Cognitive mercantilism comes for training data.

Venkatesh Rao published an essay this week, Cognitive Mercantilism, that gave me a name for something I keep running into in conversations about AI data.

It is worth reading in full. Here is the short version—and then where I think the frame fits the data market, where it does not and what it suggests Archive Signal should actually build.

Intelligence may be a market, not a hoard.

Rao draws on historian Joyce Appleby’s account of how capitalism emerged. Before it, he says, the dominant worldview was mercantilism: wealth as a fixed stock to be controlled, ordinary people as factors of production rather than consumers, and an economics focused on supply instead of demand.

In Appleby’s telling, capitalism took off partly because demand grew. Ordinary people wanted new goods, worked harder to get them, and set off a cycle of rising productivity and rising consumption.

Rao maps that shift onto AI. Cognitive mercantilism treats intelligence as a scarce stock—held first in elite minds and now in models—to be hoarded, rationed and controlled. Much of the anxiety about masses of economically redundant people quietly assumes demand for intelligence is fixed. Rao calls the mirror-image fear AI Malthusianism.

If demand instead grows as intelligence gets cheaper, the question changes. It is no longer only how much intelligence exists or who controls it. It is how many new uses people can find for it.

Rao also makes a sharper historical point: mercantilism was never really the philosophy of merchants. It belonged to people who wanted to become landlords of commerce.

Parts of the data market still look mercantilist.

Some buyers seek exclusivity and pay a premium for it. A dataset’s value is often described in terms of who else cannot have it. I have seen that instinct directly in conversations with buyers.

There are good reasons for some of it. Scarcity matters. Rights matter. A corpus competitors cannot reproduce may create a real advantage.

But the broader market is not behaving like a small group dividing a fixed pile of bullion.

It is filling with competing intermediaries: data marketplaces, wind-down specialists, expert networks and companies building evaluation environments. Forbes reported in April that SimpleClosure had processed nearly 100 deals on behalf of defunct companies in the preceding year, typically recovering between $10,000 and $100,000 per company.

Anthropic’s October 26 S-1 provides a different signal at much larger scale. The company disclosed that its vendor costs for data quality and performance-evaluation services increased by $251 million year over year in the six months through June 2026. Slator reads that as evidence of a growing commercial market for third-party data and evaluation work.

There is a limit to what the number tells us. Anthropic does not identify which activities drove the increase. It is evidence that frontier buyers pay heavily for data and evaluation services—not proof that they are buying enterprise archives.

Still, this looks less like bullion in a vault and more like buyers purchasing an expanding flow of work and evidence from many suppliers.

The supply nobody is pricing is still inside operating businesses.

The mercantilist lens is most useful on the supply side.

Some of the records most useful for training and evaluating AI agents may not be in any marketplace. They sit inside operating businesses: an inspection archive with the failure record attached, a service history where every job closed with an outcome, a sales operation where each conversation ended in a decision.

These records were not created as data products. They are the exhaust of skilled work, and they contain something many corpora lack: what an expert saw, what they did and what happened next.

To the businesses holding them, the records are usually a storage cost or a compliance burden. To a buyer thinking in terms of hoards, they are invisible. They are not listed anywhere, and they cannot be acquired by strategy memo.

Surfacing them takes owner trust, patient archive work and a serious path through privacy, trade secrets, worker expectations, customer obligations, consent and rights. Internal records are not inert raw material. They describe real people and real businesses.

That makes the supply harder to originate. It may also make a trusted supply relationship more defensible than a one-time transaction.

A corpus is a stock. A supply relationship is a current.

If Rao is right that demand for intelligence grows as its cost falls, the durable position may not be owning a static pile of data.

It may be relationships with owners who keep producing joined records—and the rights, consent and operating discipline required to use those records well.

A corpus is a stock. A supply relationship is a current.

The hoard still matters. A clean historical archive can be valuable. But the bigger business may be the pipes connecting living companies to recurring demand: new edge cases, fresh outcomes, changing equipment, corrected judgments and evidence that a system actually worked.

That is an argument, not a result. It needs testing against what buyers will pay for, what archive owners will permit and whether the economics reward the work required to make the records usable.

Where Archive Signal stands.

Archive Signal is early.

We are working on how business owners can understand what their records are, get consent and rights right and, where it makes sense, connect with buyers who can use them. We do not offer a dataset today, and nothing here is an offer of a licence.

We think the work is to be a merchant in Rao’s sense: not a landlord sitting on a hoard, but an operator finding new uses for something others treat as a cost.

So is intelligence a stock to control—or demand waiting to be discovered?

The data market will be one of the first places we find out.