NVIDIA's argument this week: in the agentic era, continuous post-training becomes a central compute workload, and the metric that matters is intelligence per dollar. Unpacked, that means models stop being trained once and shipped. They live inside a permanent loop of rollouts, reward checks, and weight updates. Training turns from an event into a process. And every stage of that loop moves light. Between GPUs, between racks, and between sites as updated weights ship out to serving fleets. NVIDIA wrote down why the demand keeps running. My last piece counted the physical invoice: what Korea's 8.4GW converts to in fiber strands, who actually makes the glass, and why faster transceivers do not mean less fiber. If training never stops, neither does demand for the road the light travels on. InP makes the light. Fiber carries it.