More than half the electricity feeding America’s largest data centers comes from gas and coal

In this articleIEA warns of natural gas trend and strained power gridsVoluntary AI climate pledge at COP31What it means Half the power…

By Vane November 18, 2026 4 min read
More than half the electricity feeding America’s largest data centers comes from gas and coal


Half the power for America’s biggest data centres comes from gas and coal

A Harvard study found roughly 400 of the largest US data centres emitted more CO2 than all of Denmark in one year. They consumed about 82 terawatt-hours, with 54% coming from fossil fuels.

The roughly 400 largest data centres in the United States produced about as much CO2 through their electricity consumption in a single year as all of Denmark. This happened before the current AI infrastructure boom could hit full force.

That is the finding of a Harvard University study published in Nature Sustainability. It is one of the first facility-level estimates of power consumption and CO2 emissions from so-called hyperscale data centres.

The research team led by Gianluca Guidi and Francesca Dominici used satellite imagery to identify 403 large-scale facilities. They matched estimated power demand between May 2024 and April 2025 with emissions data from regional grids. Depending on the assumed utilisation rate, annual consumption ranged from 68 to 99 terawatt-hours. The upper estimate matches the electricity use of entire US states such as Indiana, Michigan, or Tennessee. Under the study’s central scenario, the facilities consumed about 82 terawatt-hours, roughly 1.8 percent of total US electricity use. More than half of that power was used in just four states: Virginia, Oregon, Ohio, and Iowa.

Some of the facilities sit in regions with relatively carbon-intensive power mixes and high shares of coal generation. Overall, about 54 percent of the electricity generated for the surveyed data centres came from fossil sources. Natural gas accounted for 38 percent and coal for 15 percent. Under the Harvard study’s central scenario, that translated to about 25.7 million metric tons of CO2 per year. According to co-author Dominici, that figure matches or exceeds the annual greenhouse gas emissions of a US state like Michigan or a smaller European country such as Denmark. The emissions are also two to three times what an earlier comparable estimate found for 2018.

The researchers say the study only captures a snapshot. The observation period ended in April 2025 and the enormous energy consumption of modern AI agents has been pushing demand even higher since then. Between 2022 and 2025 alone, power consumption at the largest tech companies rose by 240 percent. This caused them to miss their own climate targets, according to a report by Climate Action Against Disinformation cited by the Associated Press.

With 2,000 more data centres planned across the US, co-author Dominici warns of a setback in efforts to curb CO2 emissions. Fossil-heavy power sourcing can also raise regional levels of air pollutants such as particulate matter, nitrogen oxides, and sulfur dioxide. These are linked to heart and lung disease and premature death. The study did not calculate specific health effects.

IEA warns of natural gas trend and strained power grids

The International Energy Agency’s (IEA) advance report Powering Artificial Intelligence also warns of growing strain on power grids and the climate. Hyperscalers are investing more than $700 billion worldwide in 2026 alone, mostly in new data centres, according to the report. In its central scenario, the IEA projects annual global CO2 emissions from data centres will rise from about 150 million metric tons of CO2 today to roughly 350 million metric tons per year by 2035. The upward revision stems largely from US operators increasingly investing in on-site natural gas power plants to bypass long waits for grid connections. Nearly 40 percent of identified AI data centre capacity is also located in areas with high or extreme water stress.

The expansion is pushing local grids to their limits. In regions including Texas, Virginia, Finland, and the United Kingdom, applications in grid interconnection queues already exceed current peak load. According to the IEA, nearly 40 percent of the 134 grid integration regulations tracked worldwide were introduced in just the past 12 months.

To shield households from rising electricity prices driven by grid expansion and to weed out speculative applications, regulators are weighing measures that vary by jurisdiction. These include minimum payment commitments, collateral, exit fees, flexible grid connections, and in some cases requirements for on-site firm power generation. Actively throttling compute loads remains expensive for operators, though, since leasing AI capacity costs $20 billion to $30 billion per gigawatt per year. This is roughly 100 times as much as comparable battery storage capacity.

Voluntary AI climate pledge at COP31

Growing criticism of the AI industry’s resource consumption has reached the United Nations. This follows local protests against new sites in the US and political backlash in recent elections. At the COP31 UN climate conference in Turkey, the International Telecommunication Union (ITU) and the host presidency introduced the “Antalya Artificial Intelligence Pledge” on Monday. It addresses the 50 percent rise in AI data centre power consumption in 2025.

The declaration calls on governments and companies to increasingly power AI infrastructure with zero- and low-emission energy. It asks for improved water and energy efficiency, disclosure of environmental impacts using standardised methods, and reuse and recycle of hardware. It also promotes AI solutions for climate action in developing countries. Separately from the pledge, experts at the conference are discussing the use of waste heat from data centres. According to the IEA, waste heat could supply district heating to about 10 percent of European households by 2030.

UN climate chief Simon Stiell warned sharply in Antalya against burdening people today with higher electricity bills and more pollution in exchange for vague promises of future AI efficiency gains. Advocates such as former US Vice President Al Gore hope AI can cut global emissions by 6 to 9 percent by 2030 through efficiency improvements. The Antalya agreement is not legally binding. All transparency and power supply measures are voluntary and are set to be reviewed for the first time at COP32.

What it means

Operators face a choice between waiting for grid connections or building their own gas plants to keep running. This choice directly increases carbon output and local pollution. The new voluntary pledge offers no legal penalties for non-compliance, leaving the industry to self-regulate without enforceable limits.


Scroll to Top