How well do these workloads parallelize? Especially over customer-tier interconnects? What's stopping someone from picking up 100 of these to setup a cool littl

How well do these workloads parallelize? Especially over customer-tier interconnects? What's stopping someone from picking up 100 of these to setup a cool little 1.6tb-vram cluster? Whole thing would probably cost less than an h100.