U.S. Data Center Electricity Demand Forecasts: LBNL, EPRI and Grid Operators

Demand forecasts for data centers come in two units, and mixing them causes most of the confusion in public debate. Energy forecasts, in terawatt-hours (TWh) per year or as a share of national electricity use, describe how much electricity the sector consumes over a year. Peak forecasts, in megawatts or gigawatts, describe the highest coincident demand a grid must serve, which is what drives new generation, transmission lines and substations. National studies such as LBNL’s and EPRI’s are mostly energy forecasts.⁵² Grid operator and utility forecasts are mostly peak forecasts, built from the load requests customers have filed.⁶⁷

The two are linked by load factor. Data centers run close to flat around the clock, so a gigawatt of data center peak brings far more annual energy than a gigawatt of residential peak. That is why a few large campuses can move a state’s total electricity use, and why our guide to data center power requirements treats sustained load, not nameplate, as the planning number.

Forecasts also differ in what they count. LBNL estimates historical consumption from server shipment data and builds scenarios forward, because, as the lab notes, the sector lacks transparency and its current consumption itself has to be estimated.¹ EPRI builds its scenarios from state-level data on operating capacity, construction in progress and announced plans.² Utilities start from signed and pending customer requests. Each method answers a different question, and none of them is a count of buildings that will exist.

The most cited U.S. figure comes from the 2024 United States Data Center Energy Usage Report, prepared by Lawrence Berkeley National Laboratory for the Department of Energy and published in December 2024.⁵ It reconstructs data center electricity use back to 2014 and offers a range of scenarios to 2028.⁵

The report found that data centers used about 58 TWh in 2014 and about 176 TWh in 2023, or roughly 4.4% of total U.S. electricity.¹ Demand more than doubled between 2017 and 2023, largely because of growth in AI servers.¹ For 2028, LBNL projects 325–580 TWh, which DOE described as about 6.7–12% of U.S. electricity, with the share depending in part on how fast the rest of the economy’s demand grows.⁸ DOE summarized the finding as data center load having tripled over the past decade and being on track to double or triple again by 2028.⁸

The width of that range is the point: the gap between the low and high scenarios is larger than the sector’s entire 2023 consumption.¹ For site work, the LBNL range is best read as a national envelope, not a forecast for any particular utility.

The Electric Power Research Institute has published two rounds of its Powering Intelligence analysis. The May 2024 edition used four growth scenarios and put data centers at about 4.6% to 9.1% of U.S. electricity generation by 2030.⁹ The February 2026 update, Powering Intelligence 2026, raised that to 9–17% by 2030, compared with roughly 4–5% today, and EPRI said the new figures are about 60% higher than its 2024 projections.²

EPRI’s 2026 scenarios are built from project stages. The low case assumes most projects under construction, plus a quarter of those in advanced planning, are running by 2030. The high case assumes everything under construction or in advanced planning is running, plus 30% of early-stage projects.² That structure is useful for land and development decisions because it makes the assumption explicit: forecast load is only as real as the project pipeline behind it.

EPRI’s state results matter more for siting than the national total. The 2026 update projects that data centers could reach 39–57% of Virginia’s electricity use by 2030, and that seven more states could exceed 20%.² States where data centers become a large share of load are the ones where regulators are most likely to tighten large-load tariffs and collateral rules.

For international context, the International Energy Agency estimates that the United States accounted for 45% of global data center electricity use in 2024 and that U.S. data centers will account for nearly half of U.S. electricity demand growth through 2030.¹⁰

Grid operators forecast peak demand for planning and capacity procurement, and since 2023 their forecasts have been rewritten around data centers. Their numbers are the ones that become transmission projects and capacity prices, which is why they matter to anyone holding land in a growth corridor.

PJM

PJM’s 2026 Long-Term Load Forecast, released in January 2026, projects summer peak growth averaging about 3.6% a year over ten years and 2.4% over twenty, reaching about 222,000 MW in 2036 (a ten-year increase of nearly 66,000 MW) and more than 253,000 MW in 2046.⁶ For comparison, PJM’s 2021 forecast had ten-year growth of about 0.3% a year.⁶ The 2026 edition also lowered near-term peaks, including the 2028 summer peak used in the capacity auction by about 4,414 MW, partly because of better vetting of data center and large-load adjustments.⁶ Our guide to PJM capacity prices explains how those forecasts feed auction results.

ERCOT

ERCOT’s preliminary 2026–2032 forecast, filed in April 2026, projects about 367,790 MW of demand by 2032, against an all-time peak of 85,508 MW set in August 2023.⁷ The total layers base growth together with medium and large loads that transmission and distribution providers submitted to ERCOT, and ERCOT’s CEO said the operator believes the forecast is higher than expected future load growth.⁷ ERCOT’s April 2026 monthly report put the large-load interconnection queue at about 445.8 GW through 2033, of which only about 5.9 GW had been energized and roughly 321 GW had no studies submitted yet.¹¹ ERCOT is moving large-load studies into coordinated batches rather than a one-at-a-time process.¹¹ By June 2026, ERCOT told the Texas Senate, the queue had grown to about 474 GW, roughly 90% of it data centers, with about 205 GW eligible for its first “Batch Zero” study.¹² See our ERCOT site selection guide and our explainer on Texas Senate Bill 6.

Grid Strategies aggregates utility filings into a national five-year view. Its November 2025 report put five-year peak load growth at 166 GW, a 3.7% annual rate, up from a 24 GW five-year forecast three years earlier, and identified data centers as the largest driver, at about 55% of forecast demand growth.³ The same report argued that the data center portion of utility forecasts is likely overstated by roughly 25 GW, based on market analysts’ estimates of what can actually be supplied and built.³

The Energy Information Administration’s Short-Term Energy Outlook is more conservative because it only looks about two years ahead. In its July 2026 outlook, EIA projected U.S. consumption rising from a record 4,195 billion kWh in 2025 to 4,269 billion kWh in 2026 and 4,399 billion kWh in 2027, with commercial-sector sales (the sector that includes most data centers) expected to pass residential sales for the first time.¹³

Every forecaster above discounts raw requests, and for good reason. Developers can approach several utilities with the same project while they look for the fastest path to power, so raw request totals can count one project more than once. Utilities have said it is very difficult to know in advance which requests will proceed, and excess requests tie up study staff, delay other customers and distort long-range resource planning.⁴ Forecast bias is not new: a 2018 LBNL study of Western utilities found most had overestimated demand growth in the mid-2000s.⁴

Utilities are responding with standardized large-load processes, larger financial commitments and requests for state policy help.⁴ Those changes are why near-term forecasts, such as PJM’s 2026 edition, have started to come down even as long-term forecasts rise.⁶ The utility load study and interconnection queue guides describe what each stage requires from a developer.

National forecasts tell you demand is real. They do not tell you whether a specific parcel can be served. For a site decision, the useful questions are local:

• What does the serving utility’s most recent integrated resource plan or load forecast assume for large loads, and how much of it is backed by signed agreements?
• Is the area’s transmission plan adding capacity near the site, or is new load being studied against an already-full system?
• Has the state or utility adopted a large-load tariff, collateral rule or study-batch process that changes how requests are prioritized?
• Is the region one where data centers are becoming a large share of total load, which tends to bring tighter rules and more scrutiny of ratepayer impacts?

Forecasts that rise faster than the grid can be built usually mean longer energization dates and more pressure on flexible load arrangements, not necessarily fewer projects. Our methodology page describes how power is screened at the parcel level. If you want a parcel checked against its utility’s current plans, get a site reviewed.