A single AI data center can now draw as much power as a mid-sized American city

A hyperscale computing campus built to train and run artificial intelligence models can now pull as much continuous electricity as a small metropolitan area, a shift that has moved data centers from a rounding error in the U.S. power system to a central planning concern for utilities and grid operators. The jump is driven less by the number of buildings going up than by what sits inside them: dense racks of graphics processing units built for AI workloads, which draw far more power per square foot than the servers that ran email and web pages a decade ago. Understanding how a single campus reached city-scale demand explains why electricity, not chips, has become the binding constraint on how fast AI infrastructure can be built.

Inside the Power-Hungry Anatomy of a Data Center

A data center is not just a warehouse of computers. According to the International Energy Agency, servers account for around 60 percent of electricity demand in a modern facility, with cooling systems, networking equipment, storage, and backup power making up the rest. Cooling alone can range from about 7 percent of consumption in an efficient hyperscale facility to more than 30 percent in a less-efficient enterprise building, because every watt spent on computation generates heat that has to be removed to keep the chips from failing.

The rise of AI has changed that equation by packing far more processing power into the same physical footprint. Accelerated servers built around specialized chips run hotter and denser than the general-purpose machines that dominated data centers before the current wave of AI investment, which is why floor space is no longer the limiting factor for a new campus. The limiting factor is how many megawatts the local grid can deliver to the site.

A Per-Capita Load That Rivals a Small Nation

The IEA’s modeling shows just how concentrated that demand has become in the United States. American data centers already have the highest electricity intensity per person of any region in the world, at around 540 kilowatt-hours per capita in 2024, a figure the agency projects will more than double to over 1,200 kilowatt-hours per capita by the end of the decade. That per-capita load is roughly ten times higher than the global average, which is why a single large campus, rather than a whole industry sector spread across a country, can register on a regional grid operator’s books the same way a new city would.

Utilities that once treated a large industrial customer as an outlier now routinely field requests from data center developers seeking hundreds of megawatts of firm power at a single site, sometimes for delivery within just a few years, against a national grid that the Department of Energy describes as having just over 1 million megawatts of total generating capacity nationwide. That timeline mismatch, more than the raw size of the request, is what has turned data center power into a planning problem rather than a routine hookup.

From 415 to 945 Terawatt-Hours by 2030

Global data center electricity consumption stood at roughly 415 terawatt-hours in 2024, about 1.5 percent of world electricity use, and had already been climbing 12 percent a year over the prior five years before the current AI buildout accelerated the trend. The IEA’s base case now projects that figure will nearly double to around 945 terawatt-hours by 2030, a 15 percent annual growth rate that is more than four times faster than electricity demand growth in every other sector combined. Electricity consumption from the accelerated servers used specifically for AI workloads is projected to grow around 30 percent annually, more than three times the pace of conventional servers, and accounts for almost half of the entire net increase in global data center electricity use through the decade.

The IEA does not treat that trajectory as fixed. Its Lift-Off Case, which assumes faster AI adoption and a more resilient supply chain for chips and cooling equipment, projects global data center demand in 2035 running about 45 percent above the base case, topping 1,700 terawatt-hours and reaching 4.4 percent of global electricity demand. A separate High Efficiency Case, built on the assumption that software and hardware efficiency gains outpace the growth in AI workloads, holds 2035 demand closer to 970 terawatt-hours even as computing capacity keeps expanding. A third Headwinds Case, reflecting slower AI adoption and tighter supply chains, sees growth plateau around 700 terawatt-hours later in the decade. The spread between those scenarios illustrates how much of the AI power story still depends on choices being made today about chip design and data center cooling, not on demand that is already locked in.

Why the U.S. Numbers Tell an Even Sharper Story

U.S. government data show the same acceleration playing out inside the country’s own electricity mix. Data centers consumed about 4.4 percent of total U.S. electricity in 2023, and the U.S. Department of Energy now expects that share to climb to somewhere between 6.7 percent and 12 percent of total national electricity consumption by 2028, citing modeling from Lawrence Berkeley National Laboratory. That is a wide range, reflecting genuine uncertainty over how quickly new AI hardware ships and how efficient the next generation of chips turns out to be, but even the low end of the projection represents a substantial reallocation of the nation’s power supply toward a single industrial use case in under five years.

Concentrated in a Handful of Counties

Unlike the electrification of vehicles or home heating, which spreads new electricity demand across millions of individual customers and thousands of communities, data center growth concentrates in specific counties and even specific substations. The United States, China, and Europe are projected to account for nearly 80 percent of global growth in data center electricity consumption through 2030, with U.S. consumption alone rising by roughly 240 terawatt-hours over 2024 levels, a 130 percent increase. That geographic clustering is precisely why a handful of counties in Virginia, Texas, and the Southwest now see data centers dominate local electricity planning discussions the way a new steel mill or refinery once did, and why the physical wires connecting those campuses to the broader grid have become as important a story as the computing happening inside them.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview