Cooling Saudi Hyperscale: Where the PUE Battle Is Won and Lost

The Saudi Arabian climate is the single largest cost penalty applied to data-center economics in the Kingdom. Annual ambient design temperatures in Riyadh routinely exceed 45°C in summer, with extended periods above 40°C across the May-September window. Compared against US-East baseline (annual design 35°C peak) or Frankfurt (32°C), the Saudi cooling penalty drives PUE upward and pulls effective $/kWh-of-IT-load higher. The cooling-architecture choices that hyperscale operators make in the Kingdom in 2026 directly determine whether the structural power advantage discussed in the power-cost analysis is captured or eroded. This analysis lays out the typical 2026 range for cooling-architecture costs and PUE outcomes across Saudi hyperscale builds.

CRAC/CRAH Legacy Architecture: $1.4M to $2.4M per MW

The legacy cooling architecture — perimeter CRAC (Computer Room Air Conditioner) or CRAH (Computer Room Air Handler) units with cold-aisle/hot-aisle containment — is still deployed in some Saudi enterprise and mid-market DCs, but it is functionally obsolete for AI-density workloads. Capex for legacy CRAC/CRAH at 8-15 kW per rack densities runs in a typical 2026 range of $1.4M to $2.4M per MW of IT load, with annual operating energy overhead pushing effective PUE into the 1.55-1.85 range in Saudi summer ambient conditions.

The practical implication is that legacy CRAC/CRAH is roughly 35-50 percent worse on PUE than modern architectures in Saudi conditions, which translates directly into effective $/kWh-of-IT-load uplift. Hyperscale tenants in the Kingdom in 2026 universally specify modern architecture for new builds; CRAC/CRAH remains relevant only for retrofit and legacy-fleet operations.

Rear-Door Heat Exchangers: $1.8M to $2.8M per MW

Rear-door heat exchanger (RDHX) architecture, in which water-cooled coils are mounted on the rack rear-doors and capture heat at the source, has emerged as the dominant air-side cooling architecture for medium-density (15-40 kW per rack) Saudi hyperscale deployments. Capex for full RDHX deployment at 25-40 kW per rack densities runs in a typical 2026 range of $1.8M to $2.8M per MW of IT load, with PUE outcomes in the 1.32-1.45 range in Saudi summer ambient.

The economic logic of RDHX is compelling for the workload mix that dominates Saudi hyperscale in 2026 — a meaningful share of standard CPU compute, mid-density GPU deployments (H100 SXM5 at 35-40 kW per rack), and the storage and networking layers that don’t justify direct-to-chip liquid cooling. RDHX delivers most of the PUE benefit of liquid cooling at lower capex, with operational simplicity that suits the broader Saudi DC-operations talent pool.

Direct-to-Chip Liquid Cooling: $2.6M to $4.2M per MW

Direct-to-chip liquid cooling (DLC) — in which coolant loops circulate through cold plates mounted directly on GPU and CPU dies — is the architecture mandated by NVIDIA’s GB200 NVL72 rack systems and increasingly specified for high-density H100/H200/B200 deployments. DLC capex at 80-130 kW per rack densities runs in a typical 2026 range of $2.6M to $4.2M per MW of IT load, with PUE outcomes in the 1.18-1.28 range even in Saudi summer ambient — meaningfully better than air-cooled architectures and approaching the best US-East benchmarks for matched workloads.

The DLC architecture stack typically includes coolant distribution units (CDUs) at row or rack scale, manifold and quick-connect plumbing, secondary loops connecting CDUs to facility cooling, and the rack-internal cold-plate-and-tubing infrastructure. The capex split across these elements is roughly 40 percent CDU and primary distribution, 30 percent rack-internal cold plates and tubing, 20 percent secondary loops and facility integration, and 10 percent specialist commissioning and operations setup.

GB200 NVL72 specifically requires factory-integrated DLC delivered as part of the NVIDIA reference rack architecture, which simplifies the deployment but ties the capex into the rack-scale silicon procurement. For non-GB200 deployments using DLC, the capex is captured as separate cooling-infrastructure spend.

Immersion Cooling: $3.4M to $5.8M per MW for Two-Phase

Immersion cooling — in which complete servers are submerged in dielectric coolant — represents the most aggressive cooling architecture available for very-high-density AI workloads. Two-phase immersion cooling at 130-180 kW per tank densities carries capex in a typical 2026 range of $3.4M to $5.8M per MW of IT load, with PUE outcomes that can reach 1.05-1.12 even in Saudi ambient conditions.

