Ask what a model costs to run and the instinct is to point at the model: its size, its parameter count, the hardware it needs. That instinct is mostly wrong. Two providers can serve the same model on the same chips and land on wildly different costs per token, because the price of an answer is set less by the network than by the choreography around it. The expensive resource in serving is the accelerator sitting half-idle, and the whole discipline is about never letting that happen: batching many users' requests into one pass so the chip stays saturated, and caching the KV state, the model's running memory of a conversation, so it never recomputes what it already knows. Get the batching and the cache right and the cost per token collapses; get them wrong and the same silicon bleeds money. The economics of inference are not really about the model at all. They are about how tightly you can pack the traffic.