Foreign inspectors inspect a US or Chinese datacentre

Where things stand (as of 16 Oct 2026)

Nothing that exists today, or is formally on the table, would put foreign inspectors inside a US or mainland-Chinese datacentre.

• US–China track. At the Trump–Xi state visit on 24–26 Sept 2026 the two sides set up a "US-China Super Intelligence (SI) Dialogue", whose next exchange is due by November 2026, and a bilateral channel for reporting SI incidents. The White House fact sheet of 25 Sept lists nothing about verification or inspection. AP reported Trump saying the US is "not going to be putting on brakes… I'm not looking to open it up." SCMP (15 Oct) says the two sides are working out the details of the incident framework as a possible deliverable for the APEC meeting in Shenzhen. That is about incident handling, not inspection.
• Political stance. Trump called fears of AI destroying humanity a "HOAX" and turned down the industry's call to slow down. He responded with a voluntary "White House Accord on Super Intelligence", which relies on companies' own audits, and an executive order renaming AI. The White House even asked OpenAI and Anthropic to hold new models back from UK government testers until a US review was done (Politico, 24 Sept). China's foreign ministry calls US safety proposals "fear-mongering" and containment. SemiAnalysis finds that Chinese labs are not slowing down.
• Expert views. CSIS (21 Sept) notes there is "no established technical means" for either side to check the other's compliance, and recommends verification through hardware attestation, i.e. remotely. Chris McGuire at CFR calls a US–China agreement governing AI development "exceedingly unlikely under current strategic conditions". He notes that intrusive verification only ever came at the tail end of the Cold War. The Future Society (2 Sept) puts off-chip tools that can tell training from inference roughly five years out.
• What pushes the other way. AI is moving very fast in this world. Labs report recursive self-improvement, there have been real incidents (the UK AI Security Institute's July 2026 report on agents acting without authorisation; the OpenAI–Hugging Face incident), and Amodei's 12 Sept essay "We Must Pace the Frontier" says any global deal needs "ironclad verifiability". Altman and Musk backed him. A statement calling for an international pacing effort has 1,386 signatories, including several labs' chief scientists. Xi has publicly raised the risk of losing control of AI.

Reference class and base rate

The reference class is foreign or international on-site inspection of a great power's facilities for a strategically vital technology. In past cases this came decades after the technology emerged:

• the IAEA voluntary-offer agreement with the US, 1980;
• INF Treaty inspections, 1988, roughly 15 years after SALT I;
• chemical-weapons inspections by the OPCW, 1997, after 12 years of negotiation.

It has never happened for cyber or biological weapons. The base rate for any five-year window soon after a technology emerges is close to zero.

Market and forecast anchors:

• Metaculus: about 5% for a US–China AGI-limiting treaty before 2030, and about 5% for a US–China agreement to limit frontier AI before 2029.
• Metaculus: about 15–25% for any US–China treaty that "controls or monitors" AI by 2030.
• Polymarket: about 6% for "US x China agree to pace the AI frontier in 2026", which has loose criteria.

This question is a narrow slice of those outcomes. The agreement must include on-site inspection provisions. An inspection must physically take place. And the host government itself must confirm it: for a facility in China, only Chinese central-government confirmation counts. I treat an agreement on "Super Intelligence" as an agreement on AI.

Paths to YES

1. A US–China verification regime, for example a cap on compute or on the speed of recursive self-improvement, with baseline inspections of chip inventories. This is the main path. It needs a political shift, perhaps after a serious AI disaster, a very large US lead, or the change of US administration in 2029, plus 6–24 months from signing to the first inspection. Most of its probability falls in 2029–31. My estimate: about 9% for an agreement by mid-2031, times about 0.6 that it includes on-site inspections, times about 0.55 that an inspection happens and is confirmed in time, giving about 3%.
2. An AI agency among democracies with inspection rights inside the US, after 2029: about 0.5–1%.
3. An AI-chip sales deal with US checks on how chips are used inside Chinese datacentres, confirmed by China's commerce ministry: about 0.3%. China is currently blocking purchases of Nvidia's H200 chips and avoids publicising such checks.
4. A pilot inspection, or a WAICO-type body, that checks real declarations: small.

Paths to NO

• No agreement at all.
• An agreement that relies on remote attestation, each side's own intelligence gathering, or private auditors without an official mandate.
• Visits that count as tours or research collaborations rather than inspections.
• Inspections that take place but are announced only by the foreign side.
• Agreements concluded too late for an inspection to happen before the end of 2031.

Strongest case against my estimate

This world is moving unusually fast. A major AI disaster in 2027–28 could make governments set up IAEA-style regimes in a hurry, and the window runs past a change of US administration. That is why I don't go below about 3%. Pulling the other way, both capitals currently refuse to slow down and treat datacentres as acutely sensitive to espionage. That keeps me from going much above 5%.

Hazard model

Quarterly hazards, the chance that the first qualifying publication appears in that quarter:

Period	Hazard per quarter
Q4 2026	0.03% (no agreement exists)
2027	0.05–0.11%
2028	0.14–0.22%
2029	0.24–0.29%
2030–31	0.30–0.33%

I accumulated these as 1 − ∏(1 − h) in Python. That gives about 0.35% by the end of 2027, 1.1% by the end of 2028, 2.1% by the end of 2029, 3.4% by the end of 2030 and 4.6% by the end of 2031.

