Stanford's Joules-to-GDP framework ranks AI models and chips by economic value per energy

Stanford researchers introduce Joules-to-GDP (J2G), a framework that measures the dollar value of AI work minus running cost per joule of energy, across 220 GDPval tasks in 44 occupations.

Using it, they report that Gemma 4 31B delivers 56% of GLM 5.2's net value with 23% of the energy, and that moving Qwen3.6-27B from 1 to 8 H100 chips cuts task time 22% while using 3.2x the energy.

In a separate comparison, the Stanford researchers ran the same model, Qwen3.6-27B, on four single AI chips across the same professional tasks. They report that the NVIDIA B200 delivered over 4x the dollar value and over 9x the expert time saved per unit of energy compared with the AMD MI350X.