Where the Surplus Goes

TAM-CMN.10 · The Common Mind · The Approximate Mind

A textile manufacturer in Tirupur, Tamil Nadu, uses an AI system to optimize his dye formulations. The system is good. It saves him four percent on chemical costs per batch, reduces waste, and shortens the trial runs that used to eat a full day of production. He pays for it through a subscription that costs less per month than the salary of the chemist he no longer employs full-time. By every measure he uses to evaluate a business decision, the system is worth what it costs.

The question he has not asked, because it is not the kind of question a textile manufacturer in Tirupur is expected to ask, is where the value goes. Not the value he captures, the four percent, the reduced waste. The value the system captures from him. His formulations, his process data, his trial results flow into a system owned and operated by a company whose servers sit in Virginia, and the patterns extracted from his data, aggregated with the data of every other textile operation using the same service, become part of the system’s intelligence, which the company sells back to the next customer at a price that reflects the accumulated knowledge of the whole. The manufacturer pays for a service. What he gives in return is not only money. It is knowledge, and the knowledge leaves Tirupur and does not come back.

This is the macroeconomics of centralized AI, and it operates at a scale that makes the individual transaction invisible. Every query sent to a rented system is a transfer. The user gets an answer. The system gets the data around the question: what was asked, in what context, how the answer was used, what followed. The surplus from this exchange, the intelligence that accrues from aggregating millions of such transfers, concentrates at the center, in the company that owns the model and the infrastructure it runs on. The center grows more capable with every interaction. The periphery, the place where the questions originate and the work happens, remains a customer.

Rent extracts. Ownership compounds.

This is the structure of a rent relationship, and it is the oldest structure in economic history. The landowner who rents to the tenant captures the surplus the tenant’s labor produces, and the tenant, however productive, remains a tenant. The colonial power that extracts raw materials from the periphery and sells finished goods back captures the value-added at the center, and the periphery, however resource-rich, remains a supplier. The pattern recurs because the mechanism is the same: the party that owns the means of transformation captures the surplus, and the party that supplies the inputs gets paid for the inputs and nothing more.

Centralized AI repeats this pattern at the cognitive layer. The inputs are data, queries, context, the texture of a billion interactions that tell the system what the world is doing. The transformation is the model’s ability to learn from those inputs and produce outputs that are worth paying for. The surplus, the growing capability that emerges from the transformation, stays at the center. Tirupur’s textile knowledge becomes part of a system that Tirupur does not own, does not govern, and cannot interrogate. The manufacturer pays his subscription and receives his four percent, and the system that sold him the four percent now knows more about textile manufacturing than any single manufacturer ever will, and that knowledge belongs to Virginia.

Locally owned inference changes the direction of the flow. When the model runs on hardware the district or the country owns, the data does not leave. The patterns extracted from that data stay local. The capability that accumulates from the aggregation of local queries enriches a system the local population controls, and the surplus, the growing intelligence of the system, compounds inside the community rather than flowing out as rent.

The difference over time is not marginal. It is the difference between a development path and a dependency. A country whose AI capability is rented will, with every passing year, owe more to the center for the same service, because the center’s capability grows and the alternative to renting shrinks. A country whose AI capability is owned will, with every passing year, become more capable itself, because the surplus from its own use compounds locally, and the distance between what it can do and what it needs to rent narrows.

Compounding is the operative word, and it is worth slowing down on because it carries the full weight of the economic argument. When the textile manufacturer’s data compounds inside a locally owned system, the system gets better at textile manufacturing in Tirupur. It learns the specific dyes available from local suppliers, the water chemistry of the region’s wells, the color tolerances of the buyers in the European markets this cluster exports to. That knowledge makes the next manufacturer who uses the system more productive, and his productivity generates data that makes the system better still. The cycle runs locally. The intelligence stays in the district. The improved capability is available to every manufacturer in the cluster, not sold back to them by a company that extracted it.

In the rented model the same cycle runs, but the compounding happens elsewhere. Tirupur’s data improves a global model that becomes better at textile manufacturing everywhere, and the improvement is sold back to Tirupur at a subscription price that reflects the global system’s growing value. The manufacturer is funding his own competition’s improvement and paying for the privilege. He does not experience this as exploitation, because the four percent is real and the subscription is affordable. But the structure is the structure, and over a decade the system that learned from Tirupur knows more about Tirupur’s own industry than Tirupur does, and that knowledge is held in Virginia.

This is where the regional coalition argument finds its constructive form. A single country, if it is small, may not generate enough queries, enough data, enough interaction to build a floor whose capability compounds meaningfully. But a coalition of countries, sharing a regional model infrastructure, pooling data across a population large enough to train on, can. West African nations sharing an agricultural model. Southeast Asian nations sharing a trade and logistics model. Caribbean nations sharing a climate and disaster-response model. The coalition does not require the members to be identical. It requires them to share enough that the aggregated intelligence serves all of them, and to govern the shared infrastructure so that the surplus stays within the coalition rather than leaking to the center.

The coalition model also answers the sovereignty question in a way that neither pure national ownership nor pure market provision can. A small country that builds alone may lack the data to build well. A small country that rents from the frontier has capability but no ownership. A small country that joins a regional coalition has both: capability that compounds through the pooled data, and ownership that is shared among peers rather than held by a distant corporation. The coalition is not a substitute for national capacity where national capacity exists. India does not need a coalition to build an agricultural model. But most countries are not India, and for them the coalition is the realistic path to a floor they actually own.

There is a role for the indigenous private champion in this, and the role is specific: the champion builds the models the floor runs on. A national AI company, privately owned and commercially motivated, develops the capability. The public floor licenses that capability, governs its deployment, and guarantees access to all. The company profits from the license and from the premium market that sits above the floor. The floor reaches the population the company would not have served on its own. The relationship is not charity and not subsidy. It is the same relationship between the grid and the power company, between the road and the freight carrier, between the rail and the private apps that the India Stack essay described. The private champion is the floor’s foundry, not its substitute.

I wonder whether the textile manufacturer in Tirupur would make the same decision about his subscription if someone showed him, in terms he could hold, where the knowledge goes after it leaves his factory.

He would probably keep the subscription. The four percent is real, and the alternative, building his own system, training his own model, maintaining his own infrastructure - is beyond what a mid-sized textile operation can do alone. This is the honest position, and the floor does not require him to become a technologist. It requires that someone, a government, a coalition, a public institution that is accountable to him rather than to shareholders in another country, builds the infrastructure that would let his data compound locally instead of compounding in Virginia. He does not need to understand the architecture. He needs the surplus to stay.

The dye formulation he optimized this morning will run through the system and join the aggregate. The four percent he saved is his. The knowledge the system gained is not. Somewhere in Virginia a model knows a little more about textile manufacturing in southern India, and the model’s owners are a little richer in a currency more durable than money, which is the intelligence extracted from a place they will never visit, about work they will never do, for people they will never meet.