What AI Data Centers Use Electricity and Water For

What AI Data Centers Use Electricity and Water For

AI data centers use electricity to run computing equipment and every system that keeps that equipment powered, connected, and cool. Water may be used directly to reject heat, indirectly to produce electricity, or not at all inside the cooling system. The amount depends on the workload, hardware, utilization, climate, facility design, power source, and reporting boundary.

That is the short answer. It is also why one gallons-per-day figure or energy multiplier cannot describe every AI service or facility.

Start with electrical work becoming heat

An AI model runs as mathematical operations on processors. Those processors receive electricity and produce useful computation, but nearly all of the electrical energy eventually becomes heat inside the building. Storage and networking equipment do the same.

The facility has to move that heat away from the chips and then out of the building. It also has to transform incoming utility power into the voltage and form the equipment can use, keep service stable during disturbances, detect fires, control access, and operate supporting systems.

flowchart LR
    A["Grid or on-site supply"] --> B["Transformers, switchgear, and power conversion"]
    B --> C["Servers, accelerators, storage, and networks"]
    C --> D["Heat collected by air or liquid"]
    D --> E["Pumps, fans, chillers, towers, or dry coolers"]
    E --> F["Heat rejected outdoors"]
    E --> G["Possible direct water use"]
    A --> H["Possible upstream water use for electricity"]

This flow exposes two different questions. How much energy reaches the information technology equipment? How much additional energy and water does the facility need to deliver and cool that computing work?

The computing load

Servers contain processors, memory, storage, power supplies, and network interfaces. AI systems often add accelerator hardware such as graphics processing units or other specialized chips. These devices can perform many operations in parallel, but a dense rack also concentrates power and heat.

Model training adjusts a model's parameters across many examples. Inference uses a trained model to produce an output. Both can use accelerators. Their resource demands vary with model size, architecture, precision, sequence length, batch size, hardware generation, memory movement, utilization, and the number of jobs or requests.

This makes a universal comparison with one web search unreliable. A small classification request, a long generative response, an image, a video, and an agent running many tool calls are different workloads. So are an early training experiment, a full training run, and repeated post-training work.

The 2024 Lawrence Berkeley National Laboratory report modeled equipment types and workloads across the US fleet. It estimated that GPU-accelerated AI servers used more than 40 terawatt-hours in 2023. That value belongs to the report's fleet model and year. It should not be divided by an assumed number of prompts to manufacture a universal per-query number.

Power delivery is not free

Electricity does not move from the grid directly into a processor. Transformers change voltage. Switchgear controls and protects circuits. Uninterruptible power systems bridge disturbances. Power distribution units and server power supplies convert electricity again. Batteries, generators, or other backup systems support continuity.

Each stage can lose some energy as heat. Redundancy also matters. A facility designed to keep operating through equipment failure may maintain spare paths or equipment that changes part-load efficiency.

The Department of Energy's data-center design guide uses power usage effectiveness, or PUE, to express the relationship between annual total facility energy and annual IT-equipment energy. A PUE of 1.2 means the facility used 1.2 units of total energy for each unit delivered to IT equipment during the stated period.

PUE is useful for locating overhead. It does not measure whether the computing work was valuable, whether total demand fell, or whether the electricity was low carbon.

Cooling moves heat through several boundaries

The first cooling boundary is at the chip or server. Air cooling moves heat into room air. Direct-to-chip liquid systems move heat into a liquid loop near the processor. Immersion places equipment in a dielectric fluid. Each approach still needs a final way to reject heat outdoors.

The next boundary may include computer-room air handlers, pumps, chilled-water loops, heat exchangers, chillers, cooling towers, dry coolers, or a hybrid. The right design depends on rack density, climate, humidity, water availability, electricity prices, reliability needs, operating temperature, and heat-reuse opportunities.

The DOE cooling-water guidance describes cooling towers and the relationship among evaporation, blowdown, water treatment, and cycles of concentration. That path is only one architecture.

