The Cost of the Floor

TAM-CMN.12 · The Common Mind · The Approximate Mind

The number people reach for when they want to argue the floor is impossible is the cost of training a frontier model. They cite figures in the hundreds of millions of dollars, headed toward billions, and they ask how a mid-income country, or a coalition of low-income countries, is supposed to compete with that. The question sounds devastating, and the answer is that nobody is asking them to.

A finance minister in Nairobi has heard the number. It appeared in a briefing prepared by her staff, sourced from a technology publication that covers the frontier labs with the kind of attention usually reserved for celebrity, and it was placed on the first page because the staff assumed the conversation was about matching it. The minister is experienced enough to recognize when a number is being used to end a discussion rather than start one. She set the briefing aside and asked a different question: not what the frontier costs, but what the floor costs.

The floor is not a frontier. It is not trying to be. Its entire economic logic rests on avoiding frontier-scale costs at every layer, and the avoidance is not a compromise that produces a worse product. It is a design choice that produces a different one, targeted at different problems, running on different hardware, funded by a different calculus. The frontier is expensive because it tries to do everything for everyone at the highest level of generality. The floor is cheap because it does specific things for specific populations and does not carry the weight of everything else.

State the costs honestly, because the honest numbers are the floor’s strongest argument.

Training a domain model, the kind that knows agriculture or health or law for a specific region in a specific language, costs a small fraction of what a frontier model costs. The data is local and finite rather than global and infinite. The model is small enough to train on hardware a university or a national research institute can acquire. The training run takes days rather than months, and if the first attempt is wrong the second is affordable. A country that cannot spend two hundred million dollars on a single training run can spend two million on a domain model that serves its farmers better than the frontier ever will, because the frontier was never going to learn what cotton costs in Warangal this week.

Inference, the ongoing cost of running the model to answer queries, is where the floor’s economics are sharpest. A large frontier model requires expensive hardware to run, burns energy at a rate that makes data-center operators nervous, and costs money on every query. A small domain model running on local hardware costs a fraction of that per query, and the fraction shrinks as the hardware improves. The district health system in Bhubaneswar, running inference on machines that cost less than a car, spends per query what a frontier system spends on the electricity alone. The cost advantage is not incremental. It is structural, because the floor is built to avoid the frontier’s inference costs rather than to match them.

The floor does not compete with the frontier on cost. It competes by making the frontier unnecessary for most of what people need.

The hardware supply chain is the dependency the floor cannot avoid, and the series will not pretend otherwise. The chips that run inference, even for small models, are manufactured in a concentrated supply chain that flows through a few companies and a smaller number of fabrication facilities. A country that owns its models and runs its own inference is still buying hardware from someone else, and a disruption at the fabrication layer reaches everyone downstream. This is a real constraint, and it is not solvable at the floor level. It is solvable, if at all, at the level of industrial policy and international trade, and the floor’s honest position is that it reduces dependency to the hardware layer and manages that dependency by diversifying suppliers and maintaining strategic reserves rather than by pretending the dependency does not exist.

The capital required to build the floor, stated as a line item a finance minister could evaluate, is within the range of what countries already spend on public infrastructure. A national AI floor for a mid-income country, including domain models for health, agriculture, education, and public services, local inference hardware deployed at the district level, a composition layer, and an epistemic monitoring system, costs roughly what a mid-sized highway project costs. It is not free. It is not trivial. But it is not the figure that makes the floor sound impossible, which is the frontier’s figure, and the frontier’s figure is the wrong comparison.

The maintenance question is harder than the build question, and it is the one that determines whether the floor survives. Public infrastructure has a well-documented failure mode: it is built with political will and fanfare, and it is maintained, or not, with budget allocations that compete against every other claim on public money. Roads are built and then not repaired. Hospitals are opened and then understaffed. Schools are wired for internet and the connection lapses when the grant period ends. The floor will face the same pressure, and the honest answer is that a floor that is built and then starved is worse than a floor that was never built, because it creates dependency in the population it serves and then withdraws, leaving people who adapted to its presence stranded.

The minister in Nairobi knows this pattern. She has seen it in her own portfolio, where a maternal health initiative launched with donor funding ran for three years, produced measurable improvements, and collapsed when the funding cycle ended and the treasury could not absorb the cost. The clinics that had come to rely on the initiative’s supply chain reverted to the chaos that preceded it, and the mothers who had learned to trust the system learned instead that the system was temporary. The floor, if it repeats this pattern, will teach the same lesson: that public provision is a promise with an expiration date.

Whether retained surplus funds the maintenance is the question that connects this essay to the previous one. If the floor generates economic value, by keeping data local, by improving agricultural productivity, by reducing the cost of public health delivery, and if that value is measurable and attributable, then the floor funds itself the way a road funds itself: through the economic activity it enables. The textile manufacturer in Tirupur whose data compounds locally, the farmer whose yields improve, the clinic whose triage is faster, these are the returns the floor produces, and if the returns are large enough and visible enough, the budget case makes itself.

If they are not, the floor depends on political will, and political will has a well-known half-life.

I wonder whether the people who say the floor is too expensive have ever calculated the cost of not having one, measured in the surplus that leaves the country every year as rent on rented intelligence.

The minister in Nairobi did that calculation. Her staff, once she redirected them from the frontier number to the floor number, produced an estimate of what the country currently spends on foreign AI services across health, agriculture, education, and public administration. The number was not large in any single line item. It was large in aggregate, and it was growing, and it was entirely outbound: money leaving the country for services that could, in principle, be provided by locally owned infrastructure at a fraction of the cost. The floor would not pay for itself on day one. But the rent it would displace was already being paid, every month, to someone else, and that rent was compounding the wrong way.

She kept the briefing with the frontier number. She wrote the floor number on the cover in red pen. It was smaller by two orders of magnitude, and it was the one that mattered.

The honest summary is this. The floor is buildable at a cost most mid-income countries can afford. It is maintainable if it generates enough visible value to survive the budget cycle. It carries a hardware dependency it cannot eliminate. It is cheaper than the frontier by design, because the design avoids frontier costs at every layer, and the avoidance is not an apology. It is the point. A floor that tried to be a frontier would be too expensive, too concentrated, and too dependent on the same supply chain it was meant to circumvent. A floor that knows it is a floor, that does the specific work for the specific population and leaves the frontier for the thin slice that genuinely requires it, is affordable, and its affordability is its most radical feature.