Researchers Introduce Joules to GDP, a Framework for AI Value Per Unit of Energy
AI is using more and more of the world’s energy. But how much economic value are we getting from AI? We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return. Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings: 1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule. 2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule. 3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value. With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations. Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh. More details below 🧵
Context
A public way to score which tasks, models and chips return the most economic value per joule gives labs and buyers a yardstick beyond raw benchmarks as AI power use climbs.
Jon Saad Falcon and collaborators released Joules to GDP, or J2G, built on OpenAI's GDPval style tasks across dozens of occupations. With Snorkel AI they also released GDPWorks for more occupations. The work argues that the highest value tasks are not always the best value per joule, and that bigger models or more GPUs are not always efficient.
Deep dive
Jon Saad Falcon (@JonSaadFalcon) introduced Joules to GDP, shortened as J2G, on Nov 4 Eastern. The framework measures how well AI turns energy into economic value across tasks, models and hardware.
Using 44 occupations from OpenAI's GDPval style work in finance, engineering, law and administrative work, the team reported three findings. Tasks with the highest economic value are not always best per joule. A smaller model can beat a larger one on value per joule. Scaling from one to eight NVIDIA H100s used 3.2 times the energy while cutting completion time by only about 22 percent, with similar value.
With Snorkel AI they also released GDPWorks, a benchmark covering 36 more occupations. Collaborators named in the post include researchers at Hazy Research and others. The lead post includes a short explainer video.
Energy use is already a live political and infrastructure story in this week's decks, so a public efficiency metric is timely even if labs debate the dollar estimates behind each task.
What to expect next
Watch for papers, code and whether labs cite J2G or GDPWorks in efficiency claims. Cloud buyers may ask vendors for value per joule style numbers. Critiques will focus on how task dollars are estimated and whether the hardware energy figures match real data center mixes. Policymakers who already argue about AI power demand may also pick up the framework when they talk about which workloads are worth the electricity.