AI is becoming an electricity problem not simply because a chatbot consumes power, but because AI concentrates large, steady, high-performance computing loads in particular places faster than grids can plan, permit, and build for them. The U.S. Department of Energy says data centers used about 4.4% of U.S. electricity in 2023. Its cited Lawrence Berkeley National Laboratory estimate puts the 2028 range at 6.7% to 12%, depending on the pace of deployment and the rest of the economy. GPUs, networking gear, cooling, and redundant power systems all matter. So do transmission lines, rate design, water availability, and the gap between a two-year data-center build and a much longer energy-infrastructure timeline. The central question is whether the benefits of AI can be matched with credible, locally fair plans for power.
The issue is speed and concentration, not a single alarming number
Yes, AI is adding a consequential new load to the power system. But the useful way to understand the problem is not to picture every prompt as a tiny appliance left on at home. It is to picture a cluster of industrial facilities asking for hundreds of megawatts, sometimes more, in a place that may have been designed around far slower growth.
DOE’s September 2026 Electricity Demand Growth Resource Hub places data-center expansion and AI applications among the main drivers of rising U.S. demand, alongside domestic manufacturing and electrification. The hub cites LBNL’s estimate that data centers consumed about 4.4% of total U.S. electricity in 2023 and could consume roughly 6.7% to 12% in 2028. That is a wide band, and it should be read as one. It is not a promise that the country will land at 12%.
The underlying LBNL report estimated 176 terawatt-hours of data-center use in 2023, with a 2028 range of 325 to 580 terawatt-hours. The spread reflects real unknowns: how many facilities are completed, how rapidly AI services spread, how efficient the hardware becomes, and how quickly other electricity uses grow. Still, even the low end changes the planning problem. A national percentage can sound abstract; a new facility’s demand arrives on a particular substation, transmission corridor, water system, and tax base.
Why AI hardware changes the data-center equation
LBNL’s report found that AI training differs fundamentally from conventional enterprise computing because it uses massive amounts of parallel computation. In measurements it reviewed, an eight-H100 AI node used an average 7.9 kilowatts during compute-saturated training, about 78% of its manufacturer-rated power. The researchers modeled annualized workload demand at 70% of rated power rather than assuming machines run at their maximum all year. That detail matters. Sensible analysis does not treat every chip as permanently maxed out, but neither should it ignore what happens when thousands of such nodes run together.
Training and inference also create different planning questions. Training a frontier model can be long, intensive, and relatively predictable at a dedicated site. Inference, the act of serving models to users and software, is more diffuse and variable. LBNL says its long-run energy impact may be larger than training but is harder to forecast because it depends on hardware, user adoption, model design, and facility characteristics.
Cooling and reliability are part of the load, too
The computer is not the whole data center. Heat must be removed, voltage conditioned, and critical services protected during outages. The International Energy Agency estimates that servers average about 60% of modern data-center demand. Cooling ranges from about 7% at efficient hyperscale facilities to more than 30% at less-efficient enterprise centers.
Location and design therefore matter. Liquid cooling can handle dense GPU racks but needs new equipment; water-based cooling can reduce some electricity needs while increasing local water concerns. Air cooling may reduce direct water use but consume more power. There is no free engineering choice, only trade-offs that should be disclosed before approval.
Data centers also need uninterruptible power supply batteries and backup generators because a momentary outage can disrupt services and damage equipment. Onsite generation and storage can reduce pressure on the grid, but their value depends on fuel, connection design, and who pays for network costs.
DOE’s proposed direction is that data centers should be capable of becoming a grid asset rather than merely a larger load. Its hub highlights onsite power, storage, and flexible demand. In plain terms, some computing work might be shifted away from an expensive peak hour, while batteries can charge at a lower-demand time and discharge later. That will not suit every workload. A live medical service or payment system cannot casually pause. But training runs, batch jobs, and some inference work may offer more flexibility than the industry’s most rigid reliability language suggests.
The grid cannot be ordered like a server rack
A data center can be developed in two or three years. The IEA notes that energy infrastructure usually requires longer planning, construction, and upfront investment. The mismatch is the heart of the bottleneck.
A large new customer needs more than a contract for electricity. It may need a substation, transformers, distribution upgrades, transmission capacity, generation adequate for peak conditions, and studies of whether its connection affects neighboring customers. Each piece has equipment lead times and permits. Transmission lines in particular cross property, jurisdictional, and environmental boundaries. Cutting review to move faster may simply move costs and conflicts downstream.
