City DC architects worry distributed training may stall even after adding GPU servers. Which AI environment component should they inventory first as a likely bottleneck alongside compute?

Inventorying the leaf-spine data center network fabric that carries east-west training traffic is the correct first bottleneck check alongside GPU compute. Distributed training depends on high-bandwidth, low-latency east-west paths between nodes; adding GPU servers without adequate fabric capacity commonly stalls jobs even when accelerators look strong. Network is a first-class AI environment component, not an afterthought.

The dedicated GPU interconnect that links GPUs inside a single training server fails because it handles intra-server GPU-to-GPU traffic rather than the inter-node east-west training traffic between servers across the fabric. The management network segment that transports telemetry from training nodes fails because it carries low-bandwidth monitoring traffic rather than the high-bandwidth east-west training traffic between GPU nodes. The batch scheduler queue policies that order queued GPU training workloads fail for the same reason: scheduler policy is not the distributed-training network path.

Exam caveat: fabric issues can look like "slow GPUs" in metrics if teams only watch accelerator utilization and ignore link saturation, PFC/ECN, or uneven job placement. Operational check: baseline east-west throughput and congestion on the leaf-spine during training, compare before/after GPU adds, and require network inventory in the AI component checklist whenever distributed training stalls after compute expansion.