Everyone remembers the gold rush. Capital remembers the railroads.
During the actual Gold Rush, thousands of men with shovels and pans showed up convinced they’d strike it rich in some river. Most of them didn’t. They went home broke, exhausted, and forgotten — while the real money accumulated elsewhere: the railroads, the logistics, the banking, the supply chains, the boring infrastructure that made the rush possible in the first place.
Leland Stanford didn’t pan for gold. He sold the rails everyone needed to reach it, and he named a university after the proceeds.
The wrapper economy
A great deal of today’s generative-AI application layer is panning for gold. A thin interface wraps a foundation model that someone else trained, on data that someone else owns, served from infrastructure that someone else operates. The product is real. The traction can be real. But the moat is fiction — because every defensible asset in the stack belongs to a layer beneath you.
If your only edge is the interface, your edge is a weekend of someone else’s engineering away.
Interfaces are the easiest thing in software to copy. The moment a wrapper demonstrates demand, the model provider, the incumbent, and a dozen funded teams can ship the same surface. What they cannot trivially copy is proprietary data — specifically, data about something the rest of the industry isn’t even collecting.
Where the rails actually are
The durable position in this cycle is not the prettiest chat window. It is ownership of a signal that models are starving for and cannot synthesize. Outcomes are commodity — everyone logs the click. Human decision behavior — the hesitation, the pivot, the prosody of doubt — is not commodity, because almost no one is capturing it at the edge and structuring it into something a model can learn from.
That is the railroad. Dominion is building it on purpose, while most of the market is still standing in the river.