Federal regulators have begun to address this. In June, the Federal Energy Regulatory Commission directed the six regional grid operators under its jurisdiction to justify or reform their rules for connecting large loads, including data centers. FERC identified five areas: efficient study processes; transparency and protection against cost shifting; co-location and behind-the-meter generation; service for flexible large loads; and study of nearby generation serving those loads.
“Behind the meter” means power equipment on the customer side of the grid meter. It may reduce withdrawals, but it does not settle who pays for backup, wires, or reserve power.
The human stakes are bills, water, work, and trust
Electricity systems socialize many costs. When a utility builds capacity or grid upgrades, regulators decide which customers pay. If a data-center project departs early, uses less power than forecast, or receives an unusually favorable rate, households and small businesses can be left supporting infrastructure built for someone else.
States are starting to set guardrails. The Public Utilities Commission of Ohio required AEP Ohio to establish a separate data-center customer class. New facilities must meet conditions including collateral, minimum demand charges, long-term contracts, and exit fees. PUCO describes the purpose plainly: prevent overbuilding and avoid shifting costs to other customers. This does not prove every large-load rate needs the same formula. It is a practical example of the principle that a customer creating large, dedicated costs should provide meaningful financial commitment.
The upside is real but uneven. Construction creates work for electricians, lineworkers, engineers, suppliers, and trades. Operating data centers employ technicians and facility specialists, though often fewer people than a manufacturing plant of comparable land and power footprint. The labor market reaches beyond the building: firms building AI products need technical talent as well as workers who can operate physical infrastructure. That is why workforce planning, including an AI Recruiting Company, belongs in the broader discussion, not as a substitute for job and training commitments from developers.
Water deserves the same specificity. A proposal should state its cooling design, expected water source, seasonal demand, wastewater plan, and what happens during drought. So should it describe backup fuel, air emissions, noise, land use, and the power source used when renewables are scarce. General pledges to procure “clean energy” are not a substitute for local disclosure. A renewable contract may improve project economics without physically placing reliable new capacity near the load.
Firm power has appeal, but no technology erases the hard parts
Solar, wind, batteries, efficiency, transmission upgrades, and demand flexibility can all help. They are not interchangeable. Solar and wind produce inexpensive energy when conditions are favorable; storage can shift that energy and provide fast response; expanded transmission can connect diverse resources. Data centers, meanwhile, often seek power that is available continuously. That is why advanced nuclear and next-generation geothermal have become central to the AI power conversation.
Nuclear has an obvious attraction: firm, low-carbon generation with a small land footprint. DOE notes that existing nuclear plants can provide round-the-clock output, but it also cautions that licensing, demonstrating, and deploying new advanced designs will take years, with widespread commercial deployment more likely in the 2030s. First-of-a-kind projects are expensive, and co-location raises metering and cost-allocation questions. Company announcements about nuclear-backed data centers should therefore be understood as commercial ambitions and contracts, not evidence that a new reactor can solve next year’s capacity gap.
Geothermal offers another form of firm power. DOE says geothermal plants generally have capacity factors around 90%, meaning they can run steadily most of the time, and that enhanced geothermal systems could reduce the resource’s historic location constraints. Yet drilling risk, site characterization, permitting, financing, and transmission still matter. It is promising infrastructure, not instant infrastructure. Geothermal thermal storage may also help with the less glamorous problem of cooling by storing cold underground and using it to reduce peak cooling demand.
Permitting is where these threads meet. Slow, duplicative processes can delay wires, generation, and storage that communities genuinely need. Fast processes that bypass environmental review, landowners, or water scrutiny can undermine the legitimacy required to build at all. The more durable path is competent review with clear timelines, early community involvement, transparent load forecasts, and enforceable cost responsibility.
What the evidence says
The evidence supports a direct answer: AI growth can be sustained, but not on autopilot. The DOE and LBNL figures show a large and uncertain rise in data-center electricity demand, not an inevitable national blackout. The IEA’s scenarios make the same broader point: efficiency, adoption, and supply bottlenecks can push outcomes in different directions.
The test is whether power planning catches up with computing ambition: credible lead times for generation, storage, and wires; truly flexible workloads where possible; and protection from avoidable rate and water burdens. The electricity problem is ultimately a governance problem. Growth earns public support when its benefits and costs are shared fairly.
