For AI workloads, inter-chip data-transfer speed often limits completion time more than raw computational speed.

“These chips have to constantly send data back and forth with each other, and in AI, the speed of that traffic, not the raw computation itself, is often what limits how fast the workload finishes.”