Stanford's Joules-to-GDP Framework Quantifies AI Energy-Economic Efficiency

Stanford's Hazy Research team (including Christopher Re and John Hennessy) has released a paper and a full suite of open-source resources for the Joules-to-GDP (J2G) framework, addressing one question: how much economic value does AI actually produce per unit of energy consumed? The framework emerges against a backdrop of energy constraints on AI and skepticism over whether AI investments will pay off, with the authors arguing the focus should be on sectors with the highest value density. The release has been shared and recommended by multiple AI news accounts.

Confirmed

• The framework quantifies AI's energy-economic efficiency with two metrics: Net Value per Joule (NVJ) (output dollars minus execution costs) and Time Saved per Joule (TSJ).
• Benchmarks cover 220 GDPval tasks across 44 occupations, tested on 5 models.
• One key finding: tasks with the highest return per joule are not the highest-value tasks.
• Specific figures relayed by @rohanpaulai: GLM 5.2 achieves $72 net value per task, topping all models.
• Another conversion: the energy used by one NVIDIA H200 running at 700W for 1 hour (about 2.52 MJ, roughly equivalent to 23 hours of laptop usage), when given to GLM 5.2, can yield about $424 in net value.

Why it matters

At a time when AI's energy consumption and return on investment face intense scrutiny, J2G offers a unified measure that converts energy directly into economic output, enabling comparisons of energy returns across tasks, models, and hardware. It provides a quantifiable analytical tool for the debate over whether AI investments are worthwhile, and suggests the industry should prioritize application scenarios with the highest value density.

2026-11-05 ~ 2026-11-05 · 5 related posts

• [source] Stanford's Joules-to-GDP framework measures how much economic value AI produces per joule of energy — HazyResearch · 2026-11-05
• Stanford AI Lab unveils Joules-to-GDP, a framework measuring economic value per joule of AI compute — KyeGomezB · 2026-11-05
• Joules-to-GDP: across 44 occupations, highest-value-per-joule AI tasks aren't the highest-value ones — sprooos · 2026-11-05
• Joules-to-GDP paper: GLM 5.2 nets $72.27/task; small models 2.5x energy efficiency — rohanpaul_ai · 2026-11-05
• One H200-hour of energy yields $424 of net value via GLM 5.2, with 13x spread across jobs — rohanpaul_ai · 2026-11-05