Cooling Data Centers in the Desert: Four Architectures and the Water They Drink

Desert data center sites look like a cooling problem and are actually a water problem. The air is hot, but it is also dry — Riyadh's ASHRAE 0.4% design wet-bulb is 18.5 °C against a 46.1 °C dry-bulb, a 27.6 K gap that makes evaporative cooling almost irresistible on paper. The IEA's October 2026 analysis found that nearly 40 percent of identified AI-dedicated data center capacity sits in areas of high or extremely high water stress [4], and the fastest-growing chunk of that capacity is in Arizona, Nevada, New Mexico, Saudi Arabia and the UAE. Every cooling architecture you can buy trades water against electricity differently, and in a desert that trade is the single most consequential — and least reversible — line item in the mechanical design.

This article walks the four architectures actually deployed in desert climates, computes a one-year water-and-energy ledger for each on a common 1 MW IT pod, and covers the sand engineering that decides whether any of them survive their first shamal season. The physics is simple enough to put in a procurement spreadsheet, which is exactly where it belongs.

Four numbers that define a desert site

Before comparing machines, pin down the climate numbers that drive every subsequent decision. All come from the ASHRAE Handbook climatic design conditions — the 0.4% cooling design dry-bulb (exceeded about 35 hours per year) and the mean-coincident wet-bulb that pairs with it [1][2][3]:

Site	0.4% design DB	Coincident WB	Wet-bulb depression	85% adiabatic outlet	What the gap means
Phoenix, AZ	43.9 °C (111 °F)	21.7 °C (71 °F)	22.2 K	25.0 °C	Huge evaporative headroom, deep water stress
Riyadh, SA	46.1 °C	18.5 °C	27.6 K	22.6 °C	The most evaporative-friendly air on Earth — and the least water
Dubai, UAE	43.5 °C	26.1 °C	17.4 K	28.7 °C	Coastal humidity eats a third of the headroom

Table 1 — ASHRAE 0.4% cooling design conditions [1][2][3]. The last column is where a direct evaporative stage at 85% effectiveness lands on the design day: DB − 0.85 × (DB − WB).

Three consequences follow from this table, and each one eliminates or elevates a whole architecture:

• The wet-bulb depression is an evaporative temptation. In Riyadh, evaporating water once across an 85%-effective media drops the air 23.5 K, from 46.1 °C to 22.6 °C. That is chiller-class supply air from a wet pad. The catch: every kWh of heat rejected by evaporation consumes a hard thermodynamic minimum of 1.5 litres of water (the latent heat of vaporization, ≈2,400 kJ/kg, converts 3,600 kJ of heat into exactly 1.5 kg of vapor). No control system, vendor or media pad changes that constant. Add the 85% latent fraction and the blowdown required to keep five cycles of concentration, and a real tower lands at ≈1.9 L/kWh — which is why the industry benchmark for evaporative plants is 1.6–2.2 L/kWh with a fleet average near 1.8 [8][9][11].
• The diurnal swing is a free battery. Desert mean daily range runs 15–25 K [12]. Riyadh drops below 24 °C for roughly 1,200–1,800 night hours a year, and adiabatic pre-cooling extends the economizer envelope by another 800–1,500 hours, for a combined 2,000–3,000 free-cooling hours and 15–25% annual cooling energy savings in UAE-class climates [7]. Phoenix offers 1,500–2,500 air-side economizer hours [10] — modest by Nordic standards (7,500+), but real money at megawatt scale.
• The dust is a tax on every architecture that breathes. Urban Middle East dust loads run above 150 µg/m³ with storm events beyond 1,000 µg/m³ [13]. Any plant that moves desert air across a heat exchanger — direct airside economizers, adiabatic pads, dry cooler fins — inherits a filtration and cleaning regime, covered in its own section below.

Option A — All-mechanical air cooling: pay in kilowatts, not litres

The architecture in Fig. 1 is the simplest desert answer: air-cooled chillers (or DX CRAC units with remote condensers) and no evaporation anywhere on site. It is the default for containerized and edge deployments, for sites with no water service at all, and for every operator who has watched a permitting fight over groundwater.

The physics penalty is unavoidable. Air-cooled condensers are rated at 35 °C ambient; at a 43–46 °C design day, manufacturers apply capacity correction factors of 0.85–0.90, so you buy roughly 10–15% more machine than the nameplate suggests [13]. Condensing temperature climbs with ambient, and the Carnot COP falls with it — in practice, warm-climate air-cooled chiller plants run PUE 1.3–1.5, about 10–15% higher energy than an evaporative plant of the same IT load [9]. Our ledger books Option A at PUE 1.45.

