A Perfect Router Can Create an Imperfect Ecosystem: Model Routing and the Direction of AI Innovation

Model routers are usually judged while holding available models fixed. We study what happens when traffic is economically material and providers choose capabilities in anticipation of future demand. Revenue, feedback, and learning then make traffic part of the private return to generalization or specialization. In a balanced two-provider model under the standard softmax routing rule, the unique equilibrium is homogeneous whenever routing responds to quality differences, while the global quality optimum specializes above an exact responsiveness threshold and approaches complete specialization. The tension persists locally with many providers, task domains, and nonseparable technologies. Capability-contingent transfers can correct investment without sacrificing current quality, but routing for future learning can remain necessary when traffic improves capability. A prespecified 45-arm QLoRA experiment finds that reallocating a fixed post-training budget produces domain-selective gains: math-heavy adaptation substantially improves MATH-500, alongside adaptation losses on GSM8K and the code benchmarks. The fitted GSM8K–MBPP frontier violates prespecified economic shape conditions, preventing an interpretable specialization threshold and illustrating the difficulty of measuring capability tradeoffs at this scale. Thus a router efficient for current queries can induce the wrong long-run direction of AI investment.