How much energy and water does using AI really take?

Every AI answer is computed in a data center, a building full of servers that need electricity to run and, often, water to stay cool. Two facts pull in opposite directions in 2026. Each simple request keeps getting cheaper in energy, because chips and models are more efficient. Yet the total keeps climbing, because more people use AI, and they use it for heavier jobs.

This article updates the earlier overview of the environmental costs of AI with the International Energy Agency’s April 2026 figures.

What a single request costs

The best measured figure for an ordinary chat comes from Google. In 2025 its researchers reported that the median text prompt in the Gemini app used 0.24 watt-hours of electricity and 0.26 milliliters of water, counting the AI chips, the servers around them, idle machines held in reserve and the building's cooling. A watt-hour is the energy a 1-watt device uses in an hour; a median is the middle value, so half of prompts used less and half used more.

The French company Mistral AI measured more widely. Its lifecycle analysis, which also counts the manufacturing of servers, puts one 400-token answer from its Le Chat assistant at 1.14 grams of CO₂ equivalent and 45 milliliters of water, about three tablespoons. (A token is a word or piece of a word.) The two numbers differ mostly because they draw the boundary in different places, not because one company is far greener.

The International Energy Agency (IEA), an intergovernmental body that tracks world energy, now puts it simply: a simple text query typically uses less electricity than running a television for the same length of time. It adds that the energy needed per AI task has fallen at least tenfold a year recently.

Why the total still grows

The IEA's April 2026 report found that data-center electricity use worldwide grew 17% in 2025, and the part used by AI-focused data centers grew about 50%. It expects data centers to go from 485 terawatt-hours in 2025 to about 950 in 2030, roughly 3% of the world's electricity, with AI-focused sites tripling their use. A terawatt-hour is a billion kilowatt-hours.

Two forces explain this. First, far more requests: the major providers reported about three times as many active users as a year earlier. Second, heavier requests. The IEA notes that video generation, reasoning models that "think" at length, and agents that carry out multi-step tasks on their own can each take hundreds or even thousands of times the energy of a short text answer.

To put the light end in perspective, the IEA calculates that if every ordinary web search in the world were replaced by a simple AI text query, it would add under 4 terawatt-hours a year, less than 1% of what data centers use today. The growth is coming from scale and from heavy tasks, not from people asking short questions.

Water: a local question

Water is used mainly to cool servers and, indirectly, by the power plants that supply them. A few drops per prompt is small for one person. The issue is where data centers are built: a large site drawing on a stressed local supply matters far more to its neighbors than a global average does. Per-prompt figures also say nothing about the training of a model, which Mistral counted at 281,000 cubic meters of water for one large model over its first 18 months.

A worked example: one week of use

Here is an illustrative estimate for one person, using Google's median for short prompts. The heavy-task rows are not measured figures; they show the IEA's range, "hundreds or thousands of times" a simple query.

Illustrative arithmetic: 150 × 0.24 Wh = 36 Wh, about what a 40-watt bulb uses in under an hour. The lesson is not that one chat is harmful. It is that a single heavy task can outweigh a week of ordinary questions.

How can you use AI more frugally?

Match the tool to the job. Most services now offer a small, fast model next to a large one, and a "thinking" or research mode next to a quick reply. Use the small, quick option for rewording, short summaries and simple facts, and save reasoning modes and agents for problems that truly need several steps. Mistral's analysis found that impact rises roughly with model size, so this choice is the biggest lever you have.

Ask once, well. A prompt that includes the context and the format you want often succeeds on the first try. Group related requests in one message rather than opening several conversations. For images and video, generate a few drafts and refine one, instead of producing dozens of variations.

Finally, skip AI when a plain search, a calculator or your own notes will do. Frugal use is less about guilt than about habit: the same choices that save energy usually save you time and money as well. For organizations, the IEA's call for more systematic energy disclosure is worth echoing to suppliers, since measured, comparable numbers are what make a greener choice possible.