What you get for that premium is a site water number that is essentially zero — 0.02–0.05 L/kWh once you count humidification make-up and domestic use [9][10] — and a mechanical plant with no water treatment program, no Legionella management, no blowdown discharge permit, and no cooling-tower plume. For a 40-ft container module carrying 50–150 kW, this is the only architecture that makes logistical sense; the units in Fig. 1 will each reject 60–130 kW of condenser heat into 46 °C air and do it all summer.

Option B — Direct-to-chip liquid cooling plus dry coolers: the AI default

Modern AI racks changed the arithmetic. A GB300-class rack moves 80–90% of its heat into the liquid loop [5][6], which means the facility's heat-rejection problem is no longer "cool 1 MW of air" — it is "pump a sealed water loop through a dry cooler." Nothing evaporates. The loop is filled once at commissioning and recirculates; Microsoft's zero-water-evaporation design, standard on all its new data centers since August 2024, works exactly this way and avoids more than 125 million litres per year per facility [6].

The reason this works in a desert is the coolant temperature class. Direct-to-chip facility water runs warm: W45 means 45 °C supply to the CDUs. A dry cooler with a 10 K approach therefore free-cools up to 35 °C ambient — and 46 °C ambients only exceed that ceiling by the 9–11 K that an adiabatic pre-cooling stage on the dry cooler's air intakes recovers. Compare the legacy W32 chilled-water class, which free-cools only to 22 °C ambient and would need mechanical trimming for most of a desert summer (Fig. 2, right panel). The 900 kW liquid loop of our 1 MW pod moves 1,077 LPM (284 gpm) at a 12 K loop rise — garden-hose hydraulics, no cooling tower anywhere.

The ledger: 90% of the heat rejects through dry coolers (fan power only), the remaining 10% air-cooled residual runs through CRAH units on a closed loop. Booked at PUE 1.15 and WUE 0.05, Option B is the cheapest plant to operate of the four — $106k per year per MW of IT — and consumes about 440 m³ of water a year, effectively the domestic and humidification baseline. Operators running this pattern report blended WUE of 0.2–0.4 L/kWh at fleet level [9]; the best air-side-dominated fleets (Meta, 2024) report 0.19 [11].

The honest caveats: dry coolers need 15–25% more heat-exchanger surface than evaporative plant for the same duty (their approach is 8–15 K against a tower's 3–5 K) [10], so the capex premium is real. And the architecture presumes your IT estate is actually liquid-ready; a hall of air-cooled legacy servers cannot route 90% of its heat into a water loop that does not exist.

Option C — Hybrid adiabatic free cooling: the Gulf workhorse

Between the two dry extremes sits the architecture the Gulf actually built its fleet on. Khazna — 30 live data centers, 673 MW across the UAE — reports air-based cooling as its primary plant, with 19 of its 30 facilities equipped with adiabatic free-cooling systems, fleet PUE of 1.5 or below, and a claimed 70% energy saving from the adiabatic stages [4][5].

The concept: dry sensible heat exchange whenever the outside air is cool enough, and water-assisted pre-cooling only for the peak hours. Night and winter hours run free; the adiabatic pads knock 8–15 K off the intake air for the afternoon block [7]. A published 2 MW retrofit in Abu Dhabi's Mussafah industrial zone puts real numbers on it: $180k project cost, PUE down from 1.62 to 1.41, 680 MWh of cooling energy saved per year ($37,400 at $0.055/kWh), 6,200 m³ of water consumed ($12,400), 4,100 adiabatic hours (47% of the year), payback 7.2 years [7].

The water bill is the architecture's defining constraint. Our ledger books the hybrid at WUE 0.79 — 6,890 m³ per year per MW, costing $13.8k — twelve times the water of the dry options and about a third of a tower plant. That ratio matches the published bands: hybrid/adiabatic systems report 0.5–1.2 L/kWh in Gulf climates against 1.6–2.2 for open towers [7][8][9]. Khazna's own 2025 ESG data shows what this means at fleet scale: 305,116 m³ of cooling water across the portfolio, of which roughly 80% was potable — which is why its DXB1 facility in Dubai runs on treated sewage effluent from Dubai Municipality, and why recycled sources now cover 19.8% of portfolio cooling water [4]. For the next density step, Khazna's 100 MW Ajman AI facility moves to air-assisted direct-to-chip liquid cooling — the Gulf's own migration path toward Option B [4].

