The UAE splits its ten-square-mile Stargate campus across the country after Iran named it a target

Somebody has already signed for capacity in Abu Dhabi this year. Stargate UAE's first phase is $30bn for one gigawatt, with 200 megawatts of it due to be running this year ^[12]. That works out at about $30m per megawatt, and the near-term 200 MW is a fifth of the phase ^[1][2]. Those numbers were drawn for one campus in the desert.

Distance is the part of the redesign that shows up inside the machine. 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 ^[19]. No figure for that loss has been published. If the job is one large training run, dispersal is a tax on it. If the job does not need every chip wired as a single machine, the same dispersal costs nothing, and the reporting does not say what share of the gigawatt falls on either side of that line.

The evidence is thinner than the redesign sounds. Reuters built the story on six people it did not name, among them a US official and a Western diplomat, along with industry executives ^[4]. Reuters reports the first phase is still going ahead, with modifications ^[14].

TNW notes that none of the protections under discussion is normal practice for a building whose job is running graphics chips ^[10]. That is the useful tell for anyone reading a regional capacity pitch: the vendor's own engineering vocabulary has changed.

G42 leads the project, controlled by Sheikh Tahnoun bin Zayed al Nahyan, with Nvidia, Cisco and SoftBank alongside ^[13]. The group has spent two years unwinding its Huawei relationship and opening its data centres to American oversight so it can keep buying advanced chips, and majority US ownership is reportedly next ^[17]. It has also been picking up capacity in the United States, including a converted office building in Minneapolis ^[18]. Abu Dhabi, meanwhile, has been telling investors that $100bn of AI capital sits in its financial centre ^[16].

Two questions sort out whether any of this lands on your roadmap. The first is whether your workload needs the whole cluster wired as one machine; you can answer that from your own job graph today, and it tells you whether a dispersed gigawatt is worth the same to you as a concentrated one. The second is who absorbs the hardening bill and the schedule slip when the masterplan is signed. Nobody outside the project can answer that one, because neither the money nor the time that hardening costs has been put in public ^[20].