The New Arbitrageurs — Summary
The earlier story about Linda, who prepared taxes for four hundred dollars a session, was about destruction. AI dissolves the information gap she stood in, the toll booth falls, and the value flows upward to whoever built the tool. This is the other half. AI does not only dissolve arbitrages. It creates them, and the entity exploiting the new ones is not a person. It is a system, moving at machine speed across thousands of markets at once, without fatigue, without ethical hesitation, without the friction that once kept arbitrage slow enough for humans to see.
Arbitrage used to require a person who could see what others could not, and that vision was expensive. The cost of human cognition kept spreads human-sized. AI removes every one of those constraints at the same time. The gap still exists. The thing exploiting it no longer has the limits that kept the gap small. That is not faster arbitrage. It is a change in what arbitrage does to the world. When the speed of exploitation crosses a threshold, the humans are not slower because they are worse. They run on different hardware, and they have been structurally excluded from territory they used to inhabit.
The harder move is the second layer. When AI closes the old gap, it simultaneously opens a new one at a level humans cannot reach. The tax tool that ends Linda’s livelihood can now see patterns across millions of returns that no individual preparer could perceive. The toll booth falls. The observatory rises. And only the entity that owns the observatory can see what it sees. These second-layer gaps are more durable than the first, because they are not about information asymmetry between people. They are about data-density asymmetry between institutions, and no amount of individual learning can close them.
When agents negotiate with agents, the arbitrage is not about information gaps between buyer and seller. It is about model gaps, whose agent holds the better model of the other party’s reservation price, and those questions map directly onto capital. The spread you did not know you paid goes somewhere specific, to whoever built the better model and captures the systematic difference across millions of transactions. At this level, arbitrage is invisible. It looks like market clearing.
The first-layer gains are real. Cheaper access to legal, medical, and financial understanding genuinely helps people long excluded from the knowledge economy. But the second-layer arbitrages do not distribute. They concentrate, not along lines of effort or contribution but along lines of data ownership and computational scale, which is to say capital, in a self-reinforcing cycle whose entry barriers rise faster than any small institution can climb. The finance sector, with the longest exposure to algorithmic arbitrage, did not become more distributed. It became more concentrated.
One wonders whether efficiency and extraction, at enough scale, become the same thing seen from different positions. Regulating the dissolution at least has a visible subject. There is a Linda. The second layer has none, only a pattern of value capture spread across millions of clearings, each looking correct. Perceiving it requires the same population-level analysis the governed have and the regulators do not. We are building the economy of machine-speed arbitrage faster than any way to understand what it is doing, to whom, and in whose interest. The new arbitrageurs are fast, patient, tireless, and invisible. They do not eat lunch at the diner. Whether what comes after looks like a market or something wearing a market’s clothes is the question this era will have to answer. We have not started.