If AI gets more efficient, why can electricity demand still rise?

The current AI energy puzzle

In its April 2026 update on AI and energy, the International Energy Agency reports falling power consumption per AI task alongside rising data-centre electricity demand. Its analysis also discusses physical constraints such as grid connections and equipment supply. This is the real-world reason to distinguish a better unit cost from a smaller total footprint.

Data centres support more than AI, and different AI tasks have different resource requirements. Training, serving a short answer, generating video, and running a long agent workflow cannot be treated as one identical query. Be careful with any claim that gives one energy number for 'using AI' without describing the task and measurement boundary.

Do the two-number calculation first

In a simplified model, total electricity equals electricity per task multiplied by task count. If consumption per task falls by half, task count can double before total consumption returns to its starting level. More than doubling increases the total; less than doubling decreases it. There is no contradiction in becoming more efficient while using more electricity overall.

This identity does not explain why demand changes. It only tells you what must be measured. You also need comparable tasks, a consistent time period, and the same system boundary. Adding cooling in one estimate but omitting it in another makes a tidy comparison misleading. Write down what the numerator includes before calculating a percentage improvement.

Where Jevons paradox enters

The Jevons paradox experiment illustrates how lower resource costs can encourage more use. Move the efficiency and demand-response controls separately. An efficiency improvement need not trigger enough extra activity to outweigh the savings; the result depends on the demand response you assume.

Rebound describes savings being partly offset by greater use. A rebound large enough to increase total consumption is sometimes called backfire. Calling every efficiency improvement 'Jevons paradox' skips the important empirical question: how much does activity actually change? A tool may become more attractive because it is cheaper, faster, more capable, or easier to embed in other products. Those mechanisms should be investigated rather than assumed.

A worked example: an internal assistant

Imagine an assistant using an arbitrary one unit of electricity per task and handling 1,000 tasks each month. A redesign cuts the unit requirement to 0.4. These are illustrative units, not measured watt-hours for a product. The redesign looks different if it changes how the organisation uses the assistant.

1. Hold activity fixed: 0.4 × 1,000 gives 400 units, compared with the original 1,000.
2. Add a new workflow: at 3,000 comparable tasks the total becomes 1,200 units, despite the better unit efficiency.
3. Evaluate the additional work: ask whether those tasks replace another process, create useful new output, or simply generate more material to inspect.

Demand is not the same as unlimited capacity

A demand curve cannot manufacture transformers, water, land, or a grid connection. Use stocks and flows to separate accumulated capacity from the rate at which capacity is added. Rapid demand growth can run into a slowly changing stock of infrastructure.

The Little's law experiment offers another lens: when arrivals exceed what a process can clear sustainably, waiting and work in progress become important. It does not model an electricity grid. It helps you ask where expansion queues form and whether extra capacity at one stage merely moves the bottleneck to another.

A better checklist for an efficiency claim

When comparing tools or reading a headline, ask for absolute consumption as well as consumption per task. Then ask whether the tasks are equivalent in quality and complexity. A cheaper answer that requires repeated retries may not be cheaper per useful result. Keep operational energy separate from embodied impacts unless the analysis explicitly includes both.

For a team deploying AI, a practical experiment is to track completed useful tasks, retries, and the total workload over a consistent period. A usage budget or a rule against unnecessary always-on processing can make the resource constraint visible. These are measurement choices; they are not a universal prescription for how every organisation should use AI.

• What is included in the energy estimate?
• Is the output comparable in quality?
• Did task count or task complexity change?
• Are actual totals reported alongside projections?

Efficiency is useful; total impact still needs evidence

The IEA's Energy and AI analysis discusses demand scenarios rather than a single inevitable future. Policy, infrastructure, adoption, and technical progress all affect the result. Electricity demand also does not translate into emissions through one universal multiplier: the generation mix and timing matter.

The useful takeaway is to keep two questions on the page at once. How much resource does one useful task require? How many useful tasks are we choosing to perform? You can celebrate a real improvement in the first while still checking whether the second overwhelms it. That habit applies to transport, computing, heating, and many other efficiency debates.