AI infrastructure is the dirty secret of the tech world, and it’s time we ripped off the bandaid. Everyone’s busy drooling over the latest chatbot that mimics human conversation; nobody wants to talk about the data centers making it possible. They burned about 415 terawatt-hours of electricity in 2024 — roughly 1.5% of everything the world generated — and the IEA expects that to more than double by 2030.

Here’s the kicker: data center demand is expected to account for almost half the growth in global electricity demand through 2030. We keep calling AI a software revolution while quietly rebuilding the grid.

And the training figures are worse than marketing admits. Training GPT-3 took 1,287 megawatt-hours and emitted about 552 tonnes of CO2 equivalent, a bit over three times a passenger jet round trip between San Francisco and New York. That “five cars over their lifetimes” line you’ve seen everywhere? Same 552 tonnes, recycled onto whichever model is trending.

The problem isn’t that we train big models. It’s that nobody meters the running cost, then everyone acts surprised when the power bill arrives.

Next time you’re wowed by a shiny demo, remember: there’s a warehouse behind it, and it’s growing.