Data-centre water is a distraction. Energy isn't.

“Every ChatGPT question uses a bottle of water.” You’ve probably read it on social media, perhaps with an invitation to feel guilty. It’s one of those sentences that contains a real problem, told with a wrong number.

How much one question uses

In August 2025 Google published a measurement of a typical (median) Gemini text request: about 0.24 watt-hours of energy and 0.26 millilitres of water, i.e. less energy than nine seconds of TV and about five drops of water. These are the company’s figures, not independently verified, and they apply to text requests, not to generating video or running an agent for hours. But the order of magnitude is clear: a single question weighs very little.

Kyle Hill, in a video I recommend, says it bluntly: focusing on the water of individual answers is a distraction.

The real problem is scale

The problem isn’t your question. It’s that billions of people and millions of agents ask questions all day, and that models keep getting bigger and reasoning longer.

The International Energy Agency estimates data centres used about 415 terawatt-hours in 2024, 1.5% of the world’s electricity, and will reach about 945 by 2030: more than double, roughly Japan’s current consumption. AI is the main driver of that growth.

That means new power plants, grids to upgrade, pressure on energy prices where data centres cluster, and competition for clean electricity between AI and the rest of an economy that needs to decarbonise.

Both things at once

They look contradictory: minimal impact per question, huge impact in total. They aren’t. It’s the same pattern as emails, video streaming and low-cost flights: each single use is negligible, the sum isn’t.

Hence two conclusions I believe are both true:

You needn’t feel guilty for using AI. A question to an assistant weighs less than many things we do without thinking. If it saves you a car trip, a printout or an hour with the computer on, the balance may even be positive.
We must demand transparency and system-level choices. How much energy new data centres use, where it comes from, who pays for grid investment. These are public decisions, not individual ones.

What you can (really) do

Use the right model for the task. A small, fast model uses a fraction of a large one reasoning for minutes. It’s cheaper too (see the guide on choosing a model).
Don’t generate videos and images endlessly for fun. Video generation is by far the most energy-hungry use.
Ask providers for data. Companies publishing verifiable measurements deserve to be rewarded.
Take an interest in your area’s energy choices. That’s where it’s decided whether AI grows renewables or gas plants.

In one line

Your question to AI isn’t the problem. The problem is an industry doubling its consumption in six years: and that’s tackled with energy policy, not guilt.

Sources

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.