Build a serving cost calculator

The calculator takes a week of traffic, 168 hourly token counts, and asks what each option would bill for it. The decision that matters most is sizing dedicated GPUs for the peak hour at your latency target and paying for them around the clock. Utilisation is then a result, not an assumption, and for office-hours traffic it comes out far lower than anyone guesses. Everything in this answer is computed in code from the lesson's prices and benchmark; no service is called.