What it consumes, and why it's so hard to quantify.

Running an AI consumes electricity and water. That's a fact. Everything else calls for caution, because the figures doing the rounds are rarely comparable with one another.

Two very different kinds of consumption

Training: building the model. A huge outlay, concentrated over a few weeks, done once.

Inference: every question asked. A tiny outlay per question, but multiplied by billions of questions a day, every day.

Conflating the two is the most common mistake. A spectacular figure about training a model tells you nothing about what your question costs, and vice versa.

Water, which catches everyone off guard

Data centres run hot, and many cool themselves with water. That water is drawn locally, often where it's already contested. Projects have been blocked for precisely that reason, not for their electricity use.

That's what makes the issue political rather than merely technical: electricity can be transported, water cannot. A data centre can draw power generated elsewhere; it can't drink another region's water.

Why the figures contradict each other

• The scope changes. Does it count manufacturing the hardware? Building the facility?
• Location changes everything. The same electricity use doesn't produce the same emissions depending on the country generating it.
• Companies publish little, and not in the same way.
• It moves fast, in both directions: models get more efficient, and we use them far more.

Practical consequence: be wary of any figure presented without its scope. "One AI query uses as much as an X" is meaningless unless it says which model, what length of answer, and what's being counted.

What can be said without getting it wrong

That usage is growing far faster than efficiency gains, so the total is rising. That the water question is local, not global. And that the order of magnitude of a single query remains small compared with many other everyday digital actions, which doesn't cancel out the total.

Do this right away: next time you read a figure about AI consumption, look for what it counts. If the scope isn't stated, the figure is worthless.