What happened

Researchers estimate that 403 US hyperscale data centres accounted for about 1.8% of national electricity consumption during the year ending April 2025. The study, published in Nature Sustainability on 16 November, puts their annual electricity use between 68 and 99 terawatt-hours across different operating scenarios. These are modelled estimates for the studied facilities, not a live reading of today's demand.S2S3R1

The central emissions estimate is about 25.7 million metric tonnes of carbon dioxide. Ars Technica reports a scenario range of 21 million to 31 million tonnes. The Guardian compares the central figure with Denmark's annual carbon emissions, while the Hill's headline describes it as larger; the underlying estimate concerns emissions attributable to electricity generation.S1S2S3R2

Why it matters

The distinction between a model and a meter reading matters here. Nature's accompanying explanation says the analysis uses assumed facility loads and annual regional electricity mixes. It does not trace the electricity actually delivered to each site, and the results apply to the 403 validated facilities.R2

The estimated generation mix was 54% fossil fuels, 25% renewables and 21% nuclear, according to the accounts of the study. Ars Technica says natural gas supplied most of the fossil-fuel share. The figures describe the electricity associated with the sampled centres, rather than every source of energy used by the wider technology industry.S1R2

There is also a boundary around what the analysis counts. Ars Technica reports that it excludes on-site generation behind the electricity meter, citing xAI's use of gas turbines in Tennessee during the study period as an example. It also notes that repeating the analysis could track growth, while cautioning that the study cannot identify AI's electricity use alone.S1

The background

The Harvard-led team used satellite imagery and Environmental Protection Agency data. Harvard's release says the available information did not separate general computing from AI-specific work. That limits what the results can establish about the carbon cost of a particular AI service.R1

Co-author Francesca Dominici told the Guardian that continued construction since the study period makes the results a conservative view of current demand. That is her assessment of an expanding industry, not a measured 2026 total. The publication date and the period analysed are therefore different dates with different meanings.S3

What happens next

The researchers say they plan further work on health effects, which this study did not directly measure. Updated evidence would also be needed to quantify demand after April 2025. Neither the study's historical estimate nor the separate pollution projections discussed in news coverage establish present-day health damage from these 403 facilities.R1S3