American households and businesses face a new strain on the power grid: data centers that consumed roughly 4.4 percent of all U.S. electricity in 2023 are on track to demand between 6.7 and 12 percent by 2028. That projection, drawn from a Lawrence Berkeley National Laboratory analysis prepared for the Department of Energy, translates to a potential jump from 176 terawatt-hours to as much as 580 terawatt-hours in just five years. The speed of that growth puts grid operators, state regulators, and ratepayers on a collision course with generation and transmission timelines that were never designed for this kind of demand surge.
Why the 12 percent ceiling matters for grid planning
The tension behind the headline is not the upper bound alone but the width of the range. A gap between 325 and 580 terawatt-hours in 2028 reflects deep uncertainty about how quickly artificial intelligence workloads, cloud computing, and cryptocurrency mining will scale, and whether efficiency gains can keep pace. The Department of Energy’s own analysis of data center demand frames this as a stress test for an electricity system already contending with aging infrastructure and rising demand from electrification of vehicles and heating.
Between 2014 and 2023, U.S. data center electricity consumption tripled, climbing from 58 terawatt-hours to 176 terawatt-hours. That growth happened while server utilization rates and facility-level efficiency, measured by metrics such as power usage effectiveness, or PUE, improved steadily. Peer-reviewed research published in Environmental Research Letters demonstrated that data center service demand can grow faster than electricity use when operators invest in better hardware and cooling. The question now is whether that decoupling pattern can hold as AI training clusters push power densities far beyond what traditional servers required.
If measured server utilization rates continue rising at the pace observed in the 2014 to 2023 LBNL data, realized 2028 electricity demand would likely land in the bottom third of the projected range, closer to 325 terawatt-hours, even without new federal efficiency mandates. That scenario depends on hyperscale operators continuing to retire older, less efficient equipment and consolidating workloads into fewer, better-cooled facilities. But the upper end of the range accounts for a world where new AI data centers proliferate faster than efficiency improvements can offset their appetite, and where smaller, less optimized facilities remain a large share of the installed base.
For grid planners, the difference between those trajectories is the difference between incremental upgrades and a scramble for new generation. A 325 terawatt-hour outcome could be met largely by accelerating already-planned renewable projects, modest gas capacity additions, and targeted transmission reinforcements. A 580 terawatt-hour outcome, by contrast, would require far more aggressive buildouts of both generation and high-voltage lines, with significant implications for permitting processes, land use conflicts, and rate design. Utilities that underestimate the load risk running short of capacity; those that overbuild could saddle customers with stranded costs if efficiency gains ultimately prevail.
LBNL data and the congressional mandate behind the numbers
The projections originate from report LBNL-2001637, prepared by Lawrence Berkeley National Laboratory researchers in response to a congressional directive under the Energy Act of 2020. That law required the Department of Energy to assess data center energy consumption and report findings to Congress. The resulting analysis builds on a methodology that LBNL researchers have refined over more than a decade, starting with foundational work published in a 2011 IEEE paper that established a building-block accounting approach: total data center energy equals IT equipment load plus infrastructure overhead for cooling, power distribution, and lighting.
The nonpartisan Congressional Research Service picked up the LBNL findings in its own CRS Report R48646, which places the electricity projections in a policy context. That CRS overview discusses implications for grid planning, reliability standards, and potential efficiency regulations. It also flags data gaps that make long-term forecasting difficult, including limited public reporting by data center operators on actual energy consumption and equipment configurations.
The LBNL methodology accounts for several variables that can swing the outcome dramatically. Server shipment volumes, average power draw per server, the share of AI-optimized hardware in new deployments, and facility-level PUE all feed into the model. Small changes in any one assumption ripple through the projections. A one-point improvement in average PUE across the national fleet, for instance, would shave tens of terawatt-hours off the 2028 total. Conversely, if AI chip power consumption keeps climbing and operators build new facilities faster than grid connections can keep up, the upper bound becomes more plausible.
Congressional staff who requested the assessment were focused not only on raw consumption but also on how that load interacts with reliability standards and climate goals. If data centers pull more power from fossil-heavy regions, they could complicate state and federal decarbonization timelines even as other sectors electrify. The CRS analysis notes that without better data, it is difficult for lawmakers to design incentives or standards that nudge operators toward cleaner grids or more efficient architectures without unintended consequences.
Gaps in the data and what to watch through 2028
Several blind spots limit confidence in any single point estimate. No primary DOE or LBNL dataset breaks down the 325 to 580 terawatt-hours range by region or utility territory, which means state regulators and grid operators cannot yet pinpoint where the heaviest load growth will land. Northern Virginia, central Ohio, and parts of Texas have attracted the largest recent data center construction pipelines, but the federal data does not map projected consumption to those corridors in a way that local planners can act on directly.
Updated primary measurements of average PUE and server utilization rates after 2023 have not yet been published. The LBNL report relies on the best available data through that year, but the AI hardware cycle is moving fast enough that conditions on the ground may already look different from the modeling inputs. Congressional hearing transcripts and utility integrated resource plans suggest that some regions are already revising load forecasts upward to account for new data center campuses, yet those localized adjustments have not been reconciled with the national projections.
Another gap is the limited visibility into smaller, distributed data centers and edge facilities. Hyperscale campuses operated by major cloud providers often have the resources and incentives to optimize efficiency, experiment with advanced cooling, and negotiate bespoke power contracts. By contrast, colocation facilities and enterprise-owned server rooms may lag in adopting best practices, pushing their PUE higher and their flexibility lower. If the growth in AI workloads spills heavily into these less efficient environments, national consumption could skew toward the upper end of the forecast band.
Regulators and utilities will be watching several indicators closely over the next few years. One is the pace of interconnection requests tied to new data centers, which provides an early signal of where large loads may materialize. Another is how quickly operators adopt measures like liquid cooling, on-site energy storage, and demand response programs that can shave peak loads or shift consumption to off-peak hours. Policy developments will matter as well: performance-based efficiency standards, enhanced disclosure requirements, or targeted tax incentives could all influence whether the sector bends toward or away from the lower end of the projected range.
For now, the 6.7 to 12 percent share of U.S. electricity use functions less as a precise forecast and more as a warning band. It underscores that data centers, once a niche load largely invisible to most consumers, are becoming a central driver of grid investment decisions. Whether the country ends up closer to 325 or 580 terawatt-hours in 2028 will depend on choices made in the next few years by cloud providers, chip designers, regulators, and utilities. The stakes are not abstract: they will show up in electricity bills, in the resilience of the power system during heat waves and storms, and in the pace at which the broader economy can decarbonize while still embracing digital growth.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
