The Jerusalem PostUS Jewish Advocacy Center, StandWithUs open lawsuit over 2024 antisemitic assault in New YorkDaily MaverickAFTER THE BELL: Durban port’s crane crisis threatens SA’s private sector logistics driveInquirerPhilHealth targets historic 100% coverage for 18,332 Bukidnon Lumads20 MinutenApple-Dienste kämpfen mit ProblemenBBC عربيالحوثيون يعلنون مقتل 14 على الأقل وإصابة آخرين في غارات سعودية على الحديدة، والتحالف يعلن إسقاط طائرة مسيرة في منطقة الرياضZDF heuteEntdecken Sie das ZDF-NachrichtenstudioPremium TimesNigeria denies Fani-Kayode was reprimanded by South AfricaDeadlineFremantle U.S. COO Suzanne Lopez Exiting After Over 20 YearsSudinfoJorn n’a pas hésité une seconde face à la scène inquiétante qui se déroulait sous ses yeux : « Si je ne fais rien, ils ne survivront probablement pas »n-tv"Geweint wie ein kleines Kind": Erpressungs-Ermittlungen bei Handball-ZweitligistTagesschauUSA gelten nicht mehr als masernfreiInteriaGlapiński wszedł do gry o TK. "To koniec kruchego pokoju"
The Daily Newsstand · Free, Always
Tuesday, November 17, 2026

America's hyperscale bit barns draw as much power as entire states

Translate

Harvard study claims emissions attributable to datacenters are two to three times higher than in 2018

Just 403 hyperscale datacenters consumed an estimated 1.8 percent of US electricity in 2024-25, and their carbon dioxide emissions were roughly two to three times an earlier forecast for 2018.

Research led by the Harvard T.H. Chan School of Public Health and the Harvard Data Science Initiative looked specifically at hyperscale facilities, the largest class of server farms that typically support cloud computing and AI processing.

The published study, "Assessing the carbon emissions and energy consumption of US hyperscale data centres," estimates their combined electricity use approached that of a single state such as Indiana, Michigan, or Tennessee under the highest-load scenario.

REG AD

The researchers identified 403 hyperscale facilities by using a larger datacenter dataset, satellite imagery, and building footprints on OpenStreetMap.

REG AD

Rather than reading electricity meters, the researchers worked out consumption from each site's electrical capacity, the hours in a year, and assumptions about how heavily the facility was running.

Across four operating scenarios, ranging from low-intensity use to AI-heavy workloads, electricity draw was believed to total 68-99 terawatt-hours (TWh). Each of those states consumed approximately 100 TWh during the study period.

The researchers also calculated carbon dioxide (CO2) emissions using eGRID2023 plant-level generation and emissions data from the US Environmental Protection Agency (EPA). Emissions ranged from 21 million to 31 million metric tons of CO2 across the scenarios, with a central estimate of 25.7 million. That compares with 10.5 million for 2018.

The study covered May 2024 to April 2025, so electricity demand and carbon emissions are likely to be even higher now, the researchers say.

"What we're seeing is already striking, but it's only a snapshot of a rapidly expanding industry," said Gianluca Guidi, lead author on the paper and postdoctoral fellow in the Department of Biostatistics at Harvard.

"The datacenter boom hasn't slowed down since our study period ended in April 2025. New facilities have been built, existing campuses have expanded, and demand for AI computing continues to grow. That means the electricity use and carbon footprint we're reporting could already be substantially larger now in 2026."

Harvard's report also considers the mix of fuels being used to generate the electricity that powers all those hyperscale data houses. The mix varies considerably between regional grids, so where a hyperscale facility is built makes a substantial difference to its attributed emissions.

Across the sites studied, roughly 54 percent of the electricity attributed to their consumption came from fossil fuels, mainly natural gas and coal. The electricity attributed to these facilities nevertheless had a lower average carbon intensity than the US grid overall.

REG AD

The researchers acknowledge limitations, including that there wasn't enough available data to study all hyperscale sites in the US, and that the data did not allow them to differentiate between power used for general-purpose compute versus AI-specialized operations at individual campuses. ®

View the original on The Register →

KioskNews shows a cleaned-up reading view extracted from the publisher’s page — the original always lives on their site, not ours.