Cutting-Edge Technologies
NextFin News - XDS Datacentres is building what it says will be the Middle East's first GPU-as-a-Service facility cooled entirely by liquid immersion, a pair of 10MW data centers in Riyadh and Jeddah where server racks are submerged in a non-conductive fluid instead of blasted with conditioned air. The wager is stark: in a region where summer air tops 50°C and industrial water now costs 8.04 riyals per cubic meter, the air-cooled data center is becoming both physically and economically unviable - and XDS, a company that does not yet operate a single facility, is betting its 15-year operating mandate on a cooling technology it believes will make air-cooled rivals the "Blockbuster of the data center market."
The Deal: Two 10MW Facilities, 368 kW Per Rack
In July 2025, XDS signed an agreement with ICS Arabia, a Saudi data center operator and manufacturer founded in 2003, to construct and deliver two 10MW immersion-cooled data centers - one in Riyadh, one in Jeddah. ICS Arabia's subsidiary, Desert Dragon Data Centers, is designing and building the facilities and will hand the Riyadh site over to XDS by June 2026, with Jeddah following in November 2026. XDS will operate the facilities for 15 years.
The design targets rack densities of up to 368 kilowatts, roughly an order of magnitude above a conventional air-cooled colocation hall, with N+N redundant power and cooling and a stated goal of Tier III certification. Clients are being offered a 99.982% uptime guarantee, GPU-as-a-Service, private cloud, and migration services that convert existing air-cooled servers for immersion deployment.
"The kingdom represents a key growth market for XDS and this initial 10MW facility will showcase the potential of immersion-cooled infrastructure to deliver both performance and sustainability."
Ghufran Hamid, the company's founder and CEO, has been blunter about the stakes:
"We are pioneering the inevitable switch from air-cooled to liquid immersion. Those companies who are still investing in just air-cooled will become the Blockbuster of the data center market."
The timing is not accidental. Saudi Arabia is in the middle of a data center construction boom tied to Vision 2030, the crown prince's plan to diversify the economy away from oil. Humain, the AI-focused subsidiary of the Public Investment Fund, has committed to an initial 18,000 Nvidia GB300 chips with "several hundred thousand" more over five years, building toward 500MW of AI data center capacity. The industry data XDS cites puts the Saudi market at roughly 1,330MW in 2024, on track to exceed 1,900MW by 2030 - a 13-14% compound annual growth rate. Every watt of that compute generates heat, and in the Gulf, heat is the binding constraint.
How Immersion Cooling Actually Works
Immersion cooling removes the middlemen between the chip and the outside world. In a conventional air-cooled hall, heat travels from the processor through a heatsink, into moving air, through a raised-floor plenum, into CRAC - computer room air conditioning - units, and finally out through chillers or cooling towers. Each step loses efficiency, and every fan and pump consumes electricity that counts against the facility's Power Usage Effectiveness ratio, the industry's measure of how much power the building itself consumes beyond the IT load.
In single-phase immersion cooling, servers sit inside sealed tanks filled with a dielectric fluid - a liquid that does not conduct electricity, so it can touch live components without shorting them. Heat transfers directly from the chip surface into the fluid by contact, which is far more efficient than air because liquid carries heat far better than gas. A pump circulates the warmed fluid through a heat exchanger, where the heat is rejected to the outside, and the cooled fluid returns to the tank. In two-phase systems, the fluid is engineered to boil at a low temperature; the vapor rises to a condenser coil, liquefies, and drips back down, carrying heat away through the latent heat of vaporization - the same physics that makes sweat cool skin.
XDS's Saudi facilities use immersion racks supplied by GRC, one of the established vendors in the space. The company's 1MW pilot hall in Sharjah, inaugurated last year, was designed to validate the thermal design, power trains, and operations under live loads before the multi-megawatt buildout.
"The one-megawatt halls validate our thermal design, power trains and operations under live loads, so clients can deploy today and scale on the same architecture into multi-megawatt footprints with consistent SLAs, sustainability metrics and network fabrics."
The efficiency gains are material. Immersion cooling vendors report data center PUE figures as low as 1.03 to 1.10, compared with 1.5 to 1.6 for typical air-cooled facilities in hot climates. Because the fluid absorbs heat directly, server fans can be removed or run at minimal speed, cutting the auxiliary power that inflates PUE. The rejected heat also exits at a useful temperature: warm-water designs return fluid at 45-65°C, hot enough that some facilities capture it for district heating or industrial processes - something the low-grade exhaust of air cooling cannot do economically. For XDS's modular customers, the same stack ships inside a 1.5-2MW container rated for up to 900 Nvidia GPUs, deployable in weeks rather than the years a conventional build can take.
Why the Gulf Makes This Structural, Not Cyclical
The case for immersion in Saudi Arabia rests on two constraints that will not mean-revert: thermodynamics and water pricing.
First, the chips have crossed a thermodynamic cliff. Nvidia's Blackwell-generation GB200 carries a thermal design power of 1,200 watts per chip, up from 700 watts for the prior H100 generation, and a single GB200 NVL72 rack draws between 120kW and 135kW. At heat fluxes of 500-600 watts per square centimeter, industry analysis concludes, air simply cannot remove heat fast enough from the chip surface - the physics of convection through air breaks down regardless of how powerful the fans become. Liquid cooling is not an efficiency upgrade here; it is a physical requirement. That is a structural shift in the industry's infrastructure stack, not a cyclical preference.
