Industrial Decarbonization
US hyperscale data centres drew 82 TWh and emitted 25.7 Mt CO₂
Harvard-led study in Nature Sustainability estimates 403 US hyperscale data centres drew ~82 TWh of power and emitted 25.7 Mt of CO₂, May 2024–April 2025, with public reproducibility limits.

Waypoints
Nature Sustainability study published 16 November 2026 estimates 68–99 TWh electricity and 21–31 Mt CO₂ for 403 US hyperscale data centres, May 2024–April 2025
Central scenario: 81.8 TWh, or 1.8% of US electricity consumption, and 25.7 million tonnes of CO₂
Electricity-weighted carbon intensity of ~314 g CO₂/kWh vs US eGRID average of 348; ~54% of attributed generation fossil-fuel based
Data could not separate AI from general-purpose computing; facility records remain restricted under a data-use agreement
Health impacts were not measured and are planned for further research
US hyperscale data centres consumed roughly 82 terawatt-hours of electricity and emitted an estimated 25.7 million metric tonnes of CO₂ between May 2024 and April 2025, according to a peer-reviewed study published in Nature Sustainability on 16 November 2026. That central scenario, led by Harvard researcher Gianluca Guidi and colleagues, represents 1.8% of US electricity consumption. Across operating scenarios, the estimate spans 68–99 TWh and 21–31 million tonnes of CO₂.
For anyone tracking compute-driven demand on the grid, that's the first facility-level historical benchmark for the infrastructure behind cloud and AI workloads — with explicit boundaries on what it does and doesn't measure.
How the numbers were built
The team started from 675 Baxtel-listed facilities, then used geocoding, OpenStreetMap footprints and manual imagery review to validate an analytical sample of 403. The authors deliberately did not extrapolate national totals from the excluded records, calling the underlying data insufficiently complete and consistent.
Power capacity data existed for 397 facilities. A gradient-boosted regression model filled six gaps using building area, climate category and balancing authority. Drop those six and the central estimates shift from 81.8 to 81.0 TWh, and from 25.7 to 25.4 million tonnes. The repository pins the central electricity figure at 81.8 TWh.
The method multiplies total facility electrical capacity — computing equipment, cooling and electrical overhead — by annual hours and a load coefficient. The central scenario assumes average annual draw equals 58% of capacity. A coefficient of 0.663 is a sensitivity scenario, not the central case; the high scenario uses 0.70. That distinction between maximum deliverable capacity and average operating demand matters a lot when you're reading a proposed campus's megawatt rating.
What does the emissions accounting show?
Emissions use EPA eGRID2023 Revision 2 with a total-output denominator: regional CO₂ divided by total net generation, including non-combustion sources. The attribution is location-based. It describes the regional electricity mix behind the demand, not which power plant fed which server.
The weighted carbon intensity works out to roughly 314 grams of CO₂ per kilowatt-hour, against a US eGRID total-output average of 348. About 54% of attributed generation is fossil-fuel based. So the footprint is large in absolute terms, but this sample's electricity is not more carbon-intensive than the national average — a point anyone quoting these figures should keep straight.
Is this an AI footprint? Not exactly
Harvard's announcement says the data could not separate general-purpose computing from AI-specialized operations at each facility, and coverage of US hyperscale centres is incomplete. The totals cannot be assigned wholesale to AI training or inference.
That boundary matters for institutions estimating their own compute carbon accounts. A sector-level total doesn't tell you what a particular university workload consumed. The study supplies regional and methodological context; allocating emissions to a specific experiment still requires data on that experiment's consumption and execution environment.
How far can you reproduce it?
The data-availability statement splits public state and balancing-authority aggregates from restricted facility records. The repository exports are the authoritative public products — not the separate visualization interface. Raw identifiers, addresses, coordinates and commercially sensitive attributes sit behind a data-use agreement.
The public repository (release v0.3.1-round3-final) supports checks of scenario totals, regional aggregates, the emissions denominator and manuscript outputs, plus a synthetic facility fixture for model validation. You can verify the published aggregate arithmetic. You cannot independently rebuild every facility observation without authorized inputs. As a template for data-availability statements, it's a concrete example of specifying what's open, what's controlled and how access works.
Independent reporting added two caveats. WIRED interviewed energy researcher Jon Koomey, who was not involved in the study; he welcomed facility-level evidence while noting some consumption information remains behind nondisclosure agreements, and identified onsite behind-the-meter generation as a gap. The Guardian reported Gina McCarthy's concern about fossil-fuel dependence.
What comes next
Harvard says health impacts were not directly measured and are planned for further research. The authors frame the window as a snapshot of an expanding industry, not a measure of today's full footprint — no updated 2026 total follows from these historical estimates.
For carbon-aware computing decisions, the practical rule stands: keep the time period, regional attribution basis and workload boundary attached to any comparison. Annual regional averages alone can't establish the emissions saved by moving a particular job.
via casrai.org (Source)
More from Elena Vasquez
Show full bio
Senior reporter covering technology and product news in Circular economy and the environmental industry at Circular Wire.
543 articles