Water demand in AI and high-density data centres: a Middle East & GCC strategy guide
Direct answer
Water demand in an AI or high-density data centre is determined primarily by the final heat-rejection method, not by an assumed number of litres per server. IT load, facility efficiency, climate, cooling mode, source-water chemistry, treatment recovery and operating schedule all influence the result.
Liquid cooling at the rack can improve heat capture and enable warmer facility-water temperatures, but it does not automatically increase or eliminate water consumption. Closed technology loops use relatively little routine make-up; significant demand arises when cooling towers, evaporative condensers or adiabatic systems reject heat through evaporation.
A credible forecast therefore begins with the hourly thermal load and traces every water flow through normal, peak, phased, maintenance and failure conditions. The output should include annual withdrawal, peak-day and peak-hour demand, consumptive use, discharge and storage refill requirements.
Start with heat rather than floor area
Water demand should be derived from the heat that must leave the facility. The starting point is measured or forecast IT power, adjusted for utilisation and the additional electrical loads that become heat inside the cooling boundary. Power usage effectiveness can support an early estimate, but detailed models should separate fans, pumps, lighting, electrical losses and other contributors.
Installed IT capacity is not the same as operating load. A 100 MW campus may energise in stages and run below nameplate during early years. Forecasts need an hourly or representative load shape, not only a final maximum.
The thermal path then determines which portion reaches each rejection system. Air-cooled halls, direct-to-chip racks and immersion zones may use different temperature tiers or plants. A single campus-average assumption can conceal a water-intensive area or overstate demand where dry rejection serves part of the load.
Heat-reuse commitments should be treated carefully. Exporting useful heat can reduce the duty reaching the primary rejection plant, but only when a receiving load is available at the same time and temperature. The base water forecast should not assume continuous export unless the offtake and backup arrangement are secured.
Separate closed-loop inventory from consumptive demand
Technology coolant loops, CDU circuits and closed facility-water systems require an initial fill and occasional make-up for sampling, maintenance or leakage. This inventory can be material during commissioning but is not normally a continuous cooling-water demand.
The forecast should schedule initial fills, flushing, cleaning and future-phase connections separately. High-purity fill water may need a polishing plant or temporary supply even when the annual volume is small. Peak fill rate should not be added automatically to simultaneous cooling make-up unless the programme requires it.
Fire-water storage is normally a separate protected inventory and should not be counted as cooling autonomy unless the applicable authority and fire strategy permit a shared arrangement. Test and refill flows may still affect the utility connection, so their timing belongs in the overall demand schedule.
Routine losses from a nominally closed system warrant investigation. A persistent make-up flow can indicate leaking valves, seals, vents or expansion arrangements. Metering closed-loop addition protects fluid quality and prevents abnormal use being hidden in the overall balance.
Convert cooling duty into evaporative demand
In an evaporative heat-rejection system, water changes phase to carry heat to the atmosphere. The theoretical evaporation depends on heat duty and the latent heat of vaporisation; actual operation also reflects equipment performance, ambient conditions and control sequence.
Cooling towers attached to water-cooled chillers reject IT heat plus chiller and pump energy. Evaporative condensers follow a different equipment arrangement but retain the same principle. Adiabatic dry coolers may use water only during hot periods or peak load.
The demand model should use supplier performance data at agreed dry-bulb and wet-bulb conditions. A rule of thumb can screen options, but it is not sufficient for utility applications, tank sizing or commercial guarantees.
Add blowdown, drift and treatment losses
Evaporation is only part of cooling-tower make-up. Blowdown removes concentrated water to keep dissolved constituents within acceptable limits. Drift carries small droplets out of the tower and should be controlled by properly selected and maintained eliminators.
Cycles of concentration influence blowdown. Higher cycles generally reduce make-up and discharge, but the safe value depends on source chemistry, materials, temperature, chemical programme and microbiological control. Assuming an ambitious cycle without analysis understates demand.
Pretreatment also uses water. Membrane reject, filter backwash, softener regeneration and cleaning must be included at their actual timing and recovery. A treatment train can improve tower cycles while still increasing total withdrawal if its recovery is poor.
Chemical storage and dosing do not usually add large volume, but they affect achievable operation. If a tower cannot maintain its intended chemistry during a supply or dosing fault, the safe response may be higher blowdown or reduced cooling capacity. Failure cases should use the demand associated with the approved operating response.
Include non-cooling uses
Domestic facilities, kitchens, showers, cleaning, humidification, laboratories, irrigation and fire-system testing can be significant, particularly in early phases with low IT utilisation. Construction water should be tracked separately but included when the utility connection and storage must support it.
Climate and landscape design influence irrigation, while air-handling strategy affects humidification or condensate production. Air-conditioning condensate may offer a low-mineral source for selected reuse, but volume and seasonality must be measured or modelled.
Water for periodic activities should be scheduled rather than averaged. Tank cleaning, cooling-loop flushing and fire testing can create short peaks or remove storage from service. Coordinating those activities with cooler weather, low IT load or alternative storage can reduce the required utility capacity.
Each use should have a source, quality requirement, flow profile and discharge route. Combining all duties into one percentage allowance makes later conservation and metering difficult.
Model climate and operating mode by hour
Annual average weather cannot capture the peak make-up requirement or the operating hours of a hybrid system. The model should combine hourly dry-bulb, wet-bulb and humidity with IT load and the cooling control sequence.
