East-west bandwidth is where AI infrastructure economics are won or lost. A fabric that looks healthy at 40% average utilization can still drop training throughput when collective communication creates synchronized bursts across hundreds of GPUs.
What under-provisioning actually costs
- GPU idle time during all-reduce and all-to-all phases — directly measurable in job runtime.
- Delayed cluster expansion because new racks cannot be absorbed without congestion.
- Emergency fabric upgrades mid-program when workloads outgrow design assumptions.
- Operations toil from chronic congestion tickets that look like application bugs.
Plan for bursts, not averages
Capacity planning for AI fabrics should model worst-case collective patterns, ECMP hashing behavior, and failure-domain headroom — not steady-state charts alone. Instrument congestion before users report slow jobs, and tie acceptance criteria to workload-representative stress tests.