The AI boom has a power bill. And it's growing faster than almost anyone predicted.
183 TWh — more than all of Pakistan
US data centres consumed 183 terawatt-hours in 2024 — over 4% of the country's total generation. In 2018 it was 1.5%.
The IEA projects this hits 426 TWh by 2030. That's a 133% increase in six years.
US data centres already consume more electricity than the entire nation of Pakistan.
Three forces stacking demand
AI training runs are the flashy driver. One frontier model training run consumes as much electricity as a small town uses in a year.
AI inference is the bigger long-term problem. Every query, every generated image, every autocomplete — all of it requires compute. Usage is growing exponentially.
The hyperscaler buildout is the capital layer underneath. Meta, Amazon, Alphabet, and Microsoft committed over $320 billion in AI infrastructure spending in 2025 alone.
$320 billion in AI infrastructure — not research budgets, but industrial-scale power-hungry construction.
Virginia: 26% of one state's entire grid
New data centres go up in 18–24 months. New power generation takes 3–5 years. That gap is a crisis waiting to happen.
In Virginia — the world's largest data centre market — data centres eat 26% of the state's electricity. Utilities say capacity is being consumed faster than generation can be built. Texas, Ireland, and the Netherlands face the same wall.
Big tech is buying nuclear plants
Hyperscalers aren't waiting for the grid to catch up:
- Microsoft restarted a reactor at Three Mile Island
- Amazon invested in small modular reactor technology
- Google is buying geothermal power at scale
The message is unambiguous: the existing grid isn't enough.
Microsoft literally restarted Three Mile Island to power its AI. That's not a metaphor — it's a grid strategy.
Jevons paradox hits compute
Individual chips are getting more efficient every generation. But total power consumption keeps climbing.
Because cheaper AI gets used for more things. Efficiency doesn't reduce demand — it expands it.
6–8% of US electricity by 2030
If buildout trajectories hold, data centres could consume 6–8% of US electricity by 2030. That makes computing infrastructure one of the largest single categories of electricity demand in the country — reshaping energy policy and carbon calculations for decades.