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

This study proposes two metrics—net dollar value and time savings—to measure the economic output of AI systems per unit of energy consumption. Based on 220 public GDPval tasks across 44 occupations, it tests 5 open-weight models and 4 types of accelerators. Results show that high-output AI occupations do not necessarily correspond to the highest economic returns per unit of energy. GLM 5.2 achieves the highest average net dollar value per task, while Gemma 4 31B delivers approximately 2.5 times the net value per unit of energy. This framework can serve as an energy-constrained optimization target for AI models, accelerators, and software, providing a new basis for evaluating the economic impact of AI on occupational work.