Is using AI bad for the environment?
Real numbers, and why they disagree.
You might have read that one ChatGPT question uses about 3 watt-hours of electricity, or ten times as much as a Google search.
In February 2025, the research group Epoch AI redid the sums with newer hardware and models. Its estimate for a typical GPT-4o question was about 0.3 watt-hours, roughly ten times less.
Watt-hour (Wh): a unit of energy. A 10-watt LED bulb left on for one hour uses 10 Wh.
In August 2025, Google published figures for a typical (median) text prompt to Gemini:
Google said the energy per prompt had fallen 33 times in twelve months.
Google said that if it had counted only the AI chips themselves, the figure would have been 0.10 Wh, less than half. It got 0.24 because it included more of the data centre: cooling, spare machines kept running, and so on.
That's the main reason the published numbers vary so much. It isn't that someone's lying. They're measuring different things.
In July 2025, Mistral published a lifecycle analysis for one of its models. That's a count of every stage: making the hardware, building the data centre, generating the electricity, as well as running the model.
Its water figure per answer is much higher than Google's because it counts water used in making the electricity and the hardware.
When you see a number, ask which box it was drawn around.
A single prompt is a tiny amount of energy. The real debate is about the total: billions of prompts a day, plus training ever-larger models, plus the new data centres being built to do it.
You don't need to feel guilty about asking a question. It's reasonable to avoid waste.
Pick the right-sized model. A small, fast model is fine for simple jobs.
Send the part that matters, not the whole 200-page document.
Start fresh chats instead of dragging a huge one along.
Be sceptical of any number without a method. Ask what was counted.