That 1,000-fold price collapse in inference is one data point, but it's a 100% valid point to bring up against the idea of market stabilization. Your point about training costs being a much larger multiple of inference costs makes the stabilization claim less than a 30% probability for me.