The quickest way to read a GPU cluster design is to ignore the GPU count and look for where it leaves the reference architecture.
NVIDIA publishes one for nearly every size. Even the names are specs. The HGX AI Factory design is built on a "2-8-9-800" server: two CPUs, eight GPUs, nine network cards, 800 Gb/s of network per GPU. It scales from 4 nodes to 128, with the switches, storage, cabling and software laid out at each step. Someone has built it at that shape and tested it.
So treat the reference as the default, and every difference from it as a decision that needs a reason.
Fewer network cards per server to save on optics. A storage platform that's already in the building. Both can be perfectly sensible. Each change still moves the design from something that was tested towards something that wasn't, and somebody ought to be able to say why that's worth it.
It works in the other direction too. A proposal that sits well above the reference for the workload it claims to serve deserves the same question. Why the extra fabric, and who asked for it?
Write the deviations down, one line each, with the reason next to it. A line with no reason is the first thing to cut.