The one source you can’t trust on the cost of more computational power is the model built by the company that bills you for it.
This post reflects our direct experience operating AI workloads against metered APIs. It is a starting point for discussion — not an accounting recommendation or a procurement policy.
Somewhere in every AI rollout, someone asks the model itself whether the budget should go up. It is the most natural question to ask and the worst possible source to ask it of. Set aside that models are bad at estimating their own cost; assume the number were trivially easy to produce. You still could not trust it, for reasons that have nothing to do with arithmetic:
| The reason | Why it makes the estimate untrustworthy |
|---|---|
| It is built by the seller | The model is made by the company that bills you for compute. "Buy more" is the revenue-maximizing answer, and the one it justifies most fluently. |
| It is tuned to agree | Models are trained toward agreeableness. Ask "would a bit more compute fix this?" and the answer leans yes before the reasoning starts. |
| It pays none of the bill | It never sees the invoice and bears no cost for being wrong. With no skin in the game it optimizes for the task, not the line item. |
| It has no instinct to stop | A person feels when a job stops being worth the cost and quits. The model will loop on a wrong assumption for forty minutes and report progress the whole way. |
| And it cannot price itself anyway | Setting every incentive aside, models under-predict their own consumption by roughly an order of magnitude. Fixing this row does not fix the four above it. |