P045 Detection thresholds for undeclared facilities

What is the smallest undeclared facility that combined national technical means and supply-chain tracking can detect with high probability and no prior knowledge of its location?

Smallest undeclared facility that can be detected.

Status

Analyzed A literature-based estimate assuming the facility is not deliberately hidden; the same source says that with active concealment facilities of 10,000 to 100,000 H100-equivalents cannot be ruled out, that the 10,000 figure rests on a single open analysis and 'does not establish a lower bound on the smallest detectable facility', and that the real bound depends on what intelligence agencies achieve. No detection probability is stated, so this survey's detection goal of 90% is not shown. Supply-chain clause: a desk bound of 'on the order of a million chips unaccounted for' from Epoch's chip tracing. Nothing built or tested; the underlying classifier (Clymer 2025) is not held.

"In conclusion, facilities at 10,000 H100-equivalents are still mostly detectable, primarily through satellite surveillance. Combining methods that are independent of each other raises the detection probability further. Below this scale, distinguishability drops sharply and detection becomes much harder."
Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements (arXiv 2607.22619v1), §5.3.1 (2026 (arXiv 2607.22619v1; stamp reads 16 Jun 2026)), L478–481

Posed by

Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Rahman (2026), Does Distributed Training Undermine Compute Governance? (MIRI); Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements.

Sub-questions

IDSub-questionWhat the sources report
P045.1What is the smallest covert facility (chips or MW) that national technical means plus supply-chain tracking detect with at least 90% probability before a 10^24 FLOP run completes, with no prior location knowledge?open
P045.2Can clusters of <= ~100 kW (one rack, a container, ~8 H100s) be detected by any thermal, power, RF or HUMINT signal at a stated false-positive rate?open
P045.3Below what per-site size do remote sensing plus energy monitoring fail to flag a site, when compute is distributed across many small sites?open
P045.4At what inter-site bandwidth does a split frontier run stay within a stated cost factor of a single-site run (so facility size gives no signal)?a source suggests an answer (not checked)
P045.5Is there a physical lower bound (MW) below which an AI facility's heat, power and RF emissions fall within commercial-building variance?open
P045.6Are satellite-classifier and load-shape-classifier errors statistically independent on a labelled site sample?implied by a source, not stated as open

Sub-questions not yet stated precisely enough to check

IDSub-question
P045.7Under active concealment, what facility sizes (e.g. 10,000-100,000 H100e) can be ruled out?

Possibly relevant work found by automated search

Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.

For P045.1: What is the smallest covert facility (chips or MW) that national technical means plus supply-chain tracking detect with at least 90% probability before a 10^24 FLOP run completes, with no prior location knowledge?

For P045.2: Can clusters of <= ~100 kW (one rack, a container, ~8 H100s) be detected by any thermal, power, RF or HUMINT signal at a stated false-positive rate?

For P045.5: Is there a physical lower bound (MW) below which an AI facility's heat, power and RF emissions fall within commercial-building variance?

Data