U.S. data centers consumed approximately 192 TWh of electricity in 2024, equating to 4.7% of total national grid demand, enough to power roughly 18 million homes [S3]. By 2026, that share is projected to climb to 5.3% of total peak summer power demand, up from 4.1% in 2025 [S6].
Longer-range scenarios point to 6.7-12% of U.S. electricity by 2028 under DOE reference cases [S1], and 9-17% by 2030 across EPRI's LowDC-to-HighDC growth bands [S5]. The acceleration is driven by AI workloads that already account for an average of 60% of electricity demand in modern data centers, though this share varies sharply by facility type [S4].
2026 Share Estimates: Where the Numbers Land
Goldman Sachs projects the U.S. data center share of peak summer demand rising from 4.1% in 2025 to 5.3% in 2026 and 8.5% by 2027 [S6]. On an annual-energy basis, scaling the 2024 baseline of 192 TWh (4.7% share) at the IEA's 10% annual baseline growth rate yields approximately 232 TWh in 2026, or roughly 5.5-6% of projected total U.S. demand in the 4,000-4,200 TWh range [S3][S5].
The CRS report flags the upside risk: "Some projections show that data center energy consumption could double or triple by 2028," with roughly one-half or greater of the electric power increase attributable to AI-driven workloads [S2]. A study modeling AI-intensive scenarios confirms baseline data-center electricity consumption growing at 10% annually, while the AI-workload share continues climbing [S8].
State-Level Concentration vs National Average
Virginia's data center share already exceeds 25% of state electricity consumption, by far the highest in the nation, with 490 facilities drawing 24 TWh per year [S3][S5]. Texas follows with 419 facilities at 17 TWh, then California (245, 11 TWh), Illinois (223, 12 TWh), and Arizona (89, 10.5 TWh) [S3].
EPRI's Medium-scenario projection sees Virginia's share rising to 39-57% by 2030, with seven additional states (Oregon, Iowa, Nebraska, Nevada, Wyoming, Arizona, and Indiana) crossing the 20% threshold [S5]. The state-level picture matters for grid planners because concentrated loads distort local capacity factors, transmission planning, and rate design in ways the national 5-7% figure does not capture.
AI Workloads as the Primary Growth Driver
AI workloads account for an average of 60% of electricity demand in modern data centers, though this share varies substantially by facility type, with training-focused hyperscale sites running far higher than inference-only edge sites [S4]. Hyperscale operators now routinely specify rack densities above 100 kW to accommodate GPU clusters, a step-change from the 8-15 kW densities typical of pre-2023 enterprise builds [S9].
The Belfer Center analysis describes a "watershed" moment in which electricity consumption growth rates are increasing across the United States, driven in part by a boom in hyperscale data center construction [S9]. Goldman Sachs concurs: the projected doubling of U.S. data center power demand by 2027 is concentrated in a small number of multi-gigawatt campuses, including the 10 GW PORTS Technology Campus in Piketon, Ohio and the 10 GW Creekstone Energy Delta Gigasite in Utah, both under construction as of 2026 [S3][S6].
Comparison: DOE, EPRI, Goldman, and IEA Projections
The major 2026 reference-case projections differ on timeframe more than on direction. DOE's 2024 baseline report (Dec 2024) projects 6.7-12% of U.S. electricity by 2028 [S1]. EPRI's February 2026 update narrows to 9-17% by 2030, then 10-20% by 2035 [S5]. Goldman Sachs focuses on peak demand specifically, projecting 5.3% of summer peak in 2026 and 8.5% in 2027 [S6]. The IEA's 10% annual baseline growth rate, applied to 2024's 192 TWh, implies roughly 232 TWh (5.5-6%) in 2026 [S3][S4].
On absolute TWh, the 2026 consensus cluster is 220-260 TWh depending on which growth curve is applied. The dispersion reflects different assumptions about hyperscale build-out pace, AI training-versus-inference mix, and grid-side response (behind-the-meter generation, demand response participation). The behind-the-meter gas build-out now under construction will cap grid-of-record growth in several of these scenarios.
What the 5-7% Share Means for Equipment Sourcing
The load profile shift forces facility engineers to re-spec switchgear, power meters, and UPS systems to handle the harmonic content and step-load transients characteristic of GPU cluster operation.
On the measurement side, modern data center power monitoring requires Class 0.2S revenue-grade power meters at the PDU level, with sub-second data logger sampling to capture transients from AI workload ramp cycles.
Limitations and Uncertainties in the 2026 Forecasts
None of the cited projections account for a full-scale grid-side response, including demand-response curtailment, behind-the-meter solar plus storage, or nuclear PPAs that could decouple data center growth from grid-of-record demand. The Goldman 5.3% summer-peak figure for 2026 assumes no major grid reliability events; EPRI's 9-17% by 2030 assumes continued Inflation Reduction Act policy support at reduced levels [S5][S6].
The Nature study's 10% annual baseline growth is explicitly labeled as a baseline, with AI-intensive scenarios diverging upward, meaning the 5-7% 2026 share is a central estimate, not a ceiling [S8]. A separate risk: 575 active restrictions across U.S. jurisdictions, including local moratoriums in Northern Virginia Loudoun County and Prince William County, could constrain growth in the highest-share state, pushing 2027-2028 expansion into Ohio, Texas, and Wyoming instead [S3].
Trackable signals for the next data point: the EPRA Q1 2027 data center load update (expected February 2027), the next Goldman peak-demand forecast revision, and the EIA Annual Energy Outlook 2027 release, which will reconcile the 192 TWh 2024 baseline with subsequent state-level filings.