If you think about how these providers deploy a cloud facility, it makes sense that the offerings in a given place are relatively static. The whole network design, thermal/mechanical design, and floor plan is built with certain assumptions and they can't just go in and rack up some new machines. It evolves pretty slowly and when a facility gets a new machine it is because they refresh the whole thing, or a large subset of it.
That said, the EPYC machine type is available in 12 zones of four different regions in the US, which isn't bad.
Usually you would have some number of enclosed aisles of racks make up a deployment pod.
You can usually customize machine configuration within a deployment pod while staying within the electrical and thermal envelope of the aisle and without changing the number of core-spine to pod-spine network links.
You could potentially build out a data hall but not fully fill it with aisles. As demand starts to trend up you can forecast two quarters into future and do the build-outs with just one quarter lead time.
I would expect very large operators to have perfected this supply chain song and dance very well.
That said, the EPYC machine type is available in 12 zones of four different regions in the US, which isn't bad.