Datacentres use 10% of US electricity

Where things stand (October 2026)

• Latest eligible source. Lawrence Berkeley National Laboratory's United States Data Center Energy Usage Report: 2025 Update came out on 18 June 2026, about six months late. RTO Insider reported on 14 June that it was "under wraps," and the lab would not say why. It puts data centres at 192 TWh, or 4.7% of US electricity, in 2024. It treats 2024 as the last historical year because shipment data ran only through Q3 2025, so its 2025 figure is partly a projection.

• Reference path. The report's reference case runs +22% in 2025 (about 234 TWh), +29% in 2026 (about 302 TWh), 464 TWh in 2028 and 649 TWh in 2030. It calls the 2030 figure 11.8% of US use (range 9.5–15.3%), but that share uses NERC's high forecast of total demand.

• The 10% line. EIA puts total US end-use consumption at 4,110 TWh in 2024 and 4,195 TWh in 2025 (Monthly Energy Review Table 7.6). Its October 2026 Short-Term Energy Outlook forecasts 4,288 TWh for 2026 and 4,356 TWh for 2027. I divided the lab's reference values by these EIA-style totals:

  • about 7.0% in 2026 and 8.6% in 2027;
  • about 10.3% in 2028;
  • about 12% in 2029.

  10% needs roughly 450–460 TWh of data-centre use in 2028–29.

• Other forecasters broadly agree. EPRI (2026) has 9–17% by 2030. BloombergNEF (July 2026) has 5.9% now and about 12% in 2030.

• Evidence for fast growth. About 66 GW of data-centre capacity is under construction in North America (JLL, mid-2026). Hyperscaler spending is about $700bn in 2026, and Goldman expects $1.2tn in 2027.

• Evidence for slower growth. NERC cut its summer-2026 load forecasts because large loads were connecting slowly. Only 50–60% of near-term data-centre capacity is expected on time (Congressional Research Service report R49326, September 2026). The most data-centre capacity ever connected to the US grid in one year is 7.1 GW. Commercial-plus-industrial power sales grew only about 50–70 TWh a year in 2025–26.

How a YES would come about

The publication lag is the main constraint. The lab's reports trail by about 12–18 months after a year ends. So 2028 data would first appear around December 2029 to late 2030, and 2029 data around December 2030 to late 2031. The 2026 and 2027 shares are very likely below 10%.

There are two main routes:

1. Further lab updates. These are not required by law: 42 U.S.C. 17112(e) mandated only one update.
2. EIA's planned mandatory national data-centre survey. EIA ran pilots in March–September 2026 and told Senators Warren and Hawley it would go nationwide. No Federal Register notice has appeared yet, and EIA's figures could come in lower than the lab's model.

DOE statements and other national labs are minor additional routes.

Model and result

I ran a Monte Carlo model in Python. It draws the data-centre growth path (centred near the lab's reference, with a 18–22% chance of an AI-spending bust), whether the lab keeps publishing and how late, when EIA's survey starts and how its figures compare with the lab's, and occasional DOE statements.

• Across variants, the probability of YES by the end of 2031 is 0.44–0.57.
• In the model's central case, about 41% of runs reach 10% by 2028 and about 74% by 2029.
• There is a qualifying publication covering 2028 or later by the end of 2031 in about 75% of runs, and one covering 2029 or later in about 66%.

The cumulative series is essentially zero until 2029, about 8% by the end of 2029, about 27% by the end of 2030, and about 48% by the end of 2031.

Strongest case against my estimate

• For higher: spending and the build pipeline point to growth faster than the lab's reference. If 10% arrives in 2027–28, there is more time for publication.
• For lower: the lab's update program is not mandated and has already been delayed for political reasons. EIA's survey may slip or meet legal challenges, as its 2024 crypto-mining survey did. Recent upward revisions have mostly been in forecasts, while the lab's own historical figures were revised down.

These roughly offset, so I land near 0.48.

