MAIN CONTENT
The electricity behind artificial intelligence
Gianluca Guidi and colleagues, including Harvard researcher Francesca Dominici, assessed 403 hyperscale facilities operating from May 2024 through April 2025. They combined information about individual sites with regional electricity generation and emissions records. This geographical approach matters because a data center’s climate footprint is influenced by the power system supplying its demand, not merely its racks of processors.
Across different assumptions about operating loads, the facilities were estimated to consume between 68 and 99 terawatt-hours and generate 21 million to 31 million metric tons of attributable carbon dioxide. The central estimate was 82 terawatt-hours and 25.7 million tons. These numbers encompass facilities supporting cloud storage and other computing alongside AI; they cannot be described as emissions caused exclusively by generative AI.
Location determines part of the footprint
According to Nature Portfolio’s summary of the published research, fossil fuels supplied roughly 54% of the electricity attributed to the facilities. Nuclear energy accounted for about 21% and renewables for approximately 25%.

















