Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · Research data infrastructure

Study maps hyperscale data-centre emissions

A Harvard-led study models US hyperscale data-centre electricity use and emissions for May 2024-April 2025, with public regional data and code.

Published 16 Nov 2026· 5 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Researchers led by Harvard published a peer-reviewed study in Nature Sustainability on 16 November 2026 estimating electricity use and carbon emissions for 403 US hyperscale data centres. The analysis covers May 2024–April 2025, giving universities and research offices a historical benchmark for the infrastructure behind cloud computing and AI, with explicit limits on what it measures.

What the study estimates

The paper by Gianluca Guidi and colleagues estimates approximately 68–99 terawatt-hours of electricity use and 21–31 million metric tonnes of attributable CO₂ across operating scenarios. Its central scenario puts demand at approximately 82 TWh, or 1.8% of US electricity consumption, and emissions at 25.7 million tonnes.

The authors’ reproducibility repository gives the central electricity figure more precisely as 81.8 TWh. It uses a facility-load coefficient of 0.58: average annual electricity draw is assumed to equal 58% of total facility electrical capacity. A coefficient of 0.663 is a separate sensitivity scenario, not the central estimate. The high scenario uses 0.70; changing the assumed load changes electricity consumption and attributable emissions together.

How the facilities and emissions were modeled

The supplementary methods describe an initial inventory of 675 Baxtel-listed facilities. Geocoding, OpenStreetMap footprints and manual imagery review produced the validated analytical sample of 403. The authors do not extrapolate national totals from the excluded or unresolved records, whose information was insufficiently complete and consistent.

Power capacity was available for 397 facilities; a gradient-boosted regression model filled six gaps using building area, climate category and balancing authority. Excluding those six facilities changes the central estimates from 81.8 to 81.0 TWh and from 25.7 to 25.4 million tonnes of CO₂. That check concerns missing capacity values; it does not turn the assumed operating loads into observed consumption.

Total facility capacity includes computing equipment, cooling and electrical overhead. The calculation multiplies that capacity by annual hours and the scenario coefficient. The supplemental analysis therefore distinguishes maximum deliverable electrical capacity from average operating demand, a distinction that matters when interpreting a proposed campus’s megawatt rating.

For emissions, the repository’s accounting notes specify EPA eGRID2023 Revision 2 and a total-output denominator: regional CO₂ emissions divided by total net generation, including non-combustion sources. The resulting attribution is location based. It describes the regional electricity mix associated with demand, rather than identifying a particular power plant supplying each server.

The repository reports an electricity-weighted carbon intensity of approximately 314 grams of CO₂ per kilowatt-hour, against a US eGRID total-output average of 348. Roughly 54% of attributed generation is fossil fuel based. The substantial absolute footprint therefore does not mean this sample’s average electricity is more carbon intensive than the national comparison. For research carbon accounting, both the amount of electricity and the emissions factor need to remain visible.

A hyperscale footprint, rather than an AI-only total

Harvard’s announcement says the available data could not separate general-purpose computing from AI-specialized operations at each facility. It also acknowledges incomplete coverage of US hyperscale centres. The figures cannot therefore be assigned wholly to AI training, inference or academic computing.

That boundary matters for an institution estimating its own compute carbon footprint. A sector-level total does not identify the electricity consumed by a particular university workload. In practical terms, this study provides regional and methodological context for such accounting; allocating emissions to a specific experiment would require information about that experiment’s consumption and execution environment.

In independent reporting, WIRED interviewed energy researcher Jon Koomey, who was not involved in the study. He welcomed facility-level evidence while noting that some consumption information remains behind nondisclosure agreements. WIRED also identified onsite, behind-the-meter generation as a gap in the analysis. The Guardian reported Gina McCarthy’s concern about fossil-fuel dependence.

What researchers can reproduce publicly

The paper’s data-availability statement distinguishes public state and balancing-authority aggregates from restricted facility records. It identifies the repository exports as the authoritative public products, rather than the separate visualization interface. Raw identifiers, addresses, coordinates and commercially sensitive attributes remain subject to a data-use agreement; research access requires the relevant approvals.

The public repository supports checks of scenario totals, regional aggregates, the emissions denominator and manuscript outputs. Its tagged release also supplies a synthetic facility fixture for model-validation workflows. The reproducibility documentation explains that synthetic split indices identify synthetic rows, not the restricted analytical facilities. Full reproduction of the facility workflow requires authorized inputs placed in the documented local folders.

For research offices reviewing reuse, the practical distinction is between checking published aggregate arithmetic and independently rebuilding every facility observation. The release enables the former without granting the latter. It offers a concrete example for a data availability statement: specify which outputs and code are open, which inputs are controlled, and how authorized access works.

Research questions that remain

Harvard says health impacts were not directly measured and are a planned subject of further research. The authors also describe the facility window as a snapshot of an expanding industry, rather than a measurement of today’s full footprint. No updated 2026 total follows automatically from the historical estimates.

For universities considering carbon-aware computing, the immediate implication is to keep the time period, regional attribution basis and workload boundary attached to any comparison. The study makes those assumptions inspectable. A comparison of annual regional averages alone cannot establish the emissions saved by moving a particular computing job.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →