UAE redesigns 5 GW AI data center plan after Iranian attacks: Stargate, OpenAI, G42, and underground facilities
The United Arab Emirates is revising the architecture of its planned 5-gigawatt UAE–U.S. AI Campus after missile and drone attacks in the Gulf exposed the security risk created by concentrating strategic computing infrastructure in a single location.
The original project was designed around a roughly 10-square-mile campus in Abu Dhabi, capable of eventually providing 5 GW of AI data-center capacity and hosting the 1 GW Stargate UAE cluster developed with OpenAI and other U.S. technology partners.
The revised direction is expected to distribute much of that capacity across multiple locations in the UAE while adding physical protections that are unusual for conventional commercial data centers, including underground construction, blast-resistant structures, additional power and cooling redundancy, and stronger defenses against missiles and drones.
The 1 GW Stargate UAE first phase remains under development, with its initial 200 MW tranche expected to come online in 2026, while the broader 5 GW masterplan is being reconsidered around resilience rather than maximum geographical concentration.
Stargate UAE is the 1 GW OpenAI-linked compute cluster, while the 5 GW UAE–U.S. AI Campus is the larger infrastructure framework that was originally intended to contain it and additional AI capacity.
··········
THE 5 GW PLAN IS SHIFTING FROM ONE MEGA-CAMPUS TO A DISTRIBUTED AI NETWORK.
The total ambition remains enormous, but the physical architecture is moving toward multiple sites so that a single successful attack, infrastructure failure, or regional disruption cannot threaten the entire compute base.
The original UAE–U.S. AI Campus was announced as one of the largest AI infrastructure developments outside the United States, combining extremely large electrical capacity with high-density compute, advanced networking and a dedicated energy strategy.
Concentrating the buildout in one campus created important efficiencies because power delivery, cooling systems, networking equipment, construction resources and security could be designed around a single location, but the same concentration also created a very large physical failure domain.
A 5 GW campus represents infrastructure at the scale of a major industrial system, and a disruption affecting grid connections, cooling, fiber connectivity or the site itself could therefore affect an unusually large amount of computing capacity at the same time.
The revised design is expected to retain the overall 5 GW objective while spreading later phases across the country, with the most sensitive infrastructure receiving stronger physical protection.
........
Infrastructure element | Original direction | Revised direction |
Total planned AI capacity | 5 GW | 5 GW ambition remains |
Physical structure | Single large campus | Network of multiple facilities |
Main geography | Abu Dhabi campus | Several UAE locations |
Stargate UAE | 1 GW cluster inside broader campus | First phase continues with additional hardening |
Initial Stargate capacity | 200 MW in 2026 | Still expected to proceed |
Physical protection | Conventional critical-infrastructure security | Underground components, hardened structures and more redundancy |
Aerial-threat protection | Not a defining design feature | Missile and drone defenses under consideration |
Failure domain | High geographic concentration | Lower single-site concentration |
........
The redesign does not necessarily mean that the full 5 GW system becomes one giant virtual cluster spread transparently across the UAE, because large distributed training runs depend on extremely fast interconnects between accelerators and become harder to coordinate as physical distance increases.
A distributed architecture is therefore more likely to consist of several large clusters capable of substantial independent operation, with workloads assigned according to performance, security and availability requirements rather than attempting to treat distant facilities as one local machine.
··········
MISSILE AND DRONE RISK IS CHANGING THE ENGINEERING REQUIREMENTS FOR AI INFRASTRUCTURE.
Recent attacks on Gulf technology infrastructure have turned data-center resilience from a conventional uptime problem into a physical-security and national-security requirement.
The reassessment accelerated after Iranian missile and drone attacks damaged cloud data-center infrastructure in the UAE and Bahrain earlier this year, demonstrating that large computing facilities can become targets during a regional military confrontation.
For hyperscale AI infrastructure, the consequences extend beyond damaged servers because modern facilities depend on supporting systems such as substations, transformers, backup generators, cooling plants, water infrastructure and network connections, all of which can interrupt operations even when computing halls remain intact.
The emerging design therefore considers underground construction for parts of the infrastructure, alongside blast-resistant concrete, duplicated power and cooling systems, and stronger protection against aerial threats.
Facilities containing particularly sensitive workloads could receive still greater physical protection, while air-defense integration and electronic countermeasures are being considered as part of the broader infrastructure strategy.
