AI and Electricity: The Efficiency Paradox

The story

Efficiency needs a denominator

The International Energy Agency's Key Questions on Energy and AI, published on 16 April 2026, estimates that data centres used 485 terawatt-hours of electricity in 2025. Its central projection is approximately 950 terawatt-hours in 2030. These totals include non-AI services; the later figure is conditional, not an observed result.

The report also describes improving efficiency per AI task: completing a task can require less electricity than before. Total consumption, however, depends on how many tasks are performed and what they involve. A more efficient request does not establish that the industry as a whole uses less power.

Imagine a hypothetical service that halves the electricity used for each identical request, while the number of requests triples. Total consumption becomes 0.5 multiplied by 3, or 1.5 times the original amount. This is illustrative arithmetic, not a measurement of any product. It shows why an efficiency claim needs both a per-request denominator and an assumption about demand.

The counterfactual changes the verdict

Efficiency improvements could slow the growth in consumption without making consumption fall. To judge that benefit, we need a counterfactual: an estimate of what would have happened without the improvement. Comparing consumption with last year's total answers a different question from comparing it with a future in which no efficiency gains occurred.

The IEA's earlier analysis also distinguishes worldwide demand from concentrated local effects. A modest global share can require substantial network expansion where a facility operates. Electricity cannot reach a particular location merely because enough exists somewhere else; transmission and connection capacity matter.

Supply and responsibility

In a hypothetical town already serving households and shops, a new data centre might place a considerable burden on local infrastructure while barely changing global totals. Those descriptions concern different scales. Whether households should help finance the expansion requires evidence about costs and benefits, rather than a presumption that innovation either always benefits or always burdens them.

Electricity use is also different from emissions. The IEA's supply analysis considers multiple generation sources, including renewables. Similar consumption can have different consequences on different grids. A clean-power contract therefore needs scrutiny of when supply is available and how unmet demand is covered, rather than an assurance based on the contract's label.

Apply the same scrutiny to benefits

AI may help energy systems operate more efficiently, but the claimed saving must be attributable to the software. Researchers need to establish what changed because of its use, whether another method could have achieved the same result, and how much energy the software itself required. Introducing AI alongside a saving does not, by itself, demonstrate causation.

For example, software that reduces cooling demand in one building would not establish savings in every building or cancel growth elsewhere. In this article's interpretation, incomplete comparisons leave claims susceptible to exaggeration. Arguments are easier to assess when they specify the system, location, electricity supply and counterfactual. Separating observed results from projections makes those assumptions testable, while leaving room for disagreement about which benefits should take priority.