Ruby (Ruishan) Zhang

Harvard University · Economics · 2026–27 job market

Fields

Industrial Organization, Innovation

Job market paper

Returns to Science of Centrally Provisioned GPUs and Optimal Allocation

Funders of research must consider where the marginal dollar is most productive and where it is least likely to displace the recipient's own investment. We study the allocation of federal computing grants to U.S. universities by estimating the returns to science of university-level computing infrastructure and modeling the crowd-out of universities' own investment. We construct an original dataset measuring Central Processing Unit (CPU) and Graphics Processing Unit (GPU) capacity over time and link it to research publications at U.S. R1 universities. Exploiting the large and discrete nature of GPU capacity increases to identify their effects, we find that AI and deep-learning publications increase more at less-resourced universities. We then estimate a structural production function where multiple scientific fields share a central computing cluster, and derive universities' investment responses to external grants. Compared to a publication-maximizing benchmark, the observed portfolio of National Science Foundation (NSF) cyberinfrastructure grants captures 39% of the publication gain attainable with the same budget. The publication-maximizing allocation shifts dollars from universities with medium-sized endowments toward both tails of the endowment distribution. However, the NSF's observed allocation is more geographically dispersed than the publication-maximizing allocation, highlighting tradeoffs in balancing efficiency versus equity.