Mapping the environmental costs of U.S. data centers

About

Data centers powering Artificial Intelligence (AI) are growing rapidly across the United States. It is projected that an average of 65 new data centers will be built per month between 2026-2030, up from 7 per month during 2010-2025. As a key part of modern information technology infrastructure, their environmental and social costs depend critically on where they locate, which is itself determined by local permits and electricity purchase agreements. However, this rapid expansion is raising concerns among policymakers and local communities about their environmental costs, particularly on their carbon emissions and air pollution.

In this project, we are using a comprehensive facility-level dataset of past, operational, and announced U.S. data centers to characterize how the environmental and community characteristics of data center locations evolve over 2010–2030. This will allow us to quantify how much of the projected growth in environmental damages can be attributed to changes in data center locations versus growth in power requirements, which can help policymakers design strategies to mitigate the environmental impact of data center expansion.

Approach

We use a comprehensive dataset of current and announced data centers planned in the future, to analyze trends in environmental costs and siting over time. We then extend our analysis of environmental costs with a formal scale-versus-composition decomposition of projected changes in carbon emissions and monetized local air pollution damages, using marginal emissions factors from Holland et al. (2024). This decomposition isolates the share of projected pollution growth attributable to growth in total data center energy consumption (scale effect) from the share attributable to shifts in the geographic distribution of data centers across locations with differing electricity grid emissions intensities (composition effect).

Key Findings

We find that in the historical period, changes in locational composition of data centers reduced carbon emissions and local air pollution damages that would have otherwise occurred by 4 and 12%, respectively. Looking ahead to projected growth from 2024–2030, we find that scale effects account for roughly 97% of the projected increase in both carbon emissions and local air pollution damages, with compositional shifts in siting contributing to the remaining 3%. Further, we find that data centers consistently locate in less densely populated areas, with planned facilities entering census tracts roughly five times less dense than tracts without data centers. Contrary to patterns documented for other disamenities, data centers are not systematically located in lower-income, higher-poverty, or higher non-White-share communities. Our results imply that policies nudging the location of data centers would do little to mitigate their aggregate environmental footprint, which is governed instead by the scale of buildout, and that permitting and community negotiations over siting will increasingly concentrate in the nation’s least population dense communities. 

Team

Paige Weber (Principal Investigator), Danae Hernandez-Cortes (Principal Investigator), Kyle Meng (Principal Investigator), and Shradhey Prasad (Project Manager).

Partners

This project is a collaboration with UC Berkeley, Arizona State University, and funded by the Alfred P. Sloan Foundation.