Found first: a primary source the press has not covered yet.
Researchers at the University of Bologna and CINECA have derived a closed-form analytical energy model for LLM inference, validated across 13 dense models from six families on NVIDIA H100 GPUs. The paper reports two findings with direct implications for inference cost and sustainability planning: agentic serving costs 13.0× more energy per request than standard serving, and Intelligence per Joule peaks at a moderate reasoning budget before falling by up to 3.14× as that budget grows 4×.
What the source says
Hiari Pizzini Cavagna, Andrea Proia, Daniele Cesarini, Zeynep Kiziltan, and Andrea Bartolini (University of Bologna and CINECA, Italy) built SIMPLE, a four-parameter model that separates prefill from decode and further divides decode into KV-cache cost and next-token computation. Tested across 13 dense LLMs from six families (OPT, LLaMA, Gemma, Falcon, Qwen2, Granite), the model achieves R²=0.969 and 11.97% MAPE with no per-model calibration, 3.1 to 3.5× below prior state of the art. Prefix caching removes 93% of prefill cost but leaves KV-cache costs intact. Cached agentic requests still run 4.8× more expensive than a non-agentic baseline, because input-length growth drives both the prefill and KV-cache terms.
Why it matters
Agentic orchestration is already the dominant serving pattern for production AI, and the 13.0× energy premium per request is its default cost structure, even with prefix caching applied. The 4.8× residual overhead on cached requests means prefix caching alone does not close the energy gap. On reasoning models, more compute budget raises accuracy but energy grows faster, so there is a peak efficiency point; heavier thinking past it loses energy efficiency by up to 3.14× across a 4× budget range. SIMPLE requires no per-model calibration, giving infrastructure and sustainability teams a usable analytical tool for fleet-scale cost projections today.