International Energy Agency (IEA)·Thomas Spencer, Siddharth Singh, Davide D'Ambrosio, Yasmine Arsalane, James Bragg, Julie Dallard, Inhoi Heo, Vincent Jacamon, Martin Kueppers, Brieuc Nerincx
A special World Energy Outlook report by the International Energy Agency examining the relationship between AI development and energy. It analyses growth in data-centre electricity consumption and capital expenditure, the physical constraints on scaling AI infrastructure, power-supply options, and implications for energy systems, prices and emissions. It also considers AI’s impact on energy and industrial efficiency, as well as demand scenarios through 2030.
Key Takeaways
The IEA expects data-centre electricity demand to increase from approximately 485 TWh in 2025 to 950 TWh in 2030. AI-focused data centres are the main driver, with their electricity consumption expected to triple over this period.
Constraints on the growth of AI data centres extend beyond electricity availability: grid connections, transformers, gas turbines, high-bandwidth memory, chips and access to capital are all critical.
The energy efficiency of models and accelerators is improving rapidly; however, the growing complexity of inference workloads—including reasoning, agentic, multimodal and video applications—could increase energy use per query many times over and offset some of these gains.
The efficiency of AI data centres is largely determined at the design stage: rack density, cooling architecture, readiness for liquid cooling, power-supply configuration and location are all costly to change later.
The rapid deployment of data centres, combined with overstated connection requests and uncertainty around peak load, creates a risk of mismatch between energy-system investment and actual demand. Better forecasting, queue management and cost allocation are needed.
Key Figures
Data center electricity consumption
485 TWh
IEA forecast for 2025.
Data center electricity consumption
950 TWh
IEA Base Case for 2030.
Data centers' share of global electricity-demand growth
just under 10 %
Share of total growth in global electricity demand through 2030.
Growth in data center electricity demand
17 %
Growth in global data center electricity demand in 2025.
Growth in electricity consumption by AI-focused data centers
50 %
Growth in 2025.
Increase in data center electricity demand
about 70 TWh
Increase between 2024 and 2025.
Growth in AI factory capacity
more than 3 times
Over the past 18 months.
CAPEX of the largest technology companies
more than 400 billion USD
Capital expenditure in 2025.
Announced CAPEX of leading hyperscalers and neo-clouds
715 billion USD
Announced amount for 2026; equivalent to 75% growth.
Cumulative investment in data centers
3,9 trillion USD
IEA estimate for 2026–2030.
Capital expenditure for advanced AI data centers
40 000–50 000 USD/kW
Including IT equipment.
Supported capacity of AI-ready servers
about 25 GW/year
The maximum that current high-bandwidth-memory production, according to analysts, can support through 2027.
Delayed or blocked data center projects in the United States
20 projects
In Q2 2025.
Investment in delayed or blocked US data center projects
98 billion USD
Estimate for 20 projects in Q2 2025.
GPU energy consumption for Medium LM text generation
0,05 Wh
Indicative amount per request; GPU consumption only.
GPU energy consumption for an agentic scenario with reasoning
50 Wh
Indicative amount per request; GPU consumption only.
Energy savings from documented AI use cases
more than 13 EJ
Potential savings by 2035 if deployment barriers are removed; equivalent to 3% of global final energy consumption.
Data center-related emissions
about 350 million tonnes
Forecast for 2035; around 2% of global power-sector emissions.
Practical Value for Data Center Owners
For a data-centre owner or investor, the report is useful as a macroeconomic and infrastructure benchmark for project validation: comparing load and CAPEX projections with the IEA baseline scenario, and assessing connection, equipment-supply and financing risks. For an AI facility concept, the key takeaway is to incorporate high power density, liquid-cooling readiness, flexible consumption and scenarios for on-site resources (BESS and gas generation) from the outset, as retrofitting these capabilities later is difficult. In discussions with the grid company, a realistic peak-load profile, phased connection and the terms of controllable or non-firm connection are important.
A Deloitte analysis of the growth in data center infrastructure for AI workloads and its implications for the U.S. power system. It uses DC Byte and Wood Mackenzie data, estimates of AI data center penetration, and projections through 2035 to assess capacity. The focus is on interconnection and generation shortfalls, the impact on utility capital expenditures, and potential sources of and approaches to accelerating power supply for large AI sites.
AI data centersdata center power supplygrid interconnectionbehind-the-meter generation