A roofline-model teardown of why MoE's sparse routing wrecks GPU utilization during decode, and MegaScale-Infer's fix: physically splitting attention and expert computation onto independently scaled GPU pools so attention replicas can pool enough requests to keep experts saturated.