Powering Artificial Intelligence: Strategies for Integrating Data Centres into Electricity Systems

Overview

The report takes stock of evolving practice in integrating data centres into electricity systems. The largest operators (hyperscalers) will spend more than USD 700 billion in 2026, and data-centre investment grew 17% in 2025 to exceed USD 550 billion, more than global oil and gas production investment. Electricity demand in advanced economies grew 2.1% in 2025, with data centres more than 20% of the increase. The report highlights public concern over electricity prices, emissions and water, and pressure on local grids. More than 34 jurisdictions are covered, about 90% of global data-centre capacity.

The policy inventory finds 134 policies and regulations guiding data-centre integration, nearly 40% announced or implemented in the last 12 months. The most common policy area is improving visibility of data-centre demand in scale, timing and location. In several jurisdictions connection queues exceed today's peak demand (Texas, Virginia, Finland, the United Kingdom), and they contain duplicate or speculative projects. Overestimating demand risks stranded grid investment, underestimating it risks bottlenecks and delay. Queue reform aims to reflect credible projects, prioritise using maturity and financial commitment, and speed up connection studies.

On connection models: because periods of system stress last relatively few hours a year, IEA analysis finds non-firm connections could enable an additional 750-900 GW of grid hosting capacity globally. Norway, South Australia, Spain and the Southwest Power Pool in the United States are testing conditional or curtailable access, time-dependent phased access, and dynamic consumption limits tied to real-time or forecast network conditions. Active management of compute load is rare; AI-optimised capacity can cost USD 20-30 billion per GW to rent for a year, about 100 times the equivalent battery capacity, making curtailment very expensive. A US event in which a data-centre load of over 3 GW abruptly disconnected and switched to backup has led some system operators to consider technical performance requirements.

On procurement and sustainability: operators signed 45 TWh of PPAs in the first half of 2026 (up nearly 20%), about 70% in the United States, with a small but growing number for advanced geothermal, SMRs, long-duration storage and virtual power plants. Ireland requires certain new data centres to build or contract new dispatchable capacity, while China and Spain have placed or are considering renewable purchase obligations. Data centres are associated with about 150 Mt of CO2 from electricity today, growing to about 350 Mt in 2035 in the central scenario (about 1.5% of global energy-related emissions). Nearly 40% of identified AI-dedicated capacity is in areas of high or extremely high water stress, and water efficiency trades off against energy efficiency.

On cost allocation and principles: the effect on bills is not clear-cut. A new large load needing network or generation investment recovered from all users can raise others' bills, while a load that fills existing capacity and pays its share can spread fixed costs and ease pressure. Large-load tariffs usually charge for reserved capacity rather than energy; minimum-payment commitments, long-term contracts matched to asset life, exit fees, collateral and phased connections address three risks: a load using less than it reserved, leaving before assets are paid off, and never connecting. The report offers six principles: transparency on location and consumption, predictable connection processes, holistic energy and water efficiency, coordinated fair tariffs, frameworks for flexibility and stability, and dialogue and experience sharing.

Key points

1. The survey covers more than 34 jurisdictions (about 90% of global data-centre capacity) and identifies 134 policies and regulations, nearly 40% in the last 12 months.
2. Hyperscalers will spend over USD 700 billion in 2026; data-centre investment grew 17% in 2025 to over USD 550 billion.
3. The most common policy area is visibility of data-centre demand in scale, timing and location; queues in some places exceed today's peak demand.
4. Queue reform: stricter entry requirements, maturity assessments, financial commitments and milestone-based capacity release.
5. Non-firm connections could enable an extra 750-900 GW of grid hosting capacity globally; models are curtailable access, phased access and dynamic consumption limits.
6. AI-optimised capacity can cost USD 20-30 billion per GW per year to rent, about 100 times the equivalent battery capacity, so compute curtailment is very costly.
7. A US disconnection of a data-centre load of over 3 GW to backup supply led operators to consider technical performance requirements.
8. BYOP is spreading: Ireland requires new dispatchable capacity; China and Spain have or consider renewable purchase obligations.
9. Operators signed 45 TWh of PPAs in H1 2026 (up nearly 20%), about 70% in the United States.
10. Data-centre CO2 from electricity rises from about 150 Mt to about 350 Mt in 2035 (about 1.5%); nearly 40% of AI capacity is in high water-stress areas.

Using it in design

• In the grid-connection module, support firm and non-firm (curtailable, phased, dynamic-limit) options and size the BESS or genset capacity needed to ride through curtailment hours.
• Capture jurisdiction BYOP rules (new dispatchable capacity, renewable purchase obligations) in the regulatory profile.
• Show in the project schedule that connection applications need maturity and financial-commitment evidence under queue reform.
• Add technical performance requirements to the electrical design checklist given the risk of abrupt large-load disconnection.
• In water-stressed regions assess the water versus energy efficiency trade-off together.
• Adjust the cost model for tariff terms (minimum payment, long-term contract, exit fee, collateral) by jurisdiction.

Who should read it

Policy makers, regulators, grid operators, developers and investors working on data-centre grid connection, tariffs and power-supply strategy.

Document contents

1. Executive summary
2. Introduction
3. Understanding where and when data centres will consume electricity
4. Load projections
5. Spatial planning
6. Enhancing the energy efficiency and sustainability of data centres
7. Where efficiency is measured
8. Reporting and facility performance
9. Long-term resilience of data centre infrastructure
10. Better management of grid connection requests by data centres
11. Permitting
12. Queue management and connection arrangements
13. Allocating costs related to grid and infrastructure upgrades for data centres
14. Securing electricity supply for data centres
15. Integrating data centres securely into the grid
16. Flexibility
17. Stability
18. Annex: Data centre and large-load policies by jurisdiction

Read the document

Publication details

Authors

International Energy Agency

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