Stanford researchers measure how much economic value AI creates per joule
The Joules-to-GDP framework shows that the strongest model or largest GPU cluster is often not the best use of a limited energy budget.
Researchers from Stanford’s HazyResearch and collaborators have introduced Joules-to-GDP, an open framework for measuring how much economic value AI produces per unit of energy. It calculates both net value per joule—estimated output value minus execution cost—and expert time saved per joule.
The team applied the method to 220 professional tasks across 44 occupations in OpenAI’s GDPval benchmark. The results challenge the assumption that the most capable model is always the most efficient choice. GLM 5.2 produced the highest average net value per task, but Gemma 4 31B delivered 56% of that value using only 23% of the energy. That translated to 2.5 times more net value per joule and 2.8 times more time saved per joule.
Hardware scaling also showed diminishing returns. Running the same model on eight Nvidia H100 GPUs used 3.2 times as much energy as a single H100 while reducing completion time by only 22%, with roughly the same economic value.
The framework gives developers and organizations a way to choose models, tasks and accelerators under an energy constraint instead of optimizing only for benchmark quality or speed. The researchers released the paper, code and GDPWorks, an additional dataset covering 36 occupations.
