> ## Content Index
> Fetch the complete content index at: https://www.implicator.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# New Harvard Study: America's Biggest Data Centers Run Mostly on Fossil Fuels
- URL: https://www.implicator.ai/harvard-study-hyperscale-data-centers-fossil-fuels/
- Published: 2026-11-16T16:00:00.000Z
- Updated: 2026-11-16T16:00:00.000Z
- Description: A Harvard-led study of 403 facilities ties 25.7 million metric tons of CO2 a year to their electricity, more than double the last estimate. The companies that own them generally do not disclose the numbers.
- Author: Marcus Schuler
- Tags: AI News, AI Research

Gianluca Guidi's team began with a list of 675 buildings. A commercial data provider had supplied it, and the [Harvard researchers](https://neverjustabuilding.org/?ref=implicator.ai) checked the records against rooftop coordinates, building footprints and satellite and aerial photographs until 403 were left that they could verify.

They worked that way because the companies that own America's largest data centers generally do not disclose what each one consumes. "Data-center operators generally do not disclose detailed infrastructure information," their paper states. Even now the building-level records stay restricted under the data provider's terms. The public gets state and regional totals.

The study, published Monday in [Nature Sustainability](https://www.nature.com/articles/s41893-026-01985-7?ref=implicator.ai), puts a number on what those 403 buildings drew from the grid and on what the regional grids burned to supply them. Between May 2024 and April 2025, the researchers estimate, they used about 82 terawatt-hours of electricity, roughly 1.8% of all the power consumed in the U.S., and more than half of it was generated from natural gas and coal. The carbon dioxide attributable to that electricity came to about 25.7 million metric tons a year, two to three times the last peer-reviewed estimate.

The Breakdown

- A Harvard-led study estimates that 403 U.S. hyperscale data centers used about 82 terawatt-hours between May 2024 and April 2025, roughly 1.8% of U.S. electricity.
- Fossil fuels generated 53.9% of the power attributed to them: natural gas 38.2% and coal 15.2%.
- Attributable emissions of about 25.7 million metric tons of CO2 are two to three times the last peer-reviewed estimate, for 2018.
- Operators generally do not disclose facility-level data, and their renewable purchases and net-zero claims sit outside the study's physical grid count.

AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/).

## What the study found

Hyperscale data centers are the largest class of the building type, the ones that run cloud services and train and serve AI models. The facilities in the sample had an average power capacity of 39.9 megawatts each.

Because the data do not show how hard each building runs, the researchers modeled four load scenarios. Electricity use across them ranges from 68 to 99 terawatt-hours a year, and attributable emissions from 21 million to 31 million metric tons. The upper end of the electricity range approaches what Indiana, Michigan or Tennessee each consume in a year, about 100 terawatt-hours.

The last peer-reviewed estimate covered 2018 and put hyperscale electricity use at 22.85 terawatt-hours and greenhouse-gas emissions at about 10.5 million tons. The new central figure for electricity is about 3.6 times that. The emissions range is two to three times as high.

"Data centers are growing incredibly fast, driven largely by AI, and decisions about where to build them and how to power them are being made right now. But there's very little facility-level information about their environmental footprint," said Guidi, the paper's lead author and a postdoctoral fellow in biostatistics at the Harvard T.H. Chan School of Public Health. "Our study helps fill that gap."

## Where the power comes from

To assign the electricity to power plants, the team assigned each data center to its balancing authority, the regional entity that matches generation to demand, and found 23 of them serving the sample. It then spread each region's data center demand across the power plants there in proportion to what each plant generated, using [Environmental Protection Agency plant data](https://www.epa.gov/egrid?ref=implicator.ai) for 2023\. Those regions held 8,802 plant records with reported generation.

Fossil fuels supplied 53.9% of the electricity attributed to the 403 facilities. Natural gas accounted for 38.2% and coal for 15.2%. Nuclear plants supplied 20.9%. Hydro, wind, solar, biomass and geothermal together supplied 25.3%.

### What 403 data centers drew from the grid

U.S. hyperscale facilities, May 2024 to April 2025, as estimated by the Harvard-led study. All figures are its central scenario.

More than double the last estimate

Electricity, terawatt-hours a year

**22.85**2018

**82**2024–25

Emissions, million metric tons a year

**10.5**2018

**25.7**2024–25

Across the study's four load scenarios the range is 68 to 99 terawatt-hours and 21 to 31 million metric tons of CO2\. The 2018 figure counted all greenhouse gases.

