The Hidden Environmental and Climate Footprint of Frontier AI: Evidence from Google’s Gemini Training and Regional Emissions ^†

Recent advances in artificial intelligence (AI) have generated widespread interest and investment across industries, yet the environmental and public health costs of large-scale model training remain poorly understood. This paper investigates those hidden impacts by examining Google’s data center in The Dalles, Oregon, which is believed to have hosted key training phases for Gemini 1.0 in mid-2023. The study addresses three central questions. First, does AI training rely primarily on clean, local energy, or does it use electricity imports from fossil fuel-heavy regions? Second, what are the environmental and human consequences of that marginal energy use? Third, how should infrastructure and energy policy respond to this growing demand? Using high-frequency data from the U.S. Energy Information Administration (EIA-930), we find a sharp rise in electricity imports into the Bonneville Power Administration (BPAT) region, which encompasses The Dalles, starting in mid-2023. This shift aligns with Gemini 1.0’s suspected training period and departs from prior trends. Economic and weather factors do not explain the increase, suggesting that AI-driven demand exceeded local capacity, which relies heavily on hydroelectricity. To trace the origin and consequences of this imported electricity, we construct a directional flow measure from PacifiCorp East (PACE) to BPAT, passing through PacifiCorp West. PACE includes several coal-fired plants in Wyoming. Our initial findings show that during the Gemini training period, emissions at two major plants—Jim Bridger and Naughton—increased substantially. At Jim Bridger, the relationship between directional flow and emissions nearly doubled for CO2 and rose by over 70 percent for SO2. These findings suggest that AI-related demand in Oregon contributed to pollution in distant, downwind communities. This case illustrates a “triple externality”: local health burdens, regional and global emissions, and public infrastructure costs. For example, the $800 million Boardman-to-Hemingway transmission line, designed to move electricity westward, is partly justified by rising demand from data centers. Subsequent analysis will examine whether these increased emissions had measurable effects on health and labor supply. I plan to link hourly plant-level emissions data from the EPA’s Continuous Emissions Monitoring System (CEMS) with worker-level labor supply data from Homebase (geo-identified to the firm and ZIP code level), and local hospital records (where available). This will allow for econometric identification of downstream impacts on productivity and respiratory illness in affected communities.