Software's old math just broke: every new user is a parts order now

For fifteen years, software had a single, generous rule: build once, sell to a million, and every new user was almost pure margin. That rule ran the public market, the venture playbook, and the products in your browser. The rule is now broken for the apps that talk back, and the new rule looks less like SaaS and more like a factory floor.

The shift is one mechanism: every interaction a user has with an AI product spends real money on a model call before it earns any revenue. That appears to convert the software business from a near-zero marginal cost engine into a hardware-style bill of materials, where each unit shipped consumes a real, priced input. The tradeoff is mechanical. Route a query to a cheap model and you protect margin while risking a worse answer. Route it to a frontier model and you protect the user while compressing the unit economics. There is no setting that gives founders both.

The repeatable lever is per-query routing, plus caching, compression, fine-tuned small models, and pricing tied to compute instead of seats. ICONIQ's 2026 State of AI survey of roughly 300 software executives puts the new average gross margin near 52 percent, well below the 65 to 85 percent band that industry benchmarks still call healthy. That gap is also why usage-based pricing has stopped being an experiment and started becoming increasingly the way the math closes for many teams.

Software is now priced like a factory, bought like a product, and routed like a network.