More than half the power came from gas and coal

- Natural gas **38.2%**
- Coal **15.2%**
- Nuclear **20.9%**
- Hydro, wind, solar, other renewables **25.3%**

Fossil fuels in total: 53.9%.

Source: Guidi et al., "Assessing the Carbon Emissions and Energy Consumption of US Hyperscale Data Centers," Nature Sustainability, Nov. 16, 2026\. Emissions are assigned from each regional grid's 2023 generation mix. The 2018 figures are from the previous peer-reviewed estimate.

The region carrying the most data center demand in the study is the balancing authority PJM. There the attributed mix was 60.0% fossil, with natural gas at 45.5% and coal at 13.8%. Renewables supplied 6.8%.

## Where the buildings sit

Virginia had 142 of the 403 facilities, more than any other state. They used about 21.4 terawatt-hours, roughly a quarter of the sample's electricity, and accounted for about 7.0 million metric tons of carbon dioxide. Virginia, Oregon, Ohio and Iowa together used more than half of the electricity.

### Four states carry most of the load, and the grid decides the carbon

State totals for the 403 hyperscale data centers in the Harvard-led study, May 2024 to April 2025, central scenario.

Virginia alone used about a quarter of the electricity

These four states: more than half of the electricity

- Virginia **21.4 TWh**, 142 facilities
- Oregon **11.1**, 56
- Ohio **8.8**, 38
- Iowa **7.2**, 26
- All other states

Less power, more carbon: Oregon against Iowa and Wyoming

StateElectricity used, TWhCO2 assigned, million metric tonsGrams per kWh **Oregon**56 facilities**11.1** **1.3**119 **Iowa**26 facilities**7.2** **3.2**445 **Wyoming**2 facilities**1.2** **1.0**848 

A kilowatt-hour drawn in Wyoming carried about seven times the carbon of one drawn in Oregon.

The ten states with the most attributed emissions

State, facilitiesPower, TWhCO2, MtGrams of CO2 per kilowatt-hour

- **Virginia**14221.47.0*325*
- **Iowa**267.23.2*445*
- **Ohio**388.82.9*325*
- **Texas**224.61.5*333*
- **Nebraska**83.51.4*395*
- **Oregon**5611.11.3*119*
- **Arizona**123.21.3*401*
- **Oklahoma**82.91.1*395*
- **Wyoming**21.21.0*848*
- **Utah**51.30.8*640*

U.S. grid average, 348 grams. Red bars are above it. About 23% of the sample's demand sits in grid regions above the average. The sample as a whole averaged 314 grams.

Source: Guidi et al., Nature Sustainability, Nov. 16, 2026\. Ranked by CO2\. Emissions are assigned from each regional grid's 2023 generation mix and do not necessarily occur in the same state.

The state figures show how much the location of a building decides the carbon attributed to it. Oregon's 56 facilities used 11.1 terawatt-hours and were assigned 1.3 million metric tons. Iowa's 26 used 7.2 terawatt-hours and were assigned 3.2 million. The grid behind the Oregon sites emitted about 119 grams of carbon dioxide per kilowatt-hour, the grid behind the Iowa sites about 445.

Wyoming has the highest carbon intensity among the ten states with the most attributed emissions. Two facilities there used 1.2 terawatt-hours and were assigned 1.0 million metric tons, at about 848 grams per kilowatt-hour. A kilowatt-hour drawn in Wyoming carried about seven times the carbon of one drawn in Oregon.

About 23% of the sample's electricity demand sits in grid regions dirtier than the national average. Weighted across all 403 facilities, the power was slightly cleaner than that average, about 314 grams per kilowatt-hour against 348 for the U.S. as a whole.

## The clean-energy claims sit outside the count

The study's numbers are physical. They describe the plants that fed each regional grid, and the paper keeps that apart from what the operators say about themselves. It notes that "some of the largest HDC operators are major private purchasers of renewable energy," then sets those purchases aside: power purchase agreements, renewable energy certificates and net-zero claims are "outside the scope of this physical grid-mix attribution unless reflected in the reported generation mix."

A certificate bought elsewhere does not change the mix on the grid a building draws from. On that mix, most of the power came from burning gas and coal.

