AI Scan

Today, margin lives at the bottom.

The bottom of the stack is priced by physics. Chips, power, and energized capacity are scarce, and scarcity sets the margin regime. Land + Power + Shell remains the most obvious and fastest path to cash-on-cash returns — and as data centers draw more pushback, energized land can explode in value.

One layer up, silicon is closed to newcomers: the performance demands are too high, manufacturing precision too complex, and supply-chain influence over adjacent components like memory is out of reach for a startup now. A lot of capital will be wasted here chasing yesterday's entry window.

The hinge is the model layer.

Models are complicated. The big open question is how much of the revenue they generate today comes from tokenmaxxing — usage inflated by poor model behavior. If it's a lot, annualized revenues diminish meaningfully even as token consumption inflects upward. Meanwhile the price of a served token keeps collapsing, with open-weights models setting the floor.

If models commoditize, margin climbs.

The scarce thing is no longer access to intelligence — it is the accumulated system around the model that makes intelligence yours. The harness hands an enterprise its alpha: its data, workflows, evals, and business rules. That creates very low, model-agnostic switching costs, and it's exactly what makes the model layer's revenue question so sharp.

Above the harness, applications go custom by design — alpha embedded, proprietary by nature — and the difference between off-the-shelf and custom software melts away.

The whole bet settles at the top.

Every layer below exists to be justified by the one above, so the macro question for the entire sector eventually resolves at the application layer: is there enough real, paid value at the top to justify the capex at the bottom? Today the gap is enormous — public estimates put AI application revenue around $19B (Menlo Ventures) against roughly $700–900B of estimated 2026 infrastructure capex, a spread on the order of 35–45x. Closing it is the job of the software and agents customers actually pay for.