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 professional tasks end to end and produce economic value. However, existing evaluations typically report economic value separately from the end-to-end energy used to produce it. Energy provides a system-level measure of resource use because it captures the combined effects of model inference, agent orchestration, tool use, repeated runs, batching, and hardware. We therefore ask: How much economic value does an AI system produce per unit of energy consumed? We define two metrics to answer this question: Dollar Value per Joule (DPJ) and Time Saved per Joule (TPJ). DPJ divides economic value by end-to-end joules consumed, while TPJ divides expert labor hours saved by end-to-end joules consumed. We apply these metrics to 220 GDPval tasks spanning 44 occupations, five AI systems, and four accelerators. Across occupation categories, DPJ and TPJ vary by factors of 8.94 and 3.75. Computer & Mathematical tasks lead DPJ at 139.1 × 10⁻⁶ USD/J, while Engineering & Architecture tasks lead TPJ at 2.22 × 10⁻⁶ hr/J. Model choice reveals a tradeoff between task value and return per joule. GLM 5.2 generates the most economic value (USD 106.72) and saves the most labor time (2.14 hours) per representative task, while Gemma 4 31B leads DPJ at 189.6 × 10⁻⁶ USD/J and TPJ at 4.19 × 10⁻⁶ hr/J. Accelerator choice further changes return per joule: NVIDIA B200 reaches 362.7 × 10⁻⁶ USD/J and 7.531 × 10⁻⁶ hr/J. Together, these results show that economic return on energy depends on both the work being performed and the complete AI system, supporting model, accelerator, and system-software choices that produce more economic value with less energy.