The buildings that power search engines, streaming and artificial intelligence are a much bigger source of climate pollution than earlier counts suggested, according to a study published on Monday, November 16.

Researchers led by Harvard University estimated that 403 of the largest data centers in the United States released about 25.7 million metric tons of carbon dioxide in a single year. That is roughly what the whole country of Denmark emits. The Associated Press reported that the figure is two to three times higher than previous studies had found.

The study was published in the journal Nature Sustainability. On the same day, negotiators at the United Nations climate talks in Turkey launched a voluntary pledge on making artificial intelligence cleaner.

What the researchers measured

A data center is a large building filled with computers that store information and run online services. Those computers run day and night, and they need more power again to keep them cool.

The study looked at 403 big U.S. facilities over the 12 months from May 2024 to April 2025. It estimated that together they used 82 terawatt-hours of electricity. According to the AP, that is enough to power nearly 7 million homes for a year.

More than half of that electricity came from burning fossil fuels, the study found. That is where the carbon dioxide comes from: not from the buildings themselves, but from the power plants that feed them.

The senior author is Francesca Dominici, a professor of biostatistics at Harvard.

Why the number is an estimate

There is an important limit to keep in mind. The figures are not readings from a meter.

Technology companies do not publish how much electricity each of their data centers uses, or how much pollution results. So the researchers had to build the picture from the data that is available and from computer modelling.

Outside experts told the AP the method was reasonable. It still means the result is an estimate, and other teams using other methods could reach a different number. The fact that this one is two to three times higher than earlier attempts shows how uncertain the field is.

It is also a snapshot of the recent past. The study period ended in April 2025.

Before the AI boom took hold

That timing is one of the most striking points in the report. The pollution measured here came before the rush to build data centers for artificial intelligence had fully got going.

AI systems need far more computing power than ordinary web services. Training a large model can keep thousands of special chips running for weeks. Answering questions from millions of users keeps them running after that.

Professor Dominici pointed to plans for about 2,000 more data centers as a reason for added concern, the AP reported.

A separate report released at the same time, from a group called Climate Action Against Disinformation, found that electricity use at the biggest technology companies rose 240 percent between 2022 and 2025, putting their own climate targets at risk.

A new pledge at the climate talks

At the UN climate conference in Antalya, Turkey, known as COP31, the host country and the UN’s telecommunications agency launched what they called the Antalya Artificial Intelligence Pledge.

It is voluntary. It asks companies and governments to run data centers on clean power and to use energy and water efficiently. It asks them to follow environmental standards when designing and managing AI, and to reuse or recycle equipment. It also asks them to assess and share the environmental impact of their data centers, and to support the use of AI to help tackle climate change.

UN News reported that the pledge had drawn 80 endorsements and 271 commitments.

The point about sharing the numbers speaks directly to the problem the Harvard team ran into. If companies shared their real figures, studies like this one would not need to rely on modelling.

Two views of the same technology

People at the talks did not agree on how worried to be.

Former U.S. Vice President Al Gore said he was less concerned about AI’s emissions than some others and was hopeful the technology could help fight climate change. He called carbon-free power for data centers essential.

An industry group leader told the AP that worries may be overblown and that AI would help solve its own energy problems.

Others were doubtful. A spokesperson for Climate Action Against Disinformation said companies were offering hope while hiding the harms.

Fengqi You, a Cornell University researcher who was not part of the study, offered a middle view. “Ultimately, the environmental footprint of AI is not fixed,” he told the AP. In other words, it depends on choices: where the buildings go, what powers them, and how efficiently they run.

Could waste heat help?

One idea getting attention is to use the heat that data centers throw away.

Computers turn almost all the electricity they use into heat. In most facilities that heat is simply vented. In some parts of Europe it is piped into district heating systems, which warm many buildings from one shared source.

The International Energy Agency estimates data centers could heat about 10 percent of European homes by 2030 if that heat were captured and fed into such networks, the AP reported. A researcher at the University of Colorado Boulder noted that liquid-cooled data centers give off steady heat at a useful temperature.

This does not cancel out the pollution from generating the electricity in the first place. It does mean less fuel needs to be burned somewhere else to heat homes.

What it means for everyday users

No single search, video or chatbot question causes much pollution. The issue is scale. Billions of small requests add up to a demand for power that has to be met every hour of the year.

That demand can also affect people who never use the services. New data centers compete for electricity and, in some places, for water used in cooling. Where power is tight, extra demand can keep older coal and gas plants running longer.

For now, individuals have little way to know how clean a particular service is, because the figures are not published.

What to watch next

The first thing to watch is disclosure: whether large companies sign the Antalya pledge and start publishing data for each facility. Follow-up studies covering 2025 and 2026, when AI-related building picked up, would show how fast emissions are growing. COP31 itself is due to end this week, and it is not yet clear whether any wording on technology and energy will make it into the final agreement.

The Harvard study does not settle how much AI will cost the climate. It does give the clearest estimate so far of where things stood before the boom, and it is a higher starting point than many had assumed.

Key Questions Answered

How much do U.S. data centers pollute? The study estimates about 25.7 million metric tons of CO2 in a year from 403 large facilities, similar to Denmark’s annual emissions.

Is that a measured figure? No. It is modelled, because companies do not publish their power use.

What is the Antalya AI Pledge? A voluntary set of commitments launched at COP31 asking companies and countries to run AI on clean energy and publish its environmental impact.

© 2026 NatureWorldNews.com All rights reserved. Do not reproduce without permission.