Software once had a superpower: build once, distribute at near-zero marginal cost, with gross margins of 75-85% that made customer acquisition seem almost too good to be true. AI is eroding that advantage. Every inference call costs money, and every user interaction now carries a compute cost that scales directly with usage.

For the application layer, this is a direct per-unit cost — a real bill of materials that was previously incidental. This creates a fundamental tradeoff: cheaper models protect margins but risk losing users, while frontier models win on quality but bleed on economics. For the first time, margin and quality conflict on a per-unit basis.

Recent survey data from ICONIQ puts average gross margins on AI products at around 52% in 2026, well below the SaaS baseline. Usage-based pricing is replacing flat subscriptions as the rational response to variable cost-to-serve. Jevons paradox also looms — cheaper inference gets consumed by more calls per user rather than banked as margin. The SaaS playbook of burning cash on growth no longer holds.

Founders who build unit economics from day one will be better positioned when the market matures.