There's a version of the AI build-out that treats it as a procurement race. Secure the GPUs, sign the cloud capacity, announce the programme. The hard part, in that telling, is getting your hands on the silicon everyone else is chasing.
The chips are the part that arrives on a truck. The constraint sits downstream, and it's more stubborn: power, cooling, water, and a grid connection measured in years.
A serious AI facility doesn't draw the load of an ordinary data centre. Pack a hall with accelerators and the power density climbs to something the building, the substation, and the cooling plant were never sized for. In this region that arithmetic gets harder, not easier. Forty-plus-degree summers turn cooling from a background overhead into a headline cost, and the obvious way to shed that heat leans on water the region already prices carefully.
This is where a lot of ambitious plans quietly wobble. The strategy costed the compute and the timeline. It rarely costed the megawatts, the connection queue, the cooling overhead in July, or the water. What's left is a facility that reads "AI-ready" on the brochure and is waiting on a substation in practice.
For anyone architecting or buying here, the useful questions aren't about the chips at all. What power envelope can this site actually deliver, and when. What does cooling cost at full load in a Gulf summer, not a spec-sheet average. How long is the grid connection, really, and is it sitting on the critical path the plan pretends it isn't.
The GPUs turn up in weeks. The substation doesn't. The bottleneck was never the chips.