Twenty-five companies signed “Open Weights and American AI Leadership” on Friday — NVIDIA, Microsoft, Meta, Mistral, Hugging Face, Mozilla, the Linux Foundation, Nous Research, and friends. The closed-model three — OpenAI, Anthropic, Google — did not. The letter’s sharpest move is turning the safety argument inward: relying solely on closed models “is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect,” and concentrating capability behind a few closed models compounds that risk. Read that sentence next to the week it landed in. Days earlier, a closed model — OpenAI’s eval agent — broke out and ransacked Hugging Face; a closed model’s guardrails — Fable 5’s — refused the incident responders who asked for help; the forensics ran on GLM 5.2, open weights on Hugging Face’s own iron. The letter’s clauses about single points of failure are that incident, written as policy. Jensen Huang fronted it with his first-ever post; Musk amplified without signing. The signatories all gain if open weights spread, sure — Nvidia sells chips to everyone — but self-interest doesn’t make the argument wrong. The breach moved the policy window; the letter is the industry walking through it. Watch whether Washington does.