The rebound

Per-query energy has fallen dramatically and total energy use has risen. Both are true, and the relationship between them is the most under-discussed thing in this whole subject.

Cheaper units, more units

When something gets cheaper per unit, people use more of it. Sometimes a little more, sometimes so much more that total consumption rises despite the efficiency gain.

This is not cynicism about efficiency — the efficiency is real and worth having. It is a statement that per-unit improvement and total consumption are separate variables, and that a claim about one is not a claim about the other. Anybody arguing from per-query efficiency to total impact has skipped the step that decides the answer.

Estimate your own rebound

Take your measured per-task energy. Imagine it drops tenfold tomorrow.

Honestly: how many tasks would you run? Not how many you should — how many you would. Most people find their usage is currently limited by cost or latency rather than by need, and that when the limit moves, usage moves with it.

Multiply your new per-task figure by your new task count. Compare to today's total.

Total energy after a tenfold efficiency gain

Your estimated new total as a percentage of today's total. Under 100 means the efficiency won; over 100 means your own usage growth ate it.

The efficiency press release

A vendor announces a model that is 90% more energy efficient per query, and concludes that this substantially reduces the environmental impact of AI.

What you can now check

You have three numbers about a subject most people argue with none.