403 hyperscale data centers used 1.8% of U.S. electricity | Magica
403 hyperscale data centers used 1.8% of U.S. electricity, study estimates
Editorial Team
••📖7 min read
Researchers estimate that 403 large U.S. data centers consumed 81.8 TWh in a year. Their regional-grid analysis reveals substantial fossil-fuel dependence, but cannot isolate AI demand or fully expose the underlying facility data.
Researchers estimate 403 hyperscale facilities used 81.8 TWh between May 2024 and April 2025, about 1.8% of U.S. electricity consumption.
Roughly 54% of their attributed electricity came from fossil fuels, although their average grid carbon intensity was below the national benchmark.
The study cannot separate AI from other computing workloads, and commercial restrictions limit scrutiny of individual facilities.
A sample of 403 U.S. hyperscale data centers used an estimated 81.8 terawatt-hours (TWh) of electricity from May 2024 through April 2025, about 1.8% of national consumption, according to a study published November 16 in Nature Sustainability. Its central estimate attributes 25.7 million metric tons of carbon dioxide to the electricity serving those facilities.
The research, led by Harvard T.H. Chan School of Public Health and the Harvard Data Science Initiative, supplies a geographically detailed baseline for infrastructure whose environmental footprint is difficult to verify publicly. An aggregate electricity total cannot show which regional grids serve that demand or which communities might face pollution from associated generation. Hyperscale data centers are the largest class of computing facilities, supporting cloud services, large language model training and deployment of AI tools. Their location shapes the regional electricity mix used to attribute their footprint. The study could not determine how much of their consumption served AI.
Magica chart using Guidi et al.’s modeled scenarios for 403 facilities. These are not confidence bounds; the AI-weighted scenario does not measure AI’s share of consumption. Source: GitHub.
A large footprint, with a wide operating range
Lead author Gianluca Guidi is a Harvard biostatistics postdoctoral fellow studying computing’s environmental impact. His earlier work includes a 2025 study of the environmental burden of U.S. bitcoin mining, cited in the new paper. Corresponding author Francesca Dominici is Harvard’s professor of biostatistics, population and data science and director of its Data Science Initiative. Their stated aim is to inform decisions about where facilities are built and how they are powered.
The researchers assembled information on 403 facilities operating between May 2024 and April 2025, using satellite imagery to help identify them. They estimated electricity use under different operating assumptions and used Environmental Protection Agency power-plant data to assign emissions according to regional generation mixes, as described in the .
South Korea is strengthening device checks and considering public Wi-Fi shutdowns before its November 19 college entrance exam, after a tax-accountant exam top scorer admitted using AI to cheat. The response still depends heavily on human screening, with no electronic device detectors planned for this year's test.
This is modeled consumption and emissions accounting, rather than a public collection of metered electricity bills. The results depend materially on how heavily the facilities run. The team's public reproducibility repository gives these scenarios for the same sample. The facility-load coefficient represents the assumed operating load relative to capacity:
Operating scenario
Facility-load coefficient
Estimated electricity use
Attributable CO₂
Low load
0.480
67.7 TWh
21.3 million metric tons
Central estimate
0.580
81.8 TWh
25.7 million metric tons
AI-weighted high
0.700
98.7 TWh
31.0 million metric tons
These are modeled operating scenarios, rather than a statistical confidence interval. The high scenario's label does not mean the researchers measured AI's share of demand: the available data could not distinguish general-purpose computing from AI-specialized operations at each facility.
Nor is this a census of all U.S. data centers, or even all hyperscale facilities. The repository describes a 675-facility initial universe and a 403-facility analytical sample; the researchers lacked enough information to study every U.S. hyperscale facility.
For a broader baseline, a Lawrence Berkeley National Laboratory report released in December 2024 estimated that the entire U.S. data-center sector consumed 176 TWh in 2023, or 4.4% of national electricity, according to the Department of Energy's announcement.
