The US and China agree to limit frontier AI training

Where things stand (13 Oct 2026)

Nothing qualifying has been agreed since the question opened, and neither government wants a training limit.

• The September summit produced talks, not limits. At the 23–25 Sept Trump–Xi state visit, the two sides set up a "US-China Super Intelligence Dialogue" to "exchange views on risks and benefits", plus a channel for reporting AI incidents (White House fact sheet, Sept 2026, https://www.whitehouse.gov/fact-sheets/2026/09/fact-sheet-president-donald-j-trump-advances-a-fair-and-reciprocal-relationship-with-china-while-hosting-historic-state-visit/; PRC Foreign Ministry readout, 24 Sept, https://www.mfa.gov.cn/eng/xw/zyxw/202609/t20260925_12031181.html). The question explicitly excludes both.
• The US side has ruled out slowing down.
  • Trump: the US is "not going to be putting on brakes"; he wants to "leave it exactly where it is", and he added "That is China's position also" (CBS, 26 Sept, https://www.cbsnews.com/news/trump-xi-us-china-ai-trade-summit/; Carnegie/Prescott, 30 Sept).
  • At the UN General Assembly he rejected "any attempt to construct a globalist scheme to control" AI.
  • His executive order 14409 (June 2026) bars mandatory licensing or pre-clearance of frontier model development.
  • His domestic answer is a voluntary industry accord.
• Beijing has rejected the idea too. It called Dario Amodei's 12 Sept "pace the frontier" essay "fearmongering" and a "Cold War playbook" (NBC, https://www.nbcnews.com/world/china/china-ai-slowdown-trump-amodei-altman-threat-cold-war-rcna597631; Global Times). SemiAnalysis (Oct 2026, https://newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier) finds China puts development first, with no compute- or capability-triggered duties on frontier labs.
• The next round agrees on nothing to do with training. It meets at APEC in Shenzhen in November, and its stated agenda is uncontrollable agents and non-state cyber and biological misuse.

Base rate and other forecasts

• History offers no precedent. The US and China have never concluded bilateral arms control that limits a strategic capability. Former Biden official Chris McGuire (CFR) notes there is no historical case of unequal rivals capping a strategically critical technology before the weaker side concedes the race (https://www.cfr.org/articles/how-to-reach-an-ai-arms-control-deal-with-china).
• Closest precedents are narrow or slow. The 2015 cyber pledge, the 2014 climate announcement and the 2024 statement on human control of nuclear weapons show that parallel political commitments are possible, but none restricted development of a strategic capability. Nuclear testing limits came 13–18 years into that arms race, after the Cuban missile crisis.
• Polymarket: about 5% (5.25% on 13 Oct, from its public price feed) for a US–China "pacing" agreement in 2026. Its criteria are broader than this question's, since conditions on releasing or deploying models also count, and small long-shot markets tend to be overpriced.
• Metaculus: about 4–9% for a US–China agreement limiting frontier AI by 2029, and about 5% for a treaty limiting advanced AI before 2030 (from search snippets; I could not open the pages directly).

Paths to YES

1. A crisis forces a deal under Trump (2027–28). Lab leaders are reportedly rehearsing for a major AI-caused cyber catastrophe within 6–12 months (Axios, 9 Oct). A US domestic pause could create demand for China to match it, and Trump is unpredictable and likes making deals with Xi.
2. A new president from 2029. Democrats are moving toward pausing or regulating AI. A deal could include restrictions that apply to one side only, or a pause until a safety condition is met.
3. A shift in the balance of power. If one side concludes it cannot catch up, or the two reach rough parity, an agreement becomes more feasible (CFR). One proposal (Lawfare) trades looser chip export controls for joint pacing.

Paths to NO

• The bar is narrow. The most feasible agreements are excluded: bans on bioweapon misuse, testing standards, incident reporting, and red lines on particular capabilities. Even Amodei calls a cap on the speed of AI building AI "just on the edge of being possible" and a full pause unlikely "any time soon".
• Other obstacles: verification is very hard, both sides fear the other will cheat, and a Taiwan crisis could intervene.