Sources

• Lab 2025 Update, 18 June 2026: eta.lbl.gov/publications/united-states-data-center-energy-2025; osti.gov/biblio/3374245
• E&E News and RTO Insider, 14 and 22 June 2026
• EIA pilot-survey press release, 25 March 2026: eia.gov/pressroom/releases/press585.php
• WIRED, 15 April 2026
• EIA Today in Energy, 19 May 2026 (id=67704)
• EIA Monthly Energy Review T07.06 CSV and Short-Term Energy Outlook PDF, October 2026
• Congressional Research Service report R49326, 1 September 2026
• JLL North America midyear 2026
• BloombergNEF via Yahoo, 22 July 2026
• EPRI Powering Intelligence 2026
• Bloomberg/Goldman, 25 September 2026

Question. By the last day of each quarter, has the EIA, the Department of Energy, or a DOE national laboratory reported that datacentres used 10% or more of US electricity for a completed calendar year?

Definitions.

• Datacentres: US facilities and dedicated spaces housing computing, storage, and networking equipment, including hyperscale, colocation, enterprise, and government datacentres. Use the reporting body's national datacentre category. Count total facility electricity use, including cooling and other supporting infrastructure, not just AI workloads. No attribution to AI is required.
• US electricity: total annual US electricity consumption across all sectors, using the national accounting scope of the qualifying source. A share of commercial-sector consumption, a state's consumption, peak demand, or generating capacity is not a national consumption share.
• Calendar year: January 1–December 31 of a year completed before publication. The Observation period convention applies; the first eligible year is 2026.
• Reported share: a retrospective measured or model-estimated share of annual electricity consumption. A preliminary estimate counts. Use an explicitly reported percentage; if none is provided, a ratio calculated from national datacentre consumption and total national electricity consumption in the same qualifying publication or accompanying dataset counts, provided both figures cover the same year and accounting scope. Do not combine incompatible series or independently select a denominator from another publication.

Resolves YES if an eligible publication reports a qualifying share of at least 10%. A designated central, reference, or best estimate at or above 10% counts even if its uncertainty interval extends below 10%. If only an interval is reported, with no designated central estimate, its lower bound must be at least 10%. Apply the threshold to the most precise reported value available in that publication and its accompanying data; do not round a value below 10% upward. An explicit statement of “10%,” “approximately 10%,” or “at least 10%” counts if no more precise value below 10% is provided.

Resolves NO / Exclusions.

• No qualifying publication by the quarter-end.
• Forecasts, scenarios, annualized partial-year figures, or estimates for a year not yet completed, even when published in an official report. A later publication must expressly report the completed year's consumption; an earlier forecast does not become historical evidence merely through the passage of time.
• An upper-bound estimate above 10% without a qualifying central estimate or lower bound; statements such as “could reach 10%” or “up to 10%.”
• Statistics for electricity generation rather than consumption, installed or requested power capacity, or a regional or sector-only denominator.
• A contractor's, industry body's, or other outside party's estimate merely linked, hosted, or quoted without adoption as a finding by an eligible reporting body. DOE funding alone does not qualify a report.

Resolving source. Official reports, statistical releases, datasets, or institutional web statements issued or coauthored by the US Energy Information Administration (EIA), the US Department of Energy (DOE), or a DOE national laboratory. Eligible laboratories are those listed in DOE's National Laboratories directory at publication, including Lawrence Berkeley National Laboratory (LBNL). Concrete qualifying publication types include LBNL's 2024 United States Data Center Energy Usage Report, United States Data Center Energy Usage Report: 2025 Update, and successors, and EIA's statistical publications or Today in Energy analyses when reporting historical national consumption. Their prospective scenarios do not qualify. This is a narrower source rule than Public confirmation (A2); third-party reporting alone cannot resolve the question. Apply Resolution date (A3), the Resolution window, and the Observation period conventions. A qualifying revision can establish YES on its publication date; a subsequent downward revision does not reverse an established YES.