AI’s next bottleneck may be electricity. The industry can order more chips and build more server halls, but power plants, substations and transmission lines operate on slower timelines—and must clear regulatory, financial and political hurdles before they can serve a new data-center campus.
That is changing the AI infrastructure race. The question is no longer simply where a cloud provider can find land, fiber and tax incentives. It is increasingly where it can obtain large quantities of dependable electricity, secure a high-voltage grid connection and persuade regulators and residents that the necessary upgrades are worth building.
The constraint is real, even if some of the largest forecasts should be treated cautiously. Data-center proposals are not the same as operating facilities, and utilities must plan for demand that may arrive later, arrive at a smaller scale or never materialize. Still, the scale of announced projects is forcing an unusually immediate confrontation between fast-moving computing investment and the deliberate pace of utility grid planning.
AI data center electricity demand is rising in concentrated places
Data centers have long been major electricity users, but AI changes the shape of demand. Training and serving large models requires dense clusters of specialized processors, networking equipment and cooling systems. A conventional enterprise data center, a cloud region and a facility built around AI computing are not identical categories. Yet the broad trend is clear: larger campuses are seeking larger and more concentrated power connections.
The International Energy Agency estimated that data centers consumed about 415 terawatt-hours of electricity globally in 2024 and projected consumption could reach roughly 945 terawatt-hours by 2030 in its 2025 electricity report. AI is one important driver of that increase, alongside conventional cloud computing and digital services. The projection is substantial, but it is a scenario-based forecast rather than a promise that every proposed facility will be built and fully utilized.
In the United States, a December 2024 Lawrence Berkeley National Laboratory report estimated data centers used about 176 terawatt-hours in 2023, or 4.4% of national electricity consumption. It projected that figure could rise to between 325 and 580 terawatt-hours by 2028, depending on how quickly facilities are constructed, equipped and used. The range itself is instructive: uncertainty is not a side note in this story. It is central to the planning problem.
The grid bottleneck is bigger than annual electricity supply
A region can have enough electricity on paper and still be unable to connect a large new customer quickly. A data center needs power at a particular place, at a particular voltage, with a level of reliability appropriate to its operations. That can require new substations, transformers, feeder lines and long-distance transmission—not merely additional generation somewhere on the system.
Utilities and regional grid operators therefore face several overlapping constraints:
- Interconnection studies: New customers and power plants must be assessed for their effect on reliability and local equipment. Studies can reveal that costly upgrades are needed before a connection is feasible.
- Transmission capacity: Existing lines may not be able to move enough power into the fast-growing load center, particularly during peak demand.
- Equipment availability: Transformers, switchgear and other high-voltage components have faced extended procurement lead times. These are specialized industrial products, not equipment a developer can replace overnight.
- Generation and reliability: A utility must ensure sufficient capacity during periods when demand is highest or renewable output is low—not just match annual energy consumption.
The distinction between a forecast and delivered power matters. A developer may reserve land, announce an investment or request an interconnection study years before its servers draw significant electricity. Conversely, once a campus is operational, its demand can grow rapidly as equipment is installed. Utilities have to make capital decisions amid both risks: underbuilding could delay economic activity, while overbuilding could leave customers paying for assets that are not fully used.
Why the geography of AI is becoming an energy question
Virginia, Texas, Georgia, Ohio, Arizona and parts of the Midwest have attracted data-center development for different combinations of fiber connectivity, land, market access, tax policy and power supply. Northern Virginia’s longstanding data-center concentration demonstrates the advantage of network density, but it also shows how success can strain local transmission and generation planning.
Newer projects are increasingly drawn to places where a utility can credibly offer a large connection. That can redirect investment toward areas with available land near high-voltage infrastructure, or toward states willing to coordinate permitting and utility expansion. It can also place major industrial-scale loads in rural communities that do not necessarily receive comparable local benefits.
Jobs, tax revenue and construction spending are part of the appeal. But residents and local officials may also face land-use disputes, noise from backup equipment, water questions associated with some cooling designs and concerns over whether grid upgrades will alter local rates or reliability. Those effects vary sharply by facility design, utility rules and local conditions. It is inaccurate to assume every data center has the same water footprint or creates the same grid burden.
The policy stakes became visible in high-profile power deals. Amazon’s agreement to obtain power for a data-center operation associated with the Susquehanna nuclear plant in Pennsylvania prompted scrutiny from federal regulators and grid stakeholders over reliability and cost allocation. Microsoft’s agreement to purchase power associated with a planned restart of Three Mile Island’s Unit 1 likewise illustrated the lengths major technology companies may go to secure long-term clean electricity. Neither case is a simple template for the wider market: nuclear arrangements are location-specific, regulated and slow to develop.
Who pays is becoming the central political dispute
Utilities traditionally spread the cost of broadly useful infrastructure across customer classes, while charging individual customers for facilities built specifically to serve them. The data-center boom tests that boundary. A new transmission line or generation project may serve one enormous campus initially but also provide wider regional value later. Deciding how much each party should pay is a regulatory judgment, not a technical calculation alone.
Consumer advocates worry that households and small businesses could absorb the costs of speculative data center construction through higher rates. Developers and utilities respond that large customers can support investment, contribute tax revenue and help finance a stronger grid. Both claims can be true in different jurisdictions. The important details are in utility rate cases, interconnection agreements and approved tariffs: demand commitments, exit fees, upfront contributions and protections if a project is delayed or cancelled.
The most useful question is not whether AI uses “too much” power in the abstract. It is whether a proposed load can be served reliably, at a transparent cost, without shifting excessive risk to other customers.
Efficiency and clean-power contracts help, but do not erase the constraint
Technology companies are pursuing more efficient chips, improved cooling, better server utilization and software that can move flexible workloads across regions or hours. Renewable-power contracts, batteries and demand-response programs can reduce emissions and help balance the system. Some operators are also exploring onsite generation.
These measures matter, but they are not instant substitutes for grid infrastructure. Batteries generally shift electricity over limited periods rather than provide continuous power for a large campus. Renewable contracts do not automatically guarantee local deliverability at every hour. Onsite gas generation can provide firm capacity but raises permitting, fuel and emissions questions. And a new nuclear plant or reactor restart remains a long-term, highly regulated undertaking.
What to watch as the power race develops
The clearest signals will come from documents less glamorous than AI model launches: utility integrated-resource plans, regional transmission plans, interconnection agreements, public utility commission decisions and local permits. Revisions to demand forecasts will be especially revealing. They show whether announced data center construction is translating into contractual commitments and actual load.
AI energy use will continue to grow, but its path will not be determined by model capability alone. It will be shaped by whether utilities can build generation and wires, whether manufacturers can supply critical equipment, and whether regulators can allocate costs credibly. In that sense, the future geography of AI may be decided as much in substations and public hearings as in data centers themselves.