2024 United States Data Center Energy Usage Report (LBNL-2001637, December 2024)

Overview

The Energy Act of 2020 asked DOE to publish an update to LBNL's 2016 study. The result estimates electricity use back to 2014 and gives a scenario range to 2028. U.S. data centers stayed near 60 TWh over 2014-2016; GPU-accelerated servers then grew enough that use reached about 76 TWh (1.9% of U.S. electricity) in 2018 and 176 TWh (4.4%) in 2023. For 2028 the scenarios span roughly 325 to 580 TWh, equivalent to 74-132 GW at 50% average capacity utilization and 6.7-12.0% of forecast U.S. electricity. The report puts compound annual growth at about 7% for 2014-2018, 18% for 2018-2023 and 13-27% for 2023-2028.

The method is a bottom-up installed-base model, chosen over top-down and extrapolation approaches because it avoids over-counting projects that have not yet picked a power provider. It models servers, storage and network equipment, then multiplies IT energy by modelled PUE and site WUE for internal, colocation and hyperscale space types, cooling systems and climates. Server assumptions include rated powers for 8-GPU nodes of 3.4 kW (A2/A40/L4 class), 6.5 kW (A100 class), 10.2 kW (H100 class) and 12.2 kW (B100 class), AI servers idling at 20% of rated power (a measured 8x H100 node idled near 18%) and conventional servers idling at 36% of maximum in 2023, falling to 27% by 2028.

Results by component show server electricity rising from about 30 TWh in 2014 to nearly 100 TWh in 2023, while GPU-accelerated servers went from under 2 TWh in 2017 to over 40 TWh in 2023. Average U.S. data center PUE fell from 1.6 in 2014 to 1.4 in 2023 and is projected at 1.15-1.35 by 2028. Site WUE stayed just above 0.36 L/kWh through 2023 and is projected to rise to 0.45-0.48 L/kWh as cooling shifts. Direct water consumption rose from 21.2 billion liters in 2014 to 66 billion in 2023, 84% of it in hyperscale and colocation. Site WUE excludes the water used to generate the electricity supplied to the site.

The indirect footprint is large: for 2023 the report estimates nearly 800 billion liters of water consumed through electricity generation, about 4.52 L/kWh, and 61 billion kg of CO2 equivalent, about 0.34 kg/kWh, against national averages of 4.35 L/kWh and 0.35 kg/kWh. These use local grid mixes and ignore power purchase agreements and behind-the-meter generation. A separate chapter estimates cryptocurrency mining electricity. The conclusions stress that the lack of direct energy data limits the analysis, that the 2028 values are wide scenarios, and that further work is needed on benchmarking, utility and grid operator collaboration and technology development.

Key points

1. U.S. data centers used about 76 TWh (1.9% of U.S. electricity) in 2018 and 176 TWh (4.4%) in 2023.
2. 2028 scenarios range from about 325 to 580 TWh, equivalent to 74-132 GW at 50% average capacity utilization and 6.7-12.0% of forecast U.S. electricity.
3. Rated 8-GPU server powers used in the model: A100 class 6.5 kW, H100 class 10.2 kW, B100 class 12.2 kW, A2/A40/L4 class 3.4 kW.
4. AI servers are assumed to idle at 20% of rated power (a measured 8x H100 node idled near 18%); conventional servers idle at 36% of maximum in 2023, falling to 27% by 2028.
5. Average U.S. data center PUE fell from 1.6 in 2014 to 1.4 in 2023 and is projected at 1.15-1.35 by 2028.
6. Average site WUE was just over 0.36 L/kWh through 2023 and is projected to rise to 0.45-0.48 L/kWh afterward.
7. Server energy grew from about 30 TWh (2014) to nearly 100 TWh (2023); GPU-accelerated servers from under 2 TWh in 2017 to over 40 TWh in 2023.
8. Direct water consumption rose from 21.2 billion liters in 2014 to 66 billion liters in 2023, with 84% in hyperscale and colocation facilities.
9. Indirect water from electricity is nearly 800 billion liters for 2023 (about 4.52 L/kWh), with 61 billion kg CO2e (about 0.34 kg/kWh).
10. Compound annual growth was about 7% for 2014-2018 and 18% for 2018-2023, with 13-27% projected for 2023-2028.

Using it in design

• Take AI rack and node power assumptions (3.4 to 12.2 kW per 8-GPU server, 20% idle) as a cross-check on IT load models, not as a substitute for vendor data.
• Use national PUE (1.4 in 2023) and site WUE (about 0.36 L/kWh) as context when presenting a design's targets, and check which definition of WUE is used.
• Report both site WUE and indirect water of electricity when comparing evaporative against dry cooling, since the report shows the two can move in opposite directions.
• Use the 2028 range with its stated capacity-utilization assumption when discussing grid capacity and phasing with utilities.
• Treat the figures as model estimates, not measurements, and avoid quoting a single 2028 value.

Who should read it

Planners, utilities, policy makers and design engineers who need current U.S. benchmarks for data center electricity, PUE, water and AI server power.

Document contents

1. Executive Summary
2. 1. Introduction
3. 2. Methodology Overview
4. Historical Best Estimates and Future Scenarios
5. 3. IT Equipment Installed Base and Power Draw
6. Server Types
7. Server Power Draw
8. Server Population
9. Storage and Network Types and Power Draw
10. 4. Data Center Infrastructure
11. Data Center Classifications
12. Data Center Infrastructure Electricity and Water Modeling
13. 5. Total Data Center Electricity/Water/Carbon Estimates
14. Water and Emission Impacts
15. 6. Total Cryptocurrency Electricity Use
16. 7. Conclusions, Limitations, and Future Work
17. Beyond 2028

Publication details

Authors

Arman Shehabi; Sarah J. Smith; Alex Hubbard; Alex Newkirk; Nuoa Lei; Md Abu Bakar Siddik; Billie Holecek; Jonathan Koomey; Eric Masanet; Dale Sartor

Publisher

Lawrence Berkeley National Laboratory

Year

2024

Document language

English

Source type

Government / public

Access

Citation only

Licence

Licence conditions not established — confirm the conditions with the publisher

Licence evidence

datacenters.lbl.gov

Citation