Analysis
China rebuffed the American AI industry's pacing argument as "fearmongering," The Information reported, responding to a week in which Dario Amodei called for slowing frontier development and Microsoft committed to building kill switches into its products.
The response was predictable and it is also the whole problem. Every voluntary-slowdown proposal contains an unstated assumption of reciprocity: if the leading labs pace themselves, the gap they hold is preserved and safety improves without ceding position. Remove reciprocity and pacing becomes unilateral disarmament in the framing that American accelerationists have used since 2023, which is precisely why the argument is politically unwinnable in Washington without some Chinese counterpart.
“Z.ai is reportedly raising $5 billion to cover compute costs.”
The capability picture makes the standoff concrete. DeepSeek's releases have repeatedly matched US frontier models on coding and reasoning benchmarks at a fraction of the inference price, and a recent Chinese filing implied a DeepSeek valuation near $52 billion. Alibaba's Qwen family and Moonshot's Kimi ship open weights that anyone can fine-tune. Z.ai is reportedly raising $5 billion to cover compute costs. None of those programs are governed by anything an American standards body would write.
Export controls were the instrument meant to solve this, and their record is mixed. Restrictions on advanced Nvidia parts slowed Chinese training capacity and simultaneously accelerated domestic substitution -- Huawei's Ascend line exists in its current form substantially because of them. The lesson generalizes: constraints that bind only one side reshape who builds what rather than whether it gets built.
Which leaves a narrow space for anything workable. The historical analogues that functioned -- the Montreal Protocol, nuclear test-ban verification -- had measurable compliance and an inspection regime. AI has neither. Compute thresholds are the only semi-verifiable proxy anyone has proposed, since large training runs are visible in power draw and chip procurement, and even that breaks down as efficiency improves and frontier capability gets cheaper to reach.
