Can We See What We Govern? Testing Whether Open Records See AI Data Centers

Decisions about AI, from export controls and grid planning to the verification of future agreements, assume that large computing facilities can be seen and counted. This paper tests that assumption against open records, the evidence a third party can actually check. In a pilot on 35 U.S. data centers, the EPA’s national facility registry found 23 of 25 sites drawn at random from Virginia’s list of data-center air permits, but only 6 of the 10 largest AI sites elsewhere. The national inventory of power plants found none of the 35, which is expected for sites without a power plant of their own. In France, data centers appear only one by one in a register of industrial sites, and the grid operator has withdrawn the coordinates of its lines for security reasons. At world scale, we combine two open lists: the Atlas of Data Center Politics (9,865 records in 126 countries) and Epoch AI’s 94 largest AI sites. The Atlas lists 63 of the 94 Epoch AI sites; 2 were checked and not found, and 29 could not be checked. Of all Atlas records, 22% cannot be placed below the country. Public statements of national data-center power in 23 countries differ more by what they count than by how they count it: in Russia, figures run from 0.84 GW for commercial data centers to about 5 GW for all data centers connected to the grid. The world and France data layers behind this paper are published as open data with DOIs (see Related works). The U.S. pilot cell matrix and the registry of national statements will follow. Every cell in the audit is in one of three states: found, checked, and not found, or not checked.