How Much Power and Water Does an AI Data Center Really Use?

A typical AI-focused data center uses as much electricity as 100,000 households, according to the International Energy Agency (IEA). The agency said the largest sites under construction in 2025 were expected to use about 20 times that. Water is harder to pin down. Some facilities use about as much as an office building and others far more. Most of the water tied to data centers is consumed at the power plants that supply them, not on site.

How much electricity do data centers use?

Worldwide, data centers consumed about 415 terawatt-hours (TWh) of electricity in 2024, the IEA estimated in its 2025 Energy and AI report. The agency put that at roughly 1.5% of global consumption. A terawatt-hour is one billion kilowatt-hours. The United States accounted for about 45% of the total.

The agency's April 2026 update put 2025 consumption at 485 TWh, up 17% from 2024. It put use by AI-focused facilities up about 50%. It projects about 950 TWh by 2030.

For the United States, Lawrence Berkeley National Laboratory estimated in June 2026 that data centers used 192 TWh in 2024. The lab put that at 4.7% of national electricity consumption. The lab's December 2024 report to Congress found that usage had held at about 60 TWh a year from 2014 to 2016. It tied the later increase to servers built around graphics processing units (GPUs), the chips used for AI.

Why do estimates of AI energy consumption vary so much?

Forecasts diverge because they use different methods and assumptions.

• Berkeley Lab projects 649 TWh for U.S. data centers in 2030 in its reference case, or 11.8% of national electricity use. The lab gives a range of 521 to 843 TWh (9.5% to 15.3%). The spread reflects uncertainty about chip shipments, how long AI chips stay in service and how intensively they run.
• The Electric Power Research Institute said in February 2026 that data centers could consume 9% to 17% of U.S. electricity by 2030. Its scenarios are based on operating, under-construction and announced projects.
• The IEA, in its 2025 report, projected about 1,200 TWh worldwide in 2035 in its base case. Its alternative cases run from about 700 TWh to about 1,700 TWh. Berkeley Lab notes that the IEA's 2025 projection for the United States in 2030, about 430 TWh, is well below its own.

Utility forecasts are disputed too. The consulting firm Grid Strategies looked at the data center demand that utilities reported to federal regulators. It found in November 2025 that this demand ran about 40% above credible industry projections, Utility Dive reported.

Do data centers use a lot of water?

Data centers consume water directly, for cooling, and indirectly, through the power plants that generate their electricity.

Berkeley Lab's 2024 report estimated direct consumption by U.S. data centers at 66 billion liters in 2023. That was up from 21.2 billion liters in 2014. The indirect footprint was much larger: nearly 800 billion liters in 2023, roughly 12 times the on-site figure. The largest facilities are known as hyperscale data centers. For 2028, the lab projected that they would consume 60 billion to 124 billion liters directly.

A peer-reviewed Nature Sustainability study from November 2025 looked at AI servers alone. It estimated they could generate an annual U.S. water footprint of 731 million to 1,125 million cubic meters between 2024 and 2030. A cubic meter is 1,000 liters. The study estimated that about 71% of that footprint would be indirect. Better siting, cleaner grids and efficiency gains could cut it by as much as 86%, the authors calculated. They called that best case extremely challenging.

Individual sites differ widely. A December 2024 review came from Virginia's Joint Legislative Audit and Review Commission. It found that most data centers in the state use about as much water as an average large office building or less. It also found that a few use substantially more and that some use less than a typical household. A 2025 paper by Berkeley Lab researchers calculated that the water used per computing workload can vary more than 10,000-fold.

Disclosure is limited. A 2021 article in the journal npj Clean Water said fewer than a third of data center operators tracked any water metrics. It was citing a 2016 Uptime Institute publication.

How much power and water does one AI query use?

Google said in August 2025 that a median text prompt in its Gemini apps used 0.24 watt-hours of electricity. It said the prompt also used about 0.26 milliliters of water, roughly five drops. Those are the company's own measurements from May 2025. They cover text prompts only and describe inference, meaning the use of an already trained model.

Small per-query numbers and large totals can both be accurate. The IEA said in April 2026 that the power needed per AI task is falling quickly while total consumption keeps rising. It gave two reasons: more people are using AI, and energy-intensive uses such as AI agents are growing.

What to watch

Three things bear on which estimates hold up. One is whether actual consumption tracks forecasts that depend on chip shipments and usage patterns still in flux. Another is disclosure. Berkeley Lab says greater transparency about data center energy use would help future research. The last is cooling design, a variable in the lab's 2024 scenarios.