Harvard: Gas and Coal Supply 54% of Electricity at the Largest US Data Centers
Harvard has traced where the 403 largest US data centers get their power: 54% of their electricity comes from gas (38%) and coal (15%). Between May 2024 and April 2025, they consumed about 82 TWh (1.8% of the country's energy) and emitted 25.7 million tonnes of CO2, as much as all of Denmark. For developers of AI products, this means energy prices and grid connection queues will become part of the cost of compute. It is better to start collecting energy-consumption metrics now.
AI-processed from The Decoder; edited by Hamidun News
Every query to an AI model and every training run burns kilowatt-hours, and a new Harvard study shows where those kilowatt-hours come from. The researchers used satellite imagery to identify the 403 largest data centers in the US and calculated that from May 2024 to April 2025 they consumed about 82 TWh (a terawatt-hour is a billion kilowatt-hours), or 1.8% of all electricity in the country.
Gas and coal supplied 54% of that energy, and emissions came to about 25.7 million tonnes of CO2 per year, the same as all of Denmark. Imagine you are building an AI service and choosing a hosting region: in Virginia or Ohio, your energy bill, carbon footprint and connection timeline all depend on the local grid.
For anyone calculating the cost of compute, this is no longer abstract environmentalism. It is a question of energy prices and future regulation.
What Harvard Calculated
A team led by Gianluca Guidi and Francesca Dominici published the work in Nature Sustainability. It is one of the first site-level estimates of consumption and emissions for hyperscale data centers. The researchers matched the estimated demand of each site against emissions data for the regional power grids. The result depends on an assumption about server utilization and ranges from 68 to 99 TWh per year. The upper bound is comparable to the consumption of entire states such as Indiana, Michigan or Tennessee.
The central scenario gives about 82 TWh. More than half of that energy goes to four states: Virginia, Oregon, Ohio and Iowa. About 54% of the electricity came from fossil fuels: natural gas supplied 38% and coal 15%. Emissions of 25.7 million tonnes of CO2 per year equal or exceed those of Michigan or of a country like Denmark. The previous comparable estimate, for 2018, was two to three times lower.
There are caveats. This is a snapshot: the period ended in April 2025, and the load from AI agents has only grown since. According to Climate Action Against Disinformation, as reported by the Associated Press, consumption at the largest tech companies rose 240% from 2022 to 2025, and they are missing their own climate targets. The researchers did not estimate health effects, even though gas and coal raise levels of fine particulates, nitrogen oxides and sulfur oxides. And about 2000 more data centers are planned in the US.
Why Operators Are Building Their Own Gas Plants
In its report Powering Artificial Intelligence, the International Energy Agency (IEA) forecasts that global data center emissions will grow from 150 million tonnes of CO2 per year today to 350 million tonnes by 2035. The forecast was revised upward mainly because of US operators, who are building gas power plants right next to their sites so they don't have to wait for a grid connection.
The queues really are huge. In Texas, Virginia, Finland and the UK, connection requests already exceed current peak demand. Almost 40% of the 134 grid connection rules tracked worldwide appeared in the past 12 months. Regulators are considering minimum payments, deposits, penalties for withdrawing from a project and flexible connections, and in some places they require "bring your own generation." The goal is to protect households from rising tariffs and to weed out speculative requests.
Operators have little incentive to cut load at peak times. Renting AI capacity costs from 20 to 30 billion dollars per gigawatt per year, roughly 100 times more than comparable battery capacity. It is easier to install your own gas generation than to shut down a cluster. In addition, almost 40% of the identified AI capacity sits in regions with high or extreme water stress. In 2026, hyperscalers will invest more than 700 billion dollars, mostly in new data centers.
The Antalya Statement: Loud but Non-Binding
At the COP31 climate conference in Turkey, the International Telecommunication Union and the host country presented the "Antalya Statement on AI." It came against the backdrop of a 50% rise in energy consumption by AI data centers in 2025. The document calls for moving infrastructure to zero- and low-emission energy, improving water and electricity efficiency, disclosing environmental impact using a common methodology, and reusing equipment. Server heat is discussed separately: by IEA's estimate, by 2030 it could heat about 10% of households in Europe.
But the statement is not legally binding. All measures are voluntary, and the first review is scheduled for COP32. Simon Stiell, the head of UN climate policy, warned that people must not be made to pay today with high bills and dirty air for vague promises of future AI efficiency. Supporters such as Al Gore hope AI will cut global emissions by 6–9% by 2030, but for now that is a bet, not a measured result.
What to Do in Practice
First, treat energy as part of your cost of goods: if your business model assumes cheap inference, watch tariffs and queues in the regions where you host. Second, requirements to disclose carbon footprint will likely reach customers of enterprise AI services too, so it is better to start collecting consumption metrics in advance. Third, efficiency (smaller models, caching, request routing) is becoming an argument not only for the budget but also for investors and regulators. The Harvard figures are estimates, based on satellite imagery and utilization assumptions, but the direction is clear: the most powerful data centers are still growing on gas.
- Nov 21, 2026Docker rebranded for AI agents, but Sandboxes shipped back in September
- Nov 21, 2026Nvidia doubled profit, but shares fell 2% over the search for a new CEO
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Google to Buy 1.99 Billion Emails from Bankrupt Spirit Airlines for AI for $10 Million
Key Questions Answered
What Harvard Calculated?
A team led by Gianluca Guidi and Francesca Dominici published the work in Nature Sustainability. It is one of the first site-level estimates of consumption and emissions for hyperscale data centers. The researchers matched the estimated demand of each site against emissions data for the regional power grids. The result depends on an assumption about server utilization and ranges from 68 to 99 TWh per year. The upper bound is comparable to the consumption of entire states such as Indiana, Michigan or Tennessee.
Why Operators Are Building Their Own Gas Plants?
In its report Powering Artificial Intelligence, the International Energy Agency (IEA) forecasts that global data center emissions will grow from 150 million tonnes of CO2 per year today to 350 million tonnes by 2035. The forecast was revised upward mainly because of US operators, who are building gas power plants right next to their sites so they don't have to wait for a grid connection. The queues really are huge. In Texas, Virginia, Finland and the UK, connection requests already exceed current peak demand.
What are the key details on The Antalya Statement: Loud but Non-Binding?
At the COP31 climate conference in Turkey, the International Telecommunication Union and the host country presented the "Antalya Statement on AI." It came against the backdrop of a 50% rise in energy consumption by AI data centers in 2025. The document calls for moving infrastructure to zero- and low-emission energy, improving water and electricity efficiency, disclosing environmental impact using a common methodology, and reusing equipment.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
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.