UC Riverside

AI's explosive growth has triggered massive data center expansion with largely unexamined public health consequences. While carbon emissions receive attention, this paper quantifies the substantial air pollution impacts throughout AI's lifecycle—from manufacturing to operation. Our analysis reveals training a Llama-3.1-scale model generates air pollutants equivalent to 10,000+ cross-country road trips. By 2030, U.S. data centers could impose annual health costs exceeding \$20 billion—twice that of coal-based steelmaking and comparable to California's vehicle emissions. These burdens disproportionately affect disadvantaged communities, where per-household impacts may be 200 times greater than in less-affected areas. We propose standardized pollution reporting, community-focused impact assessments, and health-informed AI development to advance environmental justice.