Energy and AI (World Energy Outlook Special Report)

Overview

The report starts from the premise that there is no AI without energy, and that AI also has the potential to transform the energy sector. Observing that policy makers and markets lack tools to assess the stakes, the IEA builds a new global model and dataset of data-centre electricity demand, enriched by consultation with policy makers, the technology sector and the energy industry. A typical AI-focused data centre uses as much electricity as 100,000 households, and the largest under construction will use 20 times that. Global investment in data centres has nearly doubled since 2022 to half a trillion dollars in 2024.

The current picture: data centres used about 1.5% of world electricity in 2024, or 415 TWh, with the United States at 45%, China 25% and Europe 15%. Consumption has grown about 12% a year since 2017, more than four times faster than total electricity consumption, and nearly half of US capacity sits in five regional clusters. By component, servers take about 60% of electricity, storage about 5% and networking up to 5%; cooling's share ranges from about 7% in efficient hyperscale to over 30% in less efficient enterprise facilities. PUE runs from about 2 in enterprise sites to just under 1.15 in hyperscale.

On server efficiency, idle power fell from about 60% of rated power in 2010 to about 35%, while utilisation is below 20% in enterprise sites and up to 50% in hyperscale. Accelerated servers were 24% of server electricity and 15% of total data-centre demand in 2024. In the Base Case installed data-centre capacity rises from about 100 GW today to about 225 GW in 2030; servers with eight accelerators can exceed 10 kW rated power, against under 2 kW for two-accelerator servers. A fall in global weighted-average PUE from 1.41 to 1.29 saves about 90 TWh.

Projections: the Base Case reaches about 945 TWh by 2030, slightly above Japan's current consumption, and about 1,200 TWh by 2035. Accelerated-server consumption grows about 30% a year against 9% for conventional servers, and in the United States data centres account for nearly half of demand growth to 2030. Uncertainty is handled through sensitivity cases, Lift-Off, Headwinds and High Efficiency (20% below the Base Case in 2035), giving a 2035 range of 700-1,700 TWh. Renewables meet half of the growth (over 450 TWh to 2035), natural gas adds 175 TWh and nuclear a similar amount.

On grids and impact, around 20% of planned data-centre projects could face delay unless risks are addressed: new transmission takes four to eight years in advanced economies, transformer and cable waits have doubled in three years and gas-turbine deliveries take years. In the United States 50% of data centres under development are in existing large clusters. Mitigation options are siting where grid capacity exists and operating servers or on-site generation and storage flexibly. Emissions from data-centre electricity rise from 180 Mt today to 300 Mt in 2035 in the Base Case (500 Mt in Lift-Off), staying below 1.5% of energy-sector emissions. The report also covers AI for energy optimisation and innovation.

Key points

1. Data centres used about 1.5% of world electricity (415 TWh) in 2024: United States 45%, China 25%, Europe 15%.
2. Consumption has grown about 12% a year since 2017; the Base Case reaches about 945 TWh in 2030 and about 1,200 TWh in 2035.
3. Servers use about 60% of electricity, storage about 5%, networking up to 5%; cooling ranges from about 7% (efficient hyperscale) to over 30% (less efficient enterprise).
4. PUE runs from about 2 (enterprise) to just under 1.15 (hyperscale); global average PUE falls from 1.41 to 1.29 in the Base Case, saving about 90 TWh.
5. Server idle power fell from about 60% of rated power (2010) to about 35%; utilisation is below 20% in enterprise and up to 50% in hyperscale.
6. Accelerated servers were 24% of server electricity and 15% of total demand in 2024; their consumption grows about 30% a year in the Base Case.
7. Installed capacity rises from about 100 GW to about 225 GW by 2030; eight-accelerator servers can exceed 10 kW rated power.
8. The 2035 range across cases is 700-1,700 TWh; the High Efficiency Case is 20% below the Base Case.
9. About 20% of planned data-centre projects could face grid-related delay; new transmission takes four to eight years in advanced economies.
10. Data-centre electricity emissions rise from 180 Mt to 300 Mt in 2035 in the Base Case (500 Mt in Lift-Off), below 1.5% of energy-sector emissions.

Using it in design

• Use the IEA component split (servers about 60%, storage about 5%, network up to 5%, cooling 7-30%) to sanity-check modelled energy breakdowns.
• Use IEA PUE ranges (1.15 hyperscale to about 2 enterprise) as benchmark bands in the energy model.
• Include an idle fraction (about 35% of rated power) and utilisation inputs by facility type (enterprise under 20%, hyperscale up to 50%) in the server load model.
• Reflect in rack power-density assumptions that eight-accelerator servers can exceed 10 kW rated power.
• Show grid connection lead times (transmission 4-8 years, transformer and turbine deliveries) as a schedule risk and compare siting and flexible-operation options.

Who should read it

Data-centre designers, investors, grid planners and energy-policy decision makers assessing data-centre electricity demand.

Document contents

1. Executive Summary
2. The rise of AI and its nexus with energy
3. What is AI?
4. Energy for AI and AI for energy
5. Energy for AI
6. Key definitions and concepts
7. Electricity consumption of data centres
8. Outlook for electricity consumption from data centres
9. Electricity supply to meet data centre demand
10. Data centre interactions with the electricity grid
11. AI for energy optimisation
12. AI for energy innovation
13. Emerging themes on energy and AI
14. Energy security in the age of AI
15. Implications for investment
16. An exploratory approach to determine the potential impact of AI on emissions
17. Annex A. Methodology and data tables

Related standards

• ISO/IEC 30134-2

Read the document

Publication details

Authors

International Energy Agency

Publisher

International Energy Agency

Year

2025

Document language

English

Source type

Government / public

Access

Open licence — downloadable

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.iea.org

Citation