“Despite falling per-token inference costs, AI application companies have not yet seen gross margin improvements because the simultaneous advent of reasoning capabilities has driven a massive increase in token usage per task.”