Energy, water and data centers
Can AI become more efficient while its total footprint grows? · updated 2026-09-07
Yes. A cheaper query can encourage more use, and video or long-running agents need different amounts of work from a short text answer. Global electricity totals and local water stress also measure different problems. Ask which workload, in which place, with which energy supply.
The concern
Concentrated demand can strain grid connections and local resources even with a modest global share. Longer reasoning, video and agent workloads can outweigh efficiency gains. A national average cannot tell a community who pays for infrastructure or whether its watershed can support a site.
The counter-view
Better hardware and software lower energy per task. New clean supply, storage and flexible scheduling can help; electricity-price effects depend on spare capacity and tariff design. These are options to assess locally, not a guarantee that growth is harmless.
The 2026 IEA outlook projects roughly doubled data-center electricity demand by 2030. Local outcomes remain conditional on siting, workload growth and energy-system choices.