Stanford Joules-to-GDP finds Gemma 4 returns 2.5x more value per joule than GLM 5.2
Across 220 GDPval tasks, Gemma 4 31B used 23% of GLM 5.2's energy for 56% of its output value. Eight H100s reduced completion time by 22% while using 3.24 times the energy of one.
By Ryan Merket · Published
Primary source: arxiv.org
Why it matters
Model selection is usually argued on quality, price and latency. Joules-to-GDP adds measured energy as a denominator and shows that rankings can change: a smaller model can return more dollars per megajoule, and extra GPUs can consume far more energy for a limited time saving. For operators buying inference under power constraints, that can inform routing and hardware decisions. For model builders, NVJ and TSJ give a target that penalizes failed attempts and low-value extra computation.
Jon Saad-Falcon (@JonSaadFalcon) and co-authors introduced Joules-to-GDP on November 4th in a thread on X. The framework divides the estimated economic value of AI output by the energy used to produce it. The work was submitted to arXiv on October 22nd, 2026 and last revised on November 2nd, 2026.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.