AI labs face a commodity trap at the model inference layer: models are largely undifferentiated, switching costs are low, and economic theory (Bertrand paradox) predicts margins will compress toward marginal cost. Historical parallels from railroads, telecom, and airlines show infrastructure builders rarely capture the value they create. The path to durable profitability lies in moving up the stack — from raw inference APIs toward embedded enterprise deployments, AI-native SaaS, and eventually 'digital workers.' This migration borrows from enterprise software playbooks: embedding moats (data gravity, persistent memory), ecosystem moats (flywheels, marketplaces), commercial moats (multi-year contracts, vertical integration), behavioral moats (skill erosion, relational attachment), and outcome-based pricing. While this may rescue the labs financially, it raises serious concerns about enterprise lock-in, reduced competition, and concentration of economic power — concerns regulators have so far underweighted by focusing only on the bottom layers of the stack.