A cooling tower can reduce the electricity needed for heat rejection under some conditions by evaporating water. A dry cooler can reduce direct water consumption while using more electricity or requiring more equipment under hot conditions. A hybrid can change modes. A liquid loop can improve heat collection at dense racks without deciding whether the final outdoor heat rejection uses water.

[[Data Center Cooling Options and Their Water Tradeoffs]] compares those choices in depth. The important lesson here is that "liquid cooled" does not automatically mean high water use, and "water free" does not automatically mean low total environmental impact.

Direct water use

Direct facility water can serve evaporative cooling, humidification, equipment maintenance, sanitation, fire systems, or other operations. Cooling often dominates where evaporative heat rejection is used.

The 2024 LBNL report estimated 66 billion liters of direct site water use by US data centers in 2023. It modeled a national average site water usage effectiveness just above 0.36 liters per kilowatt-hour of IT energy through 2023. The report also modeled change as hyperscale capacity and cooling choices evolved.

Those are national modeled values. They do not establish a specific campus's daily water use.

The physical source matters. A facility might receive potable water from a public utility, reclaimed water from a wastewater system, surface water, or groundwater under applicable rights and permits. The EPA describes water reuse as treating and repurposing wastewater, stormwater, or another source for a beneficial use. Reclaimed water can preserve potable supply, but it still needs treatment, pipes, pumps, contracts, backup, and a local capacity assessment.

Withdrawal is not the same as consumption

Water withdrawal is the amount removed from a ground or surface source. Consumptive use is the portion not returned to the immediate local water environment. The US Geological Survey includes evaporation, incorporation into products, and other locally unavailable water in consumptive use.

A facility can withdraw or receive more water than it consumes if part of the flow is discharged after use. An evaporative system can consume a significant portion of its makeup water. A reclaimed-water system can reduce new potable demand without eliminating consumption.

These distinctions matter because communities manage different constraints. A drinking-water system cares about treatment and peak delivery. A watershed manager cares about source withdrawals, return flow, temperature, drought, and ecology. A wastewater utility cares about discharge volume and quality. A resident may care about private wells and local aquifers.

One label called "water use" cannot answer all of them.

Source water behind electricity

Some power plants withdraw and consume water. A data center with low on-site water consumption can therefore have an indirect water footprint through the electricity it purchases.

Source WUE expands the boundary beyond facility water to include water associated with energy production under a stated method. The result depends on the generation mix, location, time, allocation method, and data source.

This is not a reason to combine direct and indirect water into one number without explanation. Direct site water affects the local water and sewer relationship. Source water may occur in another watershed. The two quantities answer different questions.

Why two facilities differ

The table is why a company average cannot be assigned to a proposed site. It is also why a design value should not be presented as measured performance.

National scale does not predict a site

LBNL's 2025 Update estimates 649 terawatt-hours of US data-center electricity use in its 2030 reference case, with a 521 to 843 terawatt-hour uncertainty range. The model is valuable because it shows how equipment shipments, utilization, cooling, and facility location can produce a major national load.

It does not tell a local utility when a campus reaches its requested demand. It does not tell a water authority which cooling system the operator will run on the hottest day. It does not identify who pays for a new substation.

Those answers need the project's phase plan, interconnection and service records, water study, permits, and operating measurements.

Read the boundary before the headline

When a data-center resource claim appears, ask what physical system it includes. Identify whether the value describes IT equipment, the facility, the campus, electricity supply, water supply, a company portfolio, or a product. Then identify the period, unit, denominator, utilization, geography, water type, and measured or modeled status.

[[How to Read Data Center Energy and Water Claims]] provides that complete method. [[The Hidden Local Costs Behind AI Data Centers]] explains how these physical flows enter public systems. For the design response, continue to E076 and [[Can AI Data Centers Be Sustainable]].

This guide was freshly written from the preserved E075 transcript and authoritative technical sources reviewed on July 27, 2026. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.