Variance-Calibrated Energy Accounting: when additional energy stops being defensible in language model inference
Year: 2026
Type: preprint
Abstract: Pre-registration and materials for a study proposing intra-model variability as the calibrator of the equivalence margin used to decide whether additional energy consumption by larger language models is statistically defensible.
Language: English
