United States Data Center Energy Usage Report: 2025 Update

Overview

This report is the update of the Lawrence Berkeley National Laboratory's (LBNL) 2024 estimate of US data center electricity use. Its authors include Sarah Josephine Smith, Alex Hubbard, Alexander Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale A. Sartor, Michael Mills and Arman Shehabi.

The method is bottom-up: it builds from IT equipment shipments, per-device energy use, cooling simulations, and facility types and locations. This structure shows that the estimate is derived from hardware- and facility-level assumptions and not from an aggregate statistic. The report also projects 2030 consumption and enriches it with sensitivity cases for AI chip shipments, chip lifetime, and AI server idle power and utilisation.

In the reference case, US data center electricity in 2030 is 649 TWh, about 11.8% of US electricity. The scenario range is 9.5%-15.3% of US electricity by 2030, compared with 6.7%-12.0% for 2028 in the 2024 report. Higher idle power and utilization for AI servers give 782 TWh, and lower installations give 578 TWh. The combined uncertainty range for 2030 is 521-843 TWh.

As a limit, the results are specific to the United States, and the 2030 projections depend on assumptions, such as AI hardware shipments, that carry large uncertainty; the wide range reflects this. The record holds only the figures in the abstract, with the detailed method and data tables in the full text of the report. The report gives sector-level context for load growth and grid demand narratives.

Key points

1. Reference case: 649 TWh of US data center electricity in 2030, about 11.8% of US electricity.
2. Scenario range 9.5%-15.3% of US electricity by 2030 (2024 report: 6.7%-12.0% for 2028).
3. Higher idle power and utilization for AI servers gives 782 TWh.
4. Lower installations give 578 TWh.
5. Combined uncertainty range for 2030 is 521-843 TWh.
6. Method is bottom-up: IT shipments, per-device energy, cooling simulations, facility types and locations.
7. Sensitivity cases: AI chip shipments, chip lifetime, AI server idle power and utilisation.

Using it in design

• Use the 2030 ranges as context in load-growth and grid-request narratives.
• The AI-server idle-power sensitivity supports modelling GPU idle draw explicitly rather than assuming proportional power.
• Label the figures as US-specific; do not apply them directly to estimates for other countries.

Who should read it

Designers, energy analysts and decision makers planning data center load growth and grid demand.

Related standards

• LBNL 2024 US Data Center Energy Usage Report

Publication details

Authors

Sarah Josephine Smith; Alex Hubbard; Alexander Newkirk; Mohan Ganeshalingam; Billie Holecek; Dale A. Sartor; Michael Mills; Arman Shehabi

Publisher

Lawrence Berkeley National Laboratory

Year

2026

Document language

English

Source type

Government / public

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation