Does it use a lot of energy and water?

One request uses a little electricity. Billions of requests, plus training the models, add up to a real share of the power grid. And every number depends on what was counted. That is the whole answer; the rest is learning to read the numbers.

Start with the totals. The International Energy Agency estimates that the world’s data centers used about 485 terawatt-hours of electricity in 2025, after growing 17 percent in a year. It expects that to roughly double by 2030, to about 950, or around 3 percent of the world’s electricity.¹ Data centers run much more than AI, but electricity use at AI-focused centers grew faster still, by 50 percent in 2025.¹ The share is larger in some places. In 2023, data centers used 4.4 percent of all electricity in the United States, a national laboratory reported.²

Now one request. Google reported in August 2025 that the median text prompt to its Gemini apps used 0.24 watt-hours of electricity and 0.26 milliliters of water, as of May 2025.³³ This is a company measuring its own service, in a paper posted without outside peer review. Its best lesson is about counting. Counted the narrower way of earlier studies, with only the AI chips in the most efficient data centers, the same prompt came to 0.10 watt-hours. Counting the whole fleet gave 2.4 times as much.³ Figure 15.1 shows both counts, and what neither one includes.³

“A bottle of water per conversation.” That phrase traces to a study by university researchers. Its 2025 version estimates that GPT-3, a 2020 model,⁴ “drinks” a 500-milliliter bottle for roughly 10 to 50 medium-length responses, depending on when and where it runs.⁵ It is an estimate for one older model, not a measurement of today’s tools. It also counts a different kind of water. Most of its total is water that power plants use to make the electricity, not water used for cooling at the data center.⁵ Google’s figure counts only water used on site.³ Both can be right, each inside its own boundary.

People disagree mostly about boundaries and about where totals are heading. Google says its figures are “one or two orders of magnitude smaller than many existing estimates” and credits measuring real traffic.³ The IEA notes that video, reasoning and agent tasks “can consume hundreds or thousands of times more energy per query than simple text generation.”¹ It adds that cheaper, more efficient AI also draws more use, so total demand is still set to rise in the near term.¹

Chapter 2 explains why every reply costs computing: each piece of text is a fresh prediction. Chapter 5 explains why size matters, since a bigger model does more arithmetic for each piece. Spot it today: when you meet an energy or water number, ask what kind of request it measures, whether it is a median or an average, and whether the water includes the power plant.