The eight design dimensions every AI DC build resolves, applied to the smallest fabric that has every problem a 30,000-GPU cluster has.
An AI data center is not a normal data center with more cables. The traffic pattern is different. The power per rack is different. The failure modes are different. The skills required to operate it are different. Posts 1 through 8 of this series established the parts: traffic patterns, fabric topologies, congestion control, telemetry. Post 9 puts them in one place and asks the design question every engineer eventually faces: when the procurement order is empty and someone needs an AI data center, where do you start.
The honest answer is to start small. The smallest sensible AI fabric is four servers and 32 GPUs. It costs about USD 1.5 million in compute and 200 thousand in network. It fits in two racks. It has every architectural problem that a 30,000-GPU cluster has, just smaller. It runs real NCCL workloads. It teaches the operator what their automation, telemetry, and procurement pipelines actually look like.