The US and China sign an AI agreement with chip or datacentre verification

Where things stand (mid-October 2026)

Nothing qualifying exists yet, and nothing under negotiation comes close. Xi's state visit to Washington (23–25 September 2026) produced only two AI outcomes:

• a "Super Intelligence (SI) Dialogue", meeting by November 2026; and
• an agreement to set up a channel for reporting AI incidents.

Both the White House fact sheet and China's "Eight Deliverables" release confirm this. Neither mentions verification, chips or inspections.

• 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/
• https://www.mfa.gov.cn/eng/xw/zyxw/202609/t20260926_12031663.html (26 September)

Even the incident channel is not finalised. Per WIRED (September 2026), it still lacks firm Chinese sign-off and officials expect it to take weeks.

Both governments' positions point away from verification:

• Trump called AI-doom fears a "hoax." He said he would "rather not integrate because we're leading by a lot" (UPI, 26 September). His science adviser told the UN Security Council the US rejects "new global governance structures" (Faegre Drinker, October 2026). Bessent's proposal is a notification system only (TNW/Axios, 3 October).
• China's foreign ministry called Amodei's "pacing" proposal "fear-mongering" (Rest of World). Beijing signals that the "low-hanging fruit" is incident reporting and shared definitions of risk, not constraints (Pekingnology, 1 October).
• Chips: Beijing has largely refused even legally offered Nvidia chips over security worries (Brookings, June 2026). That makes a chip deal with on-chip tracking or US inspections of Chinese datacentres implausible.
• Technology: Verification tools are judged years from maturity (CNAS, 22 September; The Future Society, 2 September; CSET, October 2026).

Reference class and comparable forecasts

History. No US–China arms-control agreement has ever had intrusive verification. China still refuses nuclear arms-control talks. The 1963 crisis-driven Limited Test Ban Treaty deliberately relied on national intelligence means rather than on-site inspections. Intrusive verification came only late in the Cold War, with the 1987 INF Treaty.

Existing forecasts (Metaculus, read via r.jina.ai on 15 October 2026):

• 6% for a US–China treaty limiting advanced general AI signed before 2030 (Q39149);
• 5% for an agreement limiting frontier AI before 2029 (Q38418);
• 25% for any treaty that "controls or monitors" AI by 2030 (Q21219, only 35 forecasters).

Polymarket prices a US–China "pacing" agreement in 2026 at about 6%. That bar is far looser than this question's (no signature or verification needed), and the price likely reflects a bias toward long shots.

Our question needs more than these: a signed text, plus official documentation that explicitly provides for on-chip mechanisms or datacentre inspections. I treat it as roughly half to two-thirds as likely as a "limiting treaty."

Paths and hazard model

Paths to YES:

1. Under Trump, to January 2029. This needs a crisis-driven reversal, such as a severe incident caused by a runaway AI. I put it at about 2% in total. A deal on chip sales with end-use inspections is under 0.5%, given that China rejects the chips.
2. A new administration from 2029. Polymarket gives Democrats about 64% to win in 2028. Democratic hopefuls, including Harris on 14 October, call for a treaty with China. I assume about a 60% chance such an administration negotiates seriously. Given serious talks, I give about 12% to a signed agreement with qualifying verification by the end of 2031. That works out to about 4.5%. A Republican successor adds about 1%.

Some features of the criteria make YES easier: domestic inspectors count, inspections triggered by specified events count, and verification may cover only one party.

Quarterly hazards (computed in Python, then compounded):

• about 0.1% in Q4 2026 (the November dialogue round, APEC in Shenzhen, G20 in Miami);
• 0.20–0.25% per quarter in 2027–28;
• rising to 0.45% by late 2029;
• 0.50–0.65% per quarter in 2030–31.

Cumulative results: about 1.9% by end-2028, 3.2% by end-2029, and 7.8% by end-2031.

Strongest case against

For a higher number: AI capability in this world is moving very fast, with self-improving systems and rogue-agent incidents. If transformative AI arrives by 2029–30, urgency, AI-designed verification tools and a pro-treaty US president could produce a deal like the AI-2027 scenario. That could justify 15% or more.

For a lower number: There is no precedent, trust is deep-seated low, China resists intrusive checks, and a new administration would have only about three years. Taiwan or a breakdown of the AI race into open rivalry could also close the window.

Weighting fast-AI and other worlds (about 35% versus 65%) gives roughly the same 7–8%, so I keep it.

Question. By the last day of each quarter, have the United States and Chinese governments signed an agreement on AI that provides for verification by on-chip mechanisms or by inspections of datacentres?

Definitions.

• Governments: the United States federal government and the central government of the People's Republic of China, acting through authorized representatives, including ministries or federal agencies. Agreements solely between companies, universities, subnational governments, or unofficial dialogue participants do not count.
• Signed agreement on AI: a written bilateral or multilateral instrument signed on behalf of both governments that contains provisions expressly concerning AI development, training, deployment, use, or the provision of computing resources for AI. Treaties, executive agreements, memoranda of understanding, and signed political agreements count, whether or not legally binding. Signature by both governments is required; ratification, entry into force, and implementation are not. A signed protocol or amendment to an existing agreement can qualify.
• Verification: checking compliance with at least one commitment, restriction, permission condition, or required factual declaration under the agreement. The agreement must authorize or commit to using at least one of the two methods below for that purpose. It need not prescribe every technical detail, but merely agreeing to study, discuss, or negotiate possible verification methods does not count.
• On-chip mechanisms: hardware-implemented or hardware-rooted features of computing chips that measure, record, attest to, or provide verifiable evidence about relevant hardware identity or location, compute usage, workloads, model identity, or permitted operations. Firmware or software using a chip's secure hardware features counts; ordinary software logs running on a chip do not. A hardware restriction or remote-disable capability counts only if the agreement also uses it to establish verifiable evidence of compliance, rather than solely to enforce a restriction.
• Inspections of datacentres: authorized physical, on-site examinations of facilities housing computing equipment used or intended for AI, including examination of equipment, operations, or associated records, for verification under the agreement. Inspectors may be domestic, foreign, international, or independent parties designated under the agreement. Remote document review, model testing through an API, satellite observation, and ceremonial visits alone do not count. No minimum facility size, number of inspections, reciprocity, or geographic location is required.

Resolves YES if qualifying public documentation establishes both governments' completed signatures and a provision for either qualifying verification method. Either method alone suffices. Verification may cover only a specified class of chips, facilities, models, or activities, or only one party's activities. Actual chip installation, inspection, or a finding of compliance is not required. A provision authorizing inspections upon specified triggers counts even if no trigger has occurred.

Exclusions (resolve NO). Announcements of negotiations, unsigned drafts or joint statements, unilateral verification or chip-export requirements, and private-sector arrangements without both governments' signatures do not count. General pledges of transparency, safety-evaluation exchanges, research into verification, or references to unspecified auditing do not count unless the qualifying method and its verification purpose are explicitly established. A datacentre safety, commercial, or cybersecurity inspection unrelated to verifying the agreement does not count. Resolve NO at a horizon if the required evidence has not been published by then.

Resolving source. The signed text and annexes as officially published by either government or an official treaty depositary; or official statements, readouts, or transcripts published by the White House, U.S. Department of State, another signatory U.S. federal agency, China's Ministry of Foreign Affairs, State Council, or another signatory Chinese central-government ministry or agency. Official documentation must explicitly establish signature by both governments and the qualifying verification provision; publication of the full text is not required. These are the narrower resolving sources in place of Public confirmation (A2). Apply the Resolution window and Resolution date (A3): use the publication date of the first qualifying documentation establishing all requirements, not the signature or implementation date.