403 largest U.S. data centers rivaled Denmark’s emissions
Harvard estimates 403 large U.S. data centers consumed 82 terawatt-hours in one year, with 54% fossil power. IEA expects data-center emissions to reach 350 million metric tons by 2035.

Why it matters
- Harvard’s central estimate: 403 large U.S. data centers used 82 TWh using May 2024–April 2025 data.
- Fossil fuels supplied 54% of surveyed power: 38% natural gas and 15% coal, producing 25.7 million metric tons of CO2.
- The IEA projects global data-center emissions to rise from about 150 million metric tons to 350 million per year by 2035.
- Hyperscalers will invest more than $700 billion worldwide in 2026, mostly in new data centers.
- COP31’s Antalya pledge is voluntary and faces its first review at COP32; AI data-center power use rose 50% in 2025.
A Harvard University study estimates that 403 large-scale U.S. data centers consumed about 82 terawatt-hours of electricity in one year and emitted 25.7 million metric tons of CO2, roughly matching or exceeding Denmark’s annual footprint.
That is the central scenario in a study led by Gianluca Guidi and Francesca Dominici, published in Nature Sustainability. The IEA projects global data-center emissions to climb from about 150 million metric tons today to roughly 350 million per year by 2035. The paired estimates make power supply and emissions policy central constraints on the AI buildout, rather than distant environmental risks.
What did the Harvard study measure?
The researchers used satellite imagery to identify 403 large-scale hyperscale facilities. They then matched estimated power demand from May 2024 through April 2025 against emissions data from regional grids. The Nature Sustainability paper is one of the first facility-level estimates of both electricity consumption and CO2 emissions for hyperscale data centers.
The result depends on how fully researchers assume the facilities operated. Their range ran from 68 to 99 terawatt-hours a year. At the upper end, consumption matched the electricity use of entire states such as Indiana, Michigan, or Tennessee. The central estimate of about 82 terawatt-hours equaled roughly 1.8% of all electricity used in the United States.
More than half of the facilities’ estimated electricity use occurred in just four states: Virginia, Oregon, Ohio, and Iowa. The regional mix also mattered. About 54% of the power serving the surveyed facilities came from fossil fuels. Natural gas supplied 38%, and coal supplied 15%.
Under the central scenario, those sources produced about 25.7 million metric tons of CO2 annually. Co-author Francesca Dominici said the total matched or exceeded the annual greenhouse-gas emissions of a state such as Michigan or a smaller European country such as Denmark. Harvard’s estimate also came in two to three times above an earlier comparable estimate for 2018.
Those 403 facilities already carry a measurable slice of U.S. electricity and a national-scale carbon footprint.
How much could planned growth add?
The Harvard researchers cautioned that the study captures only a snapshot, since its observation window ended in April 2025. The team said modern AI agents’ enormous energy use had already pushed demand higher. That makes the underlying data less likely to capture the full scale of AI-related electricity demand.
A Climate Action Against Disinformation report, cited by the Associated Press, found that power consumption at the largest tech companies rose 240% between 2022 and 2025. The report linked that increase to their own missed climate targets.
With 2,000 more data centers planned across the US, Dominici warned that the expansion could set back efforts to curb CO2 emissions. The researchers also said fossil-heavy power sourcing can raise regional levels of particulate matter, nitrogen oxides, and sulfur dioxide.
The Harvard study did not calculate specific health effects. The researchers linked those pollutants to heart and lung disease and premature death.
Planned capacity could add emissions and local pollution before efficiency gains catch up.
Why are grid connections becoming a bottleneck?
The International Energy Agency’s advance report, Powering Artificial Intelligence, warns that hyperscale growth strains both power grids and the climate. The IEA says hyperscalers will invest more than $700 billion worldwide in 2026 alone, mostly in new data centers.
In its central scenario, the IEA projects annual global data-center CO2 emissions to rise from about 150 million metric tons today to roughly 350 million metric tons per year by 2035. The agency attributes most of that upward revision to U.S. operators investing in on-site natural gas plants. Those plants can bypass long waits for grid connections.
The IEA also estimates that nearly 40% of identified AI data center capacity sits in areas with high or extreme water stress.
In Texas, Virginia, Finland, and the United Kingdom, grid interconnection applications already exceed current peak load, according to the IEA. The agency also tracks 134 grid integration regulations worldwide. Nearly 40% of those rules arrived in the past 12 months.
The IEA says regulators across jurisdictions are weighing ways to protect households from higher electricity prices and remove speculative grid applications. Their options include:
- Minimum payment commitments and collateral to cover new grid costs.
- Exit fees and flexible grid connections.
- On-site firm generation, a model the IEA calls “bring-your-own-power.”
Actively throttling compute loads remains expensive for operators. The IEA estimates that leasing AI capacity costs $20 billion to $30 billion per gigawatt per year, roughly 100 times the cost of comparable battery storage capacity.
Slower grid access pushes operators toward dedicated power, which can add more fossil emissions.
What can COP31 actually change?
Growing criticism of AI’s resource use has already produced local protests against new U.S. sites and political backlash in recent elections. At the COP31 UN climate conference in Turkey, the International Telecommunication Union and the host presidency introduced the “Antalya Artificial Intelligence Pledge” on Monday. The declaration addresses a 50% rise in AI data center power consumption in 2025.
The pledge calls on governments and companies to:
- Supply AI infrastructure with more zero- and low-emission electricity.
- Improve water and energy efficiency.
- Disclose environmental impacts using standardized methods.
- Reuse and recycle hardware.
- Promote AI tools for climate action in developing countries.
Experts at the conference are also discussing waste heat as a separate lever. The IEA estimates that recovered heat could supply district heating to about 10% of European households by 2030.
UN climate chief Simon Stiell warned against burdening people today with higher electricity bills and more pollution in exchange for “vague promises of future AI efficiency gains.” Former US Vice President Al Gore said AI could cut global emissions by 6% to 9% by 2030 through efficiency improvements.
The Antalya agreement does not carry legal force. Every transparency and power-supply measure is voluntary. COP32 will provide the first review, but the agreement currently sets a reporting direction rather than a binding emissions limit.
Original: nature.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 U.S. Data Centers Used as Much Electricity as Michigan
- Hyperscale data centers used 2% of US electricity in 2024-25
- Google's Emissions Jumped 50% in 2024 to Feed Generative AI
- Data Center Boom Faces a Bigger Risk Than Water or Power: Finance