Does useful intelligence require frontier compute?
The commoditization thesis depends partly on whether economically sufficient intelligence remains physically scarce. Today's frontier systems can require enormous training and inference infrastructure. That fact alone does not establish the long-run cost of sufficient deployment.
Three quantities must be separated: the cost of training the frontier, the cost of serving the frontier, and the cost of the cheapest system sufficient for a particular economic task.
Economic benchmark = cost of sufficient capability, not cost of maximum capability
A possible equilibrium contains an expensive frontier used for difficult novelty and a broad commodity layer of local or inexpensive models used for routine work. Another equilibrium contains persistent frontier dependence because tasks remain open-ended and capability improvements continue to produce large economic differences.
The paper does not assume which equilibrium wins. It identifies the load-bearing variable: if sufficient intelligence is cheap and widely deployable, intelligence rents compress. If sufficient intelligence remains scarce and frontier-dependent, the provider's moat can be much stronger.