__________________
__________________
Frontier

Sometime in the last few years, the AI industry stopped saying "foundation model" and started saying "frontier model." And it is incredibly revealing.

A foundation is a base. It promises rent from everything constructed above it.

"Frontier" describes something else entirely. Nobody owns one; you temporarily occupy it, at increasing cost, and you re-earn the position with every expedition. Stop spending and the frontier moves on without you. One word describes an asset, the other describes a treadmill.

Companies do not usually trade the durable noun for the kinetic one. A foundation sounds like the better business and implies scarcity. So why change?

Because the scarcity is not there. Profits migrate to whatever layer of the stack is still scarce, and abandon whatever layer has become good enough. For the closed-source labs, the horizon is bright and the ground is crumbling behind them as open-source labs keep pace and catch up.

"Frontier" cements this framing and justifies the capital treadmill to investors, who believe the spending is the moat, hoping for a moment of increasing returns from recursive self-improvement. A race is a business model only if the prize at the end is worth much more than the cost of running.

More subtly, the change denotes fear. In the nineties, AOL and the internet were one and the same. They were the front door to the world wide web and helped new users wrangle an unknown landscape. But as the ecosystem matured and competition increased, AOL and the internet diverged into two separate things. AOL’s distribution was no longer scarce, and as a result, its ability to seek rent vanished. The prize at the end of the race was subjected to extreme margin compression.

Right now, the frontier of agentic coding (which has led to unprecedented revenue growth) is the prize, but not forever. Falling prices continue to absorb less demanding tasks, where competition from “good-enough” alternatives destroys profits. The emergence of powerful open weight models existentially threatens the purchase of profits within agentic coding (without an obvious successor).

The labs are wagering that in AI, the product is never-good enough: demand for capability is unbounded, so the frontier never closes, so the “good-enough” deathray never catches them. If they are right, holding both the input and the demand relationship produces a monopoly of a kind the internet never allowed. If they are wrong, they are running up a tab with a payday that may never come.

Either way, the words already tell you about the odds, and perhaps what the labs think about their own standing.

"Foundation" describes a landlord. "Frontier" describes the search to become one.