From the 32-GPU pilot to the canonical production pod. What changes, what stays, and why the numbers finally resolve cleanly.
Post 9 ended at a 32-GPU pilot: 4 servers, 2 leaf switches, 2 spines, 27 kilowatts, USD 200 thousand of network. The pilot proved NCCL runs at line rate, gave the operator a clean qualification lab, and did not lock in any bad design choice.
Post 10 is what happens next. The operator’s training team asks for 8 times the compute. That is 256 GPUs. It is not “the pilot but bigger.” It is the smallest deployment where rail-optimized topology becomes worth its complexity, where the storage fabric splits off from the compute fabric, and where the operator’s automation gets its first real workout. It is also the pod that stamps out at 1024, 8192, and 32000 GPUs without changing shape.