AI's health bill: the power plants behind data centers raise questions about public costs

When the race toward artificial intelligence gets measured, the prevailing accounting tends to dwell on gigawatts required, billion-dollar investments and computing capacity. A report published on 10 September 2026 by the Environmental Protection Network (EPN) shifts the view to the indirect health consequences. EPN is a non-profit advocacy network of more than 800 former career staff and former political appointees of the US Environmental Protection Agency (EPA), drawn from both Republican and Democratic administrations; the report is part of the “Safer, Not Sicker” campaign and closes by asking readers to sign a “Data Center Health Protection Pledge”. It is not, then, a third-party document — which takes nothing away from the traceability of the data it cites. Drawing on the estimates of an academic study by researchers at UC Riverside, Caltech and the Rochester Institute of Technology, the report documents that more than 90% of the public health impact is not produced by the server halls themselves, but by the burning of coal and gas in the plants that supply them with electricity. The study estimates between $11.7 and $20.9 billion a year in US health costs by 2028, depending on the scenario; in the high-growth one, at the top of the range, that pollution would contribute to roughly 600,000 cases of asthma symptoms and around 1,300 premature deaths a year.

EPN's analysis is rigorous about the nature and the limits of the available estimates. The academic study's figures rest on impact modelling of stationary combustion and do not incorporate the effect of 30 federal actions taken since January 2025, which the report identifies as a weakening of environmental protections. EPN makes clear that no model has so far quantified the combined health impact of those regulatory measures, and that the two sets of numbers cannot be added together. The document also notes that corporate clean-energy purchasing on an annual basis does not guarantee local decarbonisation: exposure to pollution depends on when and where fossil generation is actually running, and a company can meet its annual clean-energy targets while the grid serving its data centers stays dependent on coal and gas.

The picture that emerges raises questions about how the collateral infrastructure is managed. The report flags at least 74 new gas plants proposed specifically to power data centers, and at least 188 schools located within three miles — about 4.8 kilometres — of those plants. Geographic proximity alone is not proof of direct exposure, but it calls for close, site-by-site monitoring. As for EPN's objections and its call for a health protection pledge, the sources consulted contain no public response from the EPA, the Department of Justice or the companies in the sector.

Shifting the operating costs of innovation onto local communities' health bills is not an industrial strategy but an accounting deferral. If model efficiency does not count the externality of the plants standing behind it, the convergence between technological transition and public protection remains a pure statement of intent. — Olya