The Middle East has emerged as one of the busiest places in the world when it comes to building AI data centers. Abu Dhabi, Riyadh, Dammam, Doha, and Neom are some cities that are constructing their data centers in gigawatts instead of megawatts. But building an AI data center in the Gulf is a different engineering problem than building one in Virginia or Frankfurt. The region combines some of the cheapest available capital and land in the world with some of the most severe power, water, and climate constraints. Anyone designing HPC/AI infrastructure here has to treat those constraints as first-order design inputs, not afterthoughts.
This piece looks at what is actually being built, why AI workloads draw so much more power and water than traditional enterprise computing, and what that means for how data centers in the region are designed, cooled, and connected.
Which Data Centers Are Used for AI
The fact is, not all data centers located in the area are actually HPC or AI facilities, which should be taken into account when assessing any infrastructure-related information. The difference is that AI and HPC facilities consist of densely packed GPU racks that consume significantly more electricity in an extremely small space and operate at full capacity throughout the entire process of training.
A single Nvidia GB200 NVL72 rack, for example, houses 72 Blackwell GPUs and 36 Grace CPUs and draws roughly 120 kilowatts at Nvidia’s nominal specification. Deployed systems have been observed drawing 130 to 132 kilowatts under full load. That is up from roughly 8 kilowatts per rack five years ago. Some industry analysts project rack densities approaching 1 megawatt by 2028 as GPU generations continue to scale. This is the core reason AI data centers require fundamentally different electrical and cooling infrastructure than the facilities most operators have historically built.
AI Data Center Power Consumption and Power Requirements
Power is the constraint that shapes almost every other design decision in an AI data center. The International Energy Agency indicates that the increase in global demand for electricity in data centers increased by 17 percent in 2025, while the amount of electricity consumed by facilities specialized for artificial intelligence increased by 50 percent in 2025. The projection by the IEA is that data centers will consume twice their electricity, to about 945 TWh by 2030, which is slightly greater than the present-day electricity consumption of Japan.
For the Gulf specifically, this creates a direct tension between AI ambitions and grid capacity. Wood Mackenzie has projected that UAE data center power demand will double by 2030, and has flagged regulatory gaps that constrain how operators can procure clean power. Reporting from Forbes notes that despite the UAE’s stated goal of sourcing close to a third of power demand from clean energy by 2030, current regulations prevent data center operators from signing direct corporate power purchase agreements for renewables, limiting most facilities to rooftop solar or renewable energy certificates, with gas turbines still supplying the bulk of actual load. In practice, this means an AI data center in Abu Dhabi or Riyadh or anywhere in the GCC needs a power strategy that accounts for both the physical capacity of the local grid and the regulatory pathway for how that power is contracted and certified.
At the campus level, the numbers involved are substantial. Stargate UAE alone is designed to scale to 5 gigawatts, and its initial 1-gigawatt phase is estimated to consume a meaningful share of the operating reserves on the UAE grid. Planning power infrastructure at this scale requires coordination with national utilities years in advance of any GPU shipment.
AI Data Center Power and Water Consumption: Why They Are Linked
Power and water consumption cannot be evaluated separately in data center design, because most cooling methods trade one for the other. Evaporative cooling systems need minimal power but utilize a large amount of water because, for each megawatt of heat load, one to four gallons of water are required to be evaporated every minute. In the case of air cooling, no water is used, but there is a high requirement for electricity. Liquid cooling sits in between, generally reducing both water and energy use relative to evaporative systems once deployed at scale.
Two metrics are used industry-wide to measure this. The Power Usage Effectiveness (PUE) ratio, defined by ISO/IEC 30134-2:2016 standardization, equals total power consumption within the facility to the power consumed by computing activities, with lower numbers closer to 1.0 implying more efficient usage. The Water Usage Effectiveness (WUE) is a ratio devised by The Green Grid in 2011 and represents an annual water consumption in liters used for cooling and humidification divided by total annual kilowatt hours of IT equipment power consumption. The industry average WUE stands at approximately 0.5 to 0.7 liters per kilowatt-hour, with numbers varying significantly from one type of cooling to another and depending on weather conditions.
AI Data Center Water Usage in a Desert Environment
Water usage is arguably the more acute constraint in the Middle East, more so than power. Of the seventeen most water-stressed countries in the world, eleven are in the Middle East and North Africa. A one-megawatt data center relying on water-based cooling can consume up to 25.5 million liters of water annually, roughly equivalent to the daily water use of 300,000 people, and a 100-megawatt facility can consume on the order of 2.5 billion liters a year. Scaled up to gigawatt-class campuses like Stargate UAE or HUMAIN’s planned buildout, the water implications are significant enough that they factor directly into site selection and cooling technology choice.
