Energy & environment

'One AI search boils a bottle of water!' The headline sounds precise and alarming. Is it true? The honest answer is: it depends — and that's the whole lesson.

A real cost, contested numbers

Training and running big AI models uses real electricity, and water for cooling — that part isn't in doubt. What is disputed is how much. Estimates swing wildly because they depend on which model, whether you count one-off training or everyday use, where the data centre sits, and who did the measuring. A single scary figure usually hides all of those choices.

How to read a contested claim

This is a skill you'll use far beyond energy. When you meet a striking figure, ask: what exactly does it measure, compared to what, and who's the source? A peer-reviewed range is worth more than a viral round number. The grown-up answer is often a range plus 'it depends' — less shareable than a headline, but a lot closer to true.

The takeaway

• —AI's energy and water use is real — but hard to pin to one number.
• —Estimates depend on the model, training vs use, location, and method.
• —Read the assumptions behind a claim; prefer a sourced range to a viral figure.