Energy and AI in East Asia

Overview

The report extends the 2025 Energy and AI study with three aims: to examine how AI can be applied in the energy system, to analyse how AI energy consumption affects grid planning and operation, and to offer policy recommendations. East Asia is central both to the global energy transition and to the AI industry: about 90% of the world's ten largest semiconductor companies are headquartered there (Chinese Taipei 73%, China 10%, Korea 8%) and almost six in ten chips are made in the region. Data centres use 1.5% of global electricity, projected to reach about 945 TWh (about 3%) by 2030.

Regional context: Korea aims for 29.2% renewable electricity by 2030 and Japan has a similar 2030 target, with Japan's 2024 variable-renewable share nearly twice Korea's. China met its 1,200 GW renewable target for 2030 in 2024, six years early. Lack of cross-border interconnection limits grid flexibility; Korea behaves like an island or quasi-island and needs transmission from renewables-rich Jeolla to the Seoul Metropolitan Area, which has the lowest electricity self-sufficiency.

Chapter 2 (Energy for AI) lists three determinants of data-centre energy use: PUE, coolants and current conversion. Cooling ranges from 7% to 30% of total electricity; high-density racks produce higher-temperature waste heat, and waste-heat use for local heating can be considered at early siting. In the 2024 Uptime Institute survey about 30% of data centres ran at 20 kW per rack or higher; densities above 50 kW remain a minority but have grown for six years. Air cooling can become impractical and less efficient above some density; liquid cooling (cold plate, immersion) can be combined with rear-door heat exchangers but is harder to deploy in existing facilities.

On PUE: the annual average fell from 2.5 in 2007 to 1.41 in 2024 but has plateaued since 2020. Hyperscale data centres were at 1.14 in 2024 and are expected at 1.09 by 2035; Google reported a trailing twelve-month fleet PUE of 1.09 in 2024. In the Base Case global weighted PUE falls to 1.29 in 2030, saving about 90 TWh. Asia Pacific's hot and humid climate raises cooling needs and makes free cooling less feasible. On regulation, China was first to adopt a nationwide efficiency standard (GB 40879-2021) and set a PUE of 1.5 target for 2025; the Netherlands targets 1.16, Singapore 1.3 and Japan 1.4, while the EU and Germany set no specific PUE value.

On demand and grids: in the Base Case Asia Pacific excluding China uses about 100 TWh in 2030 (11% of the global total) and Europe about 115 TWh (12%). China used 25% of the global 415 TWh in 2024 and reaches about 280 TWh, almost 30%, in 2030. At least 150 GW of data-centre projects are in advanced stages of grid-connection queues globally, some possibly placeholders. About 60% of Korea's data centres are in the Seoul Metropolitan Area, 14% in Chungcheong and only 8% in Jeolla, which has the highest renewables share. Major Korean transmission projects have been delayed about 70 months (nearly six years) on average over the past decade, mainly by local opposition.

Key points

1. Data centres use 1.5% of global electricity and are projected to reach about 945 TWh (about 3%) by 2030.
2. About 90% of the ten largest semiconductor companies are headquartered in East Asia and almost 60% of chips are made there.
3. Three determinants of AI data-centre energy use are PUE, coolants and current conversion; cooling ranges from 7% to 30% of total electricity.
4. In 2024 about 30% of surveyed data centres ran at 20 kW per rack or higher; densities over 50 kW are a minority but growing.
5. Annual average PUE fell from 2.5 (2007) to 1.41 (2024) but has plateaued since 2020; the Base Case reaches 1.29 in 2030, saving about 90 TWh.
6. Hyperscale PUE was 1.14 in 2024 and is expected at 1.09 by 2035; Google reported a fleet trailing-twelve-month PUE of 1.09.
7. China targets an average PUE of 1.5 for 2025; the Netherlands 1.16, Singapore 1.3, Japan 1.4; the EU and Germany set no specific PUE.
8. Asia Pacific excluding China uses about 100 TWh in 2030 (11%), China about 280 TWh (almost 30%) and Europe about 115 TWh.
9. About 60% of Korea's data centres are in the Seoul Metropolitan Area; Jeolla, with the highest renewables share, hosts about 8%.
10. At least 150 GW of data-centre projects are in advanced grid-connection queues globally; major Korean transmission projects were delayed about 70 months on average.

Using it in design

• In a site module for Korea and Japan, flag grid-congested metro regions and suggest renewable-rich regions.
• Include waste-heat reuse (local heating) at the early siting stage.
• Use national PUE targets in country profiles: China 1.5 (2025), Japan 1.4, Singapore 1.3, Netherlands 1.16.
• Limit free-cooling assumptions in hot, humid climates; use a higher cooling share and PUE for Asia Pacific sites.
• Compare the practical limit of air cooling with liquid options (cold plate, immersion, rear-door heat exchanger) for high-density racks.
• Add the risk of transmission delay (about 70 months on average in Korea) to the grid-connection schedule.

Who should read it

Investors, engineers and policy makers planning data-centre siting, power supply and grid connection in East Asia (Korea, Japan, China, Chinese Taipei).

Document contents

1. Introduction
2. What is AI?
3. AI in East Asia
4. How are energy and AI connected?
5. Regional context and challenges
6. Chapter 1. AI for energy
7. AI for renewable energy
8. AI for the power grid
9. AI for end users
10. AI for energy innovation
11. Chapter 2. Energy for AI
12. Energy use in data centres
13. Outlook for electricity consumption for AI in East Asia and beyond
14. Providing low-emissions electricity for data centres
15. Connecting data centres to the grid
16. Chapter 3. High-level policy recommendations

Read the document

Publication details

Authors

International Energy Agency; Korea Energy Economics Institute

Publisher

International Energy Agency

Year

2026

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