| ISSUE 07/30 · COMPUTER SCIENCE |
~3 MIN |
NOW / ENERGY + AI · WEEK 02 — MACHINES UNDER PRESSURE
Why efficient AI can still use more energy
START HERE
Every AI answer is produced on physical computer chips that use electricity. Large groups of those computers live in buildings called data centres. Efficiency means using less energy for one task. A rebound effect happens when lower cost encourages extra use that erases some or all of the expected saving. This is why a more efficient AI system can use less electricity per answer while the industry around it consumes more overall.
|
/ WHY NOW
The International Energy Agency says data-centre electricity demand rose 17% in 2025, while power used by AI-focused centres grew faster still. In the same report, it says electricity per AI task is falling rapidly. Both statements can be true. Making each task cheaper can encourage extra use that offsets the saving—a rebound effect. If total energy rises above where it began, the stronger result is called backfire or Jevons' paradox.
|
/ THE IDEA
A computer uses electricity to perform logical operations, move stored information and release waste heat. Faster chips and better software can reduce the energy needed for one answer. But total consumption equals the cost of one task multiplied by the number of tasks. If efficiency halves the energy per answer while usage quadruples, the data centre uses twice as much energy overall. Local improvement does not automatically create system-wide reduction.
THE FORMAL IDEA
total energy = energy per task × number of tasks
| energy per task = computation, memory and cooling attributable to one job | | number of tasks = how much the service is used | | both terms can change at the same time |
|
RUN THE TINY EXAMPLE
Cheaper answers, bigger total
Before: 1 unit per task × 1 million tasks = 1 million units After: 0.5 units per task × 4 million tasks = 2 million units Per-task efficiency: 2× better; total energy: 2× larger
|
The arithmetic is why an efficiency announcement cannot reveal grid impact alone. You also need total demand, where electricity is generated, and whether use lands during a clean or strained part of the day.
/ SO WHAT?
This lets you interrogate AI-energy headlines. Ask whether a number is per chip, per query, per data centre or for the whole sector. Then ask how usage changed. A smaller first number can be overwhelmed by a larger second one.
ONE CAVEAT |
| There is no single fixed energy cost for an AI request. Model size, hardware, prompt length, batching, cooling, electricity mix and what would have happened otherwise all affect the answer. |
KEEP THIS
Efficiency changes the energy per task; scale decides whether total energy falls or rises.
|
NEXT: What fusion gain actually measures
|