Research

We develop and estimate a dynamic structural model of patient migration and hospital quality investment in Taiwan's cardiac stent market, combining patient sorting across regions and hospital-product alternatives with hospitals' choices over DES portfolio breadth and pricing. The analysis shows that migration raises patient match value but can widen regional quality gaps by expanding inflow hospitals' market size and investment returns, while policy exercises point to sharp trade-offs and suggest that coordinated demand- and supply-side incentives are more promising than one-sided interventions.

Frontier AI models cost hundreds of millions of dollars to train, yet open-weight rivals can reproduce much of their capability within months by distilling their outputs. Whether this rapid imitation blunts the incentive to push the frontier, and whether the discipline runs through product-market rivalry or through appropriation of the leader's technology, has become central to the debate over open AI. We study the question with usage and posted-price data from OpenRouter, estimating a random-coefficients demand system, recovering closed-firm markups, and solving the Markov-perfect equilibrium of a quality-ladder investment game with gap-dependent catch-up. Open-weight competition does deter frontier investment, but the discipline operates almost entirely through appropriation rather than product-market rivalry: removing the open fringe speeds the frontier by only about two percent, while shutting the catch-up technology speeds it forty-four times as much. Rivalry matters so little because the open substitute is itself distilled from the frontier: as the leader advances, it is left chasing a moving target, falling further behind rather than keeping pace. The small gain from this faster frontier, however, is outweighed by what consumers lose: removing open weights is welfare-reducing, equivalent to a permanent rise of about 4.7% in the price consumers effectively pay for these models, and the verdict holds across the estimated elasticity band and at every horizon. The lesson for policy is not to restrict open models but to keep them near the frontier: the consumer-optimal lag is short, and a compute subsidy, which accelerates the frontier while raising consumer surplus, dominates restriction on the consumer-surplus margin, before its fiscal cost.