How Much Energy Does AI Use? The Real Numbers

Strip away the hype and the doom: here is what AI and the data centers behind it actually consume, from the two most credible public sources — the IEA and OpenAI itself.

One query: smaller than you think

Sam Altman put a number on it in June 2025: the average ChatGPT query uses about 0.34 watt-hours — about what an oven draws in one second, or an efficient lightbulb in a couple of minutes. A single question to an AI is a tiny amount of energy.

At scale: bigger than most industries

Billions of tiny queries — plus the far larger cost of training frontier models — add up. The IEA measured global data centre consumption at 415 TWh in 2024 (~1.5% of world electricity), growing ~12% a year since 2017. Its projection: ~945 TWh by 2030, roughly Japan's entire electricity consumption, with AI the biggest driver, and data centres accounting for nearly half of US demand growth through 2030.

Both things are true

The honest picture: individual AI use is trivially small, while the aggregate build-out is grid-shaping. That's why siting decisions — which grids, which states, which power sources — matter more than personal usage guilt. You can see exactly where the build-out lands on our facility map, and what it means for power infrastructure.

The training vs inference split

Training a frontier model is a one-off, enormous burst of compute over months; inference (answering queries) is a continuous trickle multiplied by billions of users. As AI adoption grows, inference becomes the dominant share — which is why efficiency per query, and the PUE of the facilities serving it, matter so much.

The water numbers →← Data center power