AI Data Center Water Budget: Annual Use, Peak Demand and Storage

A water budget for an AI data center has to answer two different questions: how much water the site may use over a year, and how fast the site may need water during the hottest operating condition. Water Usage Effectiveness (WUE) can help with the first question. It cannot size the service pipe, treatment plant or storage tank.

For an evaporative system, the calculation starts with heat rejected outdoors, not a generic number of liters per prompt. Evaporation is derived from that heat. Make-up water then depends on the allowable cycles of concentration. Peak demand needs a separate design-condition calculation using the coincident IT load and weather.

Budgeting rule: calculate annual withdrawal, consumption and discharge separately; then calculate peak hourly demand, design-day volume and interruption storage. A single WUE figure cannot replace those quantities.

TL;DR: AI data center water budgets

• Separate withdrawal, consumption, discharge, initial fill and indirect water; they answer different engineering and reporting questions.
• Calculate evaporative loss from rejected heat, then add blowdown and drift to obtain cooling-tower make-up.
• Use annual load, climate and operating modes for yearly volume. Use coincident design load and weather for the peak connection.
• WUE is an annual intensity metric. It cannot size the water service, treatment equipment or storage tank.
• Record every assumption—heat-rejection multiplier, latent fraction, cycles of concentration and autonomy—so the budget can be checked and updated.

Start by naming the water quantity

Several water figures are routinely called “water use,” although they describe different physical flows. A utility connection responds to withdrawal. A watershed assessment may focus on consumption. A sewer authority cares about discharge. The distinctions need to survive every table and calculation.

The site boundary should also state whether sanitary water, irrigation, humidification and fire-system testing are included. At a dry-cooled facility, those uses can exceed routine process-water top-up.

For broader context on direct, indirect and per-prompt estimates, see our AI water-use guide. This article stays with the engineering budget: annual volume, peak rate and storage.

The cooling-tower water balance

The U.S. Department of Energy states the basic mass balance as:

Make-up = Evaporation + Blowdown + Drift

The same DOE cooling-tower guidance defines cycles of concentration as the ratio between dissolved solids in the recirculating water and the make-up water. At steady state, with blowdown and drift combined as the non-evaporative loss term, the relationship can be written as:

M = E × C / (C − 1)

where M is make-up water, E is evaporation and C is cycles of concentration. Once evaporation is known, this equation converts it into the withdrawal that the site must receive. It does not divide the remaining volume between blowdown and drift; that split needs the selected tower's drift rating.

Cycles are not a free efficiency setting. The feasible value depends on the actual make-up-water chemistry, treatment regime, materials and discharge consent. DOE notes that many systems operate at two to four cycles, while six or more may be possible. Four cycles is therefore used as the base in the worked example below, with other values shown as sensitivities.

Calculate evaporation from rejected heat

A cooling tower removes heat through latent transfer, which evaporates water, and sensible transfer, which warms the air. The first-principles relationship is:

E [m³/h] = flatent × Qrejected [kW] × 0.001481

The constant uses a latent heat of vaporization of 2,430 kJ/kg at approximately 30°C. At a latent fraction of 1.0, one kilowatt of rejected heat can evaporate about 1.481 liters per hour. That is the physical ceiling per kilowatt of heat rejected.

The boundary around Qrejected matters. It may include IT heat, non-IT heat entering the loop, pump work and chiller compressor work. It is not automatically equal to IT load. A useful concept-stage relationship is:

Qrejected = kHR × PIT

The heat-rejection multiplier kHR must come from the project heat balance. An economizer condition might be close to 1.05; a compressor operating condition can be higher. Neither number should be copied into a final design without the plant model.

Worked annual water budget for a 20 MW IT design

This example demonstrates the method. It is not a prediction for a real site or a ModulEdge product. The climate hours are illustrative placeholders that must be replaced with an hourly weather file and certified equipment performance.

