The electricity appetite of U.S. data centers has grown so fast that, by 2028, these facilities could consume as much as 12 percent of all electricity generated in the country. That projection comes from a federal study released by Lawrence Berkeley National Laboratory and backed by the U.S. Department of Energy. The finding lands at a moment when grid operators and utilities are already struggling to keep pace with rising demand, and it raises hard questions about whether the nation’s power infrastructure can absorb a near-tripling of data-center load in roughly four years.
Why a 12 percent electricity share changes the grid calculus
In 2023, U.S. data centers consumed about 4.4 percent of total national electricity, according to the Energy Department. That figure already represented a steep climb from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023. The new projections push the range to 325 to 580 terawatt-hours by 2028, with the upper bound corresponding to roughly 12 percent of U.S. electricity and the lower bound sitting near 6.7 percent.
The gap between 6.7 and 12 percent is not just an academic range. It reflects different assumptions about how quickly artificial-intelligence workloads scale, how aggressively hyperscale operators build new campuses, and whether efficiency gains in chips and cooling can offset raw demand growth. If measured load growth through 2025 tracks closer to the upper-bound trajectory while efficiency metrics stay flat, the 12 percent threshold could arrive ahead of schedule rather than at the tail end of the decade. That scenario would force utilities to secure generation capacity and transmission upgrades on a timeline most have not yet planned for.
For ordinary ratepayers, the stakes are direct. When data centers claim a larger slice of fixed grid capacity, competition for electrons tightens. Wholesale power prices can rise, reliability margins can shrink during peak demand, and new industrial or residential customers may face longer waits for interconnection. Regions that are already capacity-constrained, such as parts of Virginia’s data-center corridor and central Texas, would feel those pressures first, as large new server farms compete with factories, homes, and electric vehicles for the same megawatts.
Local politics around land use and noise have long shaped where data centers get built. Now, the binding constraint in many communities is shifting from zoning to power. Counties that once welcomed server campuses for their tax base are starting to question whether hosting another multi-hundred-megawatt facility is worth the added strain on substations and transmission lines. As the national share of electricity devoted to data centers climbs, those local debates are likely to intensify.
What the LBNL report and DOE data actually show
The core evidence sits in the 2024 usage report produced by researchers at Lawrence Berkeley National Laboratory with Department of Energy backing. The report lays out multiple scenarios for 2028 consumption, each built on different growth rates for server deployments, AI training runs, and facility-level power usage effectiveness.
At the low end, the study projects 325 terawatt-hours, which would represent about 6.7 percent of national electricity. At the high end, 580 terawatt-hours would push the share to approximately 12 percent. Both figures assume that overall U.S. electricity generation grows on its current trajectory, meaning any slowdown in new power-plant construction would make the percentage even larger. The historical baseline reinforces the speed of change: consumption more than tripled between 2014 and 2023, jumping from 58 terawatt-hours to 176 terawatt-hours, and the report suggests the next tripling could happen in roughly half the time.
Berkeley Lab has tracked data-center energy use for more than a decade, and its earlier estimates were widely cited by grid planners. The new report updates those figures with post-2022 data reflecting the surge in AI-related computing, especially large training clusters and high-density inference deployments. Reuters summarized the study’s top-line finding by noting that U.S. data-center power use could nearly triple by 2028, a framing that aligns with the DOE’s own characterization of the results and underscores how unusual this pace of growth is compared with most other electricity-consuming sectors.
The DOE released the report alongside a broader push to assess how rising electricity demand from data centers, electric vehicles, and manufacturing reshoring will strain the grid. That context matters because data centers are not the only new load competing for generation and transmission. They are, however, among the fastest-growing and most geographically concentrated loads, which makes their impact on local grids disproportionate to their national share. A single hyperscale campus can draw as much power as a small city, and multiple campuses are often clustered around the same substations and high-voltage lines.
Another nuance in the LBNL scenarios is the role of efficiency. The report assumes continued improvements in server performance per watt and in cooling systems, but it also notes that these gains can be overwhelmed by sheer growth in computing demand. AI accelerators are more energy-efficient per operation than general-purpose chips, yet the volume and intensity of AI workloads can still push overall facility demand sharply higher. That dynamic helps explain how data-center electricity use can nearly triple even if individual components become more efficient.
Gaps in the forecast and what to watch next
The LBNL report supplies scenarios and methodology, but several pieces of the puzzle are still missing. Neither the DOE announcement nor the technical report includes detailed regional breakdowns of transmission constraints or specific utility interconnection queues tied to the 2028 projections. Without that granularity, it is difficult to know which parts of the country face the most acute risk of shortfalls, or how quickly new substations and lines would need to be built to accommodate expected clusters of demand.
Direct commitments from the largest data-center operators, including Amazon Web Services, Microsoft Azure, and Google Cloud, on efficiency improvements or siting changes are also absent from the official federal analysis. Those companies collectively account for a large share of new construction, and their internal forecasts for power demand would sharpen the public projections considerably. Until those figures surface, the range between 6.7 and 12 percent will remain wide, and planners will have to prepare for the upper bound while hoping that efficiency and better siting keep reality closer to the lower end.
Updated 2024 consumption figures have not yet been incorporated into the cited institutional sources, so the baseline against which growth is measured is already slightly stale. If 2024 demand significantly overshoots the 2023 level, the climb to the 2028 scenarios could be shorter and steeper than currently presented. Conversely, if efficiency measures or delays in facility buildouts slow the near-term ramp, the upper-bound scenario might prove overly conservative.
Several developments will determine which path the U.S. ultimately follows. On the technology side, the rate at which AI models grow in size and complexity will shape server requirements, as will any shift toward more efficient architectures or model-sharing strategies that reduce redundant training. On the policy side, state and regional regulators are weighing how to reform interconnection rules, encourage data centers to locate near surplus renewable generation, or require more on-site clean-power procurement to limit strain on local grids.
For now, the LBNL and DOE projections function less as a precise prediction than as a warning sign. Whether data centers reach 6.7 percent or 12 percent of U.S. electricity use by 2028, the direction of travel is clear, and the timeline is short. Utilities, regulators, and cloud providers have only a few years to align their plans, or risk discovering that the next wave of digital infrastructure is bumping up against hard physical limits of the power system.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
