Started a small position in $CIEN today. Grid interconnection queues are 3-5 years, no utility can deliver 1GW contiguous in one place, and frontier training runs cannot wait. So OpenAI, Anthropic, Google, xAI and Meta are all moving to the same playbook: secure 1-3MW pockets where power is available, deploy fast, then expand to 200-500MW slices across 5-10 campuses in a metro or region and train one model across them, exactly what $CIEN sells as scale-across: making geographically separated GPU clusters behave like a single logical supercomputer. That shift makes interconnect count grow exponentially, not linearly, 10 sites means 45 high-capacity links, each needing tens of Pb/s day-one with deterministic low latency, or training efficiency collapses. It also flips the network from best-effort DCI to part of the training fabric. $CIEN is the only Western vendor with the full stack for that.