AI's physical footprint is becoming measurable: power, water and grid capacity move into view
The IEA sees rapid growth in AI-focused electricity demand while Europe moves toward comparable data-centre disclosure. Efficiency per task is improving, but scale and heavier uses can outweigh those gains.
The short version
- What happened
- Efficiency per simple AI task is improving quickly, but demand, reasoning, agents and video can drive total electricity use upward.
- Who it affects
- Communities hosting data centres may gain investment and jobs while facing questions about grid costs, land, noise and water use.
- What remains unproven
- The IEA states that some 2026 data are estimates and that future demand depends on fast-changing technical and economic factors.
Keep in mind
The IEA states that some 2026 data are estimates and that future demand depends on fast-changing technical and economic factors.
Facility disclosure does not by itself capture the full lifecycle impact of chip manufacture, construction or electricity generation.
Multi-source analysis · 3 linked sources. Primary analysed source date: 1 January 2026. Earliest linked supporting source: 21 September 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Efficiency is rising—and so is total demand
The International Energy Agency reports that energy use per individual AI task has fallen rapidly as software and hardware improve. A simple text query now uses far less electricity than many popular comparisons imply. But the total footprint is shaped by the number and type of tasks. Reasoning, agentic work and video generation can require hundreds or thousands of times more energy per query than simple text. More capable and cheaper services can also stimulate far more use.
The IEA says global electricity demand from data centres rose 17% in 2025, while consumption at AI-focused facilities rose 50%. Its central projection takes total data-centre electricity consumption from about 485 terawatt-hours in 2025 to roughly 950 TWh in 2030, around 3% of global electricity demand. These are modelled estimates, not a fixed future. Investment, chip supply, grid connections, prices and demand could all shift the trajectory.[1]
Why local impact can be larger than the global share
A few percentage points of global electricity can sound modest, but data centres are large, concentrated loads that can arrive faster than new transmission or generation. The IEA notes rising power density and pressure on equipment such as transformers, turbines, power electronics and high-bandwidth memory. Developers may reserve more grid capacity than they initially use, making local planning harder. Where supply is tight, the cost of new infrastructure can become a political dispute over who pays.
Water impact is also site-specific. Cooling technology, climate, local water stress and the electricity mix all matter. A facility in a cool region with abundant low-carbon power has a different footprint from one competing with homes or agriculture for water during drought. Global averages should not replace local disclosure and cumulative assessment.[1]
Europe moves toward comparable disclosure
Reuters reports that the European Commission has proposed disclosure and labelling requirements covering larger data centres, including facilities at or above 500 kilowatts. Reporting would address energy and water performance and provide context such as local water stress and the reuse of waste heat. The Commission's data-centre energy pages link the underlying directive, reporting repository and studies used to develop a common rating approach.
Disclosure is not a consumption limit. Its value lies in comparability. Local authorities can assess planning claims, enterprise buyers can include efficiency in cloud procurement and investors can examine whether a portfolio depends on constrained grids or water. Metrics will need careful definitions and independent checks; operators should not be able to improve a label by moving part of the workload or omitting upstream effects without explanation.[2][3]
AI can also support the energy system
The IEA also describes potential benefits. AI can monitor equipment, predict failures and help grids use existing capacity more effectively. It estimates that well-documented applications could deliver meaningful energy savings by 2035 if barriers including fragmented data, cybersecurity and skills are addressed. Data centres may install batteries that can support grids when incentives and operating rules align.
Those benefits should be evaluated separately from the power required to run models. A useful optimisation project does not automatically offset unrelated consumption elsewhere. Organisations can improve decisions by reporting both sides: energy used by the service and verified energy or emissions avoided through the application.[1]