This working paper argues that the strategic variable in U.S. AI compute infrastructure is not the scale of capital deployed but its efficiency, the AI capability generated per dollar of investment. It develops a simple identity in which effective compute is the product of installed capacity, fleet utilization, model FLOPs utilization, and algorithmic efficiency and derives a capex-equivalence result showing that a utilization gain can substitute for new capacity in the same dollars. The paper documents the empirical landscape of hyperscaler capital expenditure, reinterprets the DeepSeek episode, and traces the macroeconomic, energy, and geopolitical channels through which capital efficiency becomes a national concern, closing with disclosure, procurement, and energy-coupled policy instruments.
Benedict Amissah-Ocran (Mon,) studied this question.