What AI actually costs the planet

AI has a real environmental cost, and it’s worth understanding plainly, without the hand-waving that says it’s nothing or the catastrophising that says it’s everything. The cost comes in two phases of very different size.

Training vs use

Training a large model is energy-intensive and occasional: a big one-off cost, amortised across everyone who later uses the model. Using it (each prompt and response, called inference) is far smaller per go, but it happens billions of times, so at scale it adds up.

There is also water. Many data centres cool with it, and that matters most in places already short of it.

Keeping it in proportion

A single chat is a small cost, roughly in the neighbourhood of other routine digital activity. Not a flight. The honest concern is aggregate and growth: a lot of people using AI a lot, and infrastructure expanding to meet them. How clean the local electricity grid is changes the answer a great deal.

The useful question isn’t “is one prompt bad?”. It’s “is all of this, growing this fast, on this grid, sustainable?” That one belongs to providers and policy as much as to users.

The growth is the story

The IEA’s reporting puts numbers on it. Global data-centre electricity demand grew around 17% in 2025, with AI-focused facilities growing considerably faster, and the agency expects data-centre consumption to roughly double by 2030. In the United States, data-centre expansion accounts for something like half of projected electricity demand growth.

Two things are true at once, and arguments about this usually go wrong by picking one. Per-prompt efficiency is genuinely improving; providers now compete on getting the same work done for fewer tokens, which is a real reduction per unit of work. Total consumption is rising fast anyway, because usage is growing faster than efficiency. Gains that get spent on more usage are not reductions.

What one person can do is use the tools deliberately. That’s covered in using AI without the waste.