Electricity: the global picture
The most-cited global source is the International Energy Agency. Its April 2026 update, Key Questions on Energy and AI, estimates that data centers used about 485 TWh of electricity in 2025, up from about 415 TWh in 2024, and projects about 950 TWh by 2030. That is around 3% of global electricity demand by 2030. Within the total, AI-focused data centers roughly triple.
Two things about that projection matter more than the headline. First, the IEA's central case barely moved from its 2025 report, even though announced spending went up sharply; it says bottlenecks across the value chain are making the more aggressive near-term scenarios less likely. Second, the United States is about half of global installed data center capacity, so the global number understates how concentrated the pressure is. Texas grid operator ERCOT alone has received interconnection requests for 160 GW of data center load, which the IEA notes is twice the existing global capacity of hyperscale and colocation data centers. Many of those requests are speculative and will never be built.
Electricity: the US picture
For the US, the reference is Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report, commissioned by the Department of Energy. It found US data centers used about 176 TWh in 2023, or 4.4% of total US electricity, and projects 325 to 580 TWh in 2028, or 6.7% to 12%. The width of that range is the honest part: it reflects real uncertainty about chip shipments, utilisation and efficiency, not sloppy modelling.
EPRI's May 2024 study, Powering Intelligence, reached a similar shape by a different route. Across four scenarios it projected data centers at 4.6% to 9.1% of US electricity by 2030. EPRI is funded by the utility industry, which is worth knowing, but its range sits inside LBNL's rather than above it.
If you remember one thing: the serious forecasts agree on direction and disagree on slope by roughly a factor of two. Anyone quoting a single number for 2030 without a range is quoting one scenario.
Why power is the constraint, not just a cost
For an AI data center, electricity is not mainly a line item. It is the thing that decides whether the building can open. A 1 GW campus needs a grid connection the size of a large power station, and interconnection queues, transformer lead times and turbine supply all run on multi-year cycles. The IEA expects 15 to 27 GW of on-site gas generation to power data centers by 2030, mostly in the US, and points out that reliable on-site generation needs 30% to 70% more capacity than the load. It also tracks a growing pipeline of small modular reactor offtake agreements, from about 25 GW at the end of 2024 to 45 GW at the end of 2025, with first projects not expected until around 2030. The reactor side of that story is in nuclear microreactors for AI data centers.
There is also a shape problem. Training jobs produce large, fast swings in power draw that a utility has to absorb. The IEA expects 20 to 25 GW of battery storage inside data centers globally by 2030, partly to smooth those swings.
Water: direct vs indirect
Water is where most public numbers go wrong, because they mix two different things.
- Direct (site) water is what the data center consumes on site, almost all of it evaporated in cooling towers or evaporative coolers. LBNL estimates US data centers consumed about 66 billion litres (roughly 17 billion US gallons) in 2023, with hyperscale and colocation facilities accounting for 84%. It projects hyperscale facilities alone at 60 to 124 billion litres in 2028.
- Indirect (source) water is what power plants consume to generate the data center's electricity. LBNL puts that at nearly 800 billion litres for US data centers in 2023, an average of 4.52 litres per kWh, based on the regional grid mix.
So for a typical US site, the power plant drinks more than ten times what the building does. That changes the levers. Moving from evaporative to closed-loop cooling cuts site water sharply; buying electricity from wind and solar cuts source water. A data center can look water-free on site and still have a large footprint through its grid.
LBNL also flags the trade-off that makes this hard: evaporative cooling is generally more energy-efficient than waterless air-cooled chillers. A site that eliminates water use often uses more electricity, which raises source water and emissions elsewhere. Lower water and lower energy do not automatically move together.
Water in an AI data center specifically
Current AI racks are designed for direct-to-chip liquid cooling, and that helps on water. Liquid carries heat out of the rack far more efficiently than air, and the coolant loop is closed. Whether the site then rejects that heat with dry coolers or with evaporative towers is a separate design choice, driven by climate and water price. Microsoft says its Fairwater Atlanta site uses a closed loop that "consumes almost zero water in its operations", with an initial fill equal to the annual use of about 20 homes. Treat operator statements like that as claims to check against the site's permits, but the engineering direction is real. The cooling options are compared in liquid cooling for AI data centers.
Per prompt: the number people ask about
The most detailed per-request figure from an operator is Google's August 2025 paper, Measuring the environmental impact of delivering AI at Google scale. It estimated the median Gemini Apps text prompt at about 0.24 Wh of energy and 0.26 mL of water, counting idle capacity and data center overhead, not only the accelerator. It also reported a 33-fold drop in energy per median prompt over one year. Two caveats: it is a median text prompt, not an image, a video or a long agentic coding session, and it is Google measuring Google. It is still more useful than estimates that divide a guessed total by a guessed number of queries.
What this means if you buy compute
- Ask for site-level figures. PUE, site WUE and the grid mix of the specific facility, not a corporate average.
- Treat power availability as a delivery risk. Capacity promised for a building that does not have its grid connection signed is a forecast.
- Count your own usage. If your board asks what your AI use costs in energy, start with your own token volume and your providers' disclosures. That conversation belongs with AI ROI.
Related
- AI infrastructure — the cluster hub
- Liquid cooling — the design choices behind site water use
- The largest AI data centers — and where their power comes from
- Nuclear microreactors — firm power for the 2030s
Frequently asked questions
How much water does an AI data center use?
It depends mostly on the cooling design and the climate. Evaporative cooling consumes water continuously; closed-loop liquid cooling consumes very little after the initial fill. Microsoft says the closed loop at its Fairwater Atlanta AI site needed an initial fill equal to what about 20 homes use in a year. Across all US data centers, LBNL estimated direct water consumption of about 66 billion litres in 2023, with hyperscale and colocation facilities accounting for 84%.
How much power does AI use?
The IEA estimates all data centers used about 485 TWh of electricity in 2025 and projects about 950 TWh by 2030, with AI-focused data centers growing fastest. At the level of a single request, Google reported in August 2025 that the median Gemini Apps text prompt used about 0.24 Wh, including idle capacity and data center overhead. Training runs and long agentic sessions use far more per task.
What share of US electricity do data centers use?
Lawrence Berkeley National Laboratory found US data centers used about 176 TWh in 2023, or 4.4% of total US electricity, and projects 6.7% to 12% by 2028 (about 325 to 580 TWh). EPRI's 2024 scenarios put data centers at 4.6% to 9.1% of US electricity by 2030.
Does AI use more water than it saves?
Nobody has a credible global answer, because the benefit side is not measured. What is measurable is the footprint, and it is local: a data center in a water-stressed basin with evaporative cooling is a real burden on that basin, while a closed-loop site on a grid powered by wind and solar uses little water at all. Ask about the specific site, not AI in general.
Why are data centers building their own power plants?
Because grid connections are slow. The IEA estimates that 15 to 27 GW of on-site natural gas may power data centers by 2030, mostly in the United States, and notes that reliable on-site generation has to be overbuilt by 30% to 70% relative to the load. With gas turbines in short supply, it is not clear this is always faster than waiting for the grid.
Sources: International Energy Agency, Key Questions on Energy and AI (April 2026); Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024); EPRI, Powering Intelligence (May 2024) via Utility Dive; Google (August 2025); Microsoft. Unit conversions are ours.