Chile

Large Data Centers In The US Already Consume As Much Energy As The Whole Of Chile

Published

on

Data centers are the backbone of the digital age. In these facilities, packed with processors, vast quantities of data that enable online services and artificial intelligence (AI) are computed, stored, and distributed. The recent AI boom has sparked a construction frenzy for data centers. Keeping them running around the clock requires huge amounts of energy, used both to power the machines themselves and to cool them. But how much? And what is their carbon footprint?

Only the companies that operate data centers know the answer to those questions. There is no public, transparent information. The industry’s opacity means the figures in circulation are estimates produced by the scientific community. A team of researchers led by Harvard University published the results of their own investigation in the journal Nature on November 16, focused on hyperscalers (the name given to the largest data centers). The study concentrates on the U.S., the country with the largest share of data centers — around half of the global total.

The numbers are huge: hyperscalers consumed between 68 and 99 terawatt-hours (TWh) of electricity. To put that in context, the highest estimate exceeds the annual consumption of a country like Chile (83.2 TWh), according to data from the Department of Energy’s statistical agency (EIA), and approaches the consumption of states such as Indiana, Michigan, or Tennessee. That is roughly 1.8% of total U.S. electricity consumption. The researchers also conclude that four states — Iowa, Ohio, Oregon, and Virginia — account for more than half of the detected energy use.

As for emissions, the team’s calculations place them at about 25.7 million metric tons of CO₂, an amount two to three times higher than a previous estimate for 2018.

The interesting part of the study is how they constructed their estimate. “There is barely any public data. The Department of Energy, for example, only provides aggregated energy consumption data for the whole country. That does not allow us to know how individual data centers behave,” says the study’s lead author, Gianluca Guidi, a postdoctoral fellow in the department of biostatistics. Because there are no public registers to check how many of these facilities exist in the U.S. or where they are located, the researchers opted for a combination of methods. First, they tried to obtain information from the operators themselves, a laborious task (company by company and state by state) that is not always successful.

Second — and here is the novelty — they cross-checked and expanded that information through direct observation: they analyzed satellite imagery to detect rooftops or plots typical of hyperscalers, which have a very particular architecture (they look like industrial warehouses but feature easily identifiable cooling systems and are often close to energy infrastructure). They estimated each data center’s installed capacity based on the square meters occupied by the facility.

Using this method, they located and mapped 403 hyperscalers in the U.S., with an average installed capacity of 40 megawatts (MW) and an average footprint of about 2,000 square meters. The authors stress that this number is an undercount, since it resulted from observations made between May 2024 and April 2025. Therefore, the consumption and emissions data are necessarily conservative. “The data center boom has not slowed since April 2025. New facilities have been built, existing campuses have been expanded and the demand for AI computing continues to grow. That means the electricity consumption and carbon footprint we report could already be considerably higher,” says Guidi.

Conservative estimates

The estimate of hyperscaler volume does not account for the activity carried out. AI training and inference processes, for example, consume far more energy than data hosting or other cloud services. In other words, a smaller data center dedicated to AI can consume more than a larger one that performs other activities. Because operators do not usually disclose what their centers are used for, the researchers overcame this obstacle by outlining three possible scenarios — low, moderate, and high consumption.

The next step was to estimate energy consumption based on each facility’s size, and emissions by taking into account the energy mix used in each locality. The authors note that energy use rose substantially in the early 2000s but remained steady through the 2010s despite industry growth (which can be explained by improvements in hardware efficiency and cooling systems). Demand has grown again in recent years, driven by the computing required to support commercial generative AI models.

The study also finds that roughly 54% of the electricity supplying hyperscalers during the observed period was generated from fossil fuels, mainly natural gas (38%) and coal (15%), and that about one quarter of the electricity came from regions with grids dirtier than the national average.

The researchers observed substantial regional differences in emissions, which vary greatly because of differing local energy mixes. Many facilities were concentrated in areas where electricity is largely generated by coal-fired plants, amplifying their carbon footprint. “Decisions about where data centers are located and how they are powered are crucial for the environment and public health, and they must be made based on reliable data,” says one of the study’s authors, Francesca Dominici, the Clarence James Gamble professor of biostatistics, population and data science and director of Harvard’s Data Science Initiative.

That view is shared by Shaolei Ren, who was not involved in the study and is an associate professor of electrical and computer engineering at the University of California, Riverside and an AI sustainability specialist. “This work illustrates why the location of data centers matters environmentally, in contrast to more aggregated analyses,” he says. Ren sees the study’s findings as a starting point for more detailed research. “If we had more information on how specific facilities actually operate, we could better identify where the biggest opportunities lie. In my own work I’m particularly interested in linking energy and carbon emissions to water use, air quality, and health impacts,” he adds.

Although the method developed by the Harvard team is interesting, it may have some gaps. “The paper assumes that all hyperscalers are connected to the grid, when that is not the case,” says Ana Valdivia, a professor of artificial intelligence, government and policy at the University of Oxford, whose recent research focuses on assessing AI’s environmental impact. “My critique of this approach is that many data centers, such as xAI’s in Tennessee, are not finding grid connections and are developing their own generation. That means the results underestimate actual data because they do not account for those that generate their own energy,” says the researcher, who did not participate in the study published in Nature.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition

Leave a Reply

Your email address will not be published.

Trending

Exit mobile version