Annual IT energy is 131,400 MWh. The mode-weighted heat rejected is 150,930 MWh_th. Applying the annual latent fraction gives 190,060 m³ of evaporation. At four cycles:

• Annual consumption from evaporation: 190,060 m³
• Annual make-up withdrawal: 253,413 m³
• Annual non-evaporative loss: approximately 63,353 m³ across blowdown and drift; the project model must split them using the selected tower's drift rating
• Withdrawal-based WUE for this stated boundary: 1.929 L/kWh of IT energy

The decimal places make the calculation auditable; they do not make it a site forecast. Weather bins, plant curves, water chemistry, IT utilization and controls will change the result.

Why cycles change withdrawal but not evaporation

Raising cycles reduces the combined non-evaporative loss required by this simplified balance. Most of that volume is normally blowdown and a much smaller part is drift, but the exact split depends on the selected tower. Changing cycles does not remove the heat that the tower must reject, so evaporation stays the same.

A higher cycle count can be useful, but only if the water analysis, treatment design and discharge route support it. Otherwise, the spreadsheet has optimized a number that the plant cannot operate at.

Why the same WUE can produce a different water bill

WUE is annual site water divided by annual IT energy. If load and water rise in proportion, WUE stays unchanged while the absolute volume changes. Using the same assumptions as the worked example:

This is why WUE is useful for intensity reporting but insufficient for a water-supply enquiry. A utility needs cubic metres per year, peak cubic metres per hour and the daily profile. For the energy-side equivalent, see our PUE guide: an annual ratio does not by itself size peak infrastructure.

Peak demand and storage require a separate calculation

Peak cooling-tower make-up should use design IT load, the rejected-heat multiplier at the design condition, a latent fraction at or near 1.0, and the cycles supported by the water chemistry. It should also use the wet-bulb condition coincident with the expected IT load, rather than combining unrelated annual extremes.

In the illustrative hot-climate case behind this model, annual-average make-up at four cycles is 29.70 m³/h. The design condition is 20 MW IT load, kHR 1.20 and a peak latent fraction of 1.0:

• Peak make-up: 47.41 m³/h, or 13.17 L/s
• 24-hour volume: 1,138 m³
• 72-hour volume: 3,413 m³

The 24-hour and 72-hour figures assume the design condition persists throughout the whole period. They are deliberate worst cases, not tank recommendations. A real diurnal weather and load profile may reduce them materially.

The ratio between peak and annual average is about 1.60. A connection sized at the average would therefore be 37% short; stated the other way, the design-hour requirement is 60% above the average. The same relationship applies to the design-day volume in this example. A tank sized from an annual-average day would cover only about 15 hours of a 24-hour worst-case interruption.

Climate matters differently for wet and hybrid plants

For a cooling tower that operates wet throughout the year, IT energy and cycles dominate annual withdrawal. In the four-climate illustrative model used for this analysis, climate moved annual tower water by about 9% through the economizer and compressor-hour split. Climate is a smaller annual variable in that case, but it is not zero.

For an adiabatically assisted dry cooler, water use switches on only when ambient conditions cross the control threshold. Climate can then dominate annual water. A location label such as “Middle East” or “Northern Europe” is not enough; the budget needs the site's hourly dry-bulb and wet-bulb data, equipment limits and operating sequence.

The cooling architecture also changes the result more than the choice between plausible cycle counts. Dry rejection can remove recurring process evaporation, while hybrid operation can reduce annual water but still require a substantial peak connection. Our liquid-cooling design guide explains the system architecture, while Does Liquid Cooling Use Water? separates closed-loop heat capture from dry, adiabatic and evaporative rejection. The extreme-heat cooling guide covers hot-weather operating choices.

What a project water budget should contain

A concept-stage water budget should be reproducible by another engineer. At minimum, record:

1. IT basis: rack schedule, design and average IT load, utilization profile and liquid-versus-air heat split.
2. Thermal boundary: every heat source entering the rejection plant, including compressor and pump work where applicable.
3. Climate basis: station, hourly or bin data, design dry-bulb and wet-bulb, and coincidence with IT load.
4. Water basis: source, laboratory chemistry, treatment, cycles of concentration, drift rating and discharge route.
5. Operating modes: dry, adiabatic, economizer and compressor hours, with the control thresholds that switch between them.
6. Outputs: withdrawal, consumption and discharge; annual volume; peak hourly and daily make-up; interruption storage; initial fill; and separately identified non-process water.
7. Evidence class: clearly label simulations, design targets, contractual guarantees and measured results.

