Regulated AI workloads here are caught between two rules that don't care about each other.
The first is about location. Sensitive and regulated data is expected to stay in-country, and increasingly the compute that touches it is expected to stay with it. So renting capacity wherever it happens to be free isn't on the table. The workload lives where the rules say it lives.
The second is about supply. The accelerators these workloads need are scarce and slow to source. Memory and packaging are the bottleneck, not the chip itself, and the largest buyers have already reserved most of what's being made. Even when you're cleared to build in-country, the hardware isn't sitting in a rack waiting for you. You queue.
The squeeze is where those two meet. The workloads with the strongest legal reason to stay local are the same ones competing hardest for compute that's in short supply everywhere. The permission and the hardware arrive on different timelines, and the gap between them is where projects stall.
In more mature markets you usually fight one of these at a time. Capacity is plentiful even if the rules are strict, or the rules are loose even if capacity is tight. Doing both at once, at this pace, is a regional speciality.
For a decision maker it reframes one question. "Why can't we just spin this up" has an honest answer: the thing you're allowed to build and the thing you can actually buy are governed separately, and neither bends for the other.
Plan for the queue, not just the architecture. Secure the capacity before you commit to the date.