Two pieces of paper sit side by side. The first is a job receipt from a computing service: one standard task, run this month, used half the energy the same task used a year ago. The second is the month's meter reading for the same service, and it is half again as large as it was a year ago. Both are accurate, and both are invented for this dispatch. The pair looks like a contradiction, the sort of thing that sends someone back to check the numbers. There is nothing wrong with the numbers. They are measuring different things.
Last Thursday, April 16, the International Energy Agency published a report, Key Questions on Energy and AI, that puts the same two facts on a single page, at the scale of the world. Measured per task, it says, the energy efficiency of AI is improving at a rate unprecedented in energy history, with energy use per AI task falling by at least an order of magnitude a year in recent years. A simple text query now typically uses less electricity than running a television for the same length of time. In the same summary, the agency estimates that global electricity demand from data centres grew 17 percent in 2025, and that demand from data centres built specifically for AI grew 50 percent.1
My argument is that these are not in tension, and that the useful lesson is about which number to watch. Energy per comparable task describes how efficiently one unit of work is done. Total electricity demand describes how much electricity all the work together requires, and it depends just as much on how many tasks are run and what kind of tasks they are. A unit can get cheaper while the total grows. This is not a claim about how much of the growth efficiency itself caused. The report does say that cheaper, more capable tools are fuelling wider adoption, but it attributes AI's energy demand to several trends at once, and nothing in it separates their shares. And it is not an argument against efficiency, which leaves demand lower than it would otherwise be. It is an argument for measuring the total directly, rather than inferring it from the unit.
Job receipt · standard task
Illustrative · abstract energy units
Meter sheet · monthly total
Same service · same task mix
Each task got cheaper.
The month got more expensive. Both are true.
The arithmetic before the paradox
Before calling anything a paradox, it helps to write down the multiplication. Total energy is the number of tasks multiplied by the energy each task uses, added up across every kind of task. In Figure 1 there is only one kind. Last year the service ran 100 tasks at 2 units each, for 200 units. This year it ran 300 tasks at 1 unit each, for 300. The unit halved; the volume tripled; the total rose by half.
That is all a "paradox" of this kind requires. It does not require that the efficiency caused the extra volume, although sometimes it does: a cheaper service invites more use. Volume might have grown for reasons that have nothing to do with price, such as new customers or a new feature. The example shows only that the two numbers can move in opposite directions at the same time. Which forces actually drove the growth is a separate question, and the answer differs from one service to the next.
The IEA report gives a sense of how fast the volume side is moving. There are no comprehensive statistics on how often or how intensively people use AI, it notes, but major model providers reported a threefold increase in active users and a fivefold increase in revenue over the past year.1 Efficiency gains of an order of magnitude a year are enormous. So is that kind of growth in use.
What counts as one task?
The example assumed every task was the same. Real workloads are not, and this is where the arithmetic gets more interesting than a simple race between efficiency and volume.
The report is specific about the range. A short text query can use a fraction of a watt-hour on modern systems. Generating a high-resolution image may take ten times more energy per request. Video generation can take hundreds to thousands of times more, depending on length and resolution, and reasoning models that produce long chains of output, agents that carry out multi-step tasks, and queries that come with large documents attached all use materially more energy per request than a brief text reply.1 The agency's summary describes AI's energy demand as the result of three fast-moving and uncertain trends: improving efficiency, surging uptake, and new capabilities that unlock new and often far more energy-intensive uses.1
This creates a trap in the word "task." If a company reports that its energy per task has fallen, the claim is only meaningful if the tasks are comparable. A short classification and a long autonomous research run are both "a task" on a usage report, and neither is a sensible unit for the other. The same goes for counting tokens, the chunks of text that models read and write: a token in a quick reply and a token in a long chain of reasoning are not the same amount of work for the same purpose. Averaging across them hides the thing most likely to drive the total, which is the mix.
Mix can push the total up even when every individual kind of work is getting cheaper. Figure 2 adds a second kind of task to the earlier example: a long run, assumed for illustration to use a hundred times the energy of a short task this year. Three of them a month double the total, even though nothing on the worksheet became less efficient.
| Month | Short tasks | Energy per short task | Long runs | Energy per long run | Total | Average per task |
|---|---|---|---|---|---|---|
| Last year | 100 | 2.0 | 0 | none | 200 | 2.0 |
| This year, same mix | 300 | 1.0 | 0 | none | 300 | 1.0 |
| This year, new mix | 300 | 1.0 | 3 | 100 | 600 | about 2.0 |
Notice the last column of the new-mix month. With three long runs added, the average energy per task is back to about where it was a year ago, even though the short tasks are twice as efficient and the long runs did not exist before. An average across unlike tasks can rise, fall or stand still for reasons that have nothing to do with efficiency. That is why a single "energy per task" figure, quoted without its mix, can mislead in either direction.