Second, the Gulf's climate turns the traditional cooling model into an economic losing proposition. When ambient air reaches 50°C, chilling water down to the 20°C that conventional systems expect requires massive mechanical refrigeration - exactly when the grid is already stressed by air conditioning demand. The ASHRAE W4 standard, which permits coolant inlet temperatures up to 45°C, changes the equation: a chip receiving 45°C water returns it at roughly 65°C, and that 65°C liquid can be cooled back to 45°C using simple dry coolers - essentially radiators - even in 50°C ambient air, with no chillers at all. Industry analysis puts the resulting PUE improvement from around 1.6 to below 1.15.
Then there is water. In December 2025, a tariff revision by utility operator Marafiq pushed industrial water to 8.04 riyals per cubic meter in the industrial cities it serves. Evaporative cooling - the cooling towers that reject heat by boiling water into the atmosphere - is the most water-intensive method available, and in a desert kingdom water is the scarcest input of all. A closed-loop immersion system consumes a fraction of the water because the fluid circulates in a sealed tank and is not lost to evaporation. The tariff change effectively put a financial floor under the immersion business case: air-side economizers and evaporative towers are not just less efficient, they are becoming regulated out of viability.
Saudi regulators are moving in the same direction. The Saudi Standards, Metrology and Quality Organization's 2023 regulation for energy efficiency in data centers requires new facilities to implement advanced energy-efficient cooling solutions, and Vision 2030's national data center strategy, launched by the communications ministry with SDAIA in 2025, targets roughly 1.5GW of capacity with sustainability thresholds attached.
The Counter-Case: Direct-to-Chip, Cost, and a Company Without a Track Record
The strongest argument against XDS's immersion bet is not that immersion is inferior - it is that direct-to-chip liquid cooling achieves most of the thermal benefit at lower cost and lower operational disruption. In a direct-to-chip system, cold plates bolt directly onto the CPU and GPU, and liquid flows through them; the rest of the server and the data hall can remain largely air-cooled. Hyperscalers and chipmakers including Nvidia have embraced direct-to-chip for their densest racks, and it avoids immersion's two biggest headaches: the cost of dielectric fluid, which must be periodically filtered or replaced, and the need to redesign or replace servers so that components can survive submersion and the fluid's chemistry.
Immersion also carries vendor and fluid risk. Two-phase systems historically relied on fluorocarbon fluids with environmental concerns, and the quick-disconnect fittings that allow tanks to be serviced are a known failure point for leaks. For an operator signing a 15-year concession, betting on one cooling architecture is a concentrated technology risk - if the industry standardizes on direct-to-chip, an immersion-built facility could face higher migration costs and a thinner resale market.
And XDS itself is an unproven operator. The company does not currently operate a data center; its only commissioned asset is a 1MW pilot hall in Sharjah. Its Saudi facilities are being designed and built by Desert Dragon, an ICS Arabia subsidiary, and handed over in 2026. The 99.982% uptime guarantee is a contractual promise, not a performance record. Arabia Abdullah Ayed Al Mazny, the general manager of ICS, framed the partnership as a shared vision, but the execution risk - thermal design under Saudi summer loads, fluid maintenance, server compatibility at 368kW per rack - sits squarely with a company that has yet to prove it can run a multi-megawatt hall at scale.
The answer to that counter-case is density. Direct-to-chip cools the chip but leaves the rest of the rack to air, which caps practical rack density well below what immersion allows. At 368kW per rack, XDS is targeting a density tier that air - even assisted air - struggles to serve in a 50°C environment. For sovereign AI customers who need maximum compute per square meter and cannot wait for grid and water infrastructure to catch up, immersion's higher upfront cost may be the price of fitting a 500MW AI buildout into a constrained site. The bet is that the total cost of ownership - power, water, and real estate per unit of compute - favors immersion in the Gulf even where direct-to-chip wins elsewhere.
What Comes Next
The near-term signal to watch is the Riyadh handover, due June 2026. If Desert Dragon delivers on schedule and XDS begins live operations without thermal throttling or fluid-contamination issues, the model becomes bankable and replicable across the Gulf - XDS already has 7MW immersion facilities in the pipeline for Dubai and Abu Dhabi, plus a 4MW retrofit project in Reading, UK, and modular 1.5-2MW container units rated for up to 900 Nvidia GPUs each.
Medium-term, the question is whether the Saudi pipeline - Humain's 500MW, the hyperscalers' regional expansions, and the 1.5GW national target - fills immersion facilities faster than direct-to-chip alternatives. A single large anchor tenant signing a multi-year GPUaaS contract at the Riyadh site would validate demand; a delay past June 2026, or a public shift by a major customer toward direct-to-chip racks, would weaken the thesis.
Long-term, the structural call stands or falls on two thresholds. If rack densities stay below roughly 50kW - the practical ceiling for advanced air cooling - immersion remains a niche for edge and sovereign AI rather than the industry default. And if water tariffs fall back or new desalination capacity floods industrial users with cheap water, the economic moat that immersion's closed loop depends on narrows. But if chip power continues climbing toward and beyond 1,200 watts and Gulf water pricing holds near current levels, the physics and the economics point the same way: in the desert, the air-cooled data center has a ceiling, and liquid is the only path through it.
XDS is not selling cooling - it is selling the right to keep building AI capacity in a country where the air is too hot and the water too expensive for the old way, and that right is only valuable if the new way actually works at scale.
Explore more exclusive insights at nextfin.ai.
