A team led by Zeyu Yang published a paper on Oct 14, 2026 that measured how much electricity 86 open-weight large language models (LLMs) consume while running seven benchmarks on six hardware platforms. The study shows most configurations use far more power than needed – a median of 2.6 to 4.8 times more, and up to 27 times more in extreme cases. Adding reasoning steps raises energy use by about 3.7×, while using low‑precision FP8 arithmetic cuts energy by roughly 25% with little loss in score.
Why it matters
Developers can cut operating costs and carbon footprints by choosing more energy‑efficient LLM setups.
