UAE가 5GW AI 파크 배치를 중앙 집중식에서 분산형 다지점 구조로 개편합니다.

Reuters, citing six informed sources, reports that the UAE is revising its original plan to construct a 5GW artificial intelligence data center park in Abu Dhabi. Rather than a single facility spanning approximately 10 square miles, the project will adopt a multi-point layout distributed across the country. As one of the largest initiatives of its kind outside the United States, the endeavor is led by G42 in partnership with NVIDIA, OpenAI, Oracle, Cisco, and SoftBank. In response to the report, G42 stated that the project is progressing as planned, noting that specific designs are continuously reviewed in accordance with the safety and resilience standards required for critical infrastructure.

The initiative was formalized in May 2025 when leaders of both nations announced the UAE-U.S. AI Accelerated Partnership and its associated campus. Designed to deliver low-latency computing to nearly half of the global population within a 3,200-kilometer radius, the 5GW capacity will draw from a mixed energy portfolio comprising nuclear, solar, and natural gas power. Within this framework, Stargate UAE comprises a roughly 1GW cluster, with its initial 200MW phase slated for delivery in 2026. According to informed sources, construction on the first phase will likely proceed but will incorporate adjusted protective measures. The remaining capacity is expected to be distributed across multiple sites, with evaluations underway for underground or mountainous locations.

This strategic shift follows heightened risk assessments regarding centralized critical facilities amid ongoing conflicts in the Gulf region. Reports indicate that officials initiated a re-evaluation of site selections following recent strikes in neighboring countries. Despite the broader revision, industry executives confirmed that the first phase—consisting of two buildings rated at approximately 100MW each—has already been topped out, with equipment currently on-site. Meanwhile, Khazna is already operating roughly 300MW of capacity within the UAE, with nearly 370MW currently under construction. While centralized parks historically streamlined power distribution and operations by co-locating computing hardware and chip inventories within the same footprint, the new approach prioritizes geographic dispersion.

The restructuring aligns with divergent priorities: U.S. hyperscale cloud and AI model developers seek to position inference workloads at nodes aligned with their respective time zones, while the UAE aims to establish computing exports as a cornerstone of its economy beyond traditional oil and gas revenues. A multi-point architecture inherently lengthens transmission and cooling networks but significantly diversifies operational risk. Crucially, existing chip export licenses and the $1.4 trillion investment commitment to the United States remain anchored to the same diplomatic framework, effectively transforming the project’s blueprint from a static map into a dynamic network.

At its core, the revision reflects a market-driven repricing of sovereign risk. On the demand side, model and cloud vendors require uninterrupted inference capacity; on the supply side, developers must ensure the 5GW portfolio remains insurable and compliant for export across host jurisdictions. Consequently, capital allocation is shifting from monolithic park construction toward multi-site development and structural reinforcement. Local operators capable of securing power across multiple emirates stand to benefit, whereas entities whose power supply and land-use strategies were predicated on a single centralized park face mounting pressure. Ultimately, this development was triggered by a Reuters disclosure rather than a new groundbreaking ceremony. The 5GW target remains unchanged, but the geography has shifted. Once computing power is classified as critical infrastructure, site selection must account for potential strike exposure rather than relying solely on electricity pricing.