Stanford's Joules-to-GDP framework measures how much economic value AI produces per joule of energy
HazyResearch · x · 2026-11-05
Stanford Hazy Research (with Christopher Re and John Hennessy) released Joules-to-GDP, an open framework measuring how efficiently AI converts energy into economic value.
It defines two metrics: Net Value per Joule (NVJ) — output dollar value minus AI execution cost per unit of energy — and Time Saving per Joule (TSJ) — expert time saved per joule. Tested on 220 GDPval tasks across 44 occupations:
- Highest-value tasks aren't the most energy-efficient: CAD design earns 10% more net value per task than building websites, but website building yields 27% more value per joule.
- The most capable model isn't always the best energy bet: of five models tested, GLM 5.2 delivers the highest net value ($72.27/task) using energy equivalent to 4 hours of active laptop use; Gemma 4 31B hits 56% of that value with 23% of the energy — 2.5x NVJ and 2.8x TSJ.
- NVJ and TSJ vary widely across accelerators for the same model, so task-model-hardware pairing matters a lot.
Paper, code, and datasets (incl. GDPWorks) are all open for reproduction on your own tasks and hardware.
Related event: Stanford's Joules-to-GDP Framework Quantifies AI Energy-Economic Efficiency(5 posts)→
More from Infra
- Decentralized RL Works: Untrusted Miners Lift LiveCodeBench 39.5→58.7 on Bittensor Subnet — bittingthembits · 2026-11-05
- Atomic Chat launches open-source Local AI Server tying models and agents together — SucceededMind · 2026-11-05
- Developers report Grok API instability: 13.58% error rate and 13.6s p90 latency in dashboards — ChrisUniverse · 2026-11-05
- Even cutting forecasted data center growth by 60% leaves Nvidia supply constrained — robleclerc · 2026-11-05
- Free live paper discussion: why speculative decoding is fast despite doing more compute — doesdatmaksense · 2026-11-05
- Airbnb: fine-tuned 2B model matches frontier classifiers at 1% of the cost — Ahmad_Al_Dahle · 2026-11-05