Microsoft, NVIDIA, Google and Meta say open-weight models lower costs and increase competition, arguing those benefits outweigh the loss of control after model weights are released.
Microsoft, NVIDIA, Google, Meta and dozens of other companies are urging U.S. policymakers not to restrict open-weight AI models, whose core files can be downloaded, modified and run on an organization’s own systems.
The July 24 statement has 77 signatories, including major AI developers, cybersecurity companies, cloud providers and technology investors. They want the federal government to expand access to computing resources and avoid broad restrictions that could limit the development or use of open-weight models.
The coalition argues that preserving access will help the United States remain competitive in AI. It also says open weights give businesses more choices about how they buy, customize, and operate AI systems.
The upside: More control for companies
Organizations can use open-weight models without training an AI system from scratch or paying for access to a provider’s most powerful model every time they perform a task, according to the statement.
Companies can instead choose smaller models for routine work and reserve more expensive systems for tasks that require greater capabilities. They can also adapt models to their own needs, run them on their preferred infrastructure and keep more control over their data.
The coalition says those options reduce the risk of becoming dependent on one AI provider. They could also increase competition among model developers, cloud companies, chipmakers, and software providers.
The downside: Open-weight models cannot be recalled once released
The statement acknowledges that open-weight models create distinct security risks. Once a developer releases the model, it loses control over who downloads it or how it is modified.
Modified versions can also be difficult to trace or reverse.
The coalition argues that restricting access would create a different set of risks. Cybersecurity teams need powerful models to identify and respond to attacks, it says. Giving researchers and developers access also allows more people to examine models for weaknesses and develop safeguards.
The companies argue that closed models are not automatically safer. Those systems can still be breached or misused. Further, concentrating advanced AI among a few providers can create common points of failure.
Distillation draws a separate policy request
The coalition also asks policymakers not to treat legitimate model distillation as theft. Distillation uses one model’s output to help train, test, or improve another model.
The statement distinguishes that practice from distilling from a closed model. It calls for targeted legal and commercial rules to address misuse without broadly restricting a technique used throughout AI development.
The coalition did not endorse a specific bill or provide a timetable for federal action. Its immediate request is for policymakers to preserve access to open-weight models while tying restrictions to demonstrated harm.
