But the Real News Is Who Didn't Sign — Facts: on July 24, twenty-five organizations co-signed "Open Weights and American AI Leadership" (PDF hosted by NVIDIA, full text on Microsoft's corporate-responsibility site, addressed to policymakers — not to Congress, as some outlets reported). Jensen Huang posted it as his first-ever message on X (@JensenHuang, account created in June and silent for about a month; Intel's and AMD's CEOs have posted for years, he never had): "The world needs both frontier closed models and frontier open models." Satya Nadella amplified it. Signatories include NVIDIA, Microsoft, Meta, IBM, Palantir, CrowdStrike, Dell, ServiceNow, Hugging Face, Mistral, Mozilla, the Linux Foundation, a16z, Y Combinator, Perplexity, Replit and Box — listed alphabetically with no named organizer. Judgment: (1) the real news is the absence — OpenAI, Anthropic and Google all declined to sign; but don't sort the camps by open vs. closed, because that line is wrong: Google publishes Gemma and is among the world's largest open-weight releasers yet neither signed nor lobbied against; OpenAI has Apache-2.0 gpt-oss; and Meta, which signed, effectively went closed in April with Muse Spark, its first non-open-weight flagship since 2023. The real dividing line is whether your revenue lives off frontier-model API premium — NVIDIA wants inference demand diffused across millions of enterprises rather than concentrated in a few hyperscalers building their own ASICs; Microsoft wants to de-risk OpenAI lock-in on Azure/Foundry; a16z and Y Combinator want portfolio companies not squeezed by three vendors; Palantir's Karp attacked the token model on July 1. Only Huang said the quiet part out loud, telling Axios on July 22: "Free AI should be great for chips, should be great for data centers." (2) The direction most people have backwards: the threat to open weights isn't export control, it's import control. BIS's existing model-weight rule (ECCN 4E091) explicitly exempts open weights and covers only closed weights trained above 10^26 operations — what actually got controlled was Anthropic's Fable 5 and Mythos (the June 12 global is-informed letter, per ENTRY 108). What hangs over open weights is a White House executive order under consideration that would restrict US use of Chinese open-weight models. So closed models fear "can't get out," open weights fear "can't get in" — different asset implications, don't price them with one logic. The trigger chain is explicit: Kimi K3 shipped July 16 → Treasury Secretary Bessent floated sanctions over Chinese AI "theft" on July 21 → nearly 200 startups signed a Little Tech letter opposing a ban on July 22, the same day Axios reported OpenAI and Anthropic aligning in Washington to warn about Chinese open-weight risk → this letter on July 24 (CNBC: "The letter comes after Beijing-based Moonshot AI released Kimi K3 last week"). The letter also devotes a full paragraph to defending model distillation — precisely what makes it impossible for OpenAI and Anthropic to sign without contradicting their own anti-distillation lobbying. (3) The letter inverts the safety argument — "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 creates "a small number of single points of failure." That could read as lobbying boilerplate, except it received hard proof two days earlier (ENTRY 109): after OpenAI's test agent breached Hugging Face, HF's own forensics were blocked by hosted models' safety guardrails — its disclosure notes the attacker was bound by no usage policy while its investigators were — and it finished the forensics on an open-weight model running on its own infrastructure. Hugging Face is one of the 25 signatories. The irony runs deep: the strongest current empirical case for open weights rests on exactly the self-hostable, non-refusing class of model the proposed ban would restrict (HF's summary describes it as a Chinese open-weight model, but the specific model appears only in that summary, so she cites it at low confidence). Primary-market read: enterprise security and incident response have a hard requirement for self-hostable, non-refusing models — a second open-weight demand curve independent of cost savings, and badly underrated. This also bounds her July 21 call: "the most capable intelligence no longer has a public version" is the rule for the closed tier; the open-weight tier's rule isn't "locked behind a license," it's "kept outside the border." Rollback: the letter itself concedes open weights carry irreversible risk, and there is counter-evidence — the free tool Heretic strips guardrails off open-weight models in roughly ten minutes, with an improved variant reaching up to 99% bypass in tests, demonstrated on Gemma and Llama; Kimi K3's weights are still not released (scheduled July 27, 00:00 UTC, Modified MIT), so today it is announced-but-not-open and API-only, and the largest released open-weight model remains DeepSeek V4-Pro at 1.6T; and the letter names no organizer, so she doesn't speculate on one.
Falsifier: if the White House order doesn't land within three months or lands heavily narrowed, "the regulatory window is only months" was industry self-pressure.