Electricity has replaced land and capital as the binding constraint on data center development. Where and how facilities get built now follows the power.
For decades, the data center map was drawn by land prices and capital availability. That map is being redrawn by a different variable.
The International Energy Agency reports that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5 percent of global consumption, and expects that figure to more than double by 2030, with AI as the primary driver. In the United States, Goldman Sachs Research has projected data center power demand rising from 31 gigawatts in 2025 to 66 gigawatts by 2027. A gigawatt is roughly the output of a large power plant, so that trajectory amounts to adding dozens of power plants of demand in only a few years.
Power is no longer a line item in the project. Power is the project. That single shift is changing how facilities are designed, where they are located and how they are financed, and each of those changes reaches every organization planning an AI deployment.
The Timeline Mismatch at the Center of Everything
The reason power now governs development is a mismatch of timelines. A data center can be built in 12 to 18 months. Connecting it to the grid can take five to seven years in constrained markets, according to Lawrence Berkeley National Laboratory and industry procurement data. The equipment that delivers that power carries queues of its own: large transformers average about two and a half years from order to delivery, and as long as five from major suppliers.
When the building is the fast part and the electricity is the slow part, everything else in the project reorganizes around the slow part. Access to firm, available power, rather than land or capital, is what decides whether a project can proceed at all.
Demand Is Rising From Inside the Rack
The pressure is not only coming from more facilities. It is coming from what each rack inside them now draws. AFCOM’s 2026 survey put average rack density at 27 kilowatts, up from 16 the year before and roughly 7 in 2021, the largest year-over-year jump in a decade of the survey’s data.
AI-specific deployments sit far above that average. Current GB200 and GB300 NVL72-class clusters draw 80 to 140 kilowatts per rack. NVIDIA’s Vera Rubin platform, entering volume production in the second half of 2026, is rated at roughly 190 to 230 kilowatts, and the Rubin Ultra rack announced for late 2027 is specified at about 600.
A single AI rack now draws more power than an entire row of legacy racks. That is why the unit that matters in planning conversations has shifted from square feet to kilowatts, and why published hardware roadmaps deserve a place in every facility discussion.
Design, Location and Financing Are All Moving at Once
Inside the facility, the order of design has inverted. Facilities are now planned around a power budget first, with the floor plan following from it. Concentrating 120 kilowatts or more in a single rack forces cooling from air to liquid, because air cannot carry away that much heat, and forces power distribution to be rebuilt to feed large loads into a small area. A hall engineered for 15 kilowatt racks cannot reach 130 by adding circuits; the change runs from the utility feed through the busway to the rack, and in practice it is a rebuild.
Outside the facility, site selection now follows power availability. Established markets with long interconnection queues, such as Northern Virginia, have become harder to build in, and some, such as Dublin, have paused new data center connections entirely. Development is moving toward regions where firm capacity can be secured on a workable timeline, and increasingly toward on-site and dedicated generation that avoids the grid queue entirely.
On the financial side, longer timelines and heavier electrical infrastructure change the shape of the investment. More capital is committed earlier and held longer before a facility produces revenue. That raises the cost of an ill-suited financing structure and increases the value of arrangements that preserve an organization’s capital during the long wait, matching the cost of a power-heavy build to the period it finally begins generating revenue.
How Should Organizations Plan Around Power?
For enterprises planning AI infrastructure, the practical response is to treat electricity as the first input rather than a downstream detail. That starts with confirming firm contracted power availability before selecting a site, not after the site is chosen. It means designing for the rack densities AI hardware is moving toward, not only the ones shipping today, since a facility built for current density can be obsolete before the next hardware generation arrives.
It also means treating interconnection and equipment lead times as part of the deployment timeline itself, rather than as separate procurement details discovered late. And it means structuring the financing to absorb long lead times, rather than assuming the fast build-and-revenue timelines of the past. A project that gets all four right moves when the opportunity appears. One that misses any of them waits, often for years.
What Power Now Decides
AI has changed what limits a data center. Land and capital still matter, but they now follow electricity rather than the other way around.
The organizations that treat power as the first question, confirming firm availability before committing to a site, are the ones that will build on a workable timeline. Those that assume the grid will follow are walking into the constraint already reshaping the industry: not enough power, not soon enough.
