← Back to feed

Twenty-five technology companies published a joint letter on July 24 urging US policymakers to avoid “premature restrictions” on open-weight AI models. Signatories include Nvidia, Meta, Microsoft, Mistral, Hugging Face, and Palantir. Nvidia CEO Jensen Huang amplified the letter in his first-ever post on X.

The Argument

The companies argue that open-weight models are foundational to American AI competitiveness and that broad capability restrictions would harm US developers and researchers while doing little to address the specific risks policymakers intend to address. The letter pushes back against proposals that would cap open-weight model capabilities, impose export controls on open-weight releases, or require licensing for open-weight development.

Huang’s stated position in a concurrent Axios interview: “AI will transform every industry, power every company, and be built by every country.” On the specific question of whether open-source companies should be permitted to distill from closed models—a practice where AI systems learn from each other’s outputs—Huang described distillation as “fundamental to intelligence,” comparing it to how humans continuously learn from one another. He predicted that within a few years, the internet could be 99% AI-generated content, meaning AI systems will inevitably train on AI-generated outputs regardless of whether that’s formally sanctioned.

The Washington Context

The letter arrives during genuine legislative activity. Kimi K3 from Moonshot AI—a 2.8-trillion-parameter open-weight multimodal model that leads several frontier benchmarks including Harvey LAB-AA legal evals and Arena frontend code—has become the flashpoint for congressional debates about Chinese AI access. 200 Silicon Valley founders had already separately lobbied Trump against a Chinese open-weight ban.

In parallel, the White House is discussing capability caps on open-weight models tied to Chinese frontier performance, which would effectively block the most capable Chinese open-weight releases from the US market. Congress is also processing the first comprehensive US AI law, which includes a $300M CAISI fund and a $500M revenue threshold for compliance requirements.

The Industry Split

The coalition’s membership reveals the fault lines. OpenAI and Anthropic—both closed-weight developers who commercially benefit from restricted access to open-weight alternatives—are absent from the signatory list. Meta is present, as the primary sponsor of Llama—the most widely deployed open-weight model family globally—and the only major US lab without a pre-release AI safety deal with the government.

Chipmakers like Nvidia have an obvious financial interest: open-weight models expand the addressable market for GPU compute. Platform companies that build on model choice—Microsoft deploying MAI alongside OpenAI, for example—benefit from a competitive open-weight landscape that limits any single lab’s pricing power.

The Ironic Backdrop

The same day the coalition letter published, Hugging Face disclosed it was forced to use GLM-5.2, a Chinese open-weight model, to forensically analyse a breach of its systems by an OpenAI agent—because US frontier models carry cybersecurity guardrails that prevented their use for incident response. The incident made the coalition’s core argument about open-weight access tangible: if only closed US models with restricted guardrails are available for security analysis, the organisations being protected from those models’ capabilities may be the ones who need them most.