The next generation of artificial intelligence companies may not be built where software firms traditionally gathered. Talent, capital, and fast internet still matter, but AI has added a physical requirement that cannot be moved through the cloud: enormous quantities of dependable electricity.
Training models and serving users require facilities packed with specialized chips, cooling equipment, networks, and backup systems. As computing campuses move toward gigawatt-scale demand, electricity can determine whether a project is affordable—or whether it can be built at all.
AI Growth Is Becoming an Electricity Planning Problem
The International Energy Agency’s 2026 assessment projects global data-center electricity consumption rising from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours in 2030, while electricity use at AI-focused facilities is expected to triple.
That growth changes the meaning of location. A conventional software startup can lease an office and expand its cloud account. An AI infrastructure company may need a dedicated substation, transmission capacity, and long-term agreements with generators or utilities.
Lawrence Berkeley National Laboratory’s 2025 update estimates that data centers could consume 11.8% of US electricity by 2030, with scenarios ranging from 9.5% to 15.3%. Its reference case reaches 649 terawatt-hours, while the modeled range extends from 521 to 843 terawatt-hours depending on chip shipments, server utilization, and other assumptions.
Not every AI company must own a data center. However, cloud providers, model developers, and businesses purchasing large amounts of computing capacity will increasingly inherit the economics of the electricity systems beneath them.
Cheap Electricity Matters, but Speed Matters More
An inexpensive tariff is valuable only when enough power can reach the site. A region with low average prices may remain unattractive if the network is congested or the utility cannot energize the project for several years.
CBRE Research identifies the cost, source, and available supply of electricity as leading considerations in data-center site selection. Its analysis notes that Atlanta, Northern Virginia, Dallas, Phoenix, and Chicago—major construction markets—have historically ranked among the less expensive US electricity markets.
The more decisive measure may be “time to power.” CBRE reports that facilities able to use existing grid capacity can have construction timelines about half as long as projects requiring new transmission or interconnection work.
A Nature Communications study describes the same mismatch from the power-system side: large data-center sites can become physically ready in one to two years, while substantial substations or transmission upgrades may require five to eight years. A slightly more expensive region can therefore be commercially superior when it offers a faster, more certain connection.
AI Campuses Can Change the Price They Came to Find
Low-cost electricity can attract data centers, but concentrated demand can eventually place upward pressure on the same market.
A 2026 Federal Reserve Bank of Dallas working paper estimates that existing data centers have raised average US wholesale electricity prices by 3% to 5%, with larger effects in major data-center corridors. Its modeled 2028 outcomes range from a 20% increase under moderate construction to 50% when proposed projects operate at high utilization.
The estimates are scenarios rather than guaranteed outcomes, but they reveal a feedback loop. Developers search for affordable power, new facilities increase demand, and utilities may need additional generation, transmission, and equipment. The location that initially offered the cheapest electricity may become costlier as the cluster expands.
Electricity policy consequently becomes industrial policy. Decisions about upgrade costs, large-customer tariffs, and generation approvals can influence whether AI investment remains attractive without shifting excessive expenses to households and existing businesses.
Flexible Computing Could Open More Locations
Not every AI workload must run at maximum power every minute. Some training, data processing, and nonurgent inference can be delayed, reduced during peak periods, or moved between facilities.
In a synthetic Texas power-system test, the Nature Communications researchers found that allowing data centers to pause or shift electricity use expanded technically feasible sites by 9% to 17% for a one-gigawatt facility and by 19% to 21% for a two-gigawatt facility.
The study estimated that pre-certified sites with agreed operating limits could achieve initial energization in 12 to 18 months, compared with roughly five to eight years under conventional processes requiring major network work.
An AI company that schedules computing around grid conditions may gain more location choices, negotiate better tariffs, and begin operating sooner. Software flexibility becomes a real-estate advantage because a responsive cluster is easier to accommodate than an equally large facility demanding uninterrupted maximum output.
The Winning Site Needs More Than the Lowest Rate
Electricity will not erase other requirements. Data-center planners must also consider water availability, noise, air permits, emissions, land constraints, construction logistics, and potential environmental-justice effects.
Energy source matters as well. Cheap power from a carbon-intensive grid may conflict with climate commitments, while renewable projects without storage or dependable backup may not provide continuous reliability.
The IEA projects that renewable energy will supply almost half of the additional electricity required by data centers through 2030, followed by natural gas and coal, with nuclear power taking a larger role later in the decade.
The strongest locations will therefore combine affordable generation with grid capacity, reliable backup, and a credible route for expansion. A market offering the lowest advertised rate may lose to one that can guarantee delivery, connection timing, and future supply.
The Next AI Hub May Form Around a Substation
The geography of technology is becoming less abstract. AI may be sold as software, but its underlying economics increasingly resemble heavy industry: large fixed investments, continuous electricity consumption, and dependence on physical infrastructure.
A prestigious address cannot compensate for a five-year connection delay. A strong talent pool may not overcome electricity costs that make every model query more expensive. A region with abundant power, quick approvals, and a flexible grid could attract investment even without a long history as a technology center.
The next major AI hub may emerge where electricity is not merely cheap, but expandable, reliable, and available on schedule. In the AI economy, the defining location advantage may no longer be proximity to venture capital. It may be proximity to a grid that can say yes.
