The race to build AI data centres is reshaping American landscapes, power grids and watersheds, largely without coordinated planning. A Cornell-led study in Nature Sustainability, headed by energy systems engineer Fengqi You with doctoral student Tianqi Xiao as first author, quantifies the stakes. At current growth rates, the analysis projects, US AI infrastructure would emit 24 to 44 million metric tons of carbon dioxide a year by 2030, comparable to adding 5 to 10 million cars to the roads, and consume 731 to 1,125 million cubic metres of water, rivalling the household use of 6 to 10 million Americans.
The team assembled roughly three years of financial, marketing and manufacturing data on the industry's expansion, combined it with location-specific power-system, resource and climate data, and modelled multiple regulatory scenarios. Siting emerged as the single most important factor: many facilities are rising in water-scarce Nevada and Arizona, while the Midwest and windbelt states such as Texas, Montana, Nebraska and South Dakota offer the best combined carbon and water profile. New York rates as a low-carbon option with water-efficient cooling.
The study's central message is that no single remedy suffices. Smart siting, faster grid decarbonisation and operational efficiency together could cut carbon impacts by about 73 per cent and water impacts by about 86 per cent against worst-case scenarios, while any one lever alone leaves the industry's net-zero targets out of reach. The work was supported by the National Science Foundation and the Eric and Wendy Schmidt AI in Science programme.