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

Efficient AI requires more than maximizing capability per unit of compute: a system should also maximize the useful work delivered by the energy it consumes. We introduce Dollar Value per Joule (DPJ) and Time Saved per Joule (TPJ), two task-value-aware efficiency metrics for complete AI systems. DPJ measures expert-comparable economic value per joule, while TPJ measures expert labor hours saved per joule. We evaluate them on 220 GDPval professional tasks spanning 44 occupations, five AI systems, and four accelerators, with four runs per task. The metrics reveal system tradeoffs hidden by generated value alone. GLM 5.2 produces the most GDP-weighted value and saves the most labor, while Gemma 4 31B leads DPJ and TPJ. In our Qwen3.6-27B accelerator comparison, NVIDIA B200 leads the tested accelerators; using one rather than eight H100s improves DPJ by 3.48x. Across occupation categories, DPJ spans 8.94x. Together, these results show that useful efficiency is a property of the full task-model-hardware-execution configuration. DPJ and TPJ provide concrete empirical targets for efficient-system prediction, system selection, and model-hardware co-design.