If AI gets more efficient, why does electricity demand grow?

Efficiency and greater consumption can coexist

AI completes a task with less electricity than before. Yet data centres use more electricity in total. These statements can seem contradictory until we ask whether the number describes one task or every task added together.

In its report published on 16 April 2026, the International Energy Agency estimates global data-centre electricity consumption at 485 TWh for 2025 and projects roughly 950 TWh in 2030. The latter is a central projection, not an observed result. The figures cover data centres overall, not just facilities devoted to AI.

This background explainer checked that report on 6 November 2026. Rather than treating a future total as a certainty, start with multiplication you can reproduce on a small scale. Then return to the headline and ask what changed.

Halve the energy per task, then triple the tasks

Imagine a fictional machine doing 100 identical tasks. Each task uses one unit of electricity, so the total is 100. An improvement cuts the energy per task to 0.5 units. For the same 100 tasks, the total falls to 50.

Now use the improved machine 300 times. The calculation is 0.5 × 300 = 150. Energy per task has halved, while total consumption has risen by half. The efficiency gain still matters: without it, those 300 tasks would have required 300 units.

Our model deliberately holds the task constant. Real requests can change from short text to long reasoning or video. There is no single energy cost that applies to every possible AI request. The IEA distinguishes improvements in efficiency, growth in use and changes in the kinds of tasks people ask models to perform. The controls above isolate only the first two ideas; they do not forecast an actual service.

Watts describe a rate; watt-hours describe an amount

A watt measures power: the rate at which energy is used or transferred. A watt-hour measures energy accumulated over time. A fictional device drawing a constant 100 W for two hours uses 200 Wh, or 0.2 kWh.

Think of the distinction between litres flowing from a tap each minute and litres collected in a bucket. This analogy illustrates a rate and an accumulated amount; it does not claim that electrical grids behave exactly like plumbing.

A proposed facility rated in megawatts and an annual total reported in terawatt-hours are not interchangeable measurements. Converting one to the other requires assumptions about operation over time. Beside the headline number, write whether it describes a maximum, an average or a yearly total. That small annotation often reveals what comparison is actually possible.

A small global share can be large in one place

The IEA’s central projection puts data centres at about 3% of global electricity demand in 2030. That worldwide share cannot describe the circumstances of every town. The report also examines the geographical clustering of data centres.

Consider three imaginary districts. Their combined electricity use is 1,000 units, including 100 units for data centres: a 10% share. Now place 50 of those data-centre units in a district whose total consumption is 100. The local share is 50%, while the combined share remains 10%. Both statements are true.

These are invented numbers for a calculation, not measurements from a real community. They show why changing the boundary changes the percentage. A global share and the practical difficulty of connecting a large new user at one location need separate investigation. Neither number cancels the other.

What is inside the box labelled “AI electricity”?

Before comparing two studies, compare their boundaries. A measurement covering a computing device alone and one covering an entire facility, including support systems, can differ without revealing which device is more efficient.

Training a model and using a trained model are also different activities. Which activity was counted? Were the outputs comparable? Under what operating conditions did the equipment run? If a source does not give those conditions, an honest comparison leaves the gap visible.

Calling the earlier 485 TWh figure “electricity for AI chat” would change the scope of the source. Dividing a broad global total into individual questions would require additional information. That is why our small model is not an electricity-bill calculator or a meter for your latest prompt. It isolates a relationship between two quantities, not the consumption of a product.

Our view: judge benefits and burdens at comparable scales

Our editorial view is that efficiency improvements and preparation for greater demand deserve attention together. A gain per task does not settle a facility’s total demand or a community’s experience. A rising total does not erase the benefit of needing less energy for the same task.

When the next report appears, we would look beyond the largest proposed number: what work is included, when equipment will operate, where it will connect and why the outlook has changed. These are questions for evaluating reporting, not predictions about a company’s returns or a particular household’s electricity bill.

Try placing three labels beside a striking headline: per task, total and place. A fourth label, observed or projected, keeps the timeline straight. The goal is not to turn every reader into an energy forecaster. It is to make the uncertainty specific enough that the next piece of evidence has somewhere to fit.

An editorial perspective This section offers comparisons and questions based on facts in the sources. It does not claim a shared underlying principle or present new research results.