The UAE is splitting up its $30bn AI data centre after Iran named it as a target

One enormous campus outside Abu Dhabi is becoming a network of smaller sites, some of them proposed underground, with air defences around them.


The flag of the United Arab Emirates flies from a pole against a blue sky with scattered cloud.
Image Credits Credit: © GHOSS via Canva.com

The United Arab Emirates has gone back to the drawing board on the biggest AI construction project in the Gulf, according to an exclusive report by Reuters.

The original idea was a single campus covering ten square miles of desert outside Abu Dhabi, roughly 26 square kilometres of computing in one place.

What is being drawn up now is a network of smaller sites spread across the country instead, on the reasoning that a target you cannot miss is a bad thing to build.

Reuters based the story on six people it did not name, among them a US official and a Western diplomat, along with executives in the industry.

We have not independently confirmed the new plan, and Reuters itself could not establish whether the revised masterplan has been signed off, or what the changes will do to the cost and the schedule.

In March, Iran fired missiles and drones at Gulf states hosting American forces. Two Amazon Web Services data centres in the UAE were damaged in that exchange, and services from them are still disrupted.

A month later, Iran’s armed forces put out a video that named Stargate a target, complete with a map showing the site sitting next to Al Dhafra Air Base, which is where American aircraft are stationed. It is difficult to think of a clearer signal, and the people building the thing appear to have taken it as one.

The protections now under discussion belong more to military engineering than to the data centre industry.

They include putting facilities underground, building the most sensitive military workloads inside mountains, ringing sites with air defences that can intercept drones and missiles or jam them electronically, pouring blast-resistant concrete, and duplicating the power and cooling so that losing one part does not take out the rest.

None of that is normal practice for a building whose main job is to run graphics chips.

Stargate UAE is the local arm of the venture OpenAI runs with Oracle and SoftBank. The first phase is a $30bn, one-gigawatt cluster, with 200 megawatts of it due to be running this year, and the full plan reaches five gigawatts.

It is led by G42, the Emirati AI group controlled by Sheikh Tahnoun bin Zayed al Nahyan, with Nvidia, Cisco and SoftBank alongside.

Reuters reports that the first phase is still going ahead, though with modifications. OpenAI has already paused Stargate UK over energy costs and regulation, so this is the second national project in the programme to hit trouble that has nothing to do with demand for the computing itself.

This has been coming for a while. When Iran released the video in April, we wrote that the threat put a question mark over the entire Gulf strategy, and the broader problem is one we have set out before: the region sold itself to the AI industry on cheap energy, capital and space, and then a war made the location itself the risk.

Abu Dhabi has been telling investors that $100bn of AI capital sits in its financial centre. Redesigning the flagship site to survive a missile is not the message that was being sold alongside it.

There is a second squeeze on the same project. G42 spent the past two years unwinding its Huawei relationship and opening its data centres to American oversight in order to keep buying advanced chips, and majority US ownership is reportedly next.

The group has also been picking up US capacity, including a converted office building in Minneapolis. Building smaller and further apart at home, while acquiring space somewhere safer abroad, looks like the same decision made twice.

Spreading a gigawatt across several sites is not a straightforward swap. Large AI training runs want their chips close together, because the cabling between them is part of the machine, and splitting a cluster across the map costs performance.

Hardening sites costs money and time. Neither of those bills has been put in public, and until the masterplan is finalised, nobody outside the project can say what the delay looks like.

Get the TNW newsletter

Get the most important tech news in your inbox each week.

Published
Back to top