Skip to content
← Back to Engineering Case Studies

Drawing No. EH–TH–053 // Data Center Engineering

How Much Water Does a 100 MW AI Data Center Use?

Reviewed August 2026

There is no single correct water consumption or water-use number for a “100 MW AI data center.” The answer can change by more than an order of magnitude depending on PUE, climate, cooling architecture, the share of heat rejected evaporatively, water-treatment recovery and the reporting boundary. The IEA notes that hyperscale AI-focused data centers can reach 100 MW or more. This worked example deliberately defines its headline case as 100 MW average IT load, not 100 MW total facility demand; at the neutral PUE of 1.15, modeled facility power is 115 MW.

This is a worked engineering comparison, not a claim about one named facility. The baseline uses 100 MW of average IT load for 8,760 h/year. Cooling presets are illustrative EngineerHub assumptions so the influence of heat rejection can be seen transparently.

100 MW data center water use: direct answer

For the transparent screening assumptions used here, a data center with 100 MW average IT load and PUE 1.15 requires about 1.71 million m³/year of external site water with full wet evaporative heat rejection. A hybrid case with a 35% equivalent wet share uses about 0.60 million m³/year. A direct-liquid system with fully dry heat rejection has essentially no routine evaporative cooling-water demand in this model and retains only 1,825 m³/year of assumed ancillary site water.

On a per-MW basis, those illustrative cases are about 47.0 m³/day/MW IT, 16.5 m³/day/MW IT and 0.05 m³/day/MW IT, respectively. These are scenario outputs, not universal industry averages; climate, PUE, operating mode, water chemistry and reporting boundary can change the result materially.

Scale the same comparison

25–1,000 MW average IT
100 MW average IT

Water demand is approximately linear with heat load in this screening model. Real facilities are more complex because weather, operating mode and load vary through the year.

Core result

same IT load, different cooling
—

Illustrative annual external site water for full wet rejection versus direct liquid cooling with dry heat rejection. The low-water case still includes 5 m³/day of other site water.

Interpretation first: “Liquid cooling” alone does not determine water use. The final heat sink does.

Five cooling scenarios

same average IT load
ScenarioPUEWet heat shareSite WUEWater inputDaily / MW ITAnnual inputAnnual consumed

“Wet heat share” is an equivalent modeled share. A 35% hybrid value can represent a system that runs dry for much of the year and uses evaporation during hotter periods; it is not a universal annual operating fraction.

Annual water comparison

external site input

What the comparison shows

three engineering takeaways

FULL WET REJECTION—Evaporation dominates consumptive loss; blowdown and treatment recovery add to external makeup.
HYBRID / SEASONAL WET—Reducing the fraction of annual heat handled evaporatively reduces water roughly in proportion to wet-path duty.
LIQUID + DRY REJECTION—A closed direct-to-chip loop paired with dry heat rejection can eliminate routine evaporative cooling-water demand.

Published WUE reference: AWS 2025

not one facility

AWS reports a 2025 global data-center WUE of 0.12 L/kWh, explicitly describing this fleet metric as water withdrawn per kWh of IT load. Applied mechanically to the same average IT energy, that corresponds to:

Annual water at 0.12 L/kWh
—
Daily equivalent
—
Do not treat this as the water use of every AWS site or as a design target. It is a published fleet metric. Climate, operating mode, water source, measurement category and infrastructure design matter. The purpose here is to show how a reported WUE translates into an understandable annual volume. AWS source. Also note that withdrawal-based and consumption-based WUE values should not be compared without checking the reporting boundary.

Why the numbers diverge

energy–water architecture
GPU / AI RACKSdirect-to-chip coolingCLOSED LIQUID LOOPsame internal conceptbut two different final heat sinksWET TOWERevaporation + blowdownDRY COOLER / CHILLERheat to air · no routine evaporation

Method and assumptions

foldable

Baseline calculation method

Each scenario is run through the same shared engine as the Data Center Water Usage & WUE Calculator. The engine calculates total facility power from IT load × PUE, assigns a fraction of heat to the wet path, then calculates circulation, evaporation, drift, blowdown, treatment reject, site water input, consumption and WUE.

The case study uses 100% utilization because the headline quantity is defined as 100 MW average IT load, not 100 MW nameplate capacity.

Illustrative scenario assumptions
  • Full wet rejection: PUE 1.15; wet share 100%; 10°C range.
  • Direct liquid + wet tower: PUE 1.15; wet share 100%; 12°C range.
  • Hybrid: PUE 1.15; equivalent wet share 35%.
  • Mostly dry + evaporative assist: PUE 1.15; equivalent wet share 10%.
  • Direct liquid + dry rejection: PUE 1.15; wet share 0%.
  • Wet cases: latent heat share 85%, cycles of concentration 4, drift 0.005%, tower-makeup pretreatment recovery 100% (no pretreatment reject in the neutral comparison).
  • All cases: 5 m³/day other site water, 50% assumed consumptive.

These are transparent comparison assumptions, not claims about typical or best-practice PUE/WUE values. Holding PUE constant isolates the effect of the modeled heat-rejection architecture; use the calculator to test architecture-specific PUE values when defensible project data are available.

Why an annualized wet share is useful

Hybrid and evaporative-assist systems may operate dry during favorable weather and use water only during hotter conditions. A single steady design point cannot represent that annual behavior. The equivalent wet-share parameter lets a first-pass annual water estimate be made, while clearly signaling that an hourly weather model would be needed for site-specific design.

Site water vs electricity-related water

This case study focuses on site water. Electricity generation can have its own water withdrawal/consumption footprint, depending on the grid mix and generation technology. That upstream quantity is intentionally kept separate because adding a generic grid-water intensity can be more misleading than useful for a site-specific project.

Current technical context

why this topic is changing quickly

IEA: hyperscale AI-focused data centers can reach 100 MW or more, so the 100 MW case used here is a realistic scale marker rather than a claim about any specific facility. IEA.
ISO/IEC 30134-9:2022: WUE is a standardized data-center KPI; the published 2022 edition remains current while a second edition is under development. ISO.
DOE FEMP: WUE is a site-based water-efficiency metric normalized to IT energy; evaporative cooling systems link water use directly to data-center heat load and cooling-system efficiency. DOE.
Microsoft, June 2026: its cooling overview distinguishes year-round cooling towers, hybrid fluid coolers, direct air, air-cooled chillers and direct-to-chip liquid systems; Microsoft says newer AI designs can operate with zero water evaporation for cooling. Microsoft.
Google: describes water cooling as an energy-efficient heat-removal option and explicitly frames cooling as an energy-water tradeoff that should be evaluated by geography and water availability. Google Data Centers.
The Green Grid, 2025: Water Usage Impact (WUI) adds local water-stress context to raw water consumption, reinforcing that liters alone do not capture location-specific impact. The Green Grid.

Related EngineerHub resources

run your own assumptions

Data Center Water Usage & WUE Calculator→Power Plant Water Balance Calculator→Cooling Tower Calculator→Heat Exchanger Calculator→Home Solar + Battery Sizing Worked Example→