Mechanism · Remote detection of data centres
Evidence & limits
On this page
R1Proposed for finding undeclared data centres above an agreed compute threshold
R1 for finding undeclared facilities, because no public work shows a systematic search that finds them.
Assessed use: finding undeclared data centres above an agreed compute threshold
Rubric assessment
- R1 met: Halstead and Larsen describe heat, imagery and other detection signals, ways to conceal a facility and the odds of detecting covert ones 4. Baker et al. place satellite imagery among supplementary verification mechanisms 5.
- R2 not met for this use: Krawec's case studies track known sites 1. Epoch AI reports that its public database finds large facilities mainly through company announcements, news, third-party databases and social media 3. Krawec states that automated data-centre detection "remains primarily conceptual at present" 1.
The level is for the primary use (There is no undeclared relevant compute). For the supporting use of estimating the capacity of known sites (Compute stock is at most a declared amount), the public Epoch database, with its published error estimates, would meet at least R2 2 3.
- Published end-to-end results on finding previously unknown large facilities over a wide area, with measured miss and false-alarm rates.
- An evaluation against a stated concealment adversary, for example disguised or underground facilities.
Assessed 2026-10-08 against rubric v1.1.
Evidence
- Krawec's May 2026 report for the Federation of American Scientists presents two case studies 1. One tracks a facility's construction over 2024 and 2025 against public announcements 1. The other counts infrastructure at an operating AI facility and checks it against utility records and permits 1.
- Epoch AI's hub covered 13 large US data centres at its launch in November 2025 2. Epoch estimated that they held about 2.5 million of the roughly 15 million H100-equivalents delivered to customers worldwide in recent years 2. The data are released under a Creative Commons Attribution licence 2.
- By September 2026 the database still mostly covered US sites, and Epoch estimated its coverage of global deployed AI computing capacity at 44% (90% CI 24–81%) 3. Epoch reports that 80% of its IT-power estimates are within a factor of 1.4 of the actual value, and that its estimates were near-perfect in the two cases where it obtained ground-truth cooling capacity 3.
- Halstead and Larsen are "relatively confident" that before a deal, intelligence services could locate every AI data centre in the country their scenario examines, other than specifically concealed ones, using methods like those of Epoch AI and commercial analysts 4. They estimate the chance that unilateral detection methods find a covert project within ten years of an international agreement 4. For a single site they give 26% at 10 MW, 35% at 100 MW and 50% at 1 GW 4. For ten sites of each size they give 37%, 67% and 85% 4. They call these estimates highly speculative, based entirely on intuition and qualitative reasoning 4.
- Krawec recommends satellite imagery as one element of verification for future agreements, and points to the IAEA's satellite imagery analysis team for nuclear safeguards 1. The report names thermal infrared, synthetic-aperture radar and multispectral imaging as directions for further research 1.
Limitations
- Facilities can be disguised as other industry or placed underground 4.
- A sufficiently small covert project could not be ruled out with confidence 4. Small data centres in existing buildings may lack distinctive signatures 1.
- Without a known location, telling data centres apart from other industrial facilities is hard, and automated detection is mainly conceptual 1.
- Chip counts, chip types and actual energy use cannot be observed 1.
- Cooling and power features vary across companies and countries, and not every site generates its own power 1.
Known flaws
Blockers
Wide-area, automated detection of data centres is not yet practical and needs large training datasets.
No measured detection or false-alarm rates for finding undeclared facilities have been published.
Recent high-resolution imagery is costly, is limited by weather and needs trained analysts.