Assessing US Hyperscale Infrastructure
A comprehensive new study published in Nature by researchers examining the expansion of artificial intelligence infrastructure has provided the first facility-level estimates for 403 US hyperscale data centers operating between May 2024 and April 2025. Reported by Ars Technica, the research addresses a critical blind spot in energy planning, revealing that these large-scale facilities consume between 68 and 99 terawatt-hours annually. This total translates to approximately 1.5% to 2.2% of total United States electricity consumption, putting the power demand of these computing hubs on par with entire states such as Wisconsin or Indiana.
The rapid proliferation of hyperscale data centers has outpaced the ability of utility providers and regional grids to plan for clean energy integration. As technology companies funnel massive investments into large language models and intensive computing workloads, collecting granular operational data has remained difficult. The research team tackled this by constructing an initial inventory of 675 facilities from commercial databases, news reports, company statements, and satellite imagery. After filtering out sites lacking verified electrical capacity or requisite cooling equipment, they focused their evaluation on 403 verified facilities with an average electrical capacity of roughly 40 megawatts.
Fossil Fuel Reliance and Emissions
According to the findings detailed in Nature, fossil fuel sources supply approximately 54% of the generation attributed to these data centers. Under the study’s central scenario—assuming an average operating capacity factor of 58%—power demand reached roughly 82 terawatt-hours, accounting for about 1.8% of national electricity consumption and generating approximately 25.7 million metric tons of carbon dioxide. Across the various facility-load scenarios modeled by the researchers, total attributable emissions ranged from 21 million to 31 million metric tons of CO2 under a location-based attribution method.
The research notes that the electricity-weighted total-output carbon intensity for these data centers sat at approximately 314 grams of CO2 per kilowatt-hour. This figure compares with a broader US total-output average of approximately 348 grams of CO2 per kilowatt-hour, reflecting the varying regional grids where these facilities cluster. Because hyperscale electricity demand is highly concentrated geographically—with four states accounting for over half of total consumption—the carbon footprint of individual hubs depends heavily on local grid mixes.
Methodological Cautiousness and Grid Stakes
The study’s authors emphasize that their figures represent a conservative baseline. By excluding 272 suspected facilities that lacked public electrical capacity numbers or verifiable infrastructure, the analysis likely undercounts total industry impact. Furthermore, calculating exact usage involves navigating operational variables, as conventional data centers typically operate at around 48 percent capacity while intensive AI training workloads push utilization closer to 70 percent.
These findings carry significant implications for regional grid planning, public health, and regulatory oversight. As grid operators contend with surging demand from technology infrastructure, the localized concentration of heavy power loads threatens to strain local transmission lines and complicate state-level carbon reduction targets. The researchers have released public aggregate data exports through a repository to aid future tracking, while underlying facility coordinates remain restricted under data-use agreements.

