Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?

Amirreza Zeinali  ⋅\cdot Jon Saad-Falcon  ⋅\cdot Avanika Narayan

Mersad Abbasi  ⋅\cdot Jonathan S. Hartley  ⋅\cdot John Hennessy Affiliation: Stanford University  ⋅\cdot University of California, Berkeley Affiliation: The University of Texas at Austin

Christopher Re  ⋅\cdot Azalia Mirhoseini

11footnotetext: Equal contribution.

1 Introduction

Existing studies assess an AI system’s economic value through proxies such as output quality and estimated worker time saving, and compare task performance with API cost or completion time (Patwardhan et al., 2025; Wan et al., 2026; Erol et al., 2026). Meanwhile, energy benchmarks measure how much energy AI systems consume, but do not quantify the economic value of the professional outputs produced with that energy (Chung et al., 2025; Tripathy et al., 2026, see Appendix A for related work). We connect these approaches by measuring the economic value of professional outputs alongside the energy required to produce them.