The property that makes fat-tree the right answer for AI training is not path count, switch cost, or even average latency. It is this: a fat-tree with full oversubscription ratio of 1:1 provides the same aggregate bandwidth for any traffic matrix.

Any pair of nodes communicating simultaneously can use their full uplink bandwidth. No traffic pattern, no matter how adversarial, can exceed the bisection. The torus has no such guarantee: cross-traffic between non-adjacent nodes shares links, and the worst-case traffic pattern saturates the torus fabric while the fat-tree remains uncongested.

This property is called non-blocking, or more precisely rearrangeably non-blocking in the fat-tree's case. For AI training where the communication pattern changes every step, having a topology that is oblivious to traffic pattern is enormously valuable.

The practical consequence for the BasePOD architecture from Chapter 7: rail-optimised design routes all AllReduce traffic for GPU-N on every DGX node through a single leaf switch. Each leaf switch's uplinks to spine provide full bisection bandwidth for that rail's AllReduce traffic. The spine provides bisection between rails for point-to-point and pipeline traffic. The fat-tree structure guarantees the bandwidth is there regardless of which GPUs are talking to which.