-weight AI models. They argue such broad restrictions could undercut America’s edge in AI development, especially as competition with China intensifies.
Coalition Urges Targeted Measures Over Sweeping Bans
The signatories, including IBM, Palantir, and the Linux Foundation, advocate for focused enforcement aimed at intellectual property theft and clear misuse rather than blanket limits on AI technologies that drive legitimate innovation. Open-weight AI models can be downloaded, customized, and operated by companies and governments on their own infrastructure, which enables better adaptability, enhanced control over data security, and lower deployment costs.
Nvidia CEO Jensen Huang highlighted the importance of coexistence between open and closed AI models. He emphasized that both approaches are key to fostering cybersecurity, safety, national control, and wider access to AI tools across industries.
Elon Musk publicly supported the letter, underscoring his alignment with Huang’s view. Musk’s company xAI, which develops the chatbot Grok, competes against offerings from OpenAI and Anthropic, though xAI was not among the letter’s signatories.
Amid Rising US-China AI Tensions
This joint statement arrives as the Trump administration weighs actions against Chinese AI firms suspected of using American technology without authorization. U.S. Treasury Secretary Scott Bessent recently said officials will investigate whether Chinese models were trained through unauthorized use of outputs from U.S.-developed systems. Measures such as sanctions and listing on the Entity List might be implemented if large-scale unauthorized replication known as industrial-scale distillation is confirmed.
Bessent also clarified that the administration supports open-source AI and aims to distinguish legitimate development from intellectual property violations.
Distillation refers to a process where a trained AI model is compressed or simplified into another model, potentially replicating performance without authorization.