The region does not have the freshwater reserves to support evaporative cooling at this scale using conventional municipal supply. This is why most new Gulf AI campuses are designed around non-freshwater sources. Neom’s Oxagon-based AI campus is explicitly designed to use Red Sea seawater for cooling, and desalinated or treated wastewater is increasingly specified in Gulf data center designs as a substitute for potable water in cooling loops. Reporting from the Middle East Council on Global Affairs also notes a regulatory gap: there is currently no GCC-wide framework setting mandatory water-use-effectiveness limits or requiring cooling designs adapted to extreme heat and water scarcity, which leaves much of the burden on individual operators and their engineering teams to self-impose those standards.
AI Data Center Cooling Systems and Design Technologies
Given the heat loads generated by modern GPU racks and the water constraints described above, the Gulf data center industry has moved toward liquid cooling faster than most other regions. Two approaches dominate current deployments.
With direct-to-chip cooling, cold plates are installed directly onto GPUs and other heat-producing hardware, through which liquid is circulated in order to extract heat directly from its source. It enjoys the fastest pace of adoption because it has higher compatibility with current rack and facility layouts, allows for densities of around 120 kW per rack, and is able to reduce the use of water by up to 20–90% in water-scarce areas relative to evaporative systems, as well as lower power requirements by 18%.
Immersion cooling involves completely immersing the server in a non-conductive liquid that will not conduct electricity but is highly efficient in absorbing heat compared to air. It is able to deal with high-density heating when compared to chip-cooling techniques and has the ability to save up to 91% of water and up to 50% of energy compared to air cooling. It requires more significant changes to server design and maintenance procedures, which has slowed adoption relative to direct-to-chip systems.
Beyond the cooling method itself, design and optimization decisions that matter in a Gulf climate include: orienting and shading buildings to reduce solar heat gain before it ever reaches mechanical systems, using treated wastewater or seawater loops instead of potable water wherever the site allows, specifying dry coolers or hybrid systems that fall back to air cooling during cooler overnight hours to reduce water draw, and building modular power and cooling capacity that can scale in phases as GPU generations and their power densities change. Facilities designed around a single fixed rack density risk becoming obsolete within a few hardware generations given how quickly GPU power draw has increased.
AI Data Center Connectivity in the Region
Power and water get most of the attention, but connectivity is the third constraint that determines whether a Gulf AI data center can actually serve global or regional customers with acceptable latency. The UAE, and Fujairah specifically, has become the largest submarine cable landing hub in the Middle East. Facilities such as Etisalat’s Smart Hub Fujairah combine a cable landing station with a carrier-neutral data center, connecting to more than 20 terrestrial and submarine cable systems and offering latency of around 30 milliseconds to a population base of more than 2 billion people across the Middle East, Europe, Africa and Asia.
That said, the region’s connectivity has a known vulnerability: an estimated 90 percent of Europe-to-Asia data traffic currently passes through Red Sea subsea cables and then overland through Egypt to reach the Mediterranean, largely bypassing the GCC countries entirely. This concentration of routes through a single geographic corridor is a resilience risk that new Gulf hubs, including Fujairah and emerging routes tied to new cable systems, are partly designed to diversify away from. Any AI data center strategy in the region needs to account for this concentration risk rather than assuming redundancy that may not actually exist at the regional level.
AI Data Center Infrastructure Bottlenecks
Put together, four bottlenecks recur across nearly every Gulf AI data center project currently under development.
Grid capacity and interconnection timelines are the most immediate constraint, since gigawatt-scale campuses require coordination with national utilities that can take years, independent of how quickly the data center shell and GPU hardware itself can be installed. Water availability ranks as the second priority because of the degree of water stress in the region and the quantity needed for cooling operations. This makes it likely that future projects will utilize seawater, desalinated water, or reclaimed water. Regulatory frameworks are a third bottleneck. According to Wood Mackenzie and other sources, however, the regulations around clean energy procurement in the UAE, as well as other similar markets, do not seem to be keeping up with the level of demand, and there are currently no GCC-wide standards for measuring water efficiency of data centers. Another challenge is that of staffing – the region requires a large number of professionals who can work on the design and maintenance of HPC/AI infrastructure and who know how to deal with both legacy data centers and GPU-intensive, liquid-cooled facilities.
What Infrastructure Is Needed for AI Data Centers
Summing up the above points, a well-functioning HPC/AI data center in the Middle East region must have the following components working together: a reliable supply of power through a contractual arrangement, in gigawatt quantities, and with a realistic possibility of having renewables or lower-carbon sources in future; a cooling system, mostly direct-to-chip or immersion-type, appropriate to the estimated density of the racks, and ideally based on the use of seawater, desalinated water, or treated wastewater rather than municipal freshwater; an ample fiber connection to the regional submarine cable landing stations to maintain competitiveness in latency for both training and inference; a scalable and phased design capable of accommodating successive generations of GPU hardware without a complete redesign; and an employment policy that addresses the current deficit in HPC/AI infrastructure engineers.
