AI energy efficiency varies sharply by model, hardware and task

November 5, 2026, 01:47Source: Snorkel AI

AI models that produce the most valuable work do not necessarily deliver the best return on energy, according to a study from researchers at Stanford, UC Berkeley and UT Austin. Gemma 4 31B generated 56% of the value of GLM 5.2 while using 23% of its energy, resulting in about 2.5 times the net value per joule.

Key points

Gemma 4 31B delivered 56% of GLM 5.2's value using 23% of its energy.
GPT-OSS 120B used 68% less energy than GLM 5.2 despite similar cost and speed.
Qwen3.6-27B produced the highest tested hardware return on an NVIDIA B200.
Using eight H100 GPUs made Qwen3.6-27B 22% faster but used 3.24 times the energy.
GLM 5.2's net value per megajoule fell 70% against AI-assisted experts.

Energy changes how AI performance is measured

The findings come from Joules-to-GDP, a paper released on November 4, 2026. Snorkel AI says its research team provided expert baselines for the 220 GDPval tasks used in the study and worked with the researchers to develop GDPWorks, which contains 83 additional tasks covering 36 occupations.

Among the five tested models, GLM 5.2 produced the highest output value, at $74.19 per task. It used 0.430 megajoules of energy, equivalent to about four hours of active laptop use, according to Snorkel AI. Gemma 4 31B produced $41.38 of output value per task, but used 0.097 megajoules and led on net value and time saved per unit of energy.

Net value subtracts the cost of running an AI system from the dollar value of its work. The study divides that result by energy use to calculate net value per megajoule, or NVJ. Snorkel AI says a high per-joule return does not necessarily mean a model completes enough work, so the paper reports total value alongside efficiency.

Similar speed and cost can hide different energy use

The study found that models with comparable cost and speed can have sharply different energy requirements. GPT-OSS 120B cost about 3.1% more than GLM 5.2 and finished about 2.1% faster, while using 68% less energy.

The preferred measure changes the result. GLM 5.2 delivered more net value per dollar of AI cost, 37.69 compared with 19.09 for GPT-OSS 120B, and more per AI hour, $191.3 compared with $102.1. GPT-OSS 120B delivered more value per megajoule, at 275.7 compared with 168.1 US dollars per megajoule.

Hardware also affected efficiency. When Qwen3.6-27B was run on a single accelerator, the NVIDIA B200 produced $535.5 of net value per megajoule, compared with $124.1 on an AMD MI350X. The NVIDIA H100 and H200 were close, at $402.2 and $391.3 respectively, showing that a newer chip was not automatically more efficient in this test.

GLM 5.2 ran on eight H200s, while the other four models used an eight-H100 node because its FP8 serving setup did not fit there. Snorkel AI says the GLM 5.2 figures therefore reflect both the model and the hardware.

More GPUs and retries reduced value per joule

Adding hardware did not produce a proportional speed increase. Moving Qwen3.6-27B from one H100 to eight made it only 22% faster, while using 3.24 times the energy. Net value per joule fell from 402.2 to 123.4 US dollars per megajoule, while output value remained about the same, at $64.79 compared with $67.81 per task.

Retries showed a similar tradeoff. Four attempts increased GLM 5.2's output value from $74.19 to $93.42, but energy use rose from 0.430 to 1.760 megajoules. Net value per joule fell from 168.1 to 48.6. The test counted all four attempts, even when an earlier attempt had succeeded. Snorkel AI says the paper did not test a system that stops after obtaining a good answer.

The comparison with AI-assisted professionals also lowered the model's measured return. Against experts working with AI, GLM 5.2's net value per megajoule fell 70%, from $168.1 to $50.0. These experts set a higher quality bar and spent about 23% less time on average, leaving less time for the AI system to save.

For users and organizations, the study points to energy as an additional factor when selecting models, accelerators and workflows. The results suggest that total output value, speed, operating cost and energy efficiency can favor different systems, so optimizing one measure alone may lead to a different choice than optimizing the full task outcome.