What a buyer can actually measure
Most organizations buying AI services do not run data centres, and they cannot see their provider's meter. But the same arithmetic applies to the things they can see: their invoices, their usage reports, and the outcomes they are buying.
The first useful habit is to pick a unit that means something to the business. Not a query or a token, but an accepted outcome: a support case resolved, a contract reviewed, a report a person actually used. Cost or energy per accepted outcome folds in the retries, the long runs and the work nobody looked at, which are exactly the things a per-query figure leaves out. The second is to watch the total workload directly, by task class, so that a shift in mix shows up as itself instead of hiding inside an average. The third is to plan for growth in volume when the unit gets cheaper, because cheaper services tend to be used more, and budgets built on last year's volume at this year's unit price will be wrong.
Location matters too, in a way a global figure cannot capture. The IEA's 485 terawatt-hours for data centres in 2025 is a world total.1 What matters to a grid operator is where the load lands and when; the report describes grid connections as a bottleneck, with developers in the United States turning to onsite gas generation because connections are slow.1 A national total does not tell you one facility's effect on its local grid, and it certainly does not tell you one customer's invoice. Electricity use, electricity prices and carbon emissions are also three different outcomes. A large load on a clean, uncongested grid and the same load on a strained, fossil-heavy one are not the same event.
The report also looks ahead. Its central projection has data centre consumption roughly doubling to about 950 terawatt-hours by 2030, around 3 percent of global electricity demand, with consumption by AI-focused centres tripling.1 That is a projection, built on assumptions about investment, bottlenecks and use; the agency itself notes that bottlenecks make the most aggressive near-term scenarios less likely and that it will keep updating the figures. It is a reason to plan for growth, not a fact to plan around.
Objections, considered
Efficiency still saves energy. It does, and this is the most important qualification. Compared with an otherwise identical world in which the same tasks were done less efficiently, efficiency reduces consumption, often dramatically. The IEA's illustration makes the point: if every conventional web search were replaced with a simple AI text query, the agency estimates it would use less than 4 terawatt-hours a year, under 1 percent of current data centre consumption.1 Nothing here argues for less efficient systems. The argument is only that efficiency describes the unit, and the total has other inputs.
Higher demand may reflect valuable new services. Very likely, in many cases. A video model or an agent that does a day's work may be worth far more energy than a quick search. Rising total demand is not by itself a sign of waste, any more than a growing factory's electricity bill is. The question for a buyer is whether the outcomes justify the workload, which requires measuring both.
Some workloads can move. The report describes operational flexibility as a real lever: data centres can shift non-urgent work to other times, move computing between locations, and use onsite storage to manage peak demand, and it sees battery storage in data centres growing enough to make them a possible grid asset if the incentives are right.1 Other work, like a customer waiting for an answer, cannot wait. Flexibility is real but uneven, and an ordinary buyer has limited influence over a provider's power arrangements beyond asking about them.
Isn't this just the Jevons paradox? The report itself uses the name: by lowering costs and making AI tools more capable and accessible, it says, efficiency gains are also fuelling broader adoption, a dynamic known as the Jevons paradox.1 That mechanism is real, and it is part of the story. But the label can explain less than it seems to. The growth in AI's electricity demand also reflects new users, new kinds of tasks and new capabilities that would have arrived at any price, and the evidence does not let anyone say how much of the increase efficiency itself caused. The mechanism in the worksheet is simpler and more useful than the name: volume and mix multiply the unit.
The unit can get smaller while the total gets larger.
Before the next AI budget
- Which unit is being compared? Ask what one "task" means in each efficiency claim, and whether last year's task and this year's are the same kind of work.
- How are volume and mix changing? Track usage by task class, including retries and long runs, so that growth in the expensive kinds shows up as itself.
- What does an accepted outcome cost? Measure cost or energy per result the business actually used, not per query sent.
- Which constraint is local? Ask providers where and when the work runs, what can be shifted, and what they report about energy and emissions for it.
Back to the two pieces of paper. Neither should be thrown away, and neither should be read alone. The receipt shows the engineering working: each task really does cost half what it did. The meter sheet shows what the organization actually pays for: all the tasks, of every kind, added up. Keep them clipped together, and when someone quotes the first, ask to see the second.