Option D — Open evaporative towers: the physics says yes, the watershed says no

On pure thermodynamics, cooling towers win the desert. The 0.4% wet-bulb is 18.5–26.1 °C across our three sites, a tower approaches it by 3–5 K, and the resulting chilled water runs PUE ≈1.18 with the lowest fan energy of any plant in this comparison — $159k per year total operating cost in our ledger, well below the $301k of the adiabatic hybrid and the $316k of the air-cooled plant.

The water ledger is why deserts don't build them anymore. At WUE 1.88 (0.85 latent fraction, five cycles of concentration), our 1 MW pod evaporates and blows down 16,474 m³ per year — 45 m³ every day, 80–90% of it potable. The Green Grid's WUE definition counts evaporation, blowdown and drift against IT energy [8][11]; industry benchmarks put open-tower plants at 1.6–2.2 L/kWh [8][9]. US fleet data frames the stakes: Microsoft's fleet WUE fell from 0.49 (2021) to 0.30 L/kWh (FY24) as it shifted away from evaporative rejection; Google reported 0.64 L/kWh for 2024 with towers still in the mix; Meta's air-side-heavy fleet hit 0.19 [6][11]. Every one of those numbers is below what a desert tower plant can physically achieve, because the 1.5 L/kWh-of-heat evaporation floor (1.76 after the 85% latent fraction, 2.2 after blowdown at five cycles) sits above them.

That is the arithmetic behind Microsoft's zero-water announcement, Meta's air-side economizer strategy, and the IEA's water-stress finding. In a desert watershed, the tower plant's 12,000+ m³ annual potable advantage over the adiabatic hybrid is not a sustainability footnote — it is the difference between a permitted site and a public hearing.

The ledger, side by side

Architecture	PUE	WUE (L/kWh IT)	Water per MW·yr	Energy cost /yr	Water cost /yr	Annual total	Best fit
A — Air-cooled chillers	1.45	0.05	438 m³	$315.4k	$0.9k	$316k	Containers, edge, no-water sites, <1 MW halls
B — DLC + dry coolers	1.15	0.05	438 m³	$105.1k	$0.9k	$106k	AI halls, 30 kW+ racks, liquid-ready IT
C — Hybrid adiabatic	1.41	0.79	6,890 m³	$287.3k	$13.8k	$301k	Hyperscale air-cooled halls with reclaimed water
D — Open towers	1.18	1.88	16,474 m³	$126.1k	$32.9k	$159k	Sites with assured non-potable water — rare in deserts

Table 2 — Computed for 1,000 kW IT, 8,760 h, hfg 2,400 kJ/kg, 85% latent fraction, 5 cycles of concentration, $0.08/kWh, $2/m³ water. WUE bands cross-checked against [7][8][9][11]. Energy overheads: A 450 kW, B 150 kW, C 410 kW (measured Abu Dhabi retrofit), D 180 kW.

Three readings of this table matter more than the rest. First, water is cheap until it isn't: at $2/m³ the tower plant's water bill ($32.9k) is dwarfed by its energy savings, which is why the trade is settled by watersheds and permits, never by utility invoices. Second, the adiabatic hybrid's energy premium is mostly the air-cooled IT estate: at 410 kW of overhead it is paying nearly chiller-plant money to move heat that direct-to-chip loops would move for a third of that. Third, Option B's $210k annual advantage over Option A is the AI build case: on a 20 MW campus that is $4.2M a year, roughly the capex of the liquid cooling fit-out itself inside five years.

Sand: the engineering that survives contact

Every architecture above moves desert air across something, and desert air carries 150 µg/m³ of dust in normal conditions and 1,000+ µg/m³ in storm events [13]. The failure mode is never dramatic; it is coil fouling that raises pressure drop month over month, fin channels bridging over, and adiabatic pads clogging with the silt the spray water leaves behind. The mitigation package is standardized across the region:

Defense	Specification	Service rhythm
Pre-filter (G4/MERV 1–4)	Coarse dust capture ahead of everything downstream	Inspect monthly; replace 1–2 months in storm season
Fine filter (F7/F8, MERV 11 minimum at coils)	Protects fins and pads; the region's minimum for coil-facing plant	Replace 3–4 months
Coil fin spacing	8–10 fins per inch (not the temperate-climate 12–14 FPI)	Epoxy/polymer coated fins; wash access both faces
Adiabatic pad water	Side-stream filtration and bleed; hard water scales pads within weeks untreated	Pad wash/replace per season; validate distribution weekly in summer

Table 3 — Sand and dust mitigation package for desert cooling plant [13]. The filtration economics only close if pre-filters actually get changed on schedule; a blinded G4 stage converts your F7 filters into consumables at 3× the rate.