Two of the largest operators have set targets that point the other way. Google set 2030 goals of net-zero emissions and carbon-free power around the clock. Amazon [pledged net-zero carbon by 2040](https://www.aboutamazon.com/news/sustainability/amazon-sustainability-report-2025?ref=implicator.ai). Their reported totals are moving away from those targets. [Google's total greenhouse gas emissions](https://sustainability.google/google-2026-environmental-report/?ref=implicator.ai) rose 18% in 2025, the largest annual increase it has reported, in a year when its electricity demand jumped 37%. Amazon's emissions rose 16% to nearly 80.9 million metric tons of carbon dioxide equivalent. The company has also bought a site in West Texas for an AI data center campus and plans to buy its power from a gas plant to be built there, whose permit allows up to 33 million tons of carbon dioxide a year, a ceiling rather than a forecast. "The world looks different now than when we co-founded the climate pledge," Amazon spokeswoman Margaret Callahan said in August, adding that the company's commitment had not changed.

Elon Musk's xAI illustrates a limit the paper names. Its authors caution that dedicated fossil generation not reflected in their grid data could cause the method to understate emissions. xAI powers its Colossus 2 data center in Memphis with gas turbines of its own, most of them set up across the state line in Southaven, Miss., and they began running after the study's window closed. Emails that xAI's consultants sent to Mississippi's environmental agency, obtained by Reuters, show 59 turbines running without a federal clean air permit, more than double the 27 the company acknowledged in January. xAI and the state argue the units are temporary and exempt. The NAACP and its Mississippi state conference, represented by Earthjustice and the Southern Environmental Law Center, [sued in April](https://earthjustice.org/case/xai-illegal-gas-power-plant-data-center-colossus?ref=implicator.ai) over 27 of the turbines. Researchers at the Massachusetts Institute of Technology and Carbon Mapper, working from a NASA satellite instrument, estimate in a study not yet peer reviewed that the site's emissions of smog-forming nitrogen oxides averaged about 870 kilograms an hour from February to August, roughly 19 times the level expected under the permit Mississippi issued in March.

The method does not assign any operator's purchases to a specific building. Nor can the paper say how much of any facility's load is AI, since "facility-level workload composition is not publicly disclosed." It had no facility-level efficiency data either.

## A buildout ahead of the reporting

The growth is recent. Data center electricity use stayed nearly flat through much of the 2010s as efficiency gains offset expansion, the paper recounts, and has climbed again in recent years, driven in large part by hyperscale facilities serving cloud and AI. In 2014 more than 60% of server electricity was consumed in companies' own internal data centers. By 2023 that share had fallen to nearly 10%. A federal laboratory estimate cited in the paper expects hyperscale and colocation sites to account for more than 90% of server electricity by 2028.

Know someone who'd find this useful? [✉️ Email it to a friend in one click](mailto:?subject=A%20newsletter%20I%20think%20you%27d%20like&body=This%20is%20one%20of%20maybe%20three%20newsletters%20I%20actually%20read.%20The%20rest%20just%20pile%20up%2C%20unread%2C%20judging%20me.%0A%0AAnd%20yes%2C%20this%20email%20mostly%20wrote%20itself%2C%20which%20is%20a%20little%20on%20the%20nose%20for%20an%20AI%20newsletter.%20Doesn%27t%20make%20it%20wrong.%20implicator.ai%20is%20good.%0A%0ASubscribe%20free%3A%20https%3A%2F%2Fwww.implicator.ai%2Fsubscribe%2F%3Futm%5Fsource%3Dnewsletter%26utm%5Fmedium%3Dforward%26utm%5Fcampaign%3Demail%5Fforward), or they can [subscribe free here](https://www.implicator.ai/subscribe/?utm%5Fsource=newsletter&utm%5Fmedium=forward&utm%5Fcampaign=forward%5Fto%5Fcolleague).

The same estimate put all U.S. data centers at about 176 terawatt-hours in 2023\. The 403 hyperscale buildings alone come to roughly 46% of that, though the authors caution that the two figures were built differently.

"Policy has historically lagged the expansion of the sector," the paper says, though it adds that state and federal regulators are increasingly focusing on data center reporting and energy use. Francesca Dominici, its corresponding author and director of the Harvard Data Science Initiative, said the aim was to give "the public, policymakers, investors, and developers a much more concrete, up-to-date picture than was available before."

"Decisions about where data centers go and how they're powered are critical environmental and public health decisions, and they should be made with good data," she said.

Many of those decisions are being made locally. On Nov. 3 voters backed restrictions or rejected pro-development measures in all but one of 41 local data center ballot measures across 10 states.