For every interval, it should identify dry, adiabatic, wet or mixed operation; cooling duty; evaporation; blowdown; treatment losses and any water recovery. The same model can estimate electrical demand, enabling a balanced comparison rather than optimising water in isolation.
Sensitivity cases should include a hotter year, degraded equipment, one cooling cell unavailable and water-restriction modes. A short extreme can determine connection capacity, while a prolonged hot period can determine storage and source reliability.
Model uncertainty should distinguish input range from calculation error. IT utilisation, weather and future cooling technology may carry wide scenarios; supplier curves and meter accuracy create different uncertainty. Showing the drivers separately helps the owner decide where better data or a design margin is worth the cost.
Reflect AI utilisation and rack-density change
AI workloads can concentrate power in fewer racks and vary with training schedules, inference demand and hardware deployment. Rack density affects local cooling distribution, while total utilised IT power drives the campus heat load.
The forecast should distinguish installed capacity, contracted load and expected utilisation. Scenarios can represent conservative, base and accelerated ramps. Where workload flexibility exists, the owner may also assess whether selected computing can shift in time or location during a utility constraint.
Future hardware may increase liquid-cooling capture or permit warmer water, changing chiller and heat-rejection performance. The model should be adaptable rather than tied to one rack type for the full asset life.
Calculate phased-campus demand
Every development phase needs its own balance. First-phase plant often operates at low turndown, causing minimum blowdown, tank turnover or treatment losses to have disproportionate impact. Later phases can exceed utility and discharge capacity even when modular equipment space is reserved.
The schedule should identify commissioning fills, overlapping construction, live cooling demand and the date each utility increment becomes available. Temporary systems and tanker deliveries need realistic capacities and mobilisation times.
Phasing also influences redundancy. One treatment module may support the early load with no standby capacity unless the project installs the second train in advance. Demand forecasts should show normal and failed-train output so resilience claims remain clear.
Test source quality and treatment recovery
Source chemistry affects both cooling blowdown and treatment yield. Higher dissolved solids, hardness or silica can limit tower cycles or reverse-osmosis recovery. Reclaimed supplies may require additional solids and biological control.
The demand calculation should use a normal and worst credible source envelope. Treatment supplier guarantees should state product capacity and recovery at those conditions. A fixed recovery percentage taken from a different water analysis is not dependable.
Blending can change demand dynamically. If potable water supplements reclaimed water during an interruption, the optimum treatment settings and tower cycles may change. Control and forecasting tools should reflect the active source rather than assume one chemistry all year.
Distinguish annual volume from connection capacity
Annual withdrawal supports sustainability and operating-cost analysis. Peak-day and peak-hour demand size the utility main, pumps, treatment and control valves. Storage autonomy and refill rate address interruption risk. These quantities should never be replaced by one average flow.
The firm connection must serve live demand under the agreed equipment-outage condition or be supported by storage. If treatment operates in batches, its peak feed and residual flow may exceed the continuous average. Sewer acceptance requires the same distinction.
Demand uncertainty should be presented as a range with identified drivers. A utility can then understand the requested reservation, expected use and possible expansion rather than receiving a single unqualified figure.
Commercial forecasts may use several probability levels. The connection should cover the approved firm requirement, while operating budgets can use an expected case and resilience planning can retain a credible upper bound. Labelling those cases prevents a conservative design allowance from being reported later as predicted consumption.
Meter and reconcile actual performance
The operating site should meter each source, treatment stage, major cooling user, blowdown, domestic use and significant reuse stream. Tank levels allow the balance to correct for storage changes. Meter accuracy and range must suit early and final phases.
Reconciliation compares total sources with measured uses, discharge and storage movement. Persistent imbalance may indicate leakage, faulty instruments or unrecorded construction demand. Data-quality flags should distinguish estimates from measured values.
Actual demand should be compared with the model after normalising for IT energy, weather and cooling mode. This validates future-phase forecasts and identifies opportunities such as excessive blowdown, low treatment recovery or unnecessary wet operation.
Forecast ownership should remain with a defined function after handover. Monthly review can incorporate revised IT deployment, meter quality and source conditions, while formal change control protects the assumptions used for permits and utility agreements. A live forecast is more valuable than a detailed commissioning model that is never updated.
Build a decision-ready demand forecast
Begin with the IT load ramp, cooling schematics, design weather, source analyses and proposed operating modes. Create separate estimates for initial fill, normal operation, peak conditions, maintenance and credible failures.
Report withdrawal, consumption and discharge by source and use. Include peak flows, storage autonomy, refill time, treatment recovery, cycles of concentration and calculation uncertainty. Link every major assumption to evidence and an owner.
Update the model when rack plans, cooling suppliers, utility allocations or water chemistry change. The approved balance should feed connection applications, equipment guarantees, sustainability reporting and commissioning tests.
The final report should include a concise calculation narrative and an input schedule, not only output tables. Reviewers should be able to trace peak make-up to cooling duty, cycles and treatment recovery, then reproduce the result. Transparent calculations remain usable when the campus changes hands or expands years later.
How Crescent Engineering can support the project
Crescent Engineering can develop phase-specific water balances, assess source chemistry, size treatment and storage, define recovery options and integrate metering. This turns the thermal concept into utility requirements that can be procured and verified.
Useful starting information includes IT capacity and utilisation scenarios, cooling-system data, hourly weather, water analyses, utility limits, development phases and resilience targets.
Crescent Engineering — Water Treatment
References and further reading
Final demand forecasts must use project-approved loads, current source analyses, supplier performance data and applicable utility requirements.
Comments