Two additions from field practice. Solar load is the quieter enemy: a dark-skinned dry cooler or condenser section in a Gulf design day runs its surface 30 K above air temperature under sol-air loading, which is why desert units ship in light gray or white and why shading structures appear on the west and south faces of condenser farms. And the same dust that blinds coils attacks the pad systems hardest — an adiabatic cell that dries out with silt in it becomes a filter nobody scheduled. The corrosion side of desert exposure (C4-to-CX coating systems, salt dust on coastal Gulf sites) is covered in detail in our container anti-corrosion piece.

Related: what airborne dust does inside the hall — ISO 14644 particle budgets and the MERV 8 vs 13 decision · C4/CX anti-corrosion coating systems for Gulf deployments

How to choose: five questions before the layout

1. What is the water, really? Potable municipal, reclaimed sewage effluent (Khazna DXB1's supply [4]), or none? The answer eliminates Option D immediately, prices Option C, and may mandate Option A/B.
2. What fraction of IT heat will actually be liquid-borne? The DLC+dry cooler ledger assumes 90%; a hall legacy air-cooled servers at 20% liquid share runs closer to Option C economics with Option A plant. Workload roadmap beats climate in this decision.
3. What is the PUE floor you can defend? A design day at 46 °C with 0.4%-percentile exceedance of 35 hours does not define the year; the bin distribution does. Demand the vendor's part-load and high-ambient performance maps, not the 35 °C rated table [13].
4. Who owns the adiabatic water treatment? Pads, silt, scale and Legionella-aware bleed management are an operating program with a budget line, not a box on the equipment schedule.
5. What does the watershed say? IEA's finding that 40% of AI capacity sits in high water stress [4] is a permitting forecast. The 12,000 m³/yr difference between the tower plant and the dry plant is the number your community-relations team will be asked about.

Eight mistakes we see in desert cooling designs

1. Selecting air-cooled plant at 35 °C nameplate and discovering the 0.85 derating during the first 46 °C week.
2. Sizing adiabatic water from annual averages instead of the peak-hour block — the pads drink in afternoons, not averages.
3. Specifying 12–14 FPI coils from a temperate-climate standard and re-coiling them after the first shamal season.
4. Booking WUE 0.0 for a closed-loop liquid plant while forgetting the dry cooler fans and the 10% air-cooled residual in the PUE.
5. Building a tower plant on potable water because the utility invoice for water looked small.
6. Treating filtration as a consumable rather than a system — G4 stages left in service until they collapse into the F7 bank.
7. Dark-colored condenser sections and zero shading, then chasing capacity losses that are actually sol-air surface temperatures.
8. Assuming the coolant temperature class. A W32 chilled-water plant free-cools to 22 °C ambient in the desert; a W45 plant free-cools to 35 °C. The class is the architecture.

Ten questions for your cooling vendor

1. Provide capacity correction curves to 50 °C ambient, not the 35 °C rating point.
2. What is the dry cooler approach at design ambient and at 80% load — and what coolant supply temperature does it presume?
3. What is the guaranteed WUE band for the adiabatic stage at our wet-bulb distribution, measured per The Green Grid definition [8]?
4. Show the free-cooling hour count for our site's actual bin data, split dry / adiabatic / mechanical.
5. What filtration stages, media classes and change-out schedule does the warranty require?
6. What pad media and water treatment program does the adiabatic section presume, and what is its annual water quality spec?
7. Can the plant accept reclaimed (non-potable) water for the evaporative stages, and what pre-treatment does that require?
8. What is the fin coating, fin spacing and wash-down provision on all outdoor coils?
9. Provide part-load PUE from 30% to 100% IT load — desert halls live at partial load overnight.
10. What is the surge/derate behavior of compressors and fans above 45 °C, and what BMS alarms protect them?

The bottom line

Desert cooling is a two-axis optimization that the industry spent a decade pretending was one-axis. The tower plant wins the energy axis and loses the watershed; the air-cooled plant wins the watershed and pays 30–45% more per year to run. The hybrid adiabatic fleet the Gulf actually built sits between, honest about its water bill and increasingly fed by reclaimed water. And the AI generation of IT — 80–90% liquid-borne heat on a W45 loop — is the first workload whose thermal architecture matches the desert's constraint set: seal the loop, reject dry, spend water only on the 35 hours that matter, and let the wet-bulb depression do the rest.