How the numbers were built

• Quarterly chances: about 0.4% now; 0.45–0.5% a quarter in 2027, when a crisis is most likely; 0.3–0.4% a quarter in the 2028 election year; rising to 0.6% a quarter from mid-2029 under a new administration.
• Running totals: about 3.5% by end-2028 (in line with Metaculus) and about 10% by end-2031. The checks were done in Python.

Strongest case for higher: AI is already improving itself, incidents are piling up, every major US lab is asking for coordination, and five years is a long time in AI. A shock on the scale of the Cuban missile crisis could produce a deal within months. I have priced that in as a tail, not the central case.

Question. By the last day of each quarter, have the US and Chinese governments announced an agreement that limits the scale or pace of frontier AI training?

Definitions.

• Governments: the United States federal government and the central government of the People's Republic of China, acting through their heads of state or government or authorized national-government representatives. Agreements only between companies, researchers, subnational governments, or political parties do not count.
• Agreement: a mutually accepted commitment announced in a joint official statement or in parallel official statements confirming the same commitment. A bilateral agreement or a multilateral agreement to which both governments expressly assent counts. A treaty, executive agreement, memorandum of understanding, or non-legally-binding political commitment may qualify; signature, ratification, entry into force, verification, and demonstrated compliance are not required.
• Frontier AI training: training of advanced general-purpose AI models described in the agreement or accompanying official statements as frontier, leading-edge, or an equivalent category, or identified by a training-compute or capability threshold expressly intended to cover such models. This includes pretraining and capability-enhancing post-training. The agreement need not use the English word “frontier,” and is not limited to covered developers.
• Limits the scale or pace: commits to a definite restriction on training, such as a maximum compute budget per run; a cap on aggregate frontier-training compute or compute capacity used for such training; a cap on model size; a maximum number or frequency of training runs; a mandatory interval between runs; or a pause or prohibition on starting, continuing, or scaling training above a specified threshold. A pause until a stated safety or authorization condition is satisfied counts. A numerical limit is not necessary for an expressly defined moratorium, but a general aspiration to slow development is insufficient.

Resolves YES if the public statements establish both governments' acceptance of at least one qualifying training restriction. The restriction may be temporary, apply only to a defined class of models or training projects, or impose asymmetric obligations, including restrictions on only one country's training. An agreed restriction with a specified future start date or activation condition counts when announced, provided the restriction itself has been agreed rather than left for future negotiation. The relevant scope and restriction must be publicly disclosed well enough to establish that they meet these criteria; the full agreement need not be published.

Exclusions (resolve NO).

• Offers, invitations, negotiating drafts, agreements to hold talks, or statements that the governments will explore or negotiate restrictions.
• General commitments to responsible development, risk reduction, cooperation, or avoiding an AI arms race without a definite training restriction.
• Evaluation, reporting, notification, licensing, or inspection requirements alone, unless they expressly prohibit or delay qualifying training pending a defined condition.
• Restrictions solely on deployment, release, inference, particular uses, or developing a particular prohibited capability, without a restriction on the scale or pace of frontier training.
• Export controls, chip-trade arrangements, datacentre policies, or energy restrictions that merely have an indirect effect on training; these count only if the agreed terms expressly restrict frontier-training scale or pace.
• Unilateral policies not accepted as part of an agreement by the other government, or a unilateral assertion of agreement without qualifying confirmation from both sides.

Resolving source. A joint government-authored statement or agreement published by either government, or parallel official government statements, including official transcripts or readouts, published by the White House, US Department of State, US Department of the Treasury, US Department of Commerce, PRC Ministry of Foreign Affairs, PRC State Council, or another national-government body expressly announcing the agreement on its government's behalf. Official government archives of these documents count. This is narrower than Public confirmation (A2): press reporting, company statements, and third-party analyses cannot resolve the question. Apply the Resolution window and Resolution date (A3). For parallel statements, the qualifying date is the publication date of the later statement needed to establish mutual assent. A joint statement requires no separate publication by the other government. A republication or reaffirmation of a pre-opening agreement alone does not count; a newly agreed extension, renewal, or amendment imposing a qualifying restriction does.