Data Centres Are Now an Electricity Story, and the Grid Is the Constraint
For twenty years the story of computing was that it got cheaper and used less power per task every year, so much so that global data-centre electricity consumption stayed roughly flat even as internet traffic multiplied. Efficiency absorbed growth. That relationship broke around 2022. The International Energy Agency's analysis of energy and artificial intelligence, published in 2025, estimated that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5 percent of global consumption, and projected that figure would reach roughly 945 terawatt-hours by 2030, more than doubling in six years.
The cause is the shape of the new workload. Training and running large AI models is dense, continuous and located in a small number of very large facilities. A single campus of the kind now being built in Texas, Virginia, Ireland or Malaysia can draw as much power as a mid-sized city, and unlike a city it wants that power every hour of the year.
That shifts the constraint from chips to the grid. A data centre can be built in eighteen months. A new high-voltage transmission line in the United States or Europe takes, on average, close to a decade from proposal to energisation, most of it in permitting. Large power transformers, the equipment that steps voltage down for delivery, have lead times measured in years because only a handful of factories make them. Utilities in the most sought-after regions have begun telling developers that interconnection is not available until the 2030s.
The response has been to go around the grid. Developers are signing contracts directly with power producers, financing new gas turbines, buying output from restarted nuclear plants and, in some cases, building generation on site. The economics work because compute is worth more per megawatt-hour than almost any other industrial use, so a data centre can pay a price for firm power that a factory or a household cannot.
That creates a policy problem regulators are only beginning to price. If the largest new electricity customers pay for their own generation and bypass the shared grid, the fixed costs of that grid fall on the remaining customers. If instead they connect to the grid and consume its spare capacity, every other new connection waits longer. Several jurisdictions, including Ireland and parts of the Netherlands, have already paused new data-centre connections in specific areas for exactly this reason.
Water is the second constraint. Cooling large facilities uses water either directly or through the power plants that supply them, and the regions with cheap land and cheap power are frequently the ones with the least water to spare. Operators have moved toward closed-loop and air cooling, at a cost in efficiency.
The efficiency lever has not disappeared. Chips get more useful work per watt each generation, and the industry's own figures show large gains per model per year. But for the first time in the history of computing, those gains are being outrun by demand, and the physical world is setting the pace. Anyone modelling the growth of artificial intelligence over the next five years is, whether they know it or not, modelling the construction schedules of utilities.
