Nvidia published a policy paper on July 24 asking Washington to protect open-weight AI models from new restrictions, and it did not sign the document alone. Seventy-six other organizations backed it, including chip rival AMD, cloud providers Google and Microsoft, and model developers OpenAI, Meta and Mistral. The paper arrives as US policymakers debate export controls and safety rules that could tilt the market toward closed, proprietary systems. Nvidia’s argument doubles as its business case: the company sells the chips that train and serve these models, open or closed, and the outcome of that debate shapes who buys them.
That commercial stake is specific, not incidental. A market where hundreds of startups, universities and enterprises build on open weights, models whose parameters anyone can download and run on their own hardware, creates broad, distributed demand for Nvidia’s chips. A market concentrated in a handful of closed labs is a harder one for Nvidia, because several of those labs, Google, Meta and OpenAI among them, are also building or buying their own custom silicon to cut their reliance on Nvidia. Keeping what the paper calls “the frontier plural” keeps Nvidia’s customer base plural too.
The paper names four concrete requests for Washington.
- Expand compute access for startups and university researchers.
- Fund shared training assets: datasets, tooling and evaluation frameworks.
- Avoid what it calls premature restrictions on open-weight models.
- Treat distillation, training a new model on another model’s outputs, as a legitimate research technique separate from unlawful extraction of a closed model’s intellectual property.
The first request tracks Nvidia’s interest most directly. Subsidized compute access for smaller developers expands the buyer pool for Nvidia hardware. The other three track the paper’s stated policy argument, that a plural, transparent ecosystem is more competitive and more secure, but they also serve the same commercial end by keeping the field from consolidating around a few vertically integrated giants.
The paper concedes a real risk on its own terms: once weights are released, they are beyond the original developer’s control, and modified copies are difficult to trace. It does not follow that concession with a mechanism for managing it, no licensing terms, disclosure rules or liability standard. It argues instead that prohibition would be worse, without specifying what would actually be better.
The distillation defense is not new. OpenAI accused engineers at DeepSeek in early 2025 of training on its models’ outputs without authorization, a dispute that never produced a clear legal line between legitimate model improvement and IP extraction. Nvidia’s paper asks Congress to settle that ambiguity in favor of the practice before a court or regulator settles it the other way.
For operators watching Washington, the number worth tracking is not open versus closed. It is whether any resulting legislation defines distillation narrowly enough to protect Nvidia’s 76 co-signers while still giving closed labs a workable claim against outright copying, since that line will decide which side of the AI stack gets to keep building on the other’s work.
Nvidia published “Open Weights and American AI Leadership” on July 24, 2026.