That 2024 report also forecast 325–580 TWh in 2028, or 6.7%–12% of U.S. electricity. The populations, periods and methods differ, so the figures cannot establish a growth rate or the sampled facilities’ share of total sector demand. The 2028 figures are a forecast issued in 2024, not an observed extension of the new study.
Magica chart using Guidi et al.’s regional-grid attribution. Shares describe modeled generation, not measured facility purchases; rounding puts their total slightly above 100%. Source: GitHub.
A cleaner average masks regional concentration
The repository attributes approximately 53.9% of the sampled facilities' electricity to fossil generation, 20.9% to nuclear power and 25.3% to renewables, with rounding accounting for the slight excess over 100%. Natural gas supplied about 38% and coal about 15%, according to the research announcement.
Yet the paper puts the facilities' electricity-weighted carbon intensity at approximately 314 grams of CO₂ per kilowatt-hour, below its U.S. total-output benchmark of 348 grams. Substantial emissions therefore coexist with a generation mix that is, on average, less carbon intensive than that benchmark.
The denominator matters. The repository's headline calculation divides regional power-plant emissions by all reported net generation, including nuclear and renewable output. It retains a combustion-only calculation as a diagnostic, rather than using it for headline emissions. The grid factors come from EPA’s eGRID2023 Revision 2, rather than measurements of the exact power mix at each facility during the study year.
National averages conceal concentration. Virginia, Oregon, Ohio and Iowa together accounted for more than half the electricity use in the sample. The researchers found that about a quarter of the attributed electricity was in grid regions with carbon intensity above the national average. Location can therefore change the emissions associated with a given amount of computing demand.
What the model cannot establish
The study assigns a footprint using regional generation. That accounting approach does not establish which power plants changed their output because a particular data center opened. Its results are an attribution of emissions, rather than a measurement of the incremental emissions caused by those facilities.
An independent critique concerns electricity generated outside the grid. Ana Valdivia, an Oxford professor of AI, government and policy whose recent research examines AI’s environmental impact, argued in an assessment of the findings that assuming grid connection can miss facilities developing their own generation. Valdivia did not participate in the study. That raises a coverage question for a footprint built around regional electricity systems; it does not supply a quantified correction to this estimate.
Shaolei Ren, a University of California, Riverside associate professor of electrical and computer engineering who studies AI sustainability and was not involved in the paper, offered a complementary assessment in the same report. He said the geographic detail helps explain why location matters, while calling for more facility information to connect energy and emissions with water use, air quality and health.
Transparency is another constraint. The paper discloses that coauthor Eric Bell works for Baxtel, which provided commercial facility data, and that coauthor Kevin Butler works for Environmental Systems Research Institute, or Esri. The other authors declare no competing interests. Funding came from the National Institute of Environmental Health Sciences and the National Institute on Aging. Facility identifiers, addresses, coordinates and commercially sensitive attributes remain protected by a data-use agreement.
Released code and aggregate state and grid-region tables allow outsiders to check scenario totals, the emissions denominator and other published outputs. Public synthetic data also support checks of the modeling workflow. Those public materials do not replace the underlying observations needed for complete independent reproduction. Access to restricted records may be considered for approved research collaborations, subject to provider authorization and an appropriate data-use agreement.
What the next power decisions need
The researchers say the footprint could already be larger because new facilities and campus expansions have followed the study period. That is a warning about an aging snapshot, rather than a measured estimate of consumption in November 2026.
In its 2024 announcement, the Energy Department identified on-site generation and storage, demand flexibility, transmission improvements and rate structures that maintain affordability as approaches to rising data-center demand. The new study does not compare their costs or establish which approach would reduce emissions most.
The next useful evidence would narrow the operating assumptions, identify electricity used for different workloads and account for on-site generation. It would also let researchers test facility-level estimates against actual consumption while protecting sensitive information.
