A US–China compute-cap treaty holds for a year without accusations

Where things stand (as of 2026-10-11)

There is no US–China treaty of any kind that limits AI training compute. Negotiations toward one have not started, and both governments have publicly rejected slowing AI down:

• Washington. At the UN General Assembly on Sept 22, 2026, Trump said the US "totally rejects any attempt to construct a globalist scheme to control" AI. After the Sept 24–26 Xi state visit he said the US "is not going to be putting on brakes… we're leading by a lot… I'm not looking to open it up" (CBS News; Carnegie, Sept 30, 2026). His main domestic step was a voluntary, non-binding "White House Accord on Super Intelligence" signed by companies (Al Jazeera, Sept 30).
• Beijing. China's Foreign Ministry called Amodei's "pace the frontier" proposal "fear-mongering." Analysts read the slowdown calls as a bid to lock in the US lead (Guardian, Sept 20; Rest of World). SemiAnalysis (Oct 9) finds that China has no frontier-risk duties tied to training compute or model capability. Its regime puts speed first.
• The only deliverables from the summit were a "Super Intelligence/AI Dialogue" due to meet in November and a channel for reporting AI incidents (White House fact sheet; China's eight-point consensus). Chinese signals say the November agenda is risk classification and incident-reporting rules, not limits on development (Pekingnology, Oct 1). Experts call even this close to the ceiling of what the relationship can bear (CFR's McGuire, quoted by VOA).
• Other governments. The most ambitious government statement, the Finland–Norway "Call for Control of Frontier AI" (28+ countries plus the EU), asks for testing, incident reporting and an international institution, not compute caps. Neither the US nor China signed it. The UN Global Dialogue on AI Governance is non-binding.

Reference class and base rate

The US and China have never concluded a bilateral arms-control treaty. China refuses quantitative limits on its nuclear forces while it is behind (Foreign Ministry, Feb 2026, cited by CFR). Historical precedents for a first binding cap on a strategic technology took 10–20+ years: the Outer Space Treaty came about 10 years after the first satellite, the Limited Test Ban Treaty 18 years after Hiroshima. The 2015 US–China cyber deal was non-binding and was followed by accusations.

Forecasting platforms point the same way:

• Metaculus: about 5–6% that a US–China AGI-limiting treaty is signed before 2030, and 4–5% for any formal agreement limiting frontier AI before 2029. These cover a broader event than this question.
• Polymarket: "US x China agree to pace the AI frontier in 2026" was at 6.9% (gamma API, 2026-10-11). That market counts non-binding announcements.

This question also requires all of the following:

• a legally binding treaty;
• a quantitative training-compute cap that binds both countries (roughly half of the conceivable "limiting" deals would use capability or testing rules instead);
• entry into force by about late 2030;
• 12 uninterrupted months with no formal accusation, suspension or expiry (a short fixed-term pause would fail);
• an official status publication after that.

Paths and timing

• To YES:
  • A severe AI catastrophe, given frequent rogue-agent incidents and labs reporting the start of recursive self-improvement, could produce a fast deal like the 1963 test ban, possibly as a binding executive agreement.
  • A new US president from January 2029 could pursue a treaty. Democrats are favoured: Polymarket gives them 90% for the House in 2026, and Trump's eventual endorsed candidate is at 31.5% to win in 2028. A symmetric cap might even appeal to China, because it would neutralise the US lead in compute.
  • Even AI-safety advocates' aspirational "AI 2040: Plan A" places the US–China deal in 2029.
• To NO: both governments' explicit opposition; verification that is still immature (CNAS); US domestic legal hurdles to binding caps on companies (Senate two-thirds, or legislation past a filibuster); and the very short window, since entry into force must come by late 2030.

The first four horizons would need a treaty already in force since 2025–Sept 2026, which does not exist. I give them a nominal floor only.

Computation

I set quarterly hazards for entry into force: about 0.03–0.07% per quarter in 2027 and 0.08% in 2028 (the Trump era, mostly via a catastrophe), then 0.1–0.3% through 2029 and 0.33–0.40% in 2030. That gives about 2.8% cumulative entry into force by the end of 2030. I then:

• shifted it by 12 months plus a publication lag (50% in the anniversary quarter, 35% one quarter later, 15% two quarters later);
• multiplied by 0.8 for an interval free of accusation, suspension or expiry.

The result is about 2.0% by 2031-12-31 (computed in Python).

Strongest case against: a thin Manifold market prices a 36% chance of an international pause by 2028, and this world is unusually volatile. I discount that market as illiquid. My estimate is consistent with Metaculus once the extra conditions here are applied.