Some operators refer to facilities designed and delivered as a complete package, power, cooling, connectivity and shell built and commissioned together, as a turnkey AI data center. Whether a project is delivered that way or built in stages by separate contractors, the underlying engineering requirements described above do not change. A workable data center AI solution in this region has to be evaluated on all of these dimensions together. A facility with abundant power but no credible water strategy, or strong connectivity but undersized electrical infrastructure, will not perform as an AI or HPC facility regardless of how it is marketed.
Frequently Asked Questions
What is HPC/AI infrastructure, and how is it different from a standard data center?
The HPC/AI infrastructure includes the power, cooling, networking, and computational capabilities that have been built for handling workloads related to high-performance computing and artificial intelligence, typically GPU-based clusters. The major difference between the two is in terms of density and the amount of heat generated. In an enterprise-level data center, a rack consumes up to 5 to 15 kW of energy and uses air cooling, while a rack like Nvidia GB200 NVL72 for HPC/AI workloads will use up over 100 kW of energy and liquid cooling.
Why is AI data center power consumption so much higher than traditional computing?
AI training and inference workloads keep the GPUs fully loaded constantly, whereas the normal enterprise applications have lower utilization rates. Moreover, the AI-specific GPUs consume much higher amounts of electricity in comparison with general CPUs. Consequently, today’s AI-optimized racks consume 15 times more electricity compared to the same hardware used five years ago. Additionally, according to the IEA report, the electricity consumption of AI-specific data centers increased by 50 percent in 2025.
How much water does an AI data center actually use?
This largely depends on the cooling approach. A one megawatt plant using water cooling can require about 25.5 million liters of water per year, which is equal to the amount of water required by 300,000 people per day. Liquid cooling approaches such as direct-to-chip cooling and immersion cooling have been found to reduce water consumption by as much as 20% – 91%, compared to older evaporation systems.
Why is AI data center design in the desert so different from other regions?
Given the extreme ambient temperature (which is above 45 degrees Celsius), together with the presence of high humidity in the coastal areas of the Gulf, coupled with severe water stress in the region, the cooling technologies used in temperate climates would not work well or would be expensive in the Middle Eastern setting. Out of the seventeen water-stressed nations, eleven of them are located in the Middle East and North Africa region, hence the reason why seawater cooling, desalination, and liquid cooling technology have been used widely in new Gulf projects.
Who is building AI data centers in the Middle East right now?
The largest current projects are Stargate UAE in Abu Dhabi, a joint venture involving G42’s Khazna, OpenAI, Oracle, SoftBank, Nvidia and Cisco, designed to scale to 5 gigawatts; and HUMAIN’s data centers in Riyadh and Dammam in Saudi Arabia, backed by an estimated 77 billion dollars in investment with a roadmap toward roughly 6 gigawatts of capacity over the next decade. Neom’s Oxagon site in Saudi Arabia is also being converted into a seawater-cooled AI data center campus following a 5 billion dollar deal with DataVolt.
Which data centers actually count as AI or HPC facilities?
Data center facilities become AI or HPC data centers due to their power densities and cooling systems, but not just by their geographic location or ownership. Those that are built upon GPU racks with power consumption greater than 50-100 kW are referred to as AI or HPC data centers since they use liquid cooling solutions for their operations and are meant to be used heavily and continuously. In contrast, those data centers with less dense racks and using air cooling can’t process massive AI trainings.
What is a turnkey AI data center?
A turnkey artificial intelligence data center is an example of a fully integrated package that includes everything from the power supply to the structure itself. A turnkey data center is normally developed and delivered by one contractor or group of contractors. The term refers to the delivery approach and not any particular technology, meaning that such a facility has to be analyzed using all of the usual criteria that would be applied to any other AI data center.
What infrastructure bottlenecks are slowing AI data center growth in the region?
The most commonly cited constraints are grid interconnection timelines, since utilities need years of lead time to deliver gigawatt-scale power; water availability, given the region’s water stress; regulatory gaps, particularly around clean energy procurement and the absence of a unified GCC standard for water-use effectiveness; and a shortage of HPC/AI infrastructure engineers with experience in liquid-cooled, GPU-dense environments.
How important is connectivity compared with power and water?
It is a separate but equally binding constraint. A data center can have abundant power and an efficient cooling design and still underperform if it cannot deliver low-latency connectivity to the markets it serves. The UAE’s Fujairah hub is the largest submarine cable landing point in the Middle East, offering roughly 30-millisecond latency to a population of more than 2 billion people, which is a significant reason it has become a preferred site for regional AI infrastructure.