Harvard study: Hyperscale data centers collectively power through more electricity than some U.S. states A Harvard-led study published in Nature Sustainability found that U.S. hyperscale data centers used between 68 and 99 terawatt-hours of electricity in a single year, comparable to the annual consumption of states such as Indiana, Michigan and Tennessee (about 100 TWh each), and emitted more than 25 million metric tons of CO2 from 2024 to 2025 — roughly three times their 2018 emissions. Researchers led by Francesca Dominici, director of the Harvard Data Science Initiative, identified 403 hyperscale facilities using satellite imagery, permit filings, remote sensing and EPA data, and found about 54% of their power came from fossil fuels, including natural gas (38%) and coal (15%). Dominici said the figures were "something so much bigger that I was actually not even expecting myself" and likely an underestimate, since the data predates a further year of construction. Hyperscale data centers across the country are using as much electricity as entire states over the course of one year, according to a new study. And with the majority of that power coming from fossil fuels, those data centers produced more than 25 million metric tons of carbon dioxide from 2024 to 2025. That’s about three times higher than the emissions from the country’s hyperscale data centers in 2018. The research adds to a growing understanding of just how much the expansion of AI https://www.fastcompany.com/section/artificial-intelligence infrastructure and cloud computing has ramped up both electricity use and fossil fuel production across the country. Though the electricity use and pollution from data centers have come under public scrutiny recently, there has been little peer-reviewed science about data center emissions, says Francesca Dominici, an author of the study, published in Nature Sustainability https://urldefense.proofpoint.com/v2/url?u=https-3A www.nature.com articles s41893-2D026-2D01985-2D7&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A CdpgnVfiiMM&r=Wrp9m1UC3xshCL184yOQ96pbP7dLuIo9cZXZn0F UsI&m=m92prc-rHEpkMM0l5ZUlPhKpdFxXb fThcn4bDjN4Z3SYyk40P9dhjA77sSdkkRT&s=H-hU8XKQBvNDcNIjkpZksZ5xsqpUTM FjNrZjP-dnCM&e= , and director of the Harvard Data Science Initiative. To her knowledge, Dominici says, the only peer-reviewed study on greenhouse gas emissions and electricity consumption of hyperscaler data centers is from 2018, years before AI caused data center development to skyrocket. Dominici and her co-researchers spent two years investigating hyperscale data centers, which often aren’t publicly transparent about their emissions. The researchers used satellite imagery to compile a list of 403 hyperscale data centers, scraped the web for information filed in permits, used remote sensing, and relied on Environmental Protection Agency EPA data to calculate emissions based on the mix of power plants around the data centers. That effort “tells you about the fact that there is a lot of secrecy about these facilities,” Dominici notes. “Why can’t we just have a database which would tell us where are the facilities, how big they are, what’s the total capacity?” The study found that the country’s hyperscale data centers together used between 68 and 99 terawatt-hours of electricity, in just one year. One terawatt hour https://www.energy.gov/cmei/systems/articles/computing-americas-offshore-wind-energy-potential can power approximately 90,000 homes. That means the hyperscale data center fleet uses as much or more electricity than entire states. Indiana, Michigan, and Tennessee each use about 100 terawatt-hours of electricity a year. Some states use far less, like South Carolina around 84 Twh , Oregon 60 Twh , and Maine 11 Twh . States like Texas, California, and Florida use much more electricity; 475, 251, and 248 TWh respectively. Those figures surprised Dominici. “It’s something so much bigger that I was actually not even expecting myself,” she says. And they’re also likely an underestimate, she adds. The paper focused on electricity use from 2024 to 2025, already making the data a year old. “So much more construction and development of hyperscaler data centers has been happening just in the last year,” she says. To supply all that electricity, these hyperscale data centers are relying mostly on fossil fuels, which make up about 54% of the power mix. That includes natural gas 38% and coal 15% . About a quarter of this electricity also came from grid regions that are dirtier than average, researchers say, where coal-fired power plants dominate. The grids on which these data centers rely “really matter” to their overall impact, Dominici says. The research also makes clear, she adds, that “there has been no attention on renewable sources,” Dominici says. Though wind and solar are the cheapest power sources to build, they still don’t offer the same stability without battery storage. Fossil fuels are still the easiest way to get quick, constant, high power. Alongside the paper, the researchers are also publishing a website, neverjustabuilding.org, https://neverjustabuilding.org/ that illustrates the findings. Through that website, the researchers hope to reach those involved in some way in data center development, from the communities where data centers are being built to policymakers to utilities to developers themselves. This study is just the first step in a series of research efforts to better understand the AI data center boom and its impacts, including not only emissions and energy use but noise, water consumption, and the health impacts of that pollution. The website allows people to sign up for updates on forthcoming research. “This gives us now a framework . . . to start not only tracking,” Dominici says, “but also to really start identifying what we can do and where we should build and how to build, so data centers will have less of an impact.”