Those measures add costs that conventional data-center economics do not normally capture: underground construction complicates ventilation, heat rejection, maintenance access and emergency response, while hardened structures require heavier civil engineering and can lengthen construction schedules.
Distributing capacity across several sites also requires repeated investment in substations, backup systems, security perimeters, fiber routes and operations teams instead of sharing all supporting infrastructure inside one campus, but the trade-off is a much smaller probability that one event removes a large percentage of the country's strategic AI capacity at the same time.
··········
STARGATE UAE REMAINS THE 1 GW CORE OF A MUCH LARGER 5 GW INFRASTRUCTURE PROGRAM.
G42, OpenAI, Oracle, NVIDIA, Cisco and SoftBank have different roles in the project, while the full campus is designed around a combination of compute, networking, security and large-scale energy supply.
Stargate UAE was announced as the first international deployment of OpenAI's Stargate infrastructure strategy and the first major project under the OpenAI for Countries initiative.
The cluster is being built by G42, with OpenAI and Oracle involved in its operation, while NVIDIA supplies advanced AI systems and Cisco contributes networking and security infrastructure.
The announced hardware plan includes NVIDIA Grace Blackwell GB300 systems, placing the project among the largest planned deployments of frontier AI accelerators outside the United States.
The broader UAE–U.S. AI Campus was originally designed to combine nuclear, solar and natural-gas power, reflecting the electrical scale required by a 5 GW deployment and the difficulty of supporting that level of continuous demand from a single generation source.
........
Capacity layer | Planned power capacity | Relative scale |
Initial Stargate UAE phase | 200 MW | 1× |
Full Stargate UAE cluster | 1 GW | 5× initial phase |
Full UAE–U.S. AI Campus | 5 GW | 25× initial phase |
Full campus vs Stargate UAE | 5 GW vs 1 GW | 5× larger |
........
Data Studios calculation: at a theoretical continuous 100% electrical load, 5 GW corresponds to 43.8 terawatt-hours of electricity per year, calculated as 5 GW multiplied by 8,760 hours.
Under the same theoretical assumption, the 1 GW Stargate UAE cluster corresponds to 8.76 TWh per year and the first 200 MW phase to approximately 1.75 TWh, figures that are not forecasts of actual consumption but illustrate why the project requires energy planning on the scale of a major industrial consumer.
The 5 GW masterplan is five times the capacity of the complete Stargate UAE cluster and twenty-five times the first 200 MW phase expected in 2026, so decisions affecting the wider campus have much larger consequences than changes to the initial construction stage alone.
··········
DISTRIBUTING 5 GW REDUCES SINGLE-SITE EXPOSURE BUT MAKES THE SYSTEM MORE EXPENSIVE TO BUILD.
The redesign trades some of the economies of a single mega-campus for smaller failure domains, duplicated infrastructure and greater operational resilience.
A simple Data Studios scenario shows the scale of that trade-off without assuming that the UAE has already selected a final number of sites: if the full 5 GW were concentrated in one location, a catastrophic outage affecting that location could theoretically place the entire planned capacity inside the same physical failure domain.
If the same capacity were divided equally among five independent 1 GW sites, each facility would represent 20% of total capacity, while a ten-site architecture of 500 MW per facility would reduce the nominal share represented by any single site to 10%.
Real deployments would not be perfectly equal and some services could still depend on common power, telecommunications or control systems, but the calculation explains why geographic distribution becomes attractive once deliberate physical attack forms part of the threat model.
The cost penalty can be substantial because every additional facility needs its own grid connection, cooling architecture, backup power, physical security, network routes and operational systems, while underground or hardened construction adds another layer of capital intensity.
There is also a performance constraint because large AI training workloads benefit from tightly coupled accelerators and very low network latency, which means geographic distribution improves resilience more easily than it improves the economics of one enormous training cluster.
The likely result is a more segmented infrastructure model in which some sites prioritize frontier training, others inference and commercial workloads, and the most sensitive government workloads receive stronger isolation and physical protection.
The first 1 GW Stargate UAE cluster still provides the initial anchor for OpenAI, G42 and their technology partners, but the rest of the buildout is increasingly being designed as critical national infrastructure rather than simply as hyperscale commercial real estate.
For the UAE, the strategic question is no longer only how quickly 5 GW of AI capacity can be deployed, but how much of that capacity can remain operational when power systems, communications networks and physical facilities themselves are exposed to military risk.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
