In Loudoun County, Virginia, the hum of server fans has become as familiar as highway traffic. The county hosts the densest cluster of data centers on Earth, and local utility Dominion Energy has warned that electricity demand in its Northern Virginia territory could double within the next several years. What is happening in Loudoun is now happening, at varying speeds, in central Texas, the Phoenix metro area, and pockets of the Midwest. Taken together, these facilities are redrawing the electricity map of the United States.
U.S. data centers consumed about 4.4% of the nation’s total electricity in 2023, drawing roughly 176 terawatt-hours, according to a congressionally mandated study by Lawrence Berkeley National Laboratory released through the U.S. Department of Energy. By 2028, that share could climb as high as 12%, reaching up to 580 terawatt-hours. To put that in perspective, the upper end of that range equals roughly the annual electricity consumption of every home in California, Texas, and Florida combined. The gap between those two numbers represents a tripling of the sector’s grid footprint in roughly five years, and as of June 2026, utilities and regulators are still scrambling to figure out who will pay for the new power plants and transmission lines needed to keep the lights on everywhere else.
Where the numbers come from
The core projections originate from the DOE report, which drew on LBNL modeling across multiple growth scenarios. The primary driver behind the acceleration is artificial intelligence. Training a single large AI model can require thousands of specialized processors running at near-full capacity for weeks or months, consuming electricity at rates that dwarf traditional cloud computing workloads. As companies race to build and deploy ever-larger models, the power demands of their facilities have surged past what grid planners anticipated even two or three years ago.
Under the DOE’s range of scenarios, data centers are expected to consume between 325 and 580 terawatt-hours by 2028, translating to roughly 6.7% to 12% of national electricity use. Even the low end of that range represents an 85% jump in five years. At the high end, the sector would add demand equivalent to about one-fifth of current total U.S. consumption. Either outcome would mark the fastest sustained load growth any single commercial category has placed on the American grid in decades.
The International Energy Agency’s Electricity 2026 outlook reinforces the domestic picture with a global lens, identifying data centers and AI as key contributors to electricity demand growth worldwide. The United States hosts a disproportionate share of the world’s hyperscale facilities, making its grid especially exposed to the acceleration. But the trend is not uniquely American: similar buildouts are underway in Ireland, the Netherlands, Singapore, and parts of northern Europe.
Why the range is so wide
The difference between 6.7% and 12% is not a rounding error. It represents two fundamentally different futures for the U.S. power system.
At the lower bound, utilities could plausibly meet incremental demand through a combination of efficiency gains, renewable energy projects already moving through interconnection queues, and modest natural gas buildout. At the upper bound, the country would likely need new baseload generation, major transmission upgrades, and possibly extended lifetimes for aging nuclear plants. Several moves in that direction are already underway: Holtec International is working to restart the Palisades nuclear plant in Michigan, and Constellation Energy has announced plans to bring Three Mile Island Unit 1 back online, in part to serve data center load.
The DOE report does not assign probabilities to specific points within its range, and it does not break down projected load by state or regional grid operator. That omission matters because data center construction is heavily concentrated geographically. Northern Virginia alone accounts for more data center capacity than most entire countries. Regional transmission organizations such as PJM Interconnection and ERCOT have published updated load forecasts that reflect rising data center demand, and PJM has flagged a surge in data center interconnection requests. However, no grid operator has yet released a public capacity plan that fully accounts for the DOE’s upper-bound scenario. Many projects in interconnection queues never reach commercial operation, making it difficult to translate pipeline volume into firm load forecasts.
The federal data also does not separate AI-specific electricity demand from conventional cloud computing and colocation services. If AI adoption accelerates faster than current models assume, the upper end of the DOE range could prove conservative. If efficiency improvements in chip design, liquid cooling, and workload scheduling outpace expectations, the lower end could hold. Hyperscalers like Microsoft, Google, Amazon, and Meta each publish sustainability reports, but none use standardized energy-use metrics on a consistent schedule, leaving analysts to piece together a patchwork picture.
What this means for electricity bills
For households and businesses that never interact with a server rack, the practical question is straightforward: will electricity rates go up?
The answer depends largely on how costs are allocated. If data center operators sign long-term power purchase agreements at market rates and fund their own grid connections, the financial burden falls primarily on those operators and, eventually, on the cloud-service customers who pay for computing capacity. Several major tech companies have taken steps in this direction, signing deals for dedicated nuclear, solar, and geothermal power.
But if utilities socialize the cost of grid upgrades, such as new substations, transmission lines, and generation capacity, across all ratepayers, residential bills could rise even in regions with little data center construction. State public utility commissions will be the key decision-makers on cost allocation, and as of mid-2026, most have not issued formal guidance on how to handle data-center-driven load growth at the scale the DOE projects. Georgia, Virginia, and Texas are among the states where the question has become politically charged, with consumer advocates pushing back against proposals that would spread infrastructure costs to households.
National averages also obscure local extremes. A national data center share of 10% could still mean that certain grid pockets see server farms accounting for a third or more of total load, straining local infrastructure and pushing wholesale prices higher for nearby industrial and residential customers.
Where cost and capacity decisions will surface next
The most concrete signals in the coming months will come from three places. First, utility integrated resource plans filed with state regulators will reveal whether power companies are building their long-term forecasts around the middle or upper end of the DOE range. Second, state commission dockets on major transmission projects will show how regulators plan to split costs between data center operators and other ratepayers. Third, corporate disclosures from cloud providers, under growing pressure from investors and local communities, will indicate how seriously tech companies are pursuing dedicated clean energy rather than drawing from the shared grid.
None of those documents will replace the federal modeling, but together they will answer the question that matters most to the roughly 130 million U.S. households connected to the grid: as the servers keep coming, who picks up the tab?
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
