Training — one run, then finished Inference — every answer, forever
Training compute is spent once, before launch, and never rises again. Inference compute starts at zero and accumulates with every answer served, crossing the total cost of training roughly a year after launch and continuing to climb indefinitely.

Training is a single bill. Inference is a bill that arrives every time anyone asks anything. The shape is the point here, not the numbers: the crossing happens, and after it nothing flattens out.