Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption (2024 White Paper)

Overview

This white paper was published by EPRI (Electric Power Research Institute) in May 2024 (document no. 3002028905). It projects US data-centre electricity use to 2030 under four growth scenarios and examines the role of AI. The report uses four 2023-2030 scenarios: low (3.7%/yr), moderate (5%), high (10%) and higher (15%) load growth. The scenarios are meant to reflect the uncertainty of AI-driven growth rather than to predict one outcome.

By the report's estimates, US data centres used about 4% of total load in 2023; the 2030 share is 4.6%, 5.0%, 6.8% or 9.1% depending on scenario (with other loads growing 1%/yr). About 80% of data-centre load was concentrated in fifteen states, led by Virginia, Texas and California. AI is estimated at 10-20% of data-centre electricity at the time of writing; a ChatGPT request is estimated at 2.9 Wh against about 0.3 Wh for a traditional Google search (about 10x).

The report also reviews efficiency levers. Average US PUE is said to have stabilised around 1.6 (2007-2023 trend). Table 4 lists emerging cooling technologies with vendor-reported TRLs; immersion cooling claims 50-95% cooling energy savings versus air. It also sets out a roadmap for collaboration between the grid and data centres. As a limit, the projections are scenario-based and vendor claims may be optimistic.

Key points

1. Four 2023-2030 scenarios: low (3.7%/yr), moderate (5%), high (10%), higher (15%) load growth.
2. US data centres used about 4% of total load in 2023; 2030 share 4.6%, 5.0%, 6.8% or 9.1% depending on scenario.
3. Fifteen states held about 80% of national data-centre load in 2023, led by Virginia, Texas and California.
4. A ChatGPT request is estimated at 2.9 Wh vs about 0.3 Wh for a traditional Google search (about 10x).
5. AI is estimated at 10-20% of data-centre electricity at the time of writing.
6. Average US PUE has stabilised around 1.6 (2007-2023 trend).
7. Table 4: immersion cooling claims 50-95% cooling energy savings vs air (vendor-reported).

Using it in design

• When a user picks a US state, show regional data-centre load share and growth scenarios as context for utility interconnection lead times.
• Use the average PUE of about 1.6 as the 'typical existing' benchmark line in PUE charts.
• Treat vendor-claimed cooling savings (Table 4) as optimistic upper bounds.
• Present the scenario range as a band rather than a single-point forecast.

Who should read it

Planners doing data-centre load forecasting and site selection, and decision makers concerned with energy policy and grid interconnection.

Related standards

• ISO/IEC 30134-2

Publication details

Authors

EPRI

Publisher

EPRI (Electric Power Research Institute)

Year

2024

Document number

3002028905

Document language

English

Source type

Expert / community

Access

Free to read

Licence

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

Citation