Health effects remain a separate unanswered question. The study did not measure them. The team plans further research into how electricity demand connects to pollution exposure, including effects on people near and downwind of fossil-fuel power plants who may live far from the data centers themselves. Establishing where those effects occur, and how power and siting decisions change them, remains work to be done.
Russia’s objections to observer participation stalled the opening of a Geneva weapons conference, where governments must decide whether to authorize negotiations on autonomous weapons and whether any resulting rules should be binding.
Apple’s iOS 27.3 beta ties higher allowances for server-based AI to existing iCloud+ plans. Basic Siri commands remain free, but undisclosed daily limits leave the value of paid access uncertain.
The partners plan to combine IBM quantum processors, NVIDIA computing and Fermilab physics expertise to study modular quantum systems. A separate decoder release promises faster error-correction research, with important limits on its performance comparisons.
Federal Reserve Chairman Kevin Warsh outlined AI scenarios spanning shared prosperity, job displacement and an investment downturn. The exercise offers no forecast or rate guidance as colleagues worry that the infrastructure buildout is adding to inflation.
OpenAI is offering approved organizations a free classifier for suspected child sexual abuse images, including AI-generated material. Early results suggest less review work, but existing detection alternatives and the need for human review qualify the launch.
Nvidia plans scientific computing racks for early 2027, extending Vera beyond its initial AI deployments. Its 2.4× performance-per-watt claim comes from one airflow simulation against AMD Turin; performance against newer rivals and delivered costs remain unresolved.
OpenAI’s CEO expressed empathy for Nvidia founder Jensen Huang but withheld judgment on reinstatement. His confidence in continued AI financing comes with a direct commercial stake and unresolved deployment risks.
Google’s 8,192-chip Ironwood submission completed the MLPerf DeepSeek-V3 benchmark in 1.70 minutes, close to NVIDIA’s 1.53 minutes at the same accelerator count. Google’s scaling data also shows the extra chip-time required for its fastest result, leaving cost and energy comparisons unresolved.
TensorWave will supply dedicated AMD infrastructure under a six-year OpenAI agreement, with delivery expected from 2027. Its operating record and financial backing sit alongside an undisclosed contract price and a recently terminated site lease.
California alleges OpenAI endangered the public and misled consumers by testing powerful AI models without adequate controls. The lawsuit follows a Hugging Face intrusion that OpenAI’s own account traced to containment failures, missing monitors and mishandled warning signs.
Perplexity has launched Frontier Research to develop AI that can formulate questions, test hypotheses and produce interpretable discoveries. The company promises more than 10,000 GPUs, but existing automated-research systems show why computing resources and generated papers alone cannot establish scientific value.
The new lab led by Stanford professors Yejin Choi and Carlos Guestrin has backing from Nvidia, Cisco and Snowflake. It promises systems that improve on company work, but research benchmarks leave production costs and customer control unproven.
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
OpenAI will open 50,000 new ChatGPT access grants on November 23, covering up to 150,000 researchers and collaborators. New recipients get Plus instead of Pro, with first-come selection replacing a lottery. Its pilot study reports perceived time savings but does not establish faster discovery.
OpenAI reports production healthcare workflows at CVS and Cigna and stronger computer-use evaluations. The results show progress across disconnected software, but leave complete-task reliability and total deployment costs unresolved.
President Donald Trump supported a return for Nvidia co-founder Jensen Huang, but the company has installed interim leadership and begun a successor search amid scrutiny of his conduct and alleged Chinese intelligence ties.
Intel disclosed the bandwidth target for its next Xeon server processors at SC26, alongside previously announced core and cache limits. Application performance, power requirements and launch timing remain unresolved.
Mercedes-Benz plans California early access to city-driving and hands-free highway assistance in December, ahead of a 2027 U.S. rollout. Mixed road tests and unannounced subscription prices leave buyers waiting for evidence of everyday value.
HSBC lowered its Zhipu share-price target to HK$500 and delayed forecast adjusted breakeven by two years. Improving API margins and substantial fundraising cushion a business whose gross profit still covers little of its research bill.