A researcher is requesting an arXiv endorsement in computer science or economics for a paper proposing a quantitative framework for AI infrastructure spending. The work argues that AI-enabled software can be analyzed using the same investment discipline applied to physical technologies, utilizing sensitivity analysis and shadow prices.
The paper addresses the lack of governing equations for software by classifying technologies based on the determinacy of their relationships. It organizes program semantics, AI scaling laws, and error statistics to align with existing technology-investment machinery. This approach aims to make AI investment decisions commensurable with other capital portfolio items.
The author presents the work as an internally coherent conjecture that challenges the intuition that "buying the biggest model" is always optimal. Empirical validation on real systems is identified as open work.