Abstract digital illustration of a data center with binary code overlay and colorful cloud patterns in the background.

Harvard-led study says U.S. AI data centers used 82 TWh and emitted 25.7 million tons of CO2 in a year

Harvard-led research finds U.S. hyperscale data centers used 82 TWh and emitted 25.7 million tons of CO2 in one year.

In short

A Harvard-led study says 403 U.S. hyperscale data centers used about 82 TWh of electricity and emitted 25.7 million metric tons of CO2 from May 2024 to April 2025. The findings sharpen the policy debate over AI, grid power and disclosure rules.

  • 403 U.S. hyperscale data centers used about 82 TWh in the study year.
  • Researchers estimated 25.7 million metric tons of CO2 from those facilities.
  • About 54% of the electricity likely came from gas and coal.
  • The carbon footprint varies sharply by local grid mix, not by AI alone.
  • Authors are building a website to track individual data centers' environmental impact.

U.S. hyperscale data centers consumed about 82 terawatt-hours of electricity and produced 25.7 million metric tons of carbon dioxide in the 12 months from May 2024 through April 2025, according to a Harvard-led study published in Nature Sustainability. The research, reported by The Verge, argues that the climate burden of AI-era computing depends less on the servers themselves than on the local power grid feeding them.

What the study found

The researchers identified 403 hyperscale data centers in the United States using satellite imagery and combined that map with EPA data to estimate electricity use and emissions over the study period. Their conclusion is blunt: these facilities are now a major industrial power load, and their environmental footprint varies sharply from place to place.

According to the study, those 403 facilities accounted for about 1.8% of total U.S. electricity consumption during the period modeled. The carbon output, 25.7 million metric tons of CO2, was compared by the researchers to Denmark’s annual carbon footprint, underscoring the scale of the emissions tied to cloud and AI infrastructure.

The paper also estimates that roughly 54% of the electricity used by hyperscale data centers likely came from gas and coal. That matters because it means the headline climate impact is being shaped not only by how much power the facilities draw, but by whether they sit in regions served by cleaner grids or fossil-heavy ones.

By the numbers

Metric Estimate Why it matters
Hyperscale data centers identified 403 Shows the scale of the facilities the researchers could confirm from satellite imagery
Annual electricity use About 82 TWh Represents a substantial share of national demand
Share of U.S. electricity consumption About 1.8% Puts hyperscale power use in national context
Estimated emissions 25.7 million metric tons of CO2 Highlights the climate cost of electricity-intensive computing
Likely electricity from gas and coal About 54% Signals heavy reliance on fossil fuels in the studied period
Modeling window May 2024 to April 2025 Defines the yearlong period analyzed with EPA data

Why it matters

The findings land at a moment when AI infrastructure is expanding quickly and utilities, regulators and communities are all debating who should pay for the added strain. The study’s central message is that AI’s environmental impact is not fixed by the technology alone. It is driven by the power system behind it.

That distinction matters for readers because the debate over AI often focuses on chips, model size or software efficiency. This study points to a different lever: where the data center is built, which utility serves it and what kind of generation mix feeds it. A facility in one state can have a much lighter carbon footprint than an otherwise similar facility in another.

Gianluca Guidi, the study’s lead author and a postdoctoral fellow at the Harvard T.H. Chan School of Public Health, summed up that point this way:

“the same computing infrastructure can have dramatically different consequences depending on the electricity system supporting it”

The practical consequence is that disclosure rules and clean-power requirements may matter as much as engineering improvements. If policymakers, investors and local residents cannot see how much electricity a proposed data center will consume, or what mix of fuels will supply it, they cannot judge the true climate cost.

How the regional grid changes the math

The study’s comparison between Wyoming and Oregon illustrates the point. The researchers found that grid emissions in coal-reliant Wyoming were roughly seven times higher per kilowatt-hour than in hydropower-rich Oregon. For a power-hungry facility, that difference can overwhelm efficiency gains inside the building itself.

That is why the authors argue that emissions are not simply a byproduct of AI growth. They are also a function of siting decisions, utility planning and the speed at which new generation comes online. The same server farm can look very different on paper depending on whether it draws from a cleaner grid or one still anchored in fossil fuels.

How did researchers reach the estimate?

The study used satellite imagery to identify 403 hyperscale data centers across the United States, then modeled their power use and emissions over the period from May 2024 through April 2025 using EPA data. That approach is notable because hyperscale facilities are not always easy to track through public records alone.

Hyperscale data centers are massive computing facilities built to support cloud services and AI workloads at large scale. They are only a fraction of all data centers in the country, which means the broader sector’s total footprint could be even larger than this study’s estimates for hyperscale sites alone suggest.

What the study did not count

The authors say the estimate does not include local air pollution from new gas infrastructure or from on-site gas turbines. That omission matters because electricity emissions are only part of the environmental story. Building or expanding gas supply to serve a large data center can add methane-related and local pollution impacts that do not appear in the carbon tally.

In other words, the 25.7 million metric tons of CO2 is likely not the full environmental cost of the facilities studied. It is a major piece of the picture, but not the whole one.

Background: why data centers are under scrutiny

For years, the climate conversation around digital infrastructure was mostly about efficiency: better cooling, smarter chips and more optimized workloads. The rise of AI has changed the scale of the debate. Training and running large models, along with the cloud systems that support them, has intensified concern about grid demand, fossil-fuel dependence and emissions.

