Will AI save the planet or heat it up? The climate balance sheet nobody has closed

THE WRONG QUESTION. Is AI good or bad for the climate? The question sounds simple, but it is badly framed. It compares two things that are not measured the same way. Data center emissions are counted every year, in tonnes, in official reports. AI's climate benefits are almost always scenarios: what could happen if many applications were adopted at scale. Our thesis is this: AI's net climate balance could be clearly positive in the long run, but today it is not proven. And it will not turn positive on its own. It will do so only if AI runs on clean power and is pointed at problems that cut emissions. If it is used to pump more oil, the balance gets worse.

WHAT IT ALREADY COSTS. Start with the negative side, because that is the measured one. In 2025 the International Energy Agency (IEA) estimated that data centers emit about 180 million tonnes of CO2 a year. That is less than 0.5% of energy-sector emissions. It is small in the global total, but it is growing fast. In its April 2026 update, the IEA reported that data center electricity use rose 17% in 2025. Global electricity demand grew 3%. The agency expects data center consumption to double by 2030, and consumption by AI-focused facilities to triple. The companies themselves confirm the trend. According to its 2026 environmental report, Google used 37% more electricity in 2025 than the year before. Its total emissions rose 18%, to about 14.5 million tonnes. Its water consumption grew 34%, to 10.9 billion gallons. Microsoft reported 20.3 million tonnes for fiscal 2025, up 25% from the previous year. Its report says it bluntly: sustainability solutions “are not scaling fast enough to meet demand.” Both companies pledged net-zero or carbon-negative status by 2030. Today they are heading the other way.

WHAT IT ALREADY DELIVERS. The positive side is real, and part of it has been measured, though at a smaller scale than the marketing suggests. The strongest case is methane, a gas that traps far more heat than CO2 in the short term. The UN's satellite alert system (MARS) uses more than 30 instruments plus AI models to find oil and gas leaks. According to the UN Environment Programme (UNEP), in July 2026 AI flagged 80% to 85% of confirmed detections before an expert reviewed them. UNEP estimates the system has helped act on sources that had emitted 1.2 million tonnes of methane. It compares this to taking almost 24 million gasoline cars off the road for a year. Weather forecasting is another real example. Google DeepMind's GenCast beat Europe's reference forecasting system on 97.2% of 1,320 targets evaluated, according to its paper in Nature. Better wind and solar forecasts help run power grids with more renewables. Fusion and new materials are still at the lab stage. Since October 2025, DeepMind has worked with Commonwealth Fusion Systems to control the plasma (the ionized gas, at millions of degrees, where fusion happens) in its SPARC experimental reactor. But that company's first commercial plant is not expected until the early 2030s. DeepMind's GNoME predicted 2.2 million crystals, about 380,000 of which the company considers stable. When the study was published, outside researchers had made only 736 of them. Several chemists have also questioned whether many of those materials are truly new.

THE PROMISE AND ITS FINE PRINT. Here lies the big misunderstanding. The most quoted figure says AI could cut global emissions by 5% to 10% by 2030. It comes from a 2023 BCG report commissioned by Google and co-written with Google. It is a modelled potential, not a forecast, and it comes from an interested party. The IEA is more cautious. It estimates that widespread adoption of AI applications that already exist could avoid about 1.4 billion tonnes of CO2 in 2035, roughly 5% of energy-related emissions. That far exceeds what data centers will emit. But the agency itself warns that it is “far smaller” than what is needed to tackle climate change. Two forces also pull the other way. The first is the rebound effect, or Jevons paradox: when something becomes more efficient and cheaper, people use more of it, and total consumption can rise. Researchers Sasha Luccioni, Emma Strubell and Kate Crawford make this point in a 2025 paper: technical efficiency alone does not guarantee less environmental harm. The second force is oil. According to Bloomberg and The Atlantic, Microsoft sells cloud and AI services to ExxonMobil to analyze oilfields. A former employee quoted by The Atlantic claimed one of those deals could generate more emissions than the company pledged to remove. A group of shareholders asked Microsoft in December 2025 to assess that risk. The company argues these deals help oil firms cut emissions. Google, for its part, has committed not to build custom AI for oil extraction.

TWO INFLATED STORIES. This debate gets exaggerated in both directions. On one side is greenwashing (looking greener than you are). Tech companies highlight “market-based” emissions, which subtract the renewable energy certificates they buy. Using that method, Google's operational emissions fell 2% in 2025. Using the “location-based” method, which reflects the actual power mix in each area, Microsoft's figures have more than doubled since 2020, according to independent analyses. On the other side is doom. The most famous example is the claim that writing an email with ChatGPT uses a bottle of water. The original figure, 519 milliliters, referred to a 100-word email that the model counted as about ten AI calls, not a single question. Even so, it spread as the cost of “every question.” Google estimates that a typical Gemini query uses 0.24 watt-hours, roughly what a TV uses in nine seconds, and 0.26 milliliters of water. These are the company's own figures, and they leave out chip manufacturing and model training. Still, they put the per-query cost in perspective. The problem is not your query. The problem is billions of queries added together, and new data centers being built at a record pace.

OUR TAKE. In the short term, AI is warming the planet, and that should be said plainly. Emissions at its biggest backers are rising and their 2030 goals are slipping away. The measured benefits, such as detected methane leaks or better weather forecasts, are real but still small next to that growth. In the long term we are optimistic, with reason. If AI speeds up fusion, batteries and carbon-capture materials, the balance could become strongly positive. The same science could help eradicate diseases and build an economy of energy abundance. But that curve will not bend by itself. We propose three tests for any AI climate promise. First: are there avoided tonnes measured by a third party, or only scenarios? Second: are emissions also reported with the location-based method, with AI broken out separately? Third: does the company sell AI to extract more fossil fuels? The implications are concrete. Regulators should require data center emissions to be reported in a standardized, audited way. Companies should run AI on real clean power, matched hour by hour, not just on certificates. And users can stop feeling guilty about every query and start holding the builders of the infrastructure to account. The climate is not decided in your chat window. It is decided by where data centers plug in, and by which problems we choose to solve with them.