403 U.S. Data Centers Used as Much Electricity as Michigan
A Harvard estimate says 403 U.S. hyperscale data centers consumed as much electricity as Michigan in a year, while a separate power pipeline could lift US emissions.

Why it matters
- Harvard's 403-facility estimate matched Michigan's annual electricity use.
- The Nature Sustainability study covered facilities operating from May 2024 through April 2025.
- Data center emissions were three times pre-AI-boom levels, while most electricity came from fossil fuels.
- The U.S. behind-the-meter pipeline reached 189 GW for data centers; full development could add 10% to overall U.S. emissions, according to BloombergNEF.
- The EIA completed three pilot surveys across 10 states and is designing a potential voluntary sector survey.
By Sophie Lindqvist
A Harvard study estimates that 403 connected U.S. hyperscale data centers used as much electricity in one year as the state of Michigan. Published Monday in Nature Sustainability, the study modeled those facilities’ power demand and carbon emissions.
The same research found that fossil fuels generated most of that electricity. It also estimates data center emissions at three times their level before the AI boom.
The finding puts AI-era compute growth at the center of U.S. grid planning, pollution accounting and local disputes over hyperscale facilities. It also lands amid projections that data centers will strain the American power grid. “Hyperscalers have been growing fast, even before the AI boom,” said Jon Koomey, an energy researcher unaffiliated with the study.
How did Harvard estimate 403 facilities?
The Harvard team assembled a list of 403 hyperscale data centers operating across the country between May 2024 and April 2025. Researchers combined industry research with satellite data to identify and characterize the facilities.
The team then modeled each facility’s energy use in a variety of scenarios. For emissions, it calculated the energy mix of each regional grid and applied those factors to the modeled demand.
That method produces an estimate rather than a meter reading for every site. It links two variables: how much electricity a facility may consume and how much carbon its regional power mix adds.
Together, data centers in Virginia, Oregon, Ohio and Iowa accounted for more than 50 percent of the energy use in the study. Virginia also contains Data Center Alley, a concentration of large facilities that has become a flashpoint for community opposition.
The four-state concentration shows why a national total can hide local pressure. Each regional grid has its own energy mix, which changes the carbon estimate attached to the same amount of electricity.
Koomey called the peer-reviewed estimate more rigorous than some back-of-the-envelope calculations. He has worked on a regular Lawrence Berkeley National Laboratory report since 2007 that approximates the industry’s energy and water use.
Koomey said some facility figures remain hidden behind nondisclosure agreements, even when companies or utilities know them. That makes public estimates harder to check and leaves researchers dependent on models when direct readings are unavailable.
In short, more than half of the study’s estimated electricity demand sits in four states.
Why is the electricity total hard to verify?
Federal and state governments do not track data center electricity use separately from the rest of the power sector. Individual facilities’ power figures often count as proprietary business information, and operators and utilities face no public-disclosure requirement.
These gaps leave outsiders with an incomplete view of a fast-growing load on the grid. They also obscure how new AI facilities affect local electricity markets, emissions and debates over large infrastructure developments.
Researchers selected hyperscale facilities, the largest category of data center. Many handle AI inference or training, but not every site in the Harvard sample has an AI-related role.
Koomey’s point narrows the causal story: AI accelerated an expansion that was already well underway. The Harvard estimate therefore captures a broader infrastructure trend, not only the current wave of AI computing.
Community opposition has nevertheless made hyperscale sites flashpoints in several states. Those local disputes now sit on top of a sector whose power use remains hard to verify.
The takeaway is that AI did not create the sector’s power demand from zero. It arrived on top of a larger expansion already hidden by sparse public data.
How current is the 40-megawatt baseline?
The Harvard sample’s average facility size is 40 megawatts. The researchers describe that as relatively small compared with campuses now under development.
Developers are building campuses that require hundreds or even thousands of megawatts. Still larger proposals would sit well above the study’s typical facility.
“Some of the massive campuses being proposed today are substantially larger than the typical facility in our dataset,” said Gianluca Guidi, the study’s lead author. He is a postdoctoral research fellow at Harvard’s T. H. Chan School of Public Health.
The dataset covers facilities operating from May 2024 through April 2025, so the study may already understate the newest buildout. Guidi’s warning points to a measurement problem that will worsen as campuses get larger.
In plain terms, 40 megawatts is the study’s starting point, not the industry’s ceiling.
What does behind-the-meter power change?
The study did not include developers who build power plants directly beside their data centers. That setup, known as behind-the-meter power, can change both a facility’s grid demand and its emissions profile.
Koomey identified onsite generation as a major uncertainty when estimating total data center power use. Developers have proposed as many as 189 gigawatts of behind-the-meter generation earmarked for data centers over the past year.
Many of those plants rely increasingly on inefficient natural gas turbines. If all queued behind-the-meter plants get built, BloombergNEF estimated last week that overall U.S. emissions could rise by 10 percent.
The practical effect is larger than the electricity counted at the grid connection. Proposed onsite generators could add substantial fuel use and carbon emissions beyond the study’s modeled totals.
In short, the 189-gigawatt pipeline creates a large blind spot for anyone estimating data center demand from grid consumption alone.
Will the EIA get better data?
The Energy Information Agency, the federal arm that tracks energy use across sectors, conducted three pilot surveys of data center power use across 10 states this year. The pilots are complete, according to EIA spokesperson Lee Tucker.
Tucker said the agency is now using the results to design a potential voluntary survey for the industry. The agency has not committed to mandatory reporting, so participation and data coverage would remain central questions.
For researchers, a consistent survey could replace one-off modeled estimates with more systematic sector-wide observations. Public officials would also gain a clearer view of the load competing for electricity in their states.
In short, better federal data could close the disclosure gap. The survey’s voluntary status means the sector may still keep its most sensitive figures out of public view.
If EIA adopts the proposed survey and keeps it current, future studies could rely on fresher, more comparable facility data. Onsite generation and voluntary participation would still limit how completely the public can see the sector’s footprint.
Source: wired.com
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- Fossil Fuels Supplied 54% of Power to US Hyperscale Data Centers
- 403 largest U.S. data centers rivaled Denmark’s emissions
- Hyperscale data centers used 2% of US electricity in 2024-25
- Google's Emissions Jumped 50% in 2024 to Feed Generative AI
- Oracle's Opower AI Cut 1 GW From North American Peak Demand in 2026