A 5 GW AI infrastructure plan can satisfy every conventional site-selection requirement and still carry a physical exposure that power studies do not capture. The UAE’s reported decision to reconsider the configuration of its planned 5 GW AI infrastructure after regional attacks puts that gap into sharper focus. The project originally took shape around a large Abu Dhabi site, while current planning reportedly considers a network of facilities distributed across the country, with planners also reviewing underground construction and additional protection measures. The significance extends beyond one project because gigawatt-scale compute brings electrical, thermal, network, and operational dependencies into a physical infrastructure footprint that can also face external physical threats. Site qualification can therefore include physical threat exposure alongside the electrical, connectivity, land, and infrastructure considerations already used to evaluate large-scale sites.

Threat Radius Becomes a Siting Primitive

Power availability, transmission access, fiber diversity, water strategy, and land characteristics have traditionally shaped the early qualification of a hyperscale site. Physical threat exposure can introduce another screening layer by adding the site’s location and surrounding risk environment to the conventional evaluation of power, connectivity, land, and supporting infrastructure. The assessment does not require predicting an attack or assigning a single universal safe distance, because the practical objective is to understand how different site locations change the physical exposure of critical infrastructure. A site with greater exposure to a credible physical threat can therefore require additional protection measures even when its electrical and connectivity characteristics outperform another candidate site. This broadens the meaning of site proximity because distance from critical electrical and network infrastructure can be evaluated alongside the physical-security conditions surrounding a candidate site.

At 5 GW, the issue becomes more material because the facility is no longer a single collection of servers operating alongside conventional supporting infrastructure. A large AI campus can concentrate substations, transformers, cooling plants, fuel systems, network entry points, control systems, and compute halls within a connected physical zone, creating several opportunities for one event to affect multiple layers simultaneously. A credible threat assessment can therefore examine the dependency chain rather than limiting the review to the primary computing halls. Critical infrastructure outside the primary computing hall can become the effective exposure point when power conversion, cooling, communications, or control systems depend on equipment that sits in the same vulnerable area. The site-selection process can consequently evaluate physical exposure before detailed architectural decisions commit the project to a particular footprint.

The Greenfield Assumption No Longer Holds

Large greenfield sites have obvious advantages for hyperscale development because they provide room for power yards, cooling systems, security setbacks, logistics routes, expansion blocks, and large compute halls within one controlled boundary. That geometry also creates a substantial visible surface footprint when a project scales toward several gigawatts of capacity. The UAE project initially called for a 10-square-mile campus, demonstrating how land availability can support an unusually concentrated infrastructure strategy at this scale. That concentration can also place multiple interdependent structures within the same physical environment, increasing the importance of evaluating how external events could affect shared infrastructure. Site selection can therefore consider whether the advantages of horizontal expansion outweigh the operational consequences of concentrating critical systems within the same physical environment. The greenfield preference remains useful, but it can no longer operate as an automatic design assumption when physical survivability carries material business value.

Surface exposure also changes the economics of what initially appears to be an efficient site. A single expansive site can reduce duplicated infrastructure, simplify maintenance routes, and support common utility systems, yet those efficiencies can increase the number of functions that depend on the same physical zone. A damaged access route, electrical yard, cooling plant, or communications path can create consequences that extend beyond the directly affected structure when systems share common dependencies. Distributed construction introduces additional cost and operational complexity, but it can reduce the concentration of vulnerable assets and create more independent recovery paths. However, the objective should not become indiscriminate fragmentation, because excessive separation can introduce new dependencies across transmission, fiber, operations, logistics, and control systems. The relevant design question becomes how much physical concentration produces useful efficiency before it creates unacceptable correlated exposure.

