The machines got more efficient, and the grid buckled

The contradiction on the grid

Every year, an AI chip does more computing for each watt it draws. By the old logic, that should ease the strain on the grid. Instead, the strain is worse than ever. US power emissions rose last year for the first time in a decade, and old coal plants that were meant to close are staying open to feed the data centres. The machines got leaner, and the appetite grew. Both are true at once, and that is the puzzle worth understanding.

The old economist’s warning

In 1865, William Stanley Jevons noticed the same thing with coal. Better steam engines used less coal for the same work, and people assumed Britain would burn less. It burned far more. Cheaper steam power made new uses worth it - in factories, ships, railways - and demand raced past the saving. This is the Jevons paradox, or the rebound effect. Make something more efficient and you make it cheaper to use. Cheaper things get used more. Often, so much more that total use goes up, not down.

The mechanism, plainly

Efficiency lowers the cost of each unit of a thing - each computation, each mile, each hour of light. A lower cost does two jobs. It shrinks the bill for the uses you already had. And it makes uses that were not worth the money suddenly worth it. That second job is the one people forget. When AI compute gets cheap, companies do not just run their old models for less. They run far more models, on far more tasks, because now they can afford to. The saving per use is real. It just gets spent on more uses.

Where you have already seen it

This is not only an AI story. Cars got much more fuel-efficient, and we drove more miles in bigger vehicles. LED bulbs use a fraction of the power of old ones, and we lit up more signs, more screens, more hours of the night. Cloud storage got so cheap we simply stopped deleting anything. In each case the engineering worked exactly as promised. The trap was assuming the amount we wanted would hold still while the price fell. It never does.

What this is not

It helps to be precise. This is not pollution being shoved across a border, where a country cleans its own ledger by importing dirty goods. And it is not a threatened ban that makes owners drill faster before the door shuts. Those move a fixed amount of harm around in space or time. The rebound effect is different: it creates new demand that did not exist before, because the thing got cheap enough to want more of.

Who is inside this

The person typing a query is one node in it. So is the family whose electricity bill carries a surcharge to build power for a data centre down the road. So is the coal plant that was going to retire and now runs another year. None of them chose the appetite. It is the sum of a billion small “why not, it’s cheap now” decisions, each sensible on its own. Efficiency was sold as the climate answer, and it is a real part of one. But on its own it changes the price, not the appetite - and appetite is the thing that sets total use.

The part to hold loosely

Efficiency still matters. Without it, the same demand would burn even more. The rebound rarely eats the whole saving, and where use is already near its ceiling, a leaner machine really does cut the total. The honest version is narrower than the slogan. Making things more efficient is necessary. It is only enough when something else holds the total in check - a price, a cap, a limit on the build. Left alone, a cheaper machine invites a bigger world to use it, and the grid feels the difference.