Where EU reporting applies, Commission Delegated Regulation (EU) 2024/1364 defines WUE as total water input divided by IT energy. The regulation is a reporting framework, not a cooling-system sizing method. Our EU data center rating-scheme guide separates binding reporting duties from the draft labels. Use the engineering budget for pipes, tanks, treatment and operating resilience.

How this applies to ModulEdge projects

ModulEdge develops the water and cooling concept from the selected IT equipment, rack schedule, site climate and required operating states. The aim is not to attach one WUE value to every location. It is to define the heat path and show what happens in normal, peak, maintenance and water-interruption conditions.

The MDC-2000 reference architecture uses warm-water direct liquid cooling for the principal AI heat load, with a separate path for residual air cooling. Its stated outdoor strategy is dry-first heat rejection with tank-fed adiabatic assistance when ambient conditions require it.

Two ModulEdge water figures need to be kept separate. The product reference's 100 m³ is a water-storage equivalent, not annual consumption. Separately, ModulEdge's current planning estimate for one roughly 2 MW IT-load configuration in temperate European or US conditions is about 100 m³ of cooling water per year. The annual figure is not a platform-wide guarantee or a substitute for an hourly water model. Very hot climates, such as Dubai, can require more adiabatic operation and higher water use.

Actual annual use, peak adiabatic rate, refill interval and required autonomy remain site-specific calculations. A project should not size its utility connection or storage tank from the 100 m³/year estimate alone.

For a project review, bring the rack schedule, site coordinates, water source and chemistry, discharge route, interruption requirement and availability objective. Those inputs allow the concept to compare annual volume, peak demand and failure behavior on the same basis.

Frequently asked questions

How much water does a 1 MW AI data center use?

There is no defensible universal figure. A 1 MW label does not state annual utilization, rejected heat, cooling method, climate, cycles of concentration or whether the number means withdrawal or consumption. Scale a documented heat-and-water balance, not a per-prompt estimate or an industry average.

What is the difference between WUE and annual water use?

WUE is an intensity: annual site water divided by annual IT energy. Annual water use is an absolute volume. Two facilities can report the same WUE but withdraw different volumes because their IT energy differs.

Can WUE size a cooling-water connection?

No. WUE does not contain a design weather condition, peak IT load, plant heat-rejection multiplier, latent fraction or required interruption duration. Calculate peak hourly make-up, design-day volume and storage separately.

Does a closed liquid-cooling loop eliminate water consumption?

Not by itself. A closed technology loop can feed a cooling tower, dry cooler, chiller or hybrid plant. Routine process-water consumption depends mainly on the outdoor heat-rejection method and its controls.

How do you calculate peak cooling-water demand?

Use the coincident design IT load, total heat entering the rejection plant, design weather, latent fraction and the selected cycles of concentration. Convert the resulting make-up rate into hourly and design-day volume. Do not divide annual water use by 8,760 and use that average to size the connection.

How large should a data center water-storage tank be?

Storage follows the peak net draw and the required interruption duration, after crediting any reliable refill that remains available during the event. It must also distinguish process-water autonomy from initial loop fill and fire-water duties. A 24- or 72-hour calculation is a scenario until the project defines the governing outage and operating response.

Does ModulEdge's 100 m³ figure mean storage or annual water use?

Both quantities currently use the same rounded number, but they are different claims. The MDC-2000 product reference uses 100 m³ as a water-storage equivalent. A newer planning estimate is about 100 m³/year of cooling water for one roughly 2 MW IT configuration in temperate European or US conditions. The annual estimate rises in hotter climates and must be replaced by project-specific simulation before design or contractual use.