AI infrastructure debt starts in the network.

Network economics are not a later-stage cleanup item.

A network sized for SaaS will throttle an AI cluster, and the cost lands on silicon you already bought.

Enterprise networks were built for a traffic pattern that is small, steady, and tolerant. Requests arrive independently. A retransmit is invisible to the user.

Distributed training is the opposite pattern. During synchronization every GPU exchanges gradients with every other GPU at once, and the step does not finish until the slowest transfer finishes. The traffic is bursty, synchronous, and east-west rather than north-south.

The tail sets the pace for every step.

On a fabric with meaningful oversubscription, those bursts collide. Congestion becomes packet loss, loss becomes retransmission, and the entire job waits on the tail.

The symptom reads as a GPU utilization problem. The cause is a fabric problem. Teams respond by buying more accelerators, which adds nodes to the same congested fabric and makes the tail worse. The invoice grows while the throughput does not.

The backend fabric is not an accessory to the cluster. It sets what fraction of the cluster you actually get to use.

Measure completion time, not only link utilization.

Size the backend fabric as non-blocking, measure completion time per training step rather than raw link utilization, and treat loss during synchronization as lost capacity rather than a network statistic.