How Much Electricity Does AI Use?

Every argument about AI infrastructure eventually lands on one question: how much electricity does it actually use? The most cited answer comes from the International Energy Agency's report Energy and AI. Here are its numbers, laid out plainly, with what they do and do not tell you.

Who is counting

The IEA is the Paris-based agency the OECD countries set up after the 1973 oil shock to track energy supply and demand. Its April 2025 report is the first time it has modelled AI's electricity use as a subject in its own right, and because it draws on utility data, chip shipments and data centre construction pipelines rather than press releases, it has become the reference point for everyone else. When a politician or a journalist quotes a global figure for data centre power, it is almost always this one.

How much data centres use now

About 415 terawatt-hours in 2024. That is roughly 1.5 per cent of all the electricity used in the world. For scale, Australia's National Electricity Market delivers a little over 200 terawatt-hours a year, so global data centres already draw about two of it.

Their use has grown by about 12 per cent a year since 2017, around four times faster than electricity consumption overall. Three regions dominate: the United States at about 45 per cent of data centre electricity, China about 25 per cent, Europe about 15 per cent. Everyone else, Australia included, shares the remaining 15 per cent.

How big one AI data centre is

The report's most repeatable line: a typical AI-focused data centre uses as much electricity as 100,000 households. The largest ones now under construction will use twenty times that, which puts a single building in the same bracket as a city of two million homes.

That is why the Australian proposals of the past year sound the way they do. A 240-megawatt site near Broome, an 800-megawatt site in South Australia's Mid North, a two-gigawatt site outside Darwin: each is a town's worth of demand arriving at one substation, in a region where the existing data centres run at five to ten megawatts.

Where it is heading

In the IEA's base case, data centre electricity use more than doubles to about 945 terawatt-hours by 2030. That is roughly what the whole of Japan uses today, and just under 3 per cent of global demand. Growth runs at about 15 per cent a year over the period.

By 2035 the base case reaches around 1,200 terawatt-hours. Across all the agency's scenarios the range is wide, from 700 to 1,700. The IEA is candid about why: nobody knows how fast AI will be taken up, how useful it will turn out to be, or how quickly the hardware gets more efficient. A thousand terawatt-hours of uncertainty is itself the finding. Somebody has to plan a grid around a guess.

Is it all AI?

Not all, but AI is the main driver. The IEA tracks what it calls accelerated servers, the GPU-based machines used for training and running models. Their electricity use is projected to grow about 30 per cent a year, against 9 per cent for conventional servers, and they account for nearly half of the net increase to 2030.

Cooling is part of the bill too. It ranges from about 7 per cent of a data centre's power in efficient hyperscale sites to more than 30 per cent in less efficient enterprise ones. That spread is the quiet reason the water and cooling rules now appearing in state policy matter: an efficient site and an inefficient one can differ by a quarter of their total load before a single query is answered.

Who feels it most

In advanced economies, data centres make up more than a fifth of electricity demand growth to 2030. In the United States, nearly half of all demand growth between 2024 and 2030 goes to data centres; per person, American data centre use was about 540 kilowatt-hours in 2024 and is projected to pass 1,200 by the end of the decade.

In emerging and developing economies outside China, data centres are about 5 per cent of demand growth. The build-out is a rich-country phenomenon, concentrated in a handful of regions, each of which is now discovering what it means to host it.

Where the power will come from

The IEA expects renewables to meet about half of the growth, around 450 terawatt-hours by 2035. Natural gas grows by about 175 terawatt-hours and nuclear adds a similar amount, with small modular reactors arriving around 2030. In practice the mix is set locally: the Darwin proposal is gas-fired by design, the Broome one is wind, solar and gas, and Victoria's new rules require operators to bring their own renewable generation and storage.

And emissions

Data centre emissions are about 180 million tonnes of CO2 now. In the base case they reach 300 million tonnes by 2035, or 500 million in the IEA's faster Lift-Off scenario. The agency notes that even the high case is still below 1.5 per cent of total energy-sector emissions, and that AI could, if applied to grids, buildings and industry, save more than it costs. That claim is the most contested line in the report, because it depends on things that have not happened yet.

So is it a lot or a little?

Both, depending on where you stand. As a slice of the world's power, it is small: under 2 per cent today, under 3 per cent in 2030. As a slice of new demand in the countries doing the building, it is large, and in the United States it is close to half. And for the town next door to a new site, it is the only number that matters, because a hundred thousand households' worth of load does not arrive as a percentage. It arrives as a substation, a pipeline and a water licence.

What Australia is building, and what it will cost in power, water and metal, is the subject of the companion piece: The Data Centre Decade.