US Data Centers Emit 25.7 Million Tons of CO2 From Their Electricity, Equal to Denmark
Harvard researchers estimated the carbon dioxide (CO2) emissions from the electricity used by the 403 largest US data centers: roughly 25.7 million tons over a year, equal to all of Denmark and about three times the figure for 2018. These facilities consumed 1.8% of the country's electricity (82 terawatt-hours), and 54% of that was generated by fossil fuels. That is why choosing a cloud region should depend not only on price and latency, but also on what the local grid runs on.
AI-processed from TNW; edited by Hamidun News
How much climate impact AI infrastructure really has has mostly been debated on the basis of gut feeling and press releases. Now there are numbers. A study led by Harvard, published in the journal Nature Sustainability, assessed the 403 largest US data centers. Over one year, from May 2024 to April 2025, the electricity powering them released roughly 25.7 million tons of carbon dioxide (CO2) into the atmosphere. That is what all of Denmark emits in a year. It is about three times the figure for 2018, when data center emissions were last estimated in a peer-reviewed study.
What was actually counted
These are so-called hyperscale data centers: the largest facilities built by cloud giants and operators of AI capacity. In the central scenario, they consumed about 82 terawatt-hours of electricity. That is roughly 1.8% of all US electricity over the year. A terawatt-hour is a billion kilowatt-hours; for comparison, an ordinary apartment uses a few thousand kilowatt-hours per year.
Importantly, the researchers did not count the data centers themselves, but the power plants that supply them. More than half, about 54%, of the electricity attributed to these facilities came from fossil fuels. Across all scenarios the range is wider: from 68 to 99 terawatt-hours and from 21 to 31 million tons of CO2.
The geography is very uneven. Virginia, Oregon, Ohio and Iowa together account for more than half of all the electricity counted. But Oregon ranks second in consumption and only sixth in emissions, because it relies heavily on hydropower. A simple illustration: the same server hall built in Oregon and in a state with coal plants will consume the same amount of energy, yet its carbon footprint will differ.
The study's senior author, Harvard professor of biostatistics Francesca Dominici, compared the result to the emissions of an entire state such as Michigan or a small European country.
What the estimate leaves out
The main caveat: the figures are modeled, not measured. Operators and utilities are not required to disclose how much each individual data center consumes. So the team compiled a list of facilities from industry research and satellite data, then estimated the emissions of power plants in regional grids.
There are also assumptions worth keeping in mind:
- Clean-energy contracts that operators buy were not counted. If a company pays for wind or solar, the model still treats its data center as drawing on the regional grid mix.
- Facilities that generate their own electricity on site were not in the sample. The practice is growing; Anthropic's purchase of gas generators is a telling example.
- Not every data center on the list is devoted to AI. Some capacity serves ordinary cloud computing, streaming and data storage.
In addition, the construction boom has continued beyond the study period. Lead author Gianluca Guidi said that today's figures, in 2026, could be "substantially higher."
The dispute and the political backdrop
Experts considered the method reasonable: given how closed companies are, modeling remains the only way to obtain an estimate. The industry disagrees. Caleb Maxx, president of the National Artificial Intelligence Association, told AP that concerns about energy and water use may be exaggerated.
The study came out as voters in the US increasingly oppose new facilities. The same day, at the COP31 climate talks in Antalya, a voluntary pledge for data centers was launched: to use clean energy and to disclose data on energy, water and emissions.
What this means in practice
For developers and founders of AI products, the takeaway is pragmatic. First, questions about the carbon footprint of your infrastructure will soon come from customers, regulators and local authorities. Second, choosing a cloud region is no longer only a question of latency and price: under the same load, Oregon and Virginia produce different footprints. Third, if disclosure becomes the norm, it is better to start measuring consumption now.
The honest limitation remains the same: this is an estimate, not a meter reading. Until operators publish real data, the debate will be conducted on models. But an order of magnitude three times higher than in 2018, on a par with Denmark, is hard to dismiss as a methodological error.
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026Claude Code Speeds Up Snowboard Kids 2 Decompilation: From 20% to 45% of Functions in Three Months
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Anthropic cuts Claude Fable 5.2 price by 20%, and on the coding test it nearly matches Opus 5.5
Key Questions Answered
What was actually counted?
These are so-called hyperscale data centers: the largest facilities built by cloud giants and operators of AI capacity. In the central scenario, they consumed about 82 terawatt-hours of electricity. That is roughly 1.8% of all US electricity over the year. A terawatt-hour is a billion kilowatt-hours; for comparison, an ordinary apartment uses a few thousand kilowatt-hours per year. Importantly, the researchers did not count the data centers themselves, but the power plants that supply them.
What the estimate leaves out?
The main caveat: the figures are modeled, not measured. Operators and utilities are not required to disclose how much each individual data center consumes. So the team compiled a list of facilities from industry research and satellite data, then estimated the emissions of power plants in regional grids. There are also assumptions worth keeping in mind: Clean-energy contracts that operators buy were not counted.
What are the key details on The dispute and the political backdrop?
Experts considered the method reasonable: given how closed companies are, modeling remains the only way to obtain an estimate. The industry disagrees. Caleb Maxx, president of the National Artificial Intelligence Association, told AP that concerns about energy and water use may be exaggerated. The study came out as voters in the US increasingly oppose new facilities.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.