Stanford AI Lab: Launches Joules-to-GDP Framework

Stanford AI Lab researchers introduced Joules-to-GDP, a framework that quantifies economic value per joule of energy spent on compute. The metric evaluates tasks from OpenAI GDPval across finance, engineering, law and administrative roles. CAD design generates roughly 10 percent more revenue per task than website building, yet the latter delivers 27 percent higher value per joule. Gemma 4 31B produces 56 percent of GLM 5.2 net value while using only 23 percent of the energy, yielding 2.5 times the return per joule. Scaling the same model from one to eight NVIDIA H100 GPUs consumes 3.2 times more energy but reduces completion time by just 22 percent with little added value. Joint work with SnorkelAI also released GDPWorks, a benchmark covering 36 additional occupations.