The Gulf is building AI compute at a scale that's hard to picture. Stargate UAE is a 1 GW cluster in Abu Dhabi, the first 200 MW due this year, on a campus planned for 5 GW. In Saudi Arabia, HUMAIN is starting with 18,000 GB300 GPUs on the way to several hundred thousand over five years.
NVIDIA calls these AI factories, and the name is fair. The product is tokens, and the whole site is built to turn a fixed supply of power into as many of them as it can.
So a CIO in Riyadh or Dubai reads the headlines and asks the obvious question. Do we need one?
Almost certainly not. These are provider builds. They exist to sell compute to a lot of tenants at once, which is why they're assembled from NVL72 racks, each one 72 GPUs in a single NVLink domain and well over 100 kW of load, then stacked into SuperPODs.
An enterprise running its own models usually lives somewhere much smaller. A handful of HGX servers, eight GPUs each. Often one. My bet is that most enterprise inference never needs to leave a single node, and the work is knowing when yours does.
In between sit the neoclouds, GPU-only providers that rent capacity to anyone who needs more than a few servers for a while, but not forever.
So the question worth asking is narrower than the headlines. What does the workload actually need, and at what point does it outgrow one box?
Leave the factories to the people selling the output. Buy for the workload you have.