The number that gets quoted most about AI’s energy use is the one attached to a single prompt. How many watt-hours does one query cost, how many bottles of water, how many times a Google search. It is a tidy frame, but perhaps it points at the wrong scale. The figure that actually matters is the one the International Energy Agency: the world’s data centres consumed around 415 terawatt-hours of electricity in 2024, roughly 1.5 per cent of global electricity consumption.
The scale most people picture is wrong
The per-query argument tends to break in both directions at once. On one side, the often-repeated claim that an AI prompt costs ten times a web search is contested. Epoch AI estimates a typical ChatGPT query running on GPT-4o uses roughly 0.3 watt-hours, about ten times less than an earlier and widely cited 3-watt-hour figure, a gap they attribute to more efficient models and more realistic assumptions about how long a typical answer runs.
Individual queries were never where the weight sat. The IEA’s framing is about aggregate demand, and that demand has been climbing for years. Data centre electricity use has grown by about 12 per cent a year since 2017, more than four times faster than total electricity consumption. The story is not one prompt — it is the building full of servers the prompt lands in, and the thousands of buildings like it.
AI is the engine, not the whole machine
The IEA names AI as the single most significant driver of growth in data centre demand, but AI is not most of the current load. Scientific American’s coverage of the report notes the agency’s estimate that AI-specific servers accounted for about 15 per cent of total data centre energy demand in 2024, with the rest going to conventional cloud computing, streaming and the ordinary machinery of digital life.
That share is contested at the edges. Alex de Vries, a researcher at VU Amsterdam and founder of Digiconomist, argues the report “is a bit vague when it comes to AI specifically”. But the wider point holds either way. As de Vries puts it, “regardless of the exact number, we’re talking several percentage of our global electricity consumption.”
What the IEA’s projection actually says
The forward-looking number carries the comparison to Japan, and it is a projection rather than a measurement. The IEA’s Base Case projects data centre electricity use to more than double to around 945 TWh by 2030, just under 3 per cent of global electricity. IEA Executive Director Fatih Birol framed it plainly: global data centre demand is set to more than double over the next five years, “consuming as much electricity by 2030 as the whole of Japan does today.”
The growth is also concentrated. China and the United States together account for nearly 80 per cent of the projected global increase to 2030. Birol noted the load falls hardest on a few national grids, with data centres on course to account “for almost half of the growth in electricity demand; in Japan, more than half; and in Malaysia, as much as one-fifth.”
The uncertainty the report builds in is worth holding. Birol’s Japan line is a forecast, not a settled fact, and the agency flags that AI adoption rates, efficiency gains and grid bottlenecks could all shift the trajectory. The IEA estimates that around 20 per cent of planned data centre capacity risks delays connecting to the grid. Its scenario range for 2035 runs from 700 to 1,700 TWh, which suggests the agency is confident about the direction and far less confident about the destination.
The open question the numbers leave behind
The tension running underneath all of this is an old one. Per-query and per-task energy is falling fast. The Brookings Institution reports that Google cut the median energy of a Gemini text prompt by a factor of 33 between May 2024 and May 2025, while its overall data centre electricity use rose 27 per cent in 2024 over the previous year. Efficiency per task improves; absolute consumption climbs anyway.
What that demand gets powered by remains unsettled. The IEA’s Base Case has most new capacity served by renewables and natural gas, but de Vries is more pessimistic, predicting that “we’re going to increase our reliance, or at least extend, our reliance on fossil fuels.” That is his view, and it sits against the report’s own finding that a large share of planned capacity is set to come from low-emissions sources.
The figure to watch is whether the efficiency curve the industry promises can outrun the demand curve the IEA models, and which fuel fills the gap while the two are still diverging.
