The Sequence Opinion - Issue 926: AI Moats in the Age of Scaling Laws

Picture an AI lab pouring several billion dollars into chips, power, researchers, and data, then shipping a model that looks unbeatable. The benchmarks rise. Developers move over. The launch turns into a moment the whole industry talks about.

And then the wall starts to look less solid. Six months later, another lab reaches roughly the same level. An open model gives away much of the same capability for far less money. Distillation trims pieces of the original behavior into smaller systems. A router quietly sends each query to whichever model is cheapest, or best, that morning.

That is the uncomfortable shift here: the castle is still real, but the moat keeps changing shape. If intelligence itself can be reproduced, what exactly is the durable value of being first to it? The article leans on Hamilton Helmer’s Seven Powers framework to draw the line between a strong product and an enduring business. Power needs both a benefit and a barrier.

AI makes castles easily. Scaling laws have made capability more predictable: more compute, more data, more engineering, better performance. Not perfectly predictable, and not in some neat vending-machine way where one billion dollars buys one unit of intelligence. But close enough that capital now matters a lot, while also being easy to mistake for protection.

That is the real warning. In AI, spending heavily can buy speed, talent, and a headline-grabbing launch. It does not automatically buy a moat. The harder question is whether anything stops the next lab, the open model, or the router from turning that expensive lead into a commodity.