AI's power bill: what the data-centre electricity numbers actually show

The scary version of the AI electricity story is easy to find. Data centres are going to eat the grid, blow past climate targets, and leave the rest of us in the dark. It makes a good headline. It is also not what the best measured numbers say.

Start with the size of the thing today. The International Energy Agency, in its 2025 Energy and AI report, put global data-centre electricity use at around 415 terawatt-hours in 2024. That is real power, roughly what a mid-sized industrial country draws, but it is about 1.5 percent of global electricity consumption. The entire category, every data centre on earth, cloud and AI and the dull ones that run payroll, is a small slice of the world's power.

The growth is real, the apocalypse is not

What makes the story feel urgent is the rate of change, not the level. The IEA found data-centre electricity demand rose about 17 percent in 2025, against roughly 3 percent growth for electricity overall, and AI-specific sites grew faster still. Its central projection has data-centre consumption roughly doubling, from about 485 terawatt-hours in 2025 to around 950 by 2030, reaching close to 3 percent of global electricity by the end of the decade.

Doubling is a lot. Three percent is not the end of the grid. Both of those are true at once, and holding them together is the whole job.

The United States is where it bites

National averages hide the problem, because data centres are not spread evenly. They cluster, on cheap land near fibre and, increasingly, wherever a utility will connect them. The United States is the clearest case. Lawrence Berkeley National Laboratory, in its December 2024 report for the US Department of Energy, found data centres used about 176 terawatt-hours in 2023, or 4.4 percent of US electricity, already more than double their share a few years earlier. Its projection is between 325 and 580 terawatt-hours by 2028, or 6.7 to 12 percent of the country's power.

Notice the range. The high end is nearly double the low end, four years out. That is not sloppiness, it is honesty about how much depends on things nobody can yet pin down: how fast the chips get more efficient, how many announced data centres actually get built, and whether the grid can connect them at all.

The bottleneck is wires, not watts

The constraint that actually decides this is unglamorous. It is the interconnection queue, the multi-year wait to plug a large new load into the grid. A data centre can be financed and designed in months. The transmission to power it takes years. That mismatch, not a global shortage of electricity, is what is slowing the buildout and pushing operators towards anything that connects faster: gas on site, deals with existing nuclear plants, and a wave of conditional agreements with small modular reactor projects that the IEA says grew from 25 gigawatts of planned capacity at the end of 2024 to 45 gigawatts a year later.

Most of those reactors do not exist yet. The agreements are bets, not megawatts.

Efficiency is the quiet variable

The other reason to distrust a straight line is that the cost of the underlying work keeps falling. Power drawn per AI task is dropping quickly, as chips, cooling and model design all improve. The catch, and the IEA is blunt about this, is that cheaper tasks invite more tasks. Agents that run for minutes instead of answering in seconds use more energy, and there are more of them every month. Efficiency pushes the per-unit number down while total demand climbs. Which force wins is genuinely unsettled.

What to actually watch

If you want to track whether AI's power demand is a manageable strain or a real crisis, the headline terawatt-hour forecast is the least useful number. Watch three things instead. First, interconnection queue times in the hotspots, because that is where a shortage shows up first. Second, how much announced capacity gets cancelled or delayed, since announcements are free and concrete is not. Third, efficiency per task, which decides whether doubling the workload doubles the power or something much smaller.

The honest summary is the boring one. AI is a large and fast-growing electricity consumer, concentrated enough to strain specific grids, and nowhere near the civilisation-level draw the scariest numbers imply. The single figure you should distrust most is the one quoted without a year, a region and an assumption attached.