In 2023, every data center in the United States combined drew about 4.4% of the nation’s retail electricity. By 2028, that figure could hit 12%, according to projections from Lawrence Berkeley National Laboratory. If that trajectory holds, the power consumed by servers, cooling systems, and networking gear would roughly triple in five years, a pace of growth the American grid has not had to absorb since the post-war manufacturing boom.
The surge is not hypothetical. Across northern Virginia, central Ohio, west Texas, and the Phoenix metro area, hyperscale operators including Amazon, Google, Microsoft, and Meta have announced or begun construction on campuses that will each require hundreds of megawatts of continuous power. Utilities that spent the last decade managing nearly flat demand are now fielding interconnection requests that, taken together, exceed the capacity of some regional grids.
Three independent analyses point the same direction
The most detailed accounting comes from the 2024 U.S. Data Center Energy Usage Report, produced by Lawrence Berkeley National Laboratory for the Department of Energy. Using retail sales totals from the Energy Information Administration as its baseline, the report pegs 2023 data center consumption at roughly 4.4% of all U.S. electricity and models a 2028 range of 6.7% to 12%, depending on the speed of AI deployment, chip efficiency gains, and new facility construction.
The EIA’s Annual Energy Outlook, most recently updated in its 2025 edition (with a 2026 edition expected but not yet confirmed as published), reinforces that picture. The agency identifies data center server demand as a primary structural driver of electricity load growth over the next decade, breaking a long stretch of nearly flat national consumption. The International Energy Agency’s Electricity report series (most recently the Electricity 2025 edition published in January 2025, with a 2026 edition anticipated) adds a global lens, estimating that data centers account for roughly half of all incremental U.S. electricity demand growth through 2030.
The Electric Power Research Institute, which works directly with U.S. utilities and has access to interconnection queue data that academic researchers rarely see in real time, published scenario analyses projecting that data centers could consume between 9% and 17% of U.S. electricity by 2030. The upper bound implies growth that would blow past Berkeley Lab’s 2028 ceiling if AI training workloads and hyperscale construction keep accelerating.
Three institutions, a national laboratory, a federal statistical agency, and an industry research body, arrived at the same core finding independently: data center electricity consumption is growing faster than any other category of commercial load in the country.
Why the range is so wide
Berkeley Lab’s 2028 spread, 6.7% to 12%, is nearly a factor of two. That gap reflects genuine unknowns. How quickly will Amazon, Google, and their peers bring new campuses online? How much will next-generation chips from Nvidia, AMD, and Intel reduce energy per computation? And can utilities actually connect new load at the pace developers are requesting, or will permitting bottlenecks act as a natural brake?
No publicly available federal dataset currently isolates data center electricity consumption from the broader commercial building stock. The EIA’s Electric Power Annual tracks total retail sales by sector, but data centers are lumped in with office buildings, hospitals, and retail stores. That makes it difficult to verify in real time whether actual consumption is tracking closer to the low or high scenario.
EPRI’s 9% to 17% range targets 2030, two years beyond Berkeley Lab’s 2028 window, and the two organizations have not published a joint reconciliation of their models. Whether their upper-bound trajectories overlap or stack is unclear from the available documentation. The 12% figure should be understood as the top of a plausible range for 2028, not a consensus forecast.
The grid bottleneck no one planned for
Permitting timelines for natural gas plants and high-voltage transmission lines typically run three to five years or longer. If the upper-bound demand trajectory materializes, several regional grids could face binding capacity shortfalls before new supply is built. PJM Interconnection, the grid operator covering 13 states from Virginia to Illinois, has already flagged a backlog of data center interconnection requests totaling tens of gigawatts, far more than its current planning cycle anticipated.
Some hyperscalers are trying to sidestep the bottleneck entirely. Amazon has signed agreements to co-locate data centers next to existing nuclear plants. Microsoft has invested in small modular reactor development. Google has signed long-term power purchase agreements for geothermal energy. These arrangements may ease pressure on the broader grid, but they also raise new regulatory questions about whether large corporate buyers should receive priority access to clean generation that might otherwise serve residential and industrial customers.
Data centers are also not the only new load competing for grid capacity. Electric vehicles, building heat pumps, and industrial electrification are all adding demand at the same time. Even if data center growth lands near the lower end of projections, utilities would still face higher overall consumption than at any point in the past decade, complicating investment decisions and rate-setting.
The data gap that makes planning harder
Berkeley Lab has emphasized that its 2024 study should be read as a set of plausible trajectories, not a single forecast. The laboratory’s researchers stress that their scenarios are designed to help policymakers and utilities plan for a wide range of outcomes, not to predict a precise percentage for any given year.
A key limitation is the absence of standardized, public reporting of data center electricity use at the facility level. Today’s estimates depend on modeled server counts and assumed power usage effectiveness ratios. Berkeley Lab argues that more transparent, anonymized reporting could sharpen future projections and help regulators distinguish between regions where demand growth is manageable and regions where it may outpace available capacity.
The 4.4% figure for 2023 is the most recent hard data point. Everything beyond it is a projection built on assumptions about construction timelines, chip efficiency, and corporate capital spending that can shift quarter to quarter. Until the EIA or another federal body publishes actual 2025 or 2026 data center consumption figures, there is no definitive way to say whether the U.S. is tracking toward the lower or upper end of the published ranges.
What this means for electricity bills and reliability
For ordinary ratepayers, the stakes are concrete. Utilities that need to build new generation and transmission to serve data centers will seek to recover those costs, and state regulators will have to decide how to split the bill between the tech companies requesting the power and the households and businesses already on the grid. In Georgia, regulators recently approved new natural gas capacity tied in part to data center load growth. In Virginia, Dominion Energy has proposed billions of dollars in transmission upgrades driven largely by the data center corridor in Loudoun County, the densest concentration of server capacity on Earth.
Reliability is equally at issue. If new demand arrives faster than new supply, grid operators may face tighter reserve margins during peak periods, increasing the risk of rolling blackouts during extreme heat or cold events. The North American Electric Reliability Corporation flagged this dynamic in its 2024 long-term reliability assessment, warning that load growth from data centers and electrification could outpace resource additions in several regions by the late 2020s.
Across the available evidence as of mid-2026, three points are solidly grounded. Data center electricity use is already material and rising. Every major institutional analysis agrees it will grow faster than the rest of the grid. And the range of plausible outcomes is wide enough that policy and investment choices made in the next two to three years will meaningfully shape where the U.S. lands within that range.
Why the next two years of grid decisions will define the decade
For now, the projections function more as a warning light than a precise gauge. They signal that data center demand is large enough and growing fast enough to matter for every major U.S. grid, but they do not yet pin down exactly how large or how fast. As updated federal statistics and refined modeling become available, including work summarized in a recent Berkeley Lab public briefing, the picture will sharpen. Until then, utilities, regulators, and the communities that depend on affordable, reliable power are making consequential infrastructure choices under conditions of real, and acknowledged, uncertainty.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
