Compute as Capital: Measuring the Macroeconomic Impact of Frontier AI Compute Constraints
A Theoretical and Empirical Framework
The Economic Institute
10 June 2025
Abstract
This paper introduces a novel theoretical and empirical framework for analyzing frontier AI compute as a distinct form of productive capital. Extending standard semi-endogenous growth models, we formalize a "compute capital" stock separate from general ICT capital and demonstrate that AI productivity is a joint function of compute capacity, data availability, and human expertise. Using a newly constructed cross-country panel dataset of effective AI compute capacity covering 42 economies from 2016 to 2024, we document substantial and widening global compute inequality: the top five economies account for over 82 percent of estimated global frontier AI compute, while the bottom quartile shares less than 1.2 percent. Exploiting quasi-natural experiments, including the October 2022 U.S. semiconductor export controls on China, regional data center openings, and cloud provider expansion events, we identify causal effects of compute access on firm-level productivity, AI-related research and development, and labor-market outcomes. A one-standard-deviation increase in effective compute access is associated with a 3.7 percentage point gain in sectoral total factor productivity and a 12 percent increase in AI-related patent filings. We then develop a structural industrial organization model of the cloud computing and semiconductor supply chain to evaluate welfare implications of market concentration, finding that the current oligopolistic structure generates deadweight losses equivalent to 0.3 to 0.8 percent of GDP in advanced economies. Policy counterfactuals suggest that targeted compute subsidies for research institutions yield higher social returns than broad-based industrial subsidies, while overly restrictive export controls accelerate indigenous chip development in targeted economies without proportionally reducing their AI capabilities. Our findings have direct implications for competition policy, industrial strategy, AI governance, and the design of international technology agreements.
1. Introduction
The rapid advancement of artificial intelligence over the past decade has transformed the global economic landscape in ways that are only beginning to be understood. Large language models, computer vision systems, protein-folding algorithms, and autonomous agents have demonstrated capabilities that were considered speculative just five years ago, and these capabilities are increasingly being deployed in commercial settings across industries ranging from healthcare and finance to manufacturing and logistics. Yet the production of frontier AI capabilities depends critically on a single, scarce, and highly concentrated input: specialized computational hardware. Graphics processing units (GPUs), tensor processing units (TPUs), and other accelerators designed for the massively parallel matrix operations that underpin modern deep learning have become the binding constraint on the pace and direction of AI progress. The cost of training a single frontier model now exceeds one hundred million dollars, with the vast majority of that cost attributable to compute, and the total global stock of frontier-capable compute hardware is controlled by a remarkably small number of firms and nations.
Despite the centrality of compute to the AI revolution, the economics profession has been slow to develop formal frameworks for analyzing compute as a distinct factor of production. Standard macroeconomic models treat information and communications technology (ICT) capital as a homogeneous aggregate, lumping together personal computers, enterprise servers, networking equipment, and AI accelerators into a single capital stock. This aggregation obscures the unique economic properties of AI compute: its extreme concentration among a handful of suppliers, its rapid obsolescence driven by Moore's Law and architectural innovation, its complementarity with large-scale data and specialized human capital, and the threshold effects that arise from the minimum viable