Batch is 50 percent off input and output at OpenAI, Anthropic and Google. Cache writes cost 25 percent (5-minute) to 100 percent (1-hour) over input list at Anthropic. Anthropic charges one rate across its full 1 million token context; OpenAI bills about 2 times input above 272,000 tokens and Google 2 times input above 200,000.

Frontier vendors do not publish cost, so their rent is bounded by open-weight hosting prices, not measured: flagship output at $20 to $30 and the top tier at $50 is 5 to 13 times a hosted open-weight large model. A seat converts the token bill into a flat $20 a month with a usage cap; no vendor publishes tokens per seat, so the seat rent cannot be measured from outside.

Three patterns. Only three rows own the chip; every other buyer pays rent to at least two layers above it. The rows with the thinnest margins, the application companies, run the most multiplicative workloads, so their cost of goods is the least forecastable. The enterprise adopter is the only buyer paying rent at all four upper layers at once, which is why its total cost of AI is three to four times its token bill.

Chips are chosen for training, batch and small-model work; terms are chosen for everything else. For every frontier workload the lever is layer seven, commercial terms. For every open-weight workload it is utilisation. Three rules of thumb: a newer generation is cheaper per token even at a higher hourly price (B200 is 2.2 times cheaper per token than H100 at index prices); a rented eight-chip H200 node beats hosted open-weight prices above about 13 percent utilisation; below that, or for any frontier model, the choice is terms.

Size is the multiple between worst and best case on that lever alone; bar length is on a log scale to 40×. The levers compound. Levers 1 to 5 and 8 to 9 need nothing from the market except measurement. Levers 6, 7, 10 and 11 pass through to buyers only where hosts compete. Lever 12 is the one no one yet provides at scale.