Saudi adoption of immersion cooling in 2026 is meaningful but still concentrated in specific niches: specialised AI training clusters, certain HPC deployments, and pilot programs at sovereign-anchor operators (Humain has piloted multiple immersion deployments). The barriers to broader adoption are operational complexity, the relative scarcity of certified maintenance tooling and procedures, and the silicon-warranty implications for some SKUs.

For purpose-built AI training facilities — where the workload is uniformly very-high-density GPU compute and the operational team is structured around the immersion architecture — the economics in Saudi conditions are attractive. The PUE advantage versus DLC translates to roughly 8-14 percent effective power-cost reduction, which on a 50MW immersion deployment over 10-year operational life represents tens of millions of USD in operating cost differential.

Water Usage and Water-Cost Dynamics

Cooling architecture choice in the Kingdom is materially shaped by water economics. Saudi industrial water — typically delivered through the Saudi Water Authority and regional desalination subsidiaries — costs substantially more per cubic meter than equivalent supply in most US or European benchmarks. Industrial water in the Kingdom in 2026 typically prices at 3.5 to 7.2 SAR per cubic meter ($0.93-$1.92 USD/m³) for high-volume tenants, with regional variation reflecting proximity to desalination capacity.

A water-cooled CRAC/CRAH or evaporative-cooling architecture at 100MW IT load typically consumes 1.4-2.2 million cubic meters of water annually (water usage effectiveness, WUE, of approximately 1.1-1.7 L/kWh of IT load). Annual water cost at this scale in the Kingdom runs $1.3M-$4.2M USD — meaningful but bounded.

DLC and immersion architectures, by contrast, achieve WUE of 0.05-0.25 L/kWh of IT load (closed-loop secondary circuits with minimal makeup water), cutting annual water consumption to 50,000-250,000 m³ for the same 100MW load. The water-cost saving is modest in absolute terms ($0.05M-$0.5M annually), but the strategic value of low water consumption in the Kingdom is meaningful from a sustainability and regulatory-positioning perspective.

The Saudi Water Authority’s water-allocation framework increasingly favours low-WUE digital-infrastructure tenants, which structurally advantages DLC and immersion architectures in the permitting and allocation process.

Adiabatic Cooling and the Hybrid Architecture

Adiabatic cooling — in which evaporative pre-cooling reduces the effective ambient temperature for downstream chiller operation — is a structurally important hybrid technology for Saudi conditions. Adiabatic pre-coolers added to a chilled-water plant typically reduce annual chiller energy consumption by 12-22 percent in Saudi summer conditions, at incremental capex of $280K-$540K per MW of IT load.

The trade-off is water consumption — adiabatic systems consume meaningful makeup water for the evaporation cycle. WUE for an adiabatic-augmented chilled-water architecture typically runs 0.6-1.1 L/kWh of IT load, which is intermediate between fully-water-cooled CRAC/CRAH and closed-loop DLC.

Most Saudi hyperscale builds in 2026 specify adiabatic augmentation as a standard element of the chilled-water plant, recognising the energy-efficiency benefit even at the water-consumption cost.

PUE-to-Cost Relationship in Saudi Climate

The structural relationship between PUE and effective $/kWh-of-IT-load in Saudi conditions is meaningfully more sensitive than in temperate climates, because the cooling load represents a larger share of total facility energy. A 0.10 improvement in PUE in Saudi conditions translates to roughly 7-9 percent reduction in effective $/kWh-of-IT-load, versus 5-7 percent in US-East baseline conditions.

For a 100MW hyperscale operating at 85 percent load factor with SEC industrial pricing, every 0.10 PUE improvement is worth approximately $3.8M-$5.2M annually in effective power-cost reduction. Over a 15-year operational lifetime, the cumulative effective value of moving from PUE 1.45 (good RDHX) to PUE 1.22 (good DLC) is on the order of $130M-$180M — comfortably justifying the incremental DLC capex.

Cooling-Architecture Mix in Practice

Most Saudi hyperscale builds in 2026 are not single-architecture facilities. The typical hyperscale floor plate hosts a mix of densities: storage and networking racks at 8-15 kW (RDHX or even legacy CRAC), CPU compute at 15-30 kW (RDHX), mid-density GPU at 30-60 kW (RDHX or DLC), and high-density GPU at 80-130 kW (DLC mandatory, immersion optional). The capex per MW of IT load varies with the mix; the typical hyperscale weighted-average cooling capex for 2026 builds runs in the $2.4M to $3.8M per MW range, with weighted-average PUE in the 1.24-1.36 range.

For pure AI training facilities — uniform high-density GPU loads — the architecture is more uniform (DLC or immersion throughout), the capex per MW is at the upper end of the range, and the PUE is at the lower end. For mixed-workload hyperscale serving general-purpose cloud and AI inference, the architecture is heterogeneous and the economics blend accordingly.

