On July 24-25, Microsoft, Nvidia, Meta, IBM, Dell, Hugging Face, Mistral, Perplexity, and 25 organizations in total published an open letter: "Open Weights and US Leadership in AI."
The ask is one line: the US government should not rush to restrict open-weight models.
But the two companies that should have signed — OpenAI and Anthropic — did not put a name on it.
What the letter says
The logic is clean: open weights let more people and businesses into the AI economy, sharpen competition, and spread gains widely instead of concentrating them in a few hands.
It specifically rebuts "open means unsafe": closed models are not inherently safe — they get breached, misused, and fail in ways you cannot see; open weights let many researchers inspect behavior, find holes, and build defenses. Locking frontier AI behind a few closed models amplifies systemic risk.
On distillation (training one model on another's outputs), the letter says it is standard industry practice and should not be conflated with illegally extracting closed-model value — handle it with "targeted legal frameworks," not a blanket ban.
Huang Jensen posted the letter as his first-ever tweet after joining X; Musk reposted "fully support"; Nadella backed it. A founder using his debut post to endorse open source carries weight.
Why OpenAI and Anthropic did not sign
The answer is not complicated: business. Both have secretly filed S-1s with the SEC, prepping IPOs near a $1 trillion valuation. Their valuation rests on "frontier closed model plus high-price API."
If open-weight models close the gap or pass closed ones soon, that pricing moat gets rerated. Both say they "support AI democratization"; neither moved the signing pen.
Not a conspiracy — a structure of interest. A closed company's moat is "you do not open"; it has no reason to dismantle it itself.
The trigger: Kimi K3 narrowed the gap to 3.3%
The letter's backdrop is Moonshot AI's July release of Kimi K3, an open-weight model whose benchmarks approach US closed frontrunners. Data shows the average performance gap between open weights and top closed models has tightened to 3.3%.
Some in the US want to restrict open releases citing "tech leakage" and "national security," even alleging Chinese teams distilled US closed capability. K3's enterprise lead Huang Zhenxin countered: K3's leap comes from original architecture, not distillation cloning.
The 25 signatories state plainly: a blanket ban on distillation and open source is throwing out the baby — it blocks no rival and breaks US domestic innovation rhythm.
What this means for you
- Choosing models: the open-closed fight is not abstract. It decides who owns your data, whether you can self-host, and where pricing goes. A stronger open camp lowers your odds of lock-in to one closed vendor.
- Cost: open weights entering the market push prices down; your leverage as a buyer rises. A frontier open model like K3 is itself a chip in your negotiation with closed vendors.
- Compliance: if you operate cross-border or in regulated industries, watch this US policy turn — a restriction, if it lands, changes which models you can use and whether data can leave.
One line: 25 giants raising the open-source flag prove open versus closed is no longer a tech-route debate but a map of interests. Who signed and who did not tells you more than the letter's text.