Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?
Existing AI evaluations treat economic value and energy consumption separately, leaving a critical gap in understanding whether more capable systems actually deliver better returns per watt. This paper formalizes Net Value per Joule ($NVJ$) and Time Saving per Joule ($TSJ$), pairing judged professional output quality with end-to-end energy measurement across 220 real-world tasks. The work matters because it shows rankings of occupations and models shift dramatically once energy is accounted for, challenging the assumption that raw capability or dollar cost alone guide efficient deployment.
This paper makes a timely and conceptually important contribution by uniting economic benchmarking with physical energy accounting through $NVJ = (G_o^b - C_o)/E_o$ and $TSJ = \sum_{i:o(i)=o} Y_i^b (H_i^b - L_i)/E_o$. Its central claim—that capability, cost, and speed are insufficient proxies for energy-efficient economic value—is convincingly demonstrated at the task level. However, the empirical edifice rests on only five tasks per occupation, and the main evaluation hinges on a single LLM judge with moderate inter-judge agreement. Consequently, the occupational rankings and cross-model frontiers are more illustrative than statistically definitive.
The strongest result is that energy-normalization materially re-ranks both occupations and models in ways that dollar cost and completion time miss. The authors show that software development delivers about 27% higher $NVJ$ than engineering once energy is included, despite yielding lower net dollar value per task. Similarly, Gemma 4 31B achieves approximately 2.5 times GLM 5.2's $NVJ$ by using 77% less energy, a finding robustly supported by direct end-to-end metering via the Intelligence per Watt framework rather than synthetic proxies.
Sample size and granularity are the primary weaknesses. With only five tasks per occupation, the 95% task-composition intervals are wide and category rankings are unstable; shifting from active-request to equal batch-share energy attribution changes nine of thirteen $TSJ$ category leaders.
The evaluation also relies on a binary win-or-tie/loss rule that ignores partial credit, and the primary LLM judge achieves only 72.3–77.2% exact agreement with alternative judges and 71.9% with human experts on a small 32-comparison sample. Additionally, the five-model comparison confounds model and hardware, since GLM 5.2 required eight H200 GPUs while the other four models were tested on eight H100s.
The paper fairly situates itself against GDPval for task design and against prior cost-centric and energy-centric benchmarks, including Intelligence per Watt for end-to-end energy measurement. We verified that GDPval indeed provides five tasks per occupation in its public gold subset, and that IPW defines intelligence efficiency as accuracy per unit power. The evidence supports the existence of tradeoffs but is too thin to support precise quantitative rankings: for example, the claim that eight H100s reduce completion time by only 22% while using 3.24x the energy is well-documented within the studied batch, but extrapolating this to typical serving requires more than 20 concurrent requests per batch. The comparison of $NVJ$ and $TSJ$ across accelerators (B200, H200, H100, MI350X) is more controlled, holding Qwen3.6-27B fixed, and therefore carries more weight.
Reproducibility is partially strong. The authors provide source URLs for code and data in the abstract, give exact vLLM configurations, context limits, and batch sizes in Appendix E Table 3, and document their active-request energy attribution rule. However, the authors acknowledge that 'end-to-end reruns require compatible hardware and serving software and may produce different outputs or energy measurements.' A critical blocker is that the GLM 5.2 versus other model comparison cannot be reproduced on identical hardware because GLM 5.2 used H200 GPUs.
An AI system can produce more economically valuable work while delivering less value for the energy it consumes. We therefore ask: How much economic value does an AI system create per unit of energy consumed? We formalize economic value in two ways: (1) net dollar value, the output's dollar value minus AI-system execution cost, and (2) time saving, the expert's completion time minus the AI system's completion time. Dividing each by measured end-to-end energy yields Net Value per Joule (NVJ) and Time Saving per Joule (TSJ), reported in USD/MJ and hours/MJ. We study 220 public GDPval tasks across 44 occupations, serving five open-weight models locally and evaluating four accelerator types. First, the occupation groups where AI systems succeed most often, deliver the greatest dollar value, or save the most time do not necessarily provide the greatest economic return per joule. For example, among the tested tasks, engineering work delivers about 10% more net dollar value per task than software development. Yet software development delivers about 27% more net value per joule. Second, GLM 5.2 delivers the highest average net dollar value, at $72.27 per task using 0.430 MJ, equivalent in energy to about four hours of active laptop use. By contrast, Gemma 4 31B delivers 56% of that net dollar value using just 23% of the energy, achieving approximately 2.5 times the NVJ and 2.8 times the TSJ. Third, across four single-accelerator deployments, Qwen3.6-27B's NVJ varies by a factor of 4.32 and TSJ by 9.27, with B200 leading both. Additionally, one and eight H100s deliver similar net dollar value, but eight reduce completion time by 22% at 3.24 times the energy. Our framework makes output value and expert time saving under an energy budget explicit objectives for improving models, accelerators, and software, while offering a new basis for assessing AI's potential economic effect on occupational work.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.