That concern is reflected in the new Harvard-led paper and in the policy response gathering around it. A bipartisan congressional bill could require greater disclosure of data centers’ electricity use, a sign that lawmakers are beginning to treat transparency as a prerequisite for managing the sector’s growth. If enacted, such a measure could make it easier to verify claims about efficiency, clean power and local impact.

As Gina McCarthy, the former White House national climate adviser and EPA administrator, put it in response to the study:

“This study is an important step.”

The paper also lands in a market where operators are racing to secure land, power and permits, often in regions that can still accommodate very large loads. That competition raises a simple question: if a company needs to power a massive facility, will it choose the cheapest path, the fastest path or the cleanest one?

How the emissions add up

The study’s estimates point to a data-center boom with real grid consequences. The 82 TWh figure is large enough to place hyperscale facilities among major electricity users in the United States, and the authors note that these facilities used more electricity than many countries individually. That comparison is a reminder that the scale of AI infrastructure is increasingly geopolitical as well as technical.

One reason the study stands out is that it does not treat hyperscale data centers as an abstract category. By identifying facilities and tying them to grid data, the researchers connect server growth to the physical reality of power plants, transmission lines and regional generation mixes.

John Sterman, director of the MIT System Dynamics group and the MIT Sloan Sustainability Initiative, underscored the unavoidable carbon issue in two separate remarks quoted by The Verge. He said,

“I think AI can be extraordinarily helpful”

and also,

“Once you allow these data centers to emit the CO2, you can never undo that.”

Those statements capture the central tension in the debate: AI may bring productivity gains and scientific benefits, but the emissions linked to its infrastructure are immediate and cumulative. Once released, carbon dioxide remains in the atmosphere for a long time, so power choices made now can shape climate outcomes far into the future.

What the comparison with Denmark means

Researchers compared the 25.7 million metric tons of emissions from the studied U.S. hyperscale data centers with Denmark’s annual carbon footprint. That does not mean the data centers and the country are interchangeable in scale or responsibility, but it does help frame the emissions in terms the public can grasp.

Climate statistics can be hard to interpret in the abstract. A country-level comparison makes the point that the carbon footprint of a relatively narrow slice of the digital economy is already national in size. That can sharpen debates over whether voluntary corporate targets are enough, or whether policy will need to move faster.

What happens next?

The most immediate next step is transparency. The study’s authors say they are building a website to track the environmental impact of individual data centers, which could give the public, journalists and policymakers a more granular look at where emissions are concentrated.

Beyond that, three questions will shape the next phase of the debate:

  • How many additional U.S. data centers are planned or already under construction?
  • Will the bipartisan disclosure bill advance, and what exact reporting would it require?
  • Will operators disclose enough information to verify or refine estimates like these?

Those questions are important because the current paper, while significant, still covers only a subset of the wider data-center universe. If hyperscale facilities account for this much electricity and carbon, the policy stakes only rise as more sites come online.

The timeline matters too. The study models the year from May 2024 through April 2025, and The Verge reported the findings on November 16, 2026. That means the numbers arrive as a snapshot of a rapidly changing sector rather than a static endpoint. If new data centers have opened since then, the footprint may already be larger.

Timeline

Date Event
May 2024 Start of the modeled electricity and emissions period
April 2025 End of the modeled electricity and emissions period
November 16, 2026 The Verge reported the Harvard-led study

The regulatory and market pressure point

The study’s most consequential insight may be its policy implication: AI’s carbon impact is partly a siting problem. That makes disclosure and utility planning central. If a company wants to build a large facility, regulators may increasingly ask not just how much power it needs, but where that power comes from and what else the grid will have to absorb to serve it.

For utilities, that could mean faster transmission upgrades, more generation investment and tougher questions from customers about who bears the cost. For operators, it could mean greater pressure to prove that promised clean power is real, additional and available at the right time, not just matched on paper elsewhere.

For communities hosting new facilities, the study gives added weight to local debates over air quality, water use, land use and grid strain. The emissions numbers are national, but the impacts are often local.

Bottom line

The Harvard-led study does not say AI is inherently doomed to be carbon-intensive. It says something more specific and more actionable: the climate cost of AI infrastructure is being determined by the electricity system around it. If policymakers want lower emissions, the place to start is not only the server room, but the grid.

Frequently asked questions

How much electricity did U.S. hyperscale data centers use in the study?

U.S. hyperscale data centers used about 82 TWh of electricity, according to the Harvard-led study reported by The Verge. The estimate covers 403 facilities and the period from May 2024 through April 2025.

How much CO2 did the data centers emit?

They emitted an estimated 25.7 million metric tons of CO2 over the year studied. The researchers compared that total with Denmark’s annual carbon footprint to show the scale of the emissions.

Why does location matter so much for data center emissions?

Location matters because the local electricity mix determines how dirty each kilowatt-hour is. The study says Wyoming’s grid emissions were about seven times higher per kWh than Oregon’s, showing that similar facilities can have very different climate impacts.

What policy changes could follow this study?

A bipartisan congressional bill could require greater disclosure of data centers’ electricity use. If that advances, it could help regulators and the public verify claims about power demand, emissions and the real climate impact of new facilities.

Share this 🚀