The Cost of the Floor — Summary

A finance minister in Nairobi receives a briefing with the cost of training a frontier model on the first page, placed there because her staff assumed the conversation was about matching it. She recognizes a number being used to end a discussion rather than start one and asks a different question: not what the frontier costs, but what the floor costs.

The floor is not a frontier and not trying to be. Training a domain model costs a small fraction of frontier costs: data local and finite, model small enough for a university, training run measured in days. Inference is where the economics are sharpest. The district health system in Bhubaneswar spends per query what a frontier system spends on electricity alone. The floor competes by making the frontier unnecessary for most of what people need.

Hardware dependency cannot be avoided: chips manufactured in a concentrated supply chain reach everyone downstream. The floor reduces dependency to this layer and manages it through diversification rather than pretending it away. The capital to build a national floor, including domain models, local hardware, a composition layer, and epistemic monitoring, costs roughly what a mid-sized highway project costs.

Maintenance is harder than building. Public infrastructure has a familiar failure mode: built with political will, maintained or not with budget allocations competing against every other claim. A floor built and then starved is worse than one never built, because it creates dependency and then withdraws. Whether retained surplus funds maintenance depends on whether the floor generates visible enough value to survive the budget cycle. The minister calculated what the country currently spends on foreign AI services. The aggregate was large, growing, and entirely outbound. She wrote the floor number on the briefing cover in red pen. It was smaller by two orders of magnitude.