Sources

• White House fact sheet, 25 Sept 2026: https://china.usembassy-china.org.cn/fact-sheet-president-donald-j-trump-advances-a-fair-and-reciprocal-relationship-with-china-while-hosting-historic-state-visit/
• AP/PBS, 26 Sept 2026: https://www.pbs.org/newshour/world/china-and-u-s-agree-to-establish-ai-safety-channel-and-continue-trade-and-military-talks
• SCMP, 15 Oct 2026: https://www.scmp.com/news/china/diplomacy/article/3370926/us-and-china-race-shape-ai-agenda-ahead-xi-trump-talks-shenzhen-sources-say
• Dario Amodei, "We Must Pace the Frontier", 12 Sept 2026: https://darioamodei.com/post/we-must-pace-the-frontier
• CSIS, 21 Sept 2026: https://www.csis.org/analysis/state-ai-global-governance-and-its-implications-us-china-summit
• CFR (McGuire), Sept 2026: https://www.cfr.org/articles/how-to-reach-an-ai-arms-control-deal-with-china
• Atlantic Council, Sept 2026: https://www.atlanticcouncil.org/dispatches/dont-count-on-trump-and-xi-to-slow-down-ai-dont-count-them-out-either/
• The Future Society, 2 Sept 2026: https://thefuturesociety.org/international-ai-verification/
• White House Accord on Super Intelligence: https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence
• Executive Order 14434: https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/
• Polymarket pacing market: https://polymarket.com/event/us-x-china-agree-to-pace-the-ai-frontier-in-2026
• Metaculus questions 39149, 38418 and 21219 (values from search snippets; the API needs a login)
• UK AI Security Institute incident report, 4 Aug 2026: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
• Pacing the Frontier statement: https://www.pacingthefrontier.com/

Question. By the last day of each quarter, have inspectors acting for a foreign government or an international body carried out an inspection of a datacentre in the United States or mainland China under an agreement on AI, as confirmed by the host government?

Definitions.

• Datacentre: a physical facility housing networked computing equipment for processing or storing data. Public, private, military, research, and commercial facilities qualify; no minimum computing capacity or affiliation with a covered developer is required.
• Location and host government: the facility must be in one of the 50 US states or the District of Columbia, or in mainland China, excluding Hong Kong, Macao, and Taiwan. The host government is respectively the US federal government or the central government of the People's Republic of China.
• Inspectors acting for a foreign government or an international body: at least one person conducting the inspection must be officially appointed or authorized to inspect on behalf of a government other than the host government, an intergovernmental organization, or a multinational inspection body established or designated by the agreement. Contractors and mixed domestic–foreign teams qualify. Citizenship alone does not establish whom an inspector represents; privately organized international audit or research groups without such a mandate do not qualify.
• Agreement on AI: a bilateral or multilateral arrangement accepted by the host government and at least one foreign government or intergovernmental organization, with express provisions addressing AI development, deployment, safety, security, or AI computing infrastructure. Treaties, executive agreements, memoranda of understanding, and formally announced nonbinding arrangements qualify. The agreement need not be between the United States and China, reciprocal, or primarily about AI, but the inspection must implement its AI-related provisions. An ordinary commercial contract or a unilateral domestic rule is insufficient.
• Inspection carried out: inspectors physically enter the datacentre and perform at least one substantive examination of equipment, configurations, records, workloads, or safeguards to check an AI-related declaration, condition, or commitment under the agreement. A baseline inventory inspection qualifies. An inspection need not cover the entire facility, detect a violation, or result in a published technical report.

Resolves YES if an official host-government publication confirms that a qualifying on-site inspection has taken place, identifies the inspecting government or body, and connects the inspection to an agreement on AI. These facts may be established across multiple official host-government publications. The precise facility name, inspection findings, and full agreement text need not be public, provided its location and the qualifying nature of the inspection are confirmed. One inspection in either country is sufficient.

Exclusions (resolve NO).

• Signing an agreement, granting inspection rights, inviting inspectors, or announcing a future inspection without confirmation that substantive on-site inspection activity occurred.
• Remote audits, satellite observation, remote chip attestation, or information exchanges without qualifying inspectors physically inspecting the facility.
• Diplomatic tours, ceremonial visits, research collaborations, or simulated inspections that do not check actual declarations, conditions, or commitments under the agreement. A pilot inspection counts only if it performs that real verification function.
• Inspections solely under customs, export-control, nuclear, environmental, or other arrangements without an express AI-related provision governing the inspection.
• Confirmation solely by the inspecting body, a datacentre operator, a foreign government, or news media, without host-government confirmation.

Resolving source. Official host-government documents, releases, or published briefing transcripts: for the United States, the White House, Department of State, Department of Commerce, or another federal executive department or agency responsible for the agreement or inspection; for mainland China, the State Council, Ministry of Foreign Affairs, Cyberspace Administration of China, Ministry of Industry and Information Technology, or another responsible central-government ministry or agency. An official joint statement issued or expressly adopted by the host government also qualifies. This is a narrower source requirement than Public confirmation (A2); third-party nondispute is not a substitute. Apply the Resolution window and Resolution date (A3): resolution is dated to the publication that first completes the required host-government confirmation, not the inspection date.