💧 How Much Water Does an AI Data Center Use? A Worked Example

There is no single water-use number for an AI data center

Water use depends on IT energy, cooling design, weather, operating schedule and the accounting boundary. A campus capacity announcement does not tell us its actual water consumption. A closed server loop does not tell us how the facility finally rejects heat. A meaningful estimate states those assumptions before giving a number.

We will calculate an illustrative 10 MW IT facility, compare two assumed water intensities and trace where water moves. Then use the Chilled Water Loop Lab and Cooling Tower Operation Lab to explore the mechanisms. The article and calculations are free; the labs are included with All-Access.

First decide what “uses water” means

• Withdrawal: water taken from a supply or source. Some may return after use.
• Consumption: the portion not returned to that source for immediate reuse, such as water evaporated during cooling. The reporting definition and geographic boundary matter.
• Circulation: water moving repeatedly through a loop. It is not automatically new withdrawal or consumption on every pass.
• Indirect water: water associated with electricity production or another upstream activity. It is outside a site-only figure unless explicitly included.

Compare like with like: annual site makeup water and regional total water consumption are different quantities. A reported volume also needs a period, source and boundary. Reclaimed-water supply, potable-water supply and evaporation should not all be collapsed into one unlabeled number.

Trace the heat beyond the server

IT equipment → room air or cold plate → facility heat transfer
             → final heat rejection → outdoor environment

Example water-cooled chiller path:
IT → chilled-water loop → chiller → condenser-water loop → cooling tower
                                                       ↑ makeup
                                                       ↓ blowdown
                                                       ↑ evaporation

Follow the final heat-rejection equipment before deciding whether evaporation occurs. A closed chilled-water or cold-plate loop may transfer heat to a separate evaporative system. A dry cooler or air-cooled condenser uses a different rejection method; hybrid or adiabatic operation can change with conditions.

The DOE Federal Energy Management Program’s data-center guidance describes the heat path through a chilled-water system, chiller and cooling tower. It also defines site water usage effectiveness as annual site water volume divided by annual IT energy. The numerator’s definition must be consistent when comparing facilities.

Worked example: 10 MW of constant IT load

These are invented sensitivity cases, not measured facilities, industry averages or predictions. Assume IT load—not total facility load—is 10 MW continuously for a non-leap year. Assume the stated WUE represents the same annual site-water boundary in each case.

IT power = 10 MW = 10,000 kW
Annual IT energy = 10,000 kW × 8,760 h = 87,600,000 kWh
Site water volume = annual IT energy × assumed site WUE

The arithmetic is straightforward; selecting a defensible WUE is the hard part. The daily figures above are annual averages, not peak-day demands. Actual load varies and hot-weather water demand can differ substantially from an annual average. Do not use these values to size a service connection or establish a real project’s environmental impact.

In this example, an intensity reduction of 1.0 L/kWh changes annual site water by 87.6 million liters at the assumed constant load. Halving actual IT energy while keeping the same intensity would halve this estimate. Neither assumption establishes how a specific facility will perform.

Where PUE belongs—and where it does not

If the hypothetical PUE is 1.3, total facility energy is 87.6 million × 1.3 = 113.88 million kWh/year. But the WUE definition used above has IT energy in its denominator. Multiplying WUE by total facility energy would inflate the estimated water volume by 30% in this example.

PUE alone cannot tell us WUE. Two designs can have similar electrical efficiency and different site-water use. For a combined electricity-and-water footprint, document the electricity source and an appropriate upstream water factor separately; do not silently add it to site WUE.

Circulating flow is not makeup water

Consider a separate illustrative loop carrying 1 MW of heat with a 10°C supply-to-return temperature rise. Taking water’s heat capacity as 4.18 kJ/(kg·°C):

Mass flow = 1,000 kW ÷ (4.18 kJ/(kg·°C) × 10°C)
          ≈ 23.9 kg/s ≈ 23.9 L/s
Hourly circulating volume ≈ 86,100 L/h (assuming 1 kg/L)

That volume passes a point in the loop repeatedly; it is not a claim that 86,100 liters of new water disappear every hour. Makeup replaces losses, not each circulation pass. Real coolant mixtures have different properties; the water approximation is only a learning calculation.

Read the companion chilled-water flow-rate explanation to connect load, flow and temperature difference.

Cooling-tower example: makeup and blowdown

For a simplified steady-state exercise, assume evaporation is 100 m³/day and ignore drift, leaks and overflow. If cycles of concentration are 4, the dissolved-solids balance gives blowdown approximately E ÷ (cycles − 1): 100 ÷ 3 = 33.3 m³/day. Makeup is evaporation plus blowdown: 133.3 m³/day.

At 6 cycles, with the same assumed evaporation, blowdown becomes 20 m³/day and makeup 120 m³/day. The modeled reduction is 13.3 m³/day. This is not permission to increase an operating tower’s concentration target: source-water chemistry, treatment, equipment limits and qualified water-management review govern the acceptable range.

DOE’s cooling-tower management guidance explains the makeup balance and the need to manage dissolved solids. In a real balance, include drift and other measured losses. Blowdown is a discharge; do not automatically label every liter of makeup as evaporative consumption.

Try two lab experiments

1. Chilled Water Loop Lab: hold heat load constant and change water flow. Observe supply/return temperature, heat removal and server inlet temperature. This demonstrates heat transport; it is not a complete site-water or WUE model.
2. Cooling Tower Operation Lab: record a baseline, change the concentration-cycles setting, then compare evaporation, makeup and blowdown-related behavior. Repeat at a different wet-bulb temperature. Its simplified model has stated assumptions; do not expect the independent 100 m³/day example above to appear as a preset.
3. Record each setting and result. Explain which change affects circulation, which affects water losses and which affects electrical input. Avoid treating one metric as a substitute for all three.

For the rest of the system, explore the Data Center Power Path Lab. It demonstrates the electrical supply path, not water consumption. Our liquid-versus-air cooling explanation covers heat capture; the final heat rejection still needs separate attention.

Questions to ask about a real proposal

Ask for measured or modeled IT utilization, annual and peak-day water demand, supply source, discharge assumptions, cooling mode by season, the precise WUE boundary and any upstream electricity-related water estimate. Ask whether a quoted megawatt figure is IT load, utility demand or a future buildout capacity. A meter-based operating record and a proposed full-build scenario should be labeled differently.

The DOE’s data-center siting FAQ notes that water needs vary with facility size and cooling technology. Use project-specific evidence rather than an unsupported “liters per AI prompt” conversion. Workload energy, allocation and cooling conditions differ too much for the assumptions here to produce such a universal number.