Two claims sit uncomfortably side by side this week, and most readers will resolve the tension in the wrong direction.
The first: two dozen American companies — Meta, Microsoft, Nvidia, IBM, Dell, Hugging Face, Mistral, Andreessen Horowitz, Y Combinator — signed a letter begging Washington to protect open-weight AI, framing it as the mechanism that spreads capability to "factories, hospitals, farms, classrooms, and main street businesses."
The second: the same week, the largest open-weight model on earth landed on Together AI with a 99% uptime SLA and a price 65% below Anthropic's Fable. It was built in Beijing.
The obvious read is that these two events reinforce each other — American industry rallying to defend the open ecosystem it built. The correct read is that the letter is fighting for a future a Chinese lab is already delivering, and that the coalition assembled to fight for it can't agree on the one clause that would actually bite.
The letter is fighting for a future a Chinese lab is already delivering, and the coalition assembled to fight for it can't agree on the one clause that would actually bite.
Start with the arithmetic, because it ends the argument before the policy debate begins.
Last week I noted that open weights had quietly come to carry most production traffic. This week the numbers are hard: open models went from 10% of tokens to 30% in a single year. Tripled.
Together's own framing is the honest one — "open + modular wins on cost, and cost is what unlocks scale." The economics have already voted. The letter is litigating a question the market closed months ago.
Now look at what's driving the tripling. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model, and it is not a fast-follower's approximation of the frontier.
On the one benchmark that runs all three flagships through a single methodology — Artificial Analysis's Intelligence Index — Fable 5 scores 60, GPT 5.6 Sol 59, and K3 57. A three-point spread on a scale where frontier-to-mid-tier gaps run fifteen points or more. On general intelligence, these are peers.
The prices are not peers. Fable charges $10 per million input tokens and $50 output. K3 charges $3 and $15. On identical workloads, Fable costs more than three times what K3 does — for a capability gap most tasks will never surface.
That is the whole game. When frontier-adjacent quality is available at a third of the price, with downloadable weights and a million-token context, the token share doesn't drift toward open. It stampedes.
This is why the coalition letter reads less like a manifesto and more like a positioning document — and why its two fault lines matter more than its rhetoric.
The first fault line is the distillation carve-out. The letter draws a line between distillation "as a legitimate technique" and "unlawful efforts to extract value from closed models," arguing the former shouldn't get swept into restrictions aimed at the latter.
This is the one clause with legal teeth. It's a direct response to the post-DeepSeek dispute, when American labs accused Chinese rivals of training on their closed models' outputs without authorization.
And it's exactly where the coalition falls apart. As Ethan Mollick put it, "the distillation carveout buried in the letter is the tell. Every signatory agreed on 'open weights good' and quietly disagreed on the clause that would matter most legally."
A coalition that can't align on its hardest edge case, in his words, "is useful for press releases and not much else."
Here's the added irony: the distillation panic that the carve-out exists to manage is itself looking obsolete. Nathan Lambert, who visited the Kimi team, argues that if adversarial distillation contributed to K3 at all, it was "to a relatively small degree" — that Chinese labs are now solving the same problems as OpenAI and Anthropic, not stealing solutions.
So the letter's most consequential clause defends against a threat model that the evidence is already undermining, and the signatories can't even agree on how to word it.
The second fault line is deeper, and it's about belief.
Mollick again: "there is no way to hold a consistent belief set where you're agi pilled and pro open source." His point is that most open-weights advocates simply don't hold the catastrophic-risk convictions that lab insiders do — they don't expect AI to present "grave semi-autonomous biosecurity or other similar risks."
This matters because restrictions on open weights only make sense if you believe the downside is catastrophic and irreversible. Once released, weights are beyond recall — the letter concedes this outright, that guardrails can be stripped and modified versions circulate untraceably.
If you genuinely believed those weights could become weapons, you'd want restrictions. The coalition wants the opposite. Which means the coalition, by revealed preference, doesn't believe the risk is catastrophic.
That's a defensible position. But it exposes the letter's security argument — the elaborate case that defenders need open models to simulate AI-equipped attackers — as motivated reasoning dressed as first principles.
Now assemble the pieces, and the asymmetry becomes uncomfortable.
The letter's stated beneficiaries are American: startups that can't afford frontier training runs, hospitals and classrooms priced out of per-token frontier fees, enterprises wanting freedom from vendor lock-in.
But the model actually delivering frontier-cheap open weights is Kimi K3, released under a modified MIT license, its full weights scheduled to drop on July 27. Xi Jinping used the same week to commit China's AI ecosystem publicly to open-source and global diffusion — the first time senior Chinese leadership has done so.
So the coalition is lobbying Washington to keep open a lane that a rival ecosystem is currently winning. The "factories, hospitals, farms" the letter invokes will, in the near term, run disproportionately on Chinese weights, because those are the weights that clear at a third of the price with a production SLA.
Which reframes who the letter is really for.
Look at the signatories with the most to gain: Nvidia, IBM, Dell. As the reporting on the letter notes plainly, a wider range of deployable models "sells more compute and services regardless of which lab produced the weights."
That's the quiet logic. The infrastructure players profit whether the winning open model is American or Chinese. Their interest isn't in an American open-model victory — it's in open-model volume, full stop. More deployable weights means more GPUs, more racks, more services sold.
Which is why the coalition could assemble two dozen names around "open weights good" and fracture the moment anyone asked what they'd actually restrict. There was never a shared strategic objective. There was a shared commercial one.
The letter presents itself as a fight to win the open-model race. Read against the token-share data and the price gap, it looks more like a fight to keep the track open — while a Beijing lab laps the field, and the companies selling the shoes collect either way.
Cost is what unlocks scale. K3 understood that first. The coalition is still drafting the memo.