## What the study cannot see

The paper is plain about its limits. It cannot separate AI workloads from ordinary cloud computing, because that split is not publicly disclosed. It covers 403 of the 675 facilities first identified, and the authors expect that gap to pull their totals below those of the full population, though they decline to call the figures a lower bound. It uses annual grid averages from 2023, not hour-by-hour dispatch. A campus with its own dedicated clean supply would emit less than the regional average suggests, and one with its own gas turbines or backup generators would emit more.

It also does not measure what the emissions do to people, which the team plans to pursue in future research. The fossil fuel plants in the regional systems tied to this demand "also release other pollutants, such as fine particulate matter, nitrogen oxides, and sulfur dioxide, that are directly linked to heart and lung disease and premature death," Dominici said. "So when data centers draw heavily on fossil fuel power, the people living near and downwind of those power plants may bear health costs, even if they're far from the data center itself."

Gina McCarthy, a former Environmental Protection Agency administrator and White House national climate adviser, responded in a statement distributed with the study. "Sadly, it doesn't surprise me that the majority of hyperscale data centers rely on fossil fuels for their power generation despite the profound and lasting climate and health impacts that could be - and should be - avoided," she said. "Communities and policymakers need reliable, independent data on what these facilities mean for our air, health, and the climate."

Vanessa Kerry, who directs global and climate health policy in the environmental health department at Harvard's public health school, put the demand to the operators. "AI could transform health care, but as this research shows, much of its growth runs on fossil fuels, which is making us sick," she said. "The companies building AI have plenty of money to do better." They "must invest in new clean power and publish the energy use and emissions of every data center."

The researchers say the numbers are likely higher today. "What we're seeing is already striking, but it's only a snapshot of a rapidly expanding industry," Guidi said. "The data center boom hasn't slowed down since our study period ended in April 2025\. New facilities have been built, existing campuses have expanded, and demand for AI computing continues to grow. That means the electricity use and carbon footprint we're reporting could already be substantially larger now in 2026."

Frequently Asked Questions

What did the Harvard study measure?

It estimated the electricity use of 403 validated U.S. hyperscale data centers between May 2024 and April 2025 and assigned that demand to the power plants in each regional grid. The central estimate is about 82 terawatt-hours and 25.7 million metric tons of CO2.

How much of the power came from fossil fuels?

About 53.9% of the electricity attributed to the facilities. Natural gas supplied 38.2% and coal 15.2%. Nuclear supplied 20.9%, and hydro, wind, solar, biomass and geothermal together 25.3%.

Which states matter most?

Virginia had 142 of the 403 facilities, about a quarter of the electricity and about 7.0 million metric tons of CO2\. Virginia, Oregon, Ohio and Iowa together used more than half of the electricity in the sample.

Do the companies' renewable energy purchases count?

No. The paper says power purchase agreements, renewable energy certificates and net-zero claims are outside the scope of its physical grid-mix attribution unless they show up in the reported generation mix.

What are the study's limits?

It cannot separate AI from other computing, covers 403 of 675 identified facilities, and uses annual 2023 grid averages. It did not measure health effects, which the researchers plan to pursue in future research.

AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/).

[Trump Calls Rejecting Data Centers a ‘Mistake’ Despite 61% OppositionPresident Donald Trump said communities that reject data centers are “making a mistake” in an interview that aired in full on Sunday, even as 61% of U.S. adults opposed a new data center in their areaThe Implicator![](https://www.implicator.ai/content/images/2026/08/20260824-114933-trump_data_centers.webp)](https://www.implicator.ai/trump-data-centers-mistake-61-percent-opposition/)

[Amazon-Backed Texas Gas Plant Could Emit 33 Million Tons of CO2 a YearAmazon is backing a gas plant in Pecos County, Texas, permitted to release up to 33 million tons of carbon dioxide a year. Amazon confirmed Friday that it acquired the GW Ranch site and plans to buy eThe Implicator![](https://www.implicator.ai/content/images/2026/08/2026-08-08-19.32.25-amazon-texas-gas-plant@2x.webp)](https://www.implicator.ai/amazon-texas-gas-plant-33-million-tons-co2/)

[Emerald AI Raises $150 Million to Make Data Centers Flexible Grid LoadsEmerald AI raised $150 million on August 25, 2026\. The round valued the data-center software company at $1.05 billion, while its system slows, pauses, caps or moves selected computing jobs when a utilThe Implicator![](https://www.implicator.ai/content/images/2026/08/20260825-064021-emerald_grid_load.webp)](https://www.implicator.ai/emerald-ai-150-million-flexible-data-center-grid-loads/)