毓秀讲堂 |【Seminar 预告】No.687--张龙天 (‌中央财经大学‌)

时 间：2026年6月4日（Thur.）

14:00 - 15:30 p.m.

地 点：同德楼111

主 讲：张龙天 (‌中央财经大学‌)

Human-generated training data is approaching exhaustion, and synthetic data produced by AI models already supplies most frontier training corpora. This reliance is self-limiting: successive generations trained on their predecessors' output drift from the real-world distribution, a phenomenon known as model collapse. Computing power is the underappreciated input through which synthetic-data fidelity can be sustained. We develop a dynamic general-equilibrium model in which consumer data and compute protect data quality while the accumulated stock of synthetic data erodes it. Our central finding is that computing power is systematically undersupplied in every decentralized regime through the compounding of three monopoly wedges, and that this underprovision, rather than the privacy or variety margin, constitutes the first-order source of welfare loss in the AI economy. Two corollaries follow: private entry into the AI sector falls short of the social optimum because entrants capture only a fraction of the love-of-variety gain they create downstream; and a binding cap on consumer-data supply can nonetheless raise welfare above the unregulated level by indirectly relaxing the free-entry margin and drawing in additional firms. Calibration shows the optimal cap closes only a small share of the welfare gap, whereas an optimal computing-investment subsidy closes most of it, with gains accruing gradually as compute accumulates. The results deliver a welfare-based case for combining data regulation with direct support for computing infrastructure.

张龙天，现任中央财经大学国际经济与贸易学院副教授、博士生导师。研究领域为宏观经济学、经济增长、数字经济。学术研究成果发表于《经济研究》、《Management Science》、《Journal of Economic Dynamics and Control》等国内外权威期刊，并出版专著1部。主持国家自然科学基金专项项目——指南引导类原创探索计划项目，以及青年科学基金项目（C类），并参与多项面上项目。