Routing across the whole model market — and absorbing every migration — is a defense the labs can't copy

A defense the labs structurally can’t copy: the vertical company picks the best model for each sub-task across the entire model market — including a competitor’s model or an open-source fine-tune for the narrow piece where it’s actually best — while a lab can only sell you its own model and tell you to migrate.

Schmidt frames two jobs here. First, absorbing migration: re-running evals on every upgrade, recalibrating prompts for the customer’s edge cases, rolling out without breaking production — “The Rest of Oz company absorbs the migration.” Second, inverting the price curve: “The labs price the floor — the least intelligence available at $X. The Rest of Oz company sells the inverse — the lowest dollar cost for the specific level of intelligence the workflow actually requires,” which is only possible if you know exactly what each sub-task needs. This is the productized, customer-facing version of AI is the computer — orchestration across 19 models is the product, not any single model; it is why Vertical models beat frontier models in their domain — specialization wins on every metric and why Open harnesses with customer-owned databases are the antidote to model-provider lock-in keep the model swappable underneath.