Sources

• guardian.com 2026-09-25
• cbsnews.com (Trump–Xi AI channel)
• carnegieendowment.org/emissary/2026/09/trump-xi-ai-safety-no-deal
• darioamodei.com/post/we-must-pace-the-frontier
• cfr.org/articles/how-to-reach-an-ai-arms-control-deal-with-china
• restofworld.org/2026/china-rejects-dario-amodei…
• newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier
• pekingnology.com/p/china-signals-what-it-wants-from
• presidentti.fi/en/a-call-for-control-of-frontier-ai-models
• ielp.worldtradelaw.net/2026/09/prospects-for-an-ai-treaty
• cnas.org (Verification, Not Trust, 2026-09-22)
• Metaculus questions 39149, 38418, 21219
• Polymarket gamma API

Question. By the last day of each quarter, has a treaty that caps the compute used to train AI models, with both the United States and the People's Republic of China as parties, been in force for twelve consecutive months without either government formally accusing the other of violating it?

Definitions.

• Treaty: a written international agreement legally binding under international law, whether bilateral or multilateral and regardless of its title or domestic approval procedure. A binding executive agreement counts; a political declaration, voluntary pledge, or nonbinding memorandum does not.
• Caps training compute: imposes an operative, legally binding quantitative upper limit on computation used for AI-model training, either per training run/model or in aggregate over a specified period. The limit may be expressed in computational operations or an equivalent defined compute-accounting unit. A prohibition on training above a specified compute threshold counts, as does a zero-compute moratorium. Limits may cover a specified class of models and may differ between the parties, but must constrain training in both countries. A reporting, evaluation, or licensing threshold that permits training above it without any binding upper limit does not count. Chip export controls, limits on chip stocks or electrical capacity alone, and limits solely on inference or deployment do not count.
• In force for twelve consecutive months: the treaty and its qualifying compute limits have been legally in effect for both countries throughout an uninterrupted twelve-calendar-month interval, ending on the corresponding anniversary date (February 28 if the start date was February 29). Signature or provisional application pending entry into force is insufficient. Reservations exempting either country from all qualifying limits, suspension of those limits, withdrawal becoming effective, or expiry interrupt the interval. Amendments changing the numerical ceiling do not interrupt it if qualifying limits remain continuously binding on both countries.
• Formally accusing: an affirmative assertion, made on behalf of either national government in an official statement, compliance report, diplomatic communication, or treaty-body submission, that the other country has breached any obligation of this treaty, not only its compute cap. Statements by heads of government, ministries, executive agencies, or authorized diplomatic representatives count when recorded in an official publication. An assertion that private actors violated the treaty counts only if the government also attributes treaty noncompliance to the other state. Questions, requests for clarification or inspection, expressions of concern or suspicion, and statements that compliance cannot be verified do not count. Legislative criticism, campaign statements, anonymous briefings, and independent accusations by inspectors or third countries do not count unless adopted by one of the two governments.
• Without accusation: no qualifying accusation was made during the interval on the official public record available at resolution. A subsequently withdrawn accusation still interrupts that interval. Any later accusation-free twelve-month interval may qualify; the treaty need not have been accusation-free since entry into force. Private diplomatic accusations count only if an authorized official source publicly confirms their content and date before resolution.

Resolves YES when an eligible official publication issued on or after completion of a qualifying twelve-month interval establishes that the treaty and qualifying limits remained in force for both countries through that interval, and a review of the official sources finds no qualifying accusation made during it. No affirmative declaration that there were no accusations is required. The interval may begin before the question's opening date, but must end on or after it. Actual undisclosed violations do not prevent YES. Later accusations or later disclosures do not reverse an established YES, under the Resolution window convention.

Resolves NO at a horizon if no qualifying interval has been established by that horizon. Negotiations, signing without entry into force, parallel unilateral pauses, and fewer than twelve completed consecutive months do not suffice.

Resolving source. The authenticated treaty text, annexes and amendments, and entry-into-force/status notices published by its designated depositary or treaty secretariat; the United Nations Treaty Collection (treaties.un.org); the U.S. Department of State's Treaties and Other International Acts Series and Treaties in Force (state.gov); and China's Ministry of Foreign Affairs treaty records (mfa.gov.cn). For a bilateral treaty without a depositary, official entry-into-force notices from either foreign ministry suffice absent conflicting official records. Accusations are established by official publications or transcripts of the White House, U.S. executive departments and diplomatic missions, China's State Council, Ministry of Foreign Affairs, other central-government ministries and diplomatic missions, or official treaty-body records reproducing a government's communication. These are the exclusive resolving sources, rather than the broader Public confirmation convention A2. Apply the Resolution window convention and Resolution date convention A3: resolution is on publication of the qualifying post-interval status confirmation, not automatically on the anniversary. Earlier records may establish entry into force and intervening accusations but cannot alone trigger YES.