From Centralized Scale to Distributed Mass

A single massive hall offers powerful economies of scale for construction, electrical distribution, cooling, operations, and equipment deployment. Those efficiencies become less compelling when one physical incident can affect a large percentage of the site’s available compute or its supporting infrastructure. A distributed model divides capacity across multiple structures or sites so that one event does not automatically translate into proportional loss of total capability. Reports on the UAE project indicate that planners are considering a network of facilities across the country rather than maintaining the original concentration in one large campus. This approach changes the unit of resilience from the building to the compute fleet, allowing operators to evaluate how much capacity can remain functional when one physical zone becomes unavailable. The architecture can instead use a set of operationally meaningful blocks rather than relying entirely on one dominant physical footprint.

Distributed mass does not mean simply placing identical buildings at random locations and connecting them to the same dependencies. Each block can undergo evaluation for electrical independence, cooling autonomy, network diversity, operational access, and control capability so that the consequences of losing one physical area remain bounded. Geographic separation only creates resilience when the systems that connect the separated blocks do not recreate the same point of failure through shared substations, common fiber corridors, centralized controls, or exposed utility infrastructure. The design target should therefore focus on independent failure domains rather than physical distance alone. Therefore, a distributed architecture can preserve more useful compute when each block has defined boundaries around the infrastructure that must survive with it. The result is a more granular resilience model in which operators can lose one block without necessarily losing the workload capacity attached to the entire project.

Below Grade as the Primary Protection Layer

Subsurface construction changes the physical relationship between critical equipment and external threats by placing structural mass between sensitive systems and the surrounding environment. Depth, surrounding material, structural reinforcement, controlled access points, and compartmentalization can reduce exposure to certain external events, although engineers must evaluate the protection level against geology, construction, ventilation, utilities, entrances, and the specific threat scenario. Existing underground data facilities demonstrate that subterranean environments can support high-security computing while also providing stable thermal conditions in suitable geological settings. The relevant design shift in the UAE project is that planners are considering below-grade construction as a physical-protection measure rather than simply an alternative location for equipment. Critical electrical, cooling, network, and control systems can also require corresponding protection analysis if the project intends underground placement to preserve computing operations during an external disruption.

A below-grade strategy also introduces constraints that can undermine its value if engineers treat depth as a standalone solution. Excavation, geological conditions, ventilation, heat rejection, equipment logistics, maintenance access, and expansion paths all require dedicated engineering when critical infrastructure moves below grade. Underground facilities can provide physical protection, but they can also create operational bottlenecks. Those constraints can become more significant when large equipment requires specialized access routes or when essential supporting systems cannot receive the same level of physical protection as the underground space. Compartmentalization therefore becomes an important complement to depth because protected compute can still depend on power conversion, cooling, communications, or access systems that sit in less protected areas. Meanwhile, a balanced architecture can place the most critical functions below grade while retaining maintainable surface infrastructure where teams can control exposure.

Survivability Measured by Miss Distance

Resilience at gigawatt scale can also address physical exposure that uptime calculations alone do not capture when an external event can affect multiple infrastructure dependencies at once. A conventional availability model can describe redundant power paths, backup generation, cooling capacity, and network diversity without fully describing the consequences of a kinetic event that affects several systems simultaneously. The UAE’s reported redesign discussions show how quickly a concentrated AI infrastructure strategy can face reconsideration when regional physical threats become more credible. Physical separation from credible threats can therefore serve as a useful planning consideration because it connects geographic exposure with the amount of critical function that may remain available after an external disruption. The metric does not replace electrical or network resilience calculations, but it adds a physical dimension to the question of how much capacity a site can realistically protect.

For C-level infrastructure decisions, the implication is less about predicting conflict and more about recognizing that physical exposure can become a business constraint before construction begins. Site qualification can incorporate physical threat exposure, geographic separation, protected infrastructure, compartmentalization, and recoverable compute alongside power, fiber, cooling, and expansion capacity. The most resilient 5 GW architecture may not be the one that places the greatest amount of compute on the most convenient site, because concentration can convert construction efficiency into correlated operational exposure. A distributed and selectively hardened design can require more capital and more complex operations while providing a more controlled path to preserving critical workloads during localized disruption. The new design question is consequently straightforward: not only how much capacity can a site host, but how much critical function can remain when disruption occurs within its surrounding threat environment.