The 2027-2030 Trajectory

The cooling-architecture landscape will continue evolving rapidly through 2027-2030. The Rubin generation of NVIDIA silicon will push rack densities further, with reference architectures suggesting 200+ kW per rack for some configurations. AMD’s MI400 series and the various startup AI silicon entrants will similarly drive density. The implication is that DLC will continue gaining share against RDHX, and immersion will move from niche to mainstream for purpose-built AI facilities.

Saudi hyperscale builds breaking ground in 2026-2027 should specify cooling architecture with explicit density-flexibility provisioning — the ability to upgrade from RDHX to DLC, and from DLC to immersion, without major facility refit — to manage the multi-generation density progression.

These ranges are analytical estimates synthesised from observed hyperscale build economics, vendor pricing for cooling subsystems, and PUE benchmarking through 2025-2026. They should not be treated as committed price quotes; specific cooling-architecture economics vary based on facility design, climate-zone specifics, water-allocation framework, and operator preference.

Operational Cost: Maintenance, Coolant, and Replacement

Beyond capex, the operational cost of cooling architectures is a meaningful component of TCO. Annual maintenance cost for legacy CRAC/CRAH typically runs 3-5 percent of capex annually, dominated by filter replacement, refrigerant top-up, and compressor service. RDHX maintenance cost runs 2-4 percent of capex annually, with the heat-exchanger coils requiring periodic descaling but otherwise operationally simple.

DLC operational cost is more complex. Annual maintenance runs 4-7 percent of capex annually, covering coolant chemistry maintenance (corrosion-inhibitor monitoring, biocide management, pH stability), CDU service, manifold and quick-connect inspection, and the specialist labour for liquid-cooling operations. Coolant replacement at typical 5-7 year intervals adds another $120K to $280K per MW at replacement events. For Saudi operators, the maintenance-skill-pool development cost is a real but typically transient overhead during the 2024-2026 ramp.

Immersion cooling operational cost is the highest of the architectures, with annual maintenance running 5-9 percent of capex annually, dominated by coolant chemistry, dielectric-fluid quality monitoring, and the specialist labour for tank-immersion operations. Dielectric coolant replacement at 7-12 year intervals adds another $220K to $480K per MW at replacement events.

Vendor Landscape and Procurement Counterparties

The cooling-vendor landscape for Saudi hyperscale builds in 2026 includes both global specialists and emerging Saudi-native capability. Global specialists including Vertiv, Schneider Electric, Stulz, Munters, and the various RDHX specialists (CoolIT, Chilldyne, ZutaCore for two-phase) maintain substantial Saudi presence and serve the bulk of new hyperscale specifications. Liquid-cooling specialists including CoolIT Systems, Asetek, JetCool, and the various GB200-NVL72-aligned providers serve the AI-density tier specifically.

Saudi-native and regional capability has emerged during 2024-2026 with several local engineering firms developing competence in cooling-system integration and operations support. This capability is still maturing but is structurally important for the long-term operational independence of Saudi hyperscale operations.

Climate-Resilience and Extreme-Weather Provisioning

Saudi cooling-architecture design must account for climate-resilience considerations beyond the standard summer ambient. Sandstorm events specifically — the haboobs that affect inland Saudi sites particularly during spring — drive specific design requirements for outdoor air-handling intake filtration and cleaning protocols. Coastal sites (Jeddah, Dammam, NEOM) face salt-air corrosion implications for outdoor heat-rejection equipment that drive specific material specifications and maintenance protocols.

The cost premium for climate-resilient cooling design in Saudi conditions typically runs 4-9 percent above standard hyperscale specifications for inland sites, and 6-14 percent above standard for coastal sites. This is built into the typical hyperscale capex ranges quoted earlier but is worth understanding as a structural component.

Heat Reuse and District-Cooling Integration

A growing area of Saudi cooling-architecture innovation is heat reuse — capturing waste heat from data-center operations for productive use in adjacent applications. NEOM specifically has incorporated heat-reuse provisions into its master planning, with data-center waste heat targeted for use in district heating, desalination plant pre-heat, and certain industrial process heating applications.

The economic value of heat reuse is small in absolute terms (typically $0.5M-$2M annually per 100MW of IT load) but the strategic and sustainability-positioning value is meaningful for tenants serving customers with strict scope-3 emission targets. Saudi hyperscale builds with explicit heat-reuse provisioning are still a minority of total deployments but are growing as a share, particularly among NEOM-region builds and the most sustainability-positioned major-enterprise tenants.

For deeper reading: