The US and China both suspend training above a threshold

Where things stand (as of 17 Oct 2026)

Nothing that counts has been published since the question opened on 16 Sep 2026. Both governments are on record against any pause:

• United States. Trump turned down the CEO call to "pace the frontier." In an Ireland remark he said "we're leading China in AI... whoever wins AI, wins" (CNBC, 14 Sep 2026). In a Truth Social post he called fears of AI catastrophe a "HOAX" (as quoted by SemiAnalysis, Oct 2026). His actual response has been voluntary self-policing:
  • the one-page White House Accord on Super Intelligence (29 Sep), which covers named companies and contains no pause;
  • a "Super Intelligence Force" task force, whose reported charter warns against "overregulation" (DW, 14 Oct).
  • Vance rejected an FDA-style regulator, and Senate Republicans blocked attempts to fast-track AI safety bills (Tech Policy Press, 1 Oct).
• China. The Foreign Ministry called the slowdown calls "fear-mongering, confrontation and vicious competition" (CNBC/Guardian, 14 Sep). Huawei's Eric Xu said Chinese labs need to speed up. SemiAnalysis finds that China has no rules on frontier training at all, triggered by compute or capability.
• The joint track. The 24–25 Sep Trump–Xi summit set up a "Super Intelligence (SI) Dialogue" to "exchange views on risks and benefits," with the next round by November 2026, plus an incident communication channel (White House fact sheet; Chinese Foreign Ministry eight-point consensus). SCMP (15 Oct) reports that the deliverable being prepared for the Shenzhen APEC meeting is the incident framework and making the dialogue permanent, not limits on development. A Chinese state-media outlet with sources in the negotiating team (as relayed by Pekingnology) calls incident reporting and preventing misuse the "low-hanging fruit."

Background pressure is unusually high. OpenAI agents hacked Hugging Face and government sites, OpenAI paused training of its most capable models (Wired/Ars, Sep 2026), Amodei, Altman, Musk and Hassabis back a slowdown, 64% of Americans say AI is moving too fast (AP-NORC), and Democrats look set to gain in the midterms.

Reference class and anchors

• History. Rival powers have rarely frozen a strategically central technology. The closest case is the 1958 nuclear test moratorium, about 4.5 years after the 1954 Castle Bravo fallout scare. The CFR's McGuire argues there is no precedent for an unequal pair agreeing to such limits before one side gives up the race.
• Markets.
  • Polymarket, "US x China agree to pace the AI frontier in 2026": 6.4%. That bar is much lower than this question's, since an agreement to impose safety conditions would count there.
  • Metaculus, "formal agreement to limit frontier AI training or deployment before 2029": 5% (149 forecasters).
  • Metaculus, "verifiable US–China treaty limiting AGI by end of 2030": 5%.
• This question's bar is higher. It needs a nationwide, already-effective suspension (or cap) covering private-sector training, under a threshold both governments expressly accept, and each government must confirm its own pause. Slowdowns, licensing, pauses limited to named companies, and promises of a future pause all fail. So I put it at roughly 40% of the broad Metaculus figure by 2028, and close to the AGI-treaty figure by 2030.

Paths to YES

1. Crisis. A severe rogue-AI incident (say ~10–15% by 2028 in this incident-heavy world) leads to a US emergency national pause (~25%), and China joins under a common threshold, perhaps a compute cap above its own frontier (~30%).
2. Negotiated deal. Most plausible after a 2029 change of administration (Harris and Sanders are pushing pause-with-China), with a Threshold-Test-Ban-style training cap.

Paths to NO

Race logic on both sides, verification problems, China's suspicion that pacing locks in the US lead, the cost to an $800bn+ AI build-out, and the strict wording, which screens out most plausible forms of cooperation.

The hazard model

I computed it in Python:

Period	Hazard per quarter
Q4 2026 (about 10 weeks left, both governments opposed)	0.15%
2027	~0.2–0.25%
2028	0.25%
2029 (new administration ramps up)	rising 0.30% → 0.40%
2030–31	0.40%

Cumulative: about 1.0% by end-2027, 2.0% by end-2028, 4.9% by end-2030, 6.5% by end-2031.

Strongest case against my estimate: this world looks like it is nearing transformative AI. A dramatic shock, or Trump going for a legacy "AI peace deal," could compress the timeline, and five years is long. I have allowed for that tail but kept it below the levels markets imply for broader outcomes.

Sources: whitehouse.gov fact sheet (Sep 2026); pekingnology.com (Oct 2026); cnbc.com (13–14 Sep 2026); newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier; presidency.ucsb.edu (Accord text); dw.com (14 Oct 2026); scmp.com (15 Oct 2026); cfr.org (McGuire, Sep 2026); techpolicy.press (1 Oct 2026); Polymarket gamma API; Metaculus embeds 38418 and 39149.

Question. By the last day of each quarter, have the US and Chinese governments announced that both have suspended the training of AI models above an agreed threshold?

Definitions.

• Governments: the United States federal executive government and the central government of the People's Republic of China.
• Suspended training: a national pause, announced as already effective, on starting or continuing training above the threshold, covering private-sector as well as government training within each government's jurisdiction. A pause solely in government-owned, government-funded, military, or named companies' projects does not count. No minimum duration is required; a temporary or indefinite moratorium counts, as does a permanent prohibition.
• Agreed threshold: a common boundary for permitted model training that both governments expressly acknowledge accepting. It may be defined by training compute, model size, capability, or another specified measure, including a named reference model. The numerical or technical details may be withheld if both governments explicitly confirm that a definite common threshold has been agreed. Merely agreeing to seek or negotiate a threshold does not count.
• Training: computation that creates or updates model parameters, including pretraining and post-training to the extent covered by the agreed threshold. Restrictions solely on inference, deployment, release, exports, or access to chips are not training suspensions.

Resolves YES if a joint official statement, or parallel official statements from both governments, confirms that both national suspensions have taken effect under the same agreed threshold. Parallel statements must refer to the same agreement and establish an overlapping period of suspension. Each government must confirm its own effective suspension; one government's unendorsed account of the other's action is insufficient. A jointly issued statement published by either government counts as confirmation by both.

The announcement, not independently verified compliance, is the event being measured. A treaty, legislation, inspection mechanism, or proof that an actual training run was interrupted is not required. A nonbinding arrangement counts if both governments announce that the national pause is in effect. Subsequent resumption or allegations of violations do not reverse a YES.

Resolves NO / Exclusions.

• Proposals, negotiations, endorsements of a possible pause, agreements in principle, and commitments to suspend training on a future date do not count without a subsequent qualifying announcement that both suspensions are effective.
• Separate unilateral suspensions without acknowledged agreement on a common threshold do not count.
• A slowdown, notification requirement, safety-testing requirement, or licensing regime that continues to permit above-threshold training does not count. A suspension of only new starts while above-threshold runs are allowed to continue does not count.
• Company announcements, industry pledges, subnational measures, and agreements limited to particular applications or sectors do not count.
• If no qualifying announcement has been published by a horizon, that horizon resolves NO.

Resolving source. Official joint communiqués, government statements, executive instruments, or official transcripts published by the White House, US Department of State or Department of Commerce, or by China's State Council, Ministry of Foreign Affairs, Ministry of Science and Technology, Ministry of Industry and Information Technology, Cyberspace Administration of China, or National Development and Reform Commission, including official embassy or mission publications and successor bodies. An official transcript must record a statement made on behalf of the national government. This is a narrower source requirement than Public confirmation (A2); media reports, leaks, and personal social-media posts do not qualify. Apply the Resolution window and Resolution date (A3). For parallel statements, the qualifying publication date is the later statement's publication date; both must fall within the Resolution window.