For years the constraint on building a new data center was land, fiber, and capital. In 2026 the binding constraint has shifted to something more fundamental: electricity. The generative-AI boom has produced facilities so power-hungry that the ability to secure enough megawatts from the local grid, rather than the availability of real estate, now decides where and whether a campus gets built.
That shift has put technology companies and electric utilities on a collision course. AI workloads demand vast, steady, and rapidly growing amounts of power, while much of the American grid was engineered decades ago for a slower, more predictable pattern of demand. The result is a scramble to add generation, upgrade transmission, and rethink how large customers get connected, all while keeping the lights on for everyone else.
Racks that draw ten times the power they used to
The heart of the problem is how much electricity modern AI hardware consumes in a small footprint. Traditional server racks drew something like 5 to 15 kilowatts each. AI-optimized racks packed with high-end accelerators now demand anywhere from 30 kilowatts to well over 110 kilowatts, an order-of-magnitude jump that concentrates enormous power draw and heat into the same physical space. New AI-ready sites are being designed to consume 100 to 300 megawatts, and some hyperscale campuses are being planned at a full gigawatt, roughly the appetite of a small city. Industry analysts tracking the buildout describe a landscape in which, as one 2026 outlook put it, power has become the defining variable of the entire sector. Those densities also force a rethink of cooling, since air alone cannot remove that much heat and liquid cooling is becoming standard.
Why the grid was not built for this
The strain is not only about the size of any single facility but about how fast aggregate demand is climbing against infrastructure that changes slowly. Electricity demand from data centers is rising faster than the power grid, much of it built generations ago, was designed to absorb. By one widely cited estimate, global data-center electricity consumption could approach 1,050 terawatt-hours in 2026, a level that would rank the sector among the largest electricity consumers in the world if it were a country. A Brookings analysis of the AI energy landscape frames the challenge as a mismatch between the multi-year timelines required to build transmission lines and generation capacity and the far faster pace at which AI companies want to switch new compute online. Transmission projects can take the better part of a decade to permit and construct; a data center can be built in a fraction of that time, which leaves utilities perpetually chasing demand.
Grid availability replaces real estate as the bottleneck
The clearest sign of how much the calculus has changed is where developers now look first. Securing sufficient power from a local utility has become the critical-path item in siting a data center, displacing the traditional considerations of cheap land and network connectivity. Analysts at Gartner have projected that power shortages could restrict some 40 percent of AI data centers by 2027, a striking figure that captures how quickly demand is outrunning available supply in the most congested regions. Some utilities have accumulated interconnection queues stretching years into the future, and in a handful of markets the sheer concentration of proposed projects has forced regulators and grid operators to weigh how to add such large loads without destabilizing service or shifting costs onto ordinary ratepayers.
Tech companies start building their own power
Faced with grids that cannot keep pace, the largest technology firms have stopped waiting in line. Rather than simply signing contracts to buy electricity, companies are moving into direct ownership and development of generation assets, from long-term deals for nuclear output to investments in new plants and on-site generation. Reporting on the sector’s turn toward dedicated supply, including coverage of nuclear power deals tied to AI demand, describes a strategy in which firms lock up firm, around-the-clock power sources that can run a data center regardless of what the surrounding grid can spare. Nuclear is particularly attractive because it delivers steady baseload output without the intermittency of wind and solar, and because a single reactor can supply the enormous, constant draw an AI campus requires. Some operators have gone so far as to explore restarting shuttered plants or building small modular reactors adjacent to their facilities.
What the crunch means for ordinary customers
The power wall is not only an industry problem; it has consequences for households and businesses that share the same wires. When a utility must rapidly expand capacity to serve a cluster of data centers, the cost of new generation and transmission can flow through to rates unless regulators structure the deals to make the large customers bear their share. Rising overall demand can also tighten reserve margins, the cushion of spare capacity that keeps the grid stable during heat waves and cold snaps, raising the stakes of every future demand spike. Regulators, utilities, and AI companies are now negotiating the terms of that relationship, from special tariffs for very large loads to requirements that data centers bring their own generation or curtail usage when the grid is stressed.
A test of how fast the grid can adapt
The underlying tension is a mismatch of clocks. AI compute demand is doubling on a timescale measured in months, while the machinery of building power plants and high-voltage lines moves in years and sometimes decades. Whether the industry hits a hard ceiling or engineers its way through will depend on how quickly generation can be added, how creatively utilities and developers share the burden, and whether efficiency gains in chips and cooling can blunt the growth in demand. For now, the phrase that best captures the moment is the one utilities keep repeating: the challenge is no longer finding a place to build, but finding the power to run it.
This article was researched and written with the assistance of AI and reviewed by an editor prior to publication.
More from Morning Overview
