This Fall's AI Rulebooks All Stop at the U.S. Border

In May, DeepSeek cut the price of its V4-Pro model by 75 percent, permanently, leaving it 12 to 19 times cheaper than OpenAI's GPT-5.5 and Anthropic's Opus 4.7 for equivalent work ^[1]. The cut was made possible by an unlikely enabler: U.S. export restrictions that cut the Chinese lab off from Nvidia's chips and pushed it onto cheaper Huawei Ascend processors. The one policy Washington managed to enforce on Chinese AI is the one that made it cheaper.

It is an efficiency gain being passed through. — Sanchit Vir Gogia

That is the competitor this fall's rulebooks never quite name. Over six weeks, the American labs — the ones building the most advanced models, called "frontier" models — proposed four ways to slow the whole enterprise down. Read each one for where it reaches, and the same boundary shows up in every draft. Dario Amodei was clearest about the mechanics, so read him for the jurisdiction.

The most effective method of pacing is via regulation that targets all U.S. frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily. — Dario Amodei

The load-bearing word is "U.S." His preferred instrument is "regulation that targets all U.S. frontier AI companies." The 1,367 researchers and executives who signed the August letter — from OpenAI, Anthropic, and Google DeepMind — asked Washington to lead the world in "pacing" development, with a licensing process and registration of AI systems, and signaled they'd halt only if competitors accepted the same restrictions ^[2]. OpenAI took its own version to Congress: mandatory, capability-based national rules before the December adjournment, with pre-deployment testing and reporting gates ^[3]. And Google, OpenAI, and Anthropic spent months negotiating a self-regulatory standards body modeled on Wall Street's FINRA ^[4]. Sam Altman told the United Nations that such decisions could not be left to a few companies — while sitting at the table where those three companies were drafting the rules.

in order for AI to be democratic, decisions cannot be made by companies in San Francisco alone. — Sam Altman

Now slow down, because the quarrel over these rules is itself the story, and it is conducted entirely among Americans. Cohere's CEO Aidan Gomez called the standards body a cartel.

The dispute is over who writes them, who gets to participate and whose interests the rules are protecting. — Aiden Gomez

Palantir's Joe Lonsdale was blunter: OpenAI and Anthropic, he charged, were funding outside groups to push regulation and lock in an oligopoly ^[5].

I think that what OpenAI and Anthropic are pushing right now is very dangerous, and we don’t want an oligopoly that controls all of our policy here with the government. — Joe Lonsdale

Meta advised the White House against the body, and Mark Zuckerberg, Nvidia's Jensen Huang, and Elon Musk pressed Trump to reject the proposed regulator — which the administration did, on the stated ground that a one-sided slowdown hands the race to China ^[6]. Amodei has an answer to the entrenchment charge: he calls the choice between concentrated regulation and wide distribution of AI a "false choice" ^[7].

the choice between concentrating AI power through regulation and distributing it widely was a "false choice". — Dario Amodei

Neither side's motive is provable here. What the drafts themselves say is enough: they bind only American frontier firms. The letter's offer to halt runs only "if competitors agreed to similar restrictions" — restrictions no proposal on the table can enforce on a lab in Hangzhou ^[4][2]. And the frontier scope cuts the other way too, because the rivals are no longer behind it. Alibaba's Qwen and Moonshot's Kimi are high-performance models, released open-weight — their code downloadable by anyone, not sold through subscriptions — and closing the gap with the U.S. labs ^[8]. They're the competition eroding the labs' business, and they fall outside every rule the labs drafted. Where the labs do meet those rivals is the one front no proposal touches: price. OpenAI slashed fees on its Luna model by 80 percent in August, explicitly to hold share against Alibaba's Qwen ^[9]. Anthropic shipped a half-price enterprise model, Sonnet 5.5, weeks after its own CEO called for slowing the most advanced systems — and was careful to note it "does not advance the frontier of AI capabilities" ^[10]. Meanwhile the Chinese labs built the mirror image: Qwen and Kimi free for students and startups, fees only above $50 million and $20 million in revenue ^[11]. The mechanism runs one direction. Export controls took Nvidia's chips away from the Chinese labs, they rebuilt on cheaper domestic silicon, the savings showed up in a permanent price cut, and the Western labs were pulled into value pricing in response ^[1]. That is the single U.S. policy that actually reached the Chinese models — and it made them cheaper, not slower. So the operational present is plain. The regulator is rejected ^[6]. The Congressional ask is pending, with no bill ^[3]. And the only front currently running is price — the contest none of the four rulebooks reaches.