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

AI systems are increasingly evaluated on their ability to complete real occupational tasks end to end and produce economic value. Existing evaluations, however, typically measure economic value separately from the end-to-end energy it takes to produce that value. This raises the question: How much economic value does an AI system create per unit of energy consumed? To answer this question, we define Dollar Value per Joule (DPJ) and Time Saved per Joule (TPJ) by dividing task dollar value and human labor hours saved, respectively, by end-to-end system energy. We apply them to 220 GDPval tasks across 44 occupations and compare models and accelerators. The analysis yields three main results. First, DPJ and TPJ vary across occupation categories by factors of 8.94 and 3.75: Computer and Mathematical tasks lead DPJ at $139.1 \times 10^{-6}$ USD/J, while Engineering tasks lead TPJ at $2.22 \times 10^{-6}$ hr/J. Second, model choice reveals a tradeoff between task dollar value and economic return per joule. GLM 5.2 produces the highest task dollar value (USD 106.72) and saves the most labor time (2.14 hours) per representative task, but ranks fourth on both efficiency metrics; Gemma 4 31B ranks first at $189.6 \times 10^{-6}$ USD/J and $4.19 \times 10^{-6}$ hr/J. Third, among the tested accelerators, NVIDIA B200 performs best at $362.7 \times 10^{-6}$ USD/J and $7.531 \times 10^{-6}$ hr/J. Together, these findings show that economic return depends on both the work and the AI system, informing model, accelerator, retry, and inference-pricing decisions.