Planning a desert or high-ambient deployment? Send us your IT load, site climate and water constraints — we will work the same four-architecture ledger on your numbers before you commit a mechanical layout to a 15-year water decision.

References

1. LAA Architects, "Dark aluminium window frames in Dubai" — tabulates the ASHRAE Handbook–Fundamentals 2025 climatic design conditions for Dubai International (0.4% cooling DB 43.5 °C, annual extreme 46.4 °C) with clear-sky solar constants. laa.ae
2. WorkZen Academy, "Climate Data & Design Conditions" — ASHRAE design temperature tables; Phoenix 0.4% cooling 111 °F DB / 71 °F MCWB. academy.workzen.io
3. Dar Anan, "Hotel HVAC Design Requirements in Saudi Arabia" — Riyadh summer design conditions per ASHRAE: 46 °C cooling DB, 18 °C coincident WB. daranan.com
4. Khaleej Times, "How UAE prepares AI data centers for heavy rains, extreme heat, water shortages" (Nov 2026) — Khazna ESG data: 305,116 m³ cooling water (2025, ~80% potable), 19 of 30 facilities on adiabatic free cooling, DXB1 treated sewage effluent, Ajman 100 MW air-assisted D2C, IEA finding that ~40% of AI capacity sits in high water stress. khaleejtimes.com
5. Khazna Data Centers, "Wired For Impact" — PUE ≤1.5 in Middle East climates, 70% energy savings via adiabatic free cooling, treated-sewage-water use. khazna.ae
6. Microsoft, "Sustainable by design: Next-generation datacenters consume zero water for cooling" (Dec 9, 2024) — chip-level closed-loop design from Aug 2024, >125M L/yr per datacenter avoided, fleet WUE 0.30 L/kWh FY24 vs 0.49 in 2021, Phoenix pilot. microsoft.com
7. RAX Engineering, "Free Cooling & Adiabatic Systems for Hot-Climate Data Centers" — Abu Dhabi 2 MW adiabatic retrofit (PUE 1.62→1.41, 680 MWh saved, 6,200 m³ water, 4,100 adiabatic hours), UAE free-cooling hour bands, 1 MW adiabatic water draw 8,000–15,000 m³/yr, WUE 0.5–1.2. rax.ae
8. Reynolds & Bauhm, "Data Centre Water Usage Effectiveness (WUE) Calculator" — Green Grid WP#35 definition, evaporation/blowdown/make-up balance (hfg 2,400 kJ/kg, 85% latent fraction, CoC), WUE benchmark bands (dry 0–0.1, hybrid 0.2–0.9, efficient evaporative 1.0–1.6, open towers 1.6–2.2). reynoldsbauhm.co.uk
9. Datacentres.com, "The WUE Divide: Why Cooling Technology Choice Swings Data Centre Water Use by 18x" (Sep 2026) — towers 1.6–2.0 L/kWh at PUE 1.1–1.2; air-cooled ≈0 L/kWh at PUE 1.3–1.5 warm-climate; adiabatic 0.5–1.0; DLC closed-loop 0.2–0.4. datacentres.com
10. Datacentres.com, "Data Centre Water Usage: WUE, Cooling Consumption, and Drought Risk" — Phoenix air-side economizer hours 1,500–2,500/yr, dry cooler approach 8–15 K vs tower 3–5 K, closed-loop capex +15–25%. datacentres.com
11. RPM Water Equity, "What Is WUE? Water Usage Effectiveness Explained" — The Green Grid White Paper #35 definition; Meta fleet 0.19 L/kWh (2024), Microsoft 0.27 (FY25), industry averages ≈1.8–1.9. rpmwes.com
12. HVAC Systems Encyclopedia, "Hot-Dry Desert Climate HVAC Design Considerations" — design-condition parameter ranges (DB 40–50 °C, WB 18–24 °C, diurnal swing 15–25 K), evaporative system effectiveness table (direct 70–90%, indirect 50–75%, multi-stage 100–130%). ingener.by
13. HVAC Systems Encyclopedia, "Middle East HVAC Practices and Desert Climate Design" — dust concentrations (>150 µg/m³ urban, 1,000+ storm), G4/F7/F9 filtration table, 8–10 FPI fin spacing, MERV 11 coil minimum, 0.85–0.90 high-ambient capacity correction factors. ingener.by