AI and Climate: Data Centers Use 415 TWh a Year, While the Benefits Remain Mostly Forecasts

The question sounds almost like a political slogan: will AI burn the planet or help save it? The numbers give a less dramatic but more useful answer. According to the International Energy Agency (IEA), data centers consumed about 415 TWh in 2024, roughly 1.5% of the world's electricity. By 2030, under the IEA's base-case forecast, that figure could grow to 945 TWh. That is comparable to Japan's annual consumption. At the same time, data centers account for less than 1% of energy-related CO2 emissions, so today the problem is local rather than global. The concern is the speed of growth.

How much power AI actually consumes

The most visible figures come from model training and, especially, inference: with billions of queries, small amounts add up to megawatts. Per-query estimates vary widely and depend on methodology. In 2025, Google reported that the median Gemini query consumes 0.24 Wh, while OpenAI cited about 0.34 Wh for ChatGPT. For a single user, that is negligible. But heavy workloads such as video generation, long reasoning chains, and agents that make dozens of model calls for a single task push the average higher.

The benchmarks the debate relies on:

• Global data centers, 2024: about 415 TWh, roughly 1.5% of global electricity (IEA).
• 2030 forecast: about 945 TWh in the base-case scenario.
• United States, 2023: 4.4% of national electricity consumption, rising to 6.7–12% by 2028, according to Lawrence Berkeley National Laboratory.
• Google's emissions in 2024 were 48% above 2019 levels; Microsoft's have grown about 29% since 2020.
• Median Gemini query: 0.24 Wh, according to Google itself.

The individual model query has become noticeably more efficient. The problem is that there are vastly more of them, and the load is concentrated in a handful of regions: Virginia, Texas, Ireland, and the area around Singapore. There, grids hit their limits before new generation comes online.

What AI can offer the climate

The other half of the question has real grounding too. Neural-network weather forecasting models such as GraphCast and similar systems produce a global forecast in minutes on a single accelerator, whereas classical numerical methods need supercomputers. Algorithms help balance load on grids with a high share of solar and wind, detect methane leaks in satellite imagery, optimize heating and cooling in buildings, and screen materials for batteries. The IEA notes that, if widely adopted, such applications could potentially cut emissions by more than data centers add. The key word here is "potentially": there is no automatic effect, and adoption is slow and uneven.

There is also a flip side. Oil and gas companies use the same models to find and develop fields, and comparative calculations often leave that contribution to emissions out. Efficiency gains also frequently lead to higher demand: when computing gets cheaper, people use more of it. Economists have known this effect since the 19th century as the Jevons paradox.

What to watch next

To avoid drowning in headlines, it is worth tracking a few things. First: where the electricity for new campuses comes from. Contracts for nuclear and geothermal generation have a lasting effect, while extending the life of coal and gas plants wipes it out.

Second: whether companies disclose consumption by hour and by site, rather than only annual averages paired with green certificates. Third: whether reporting counts water and hardware emissions, not just electricity. For businesses, the takeaway is practical: choose the region and timing for heavy workloads, account for the grid's carbon intensity, and measure energy per unit of output rather than per query.

For regulators, transparent data matters most. Without it, any talk of "good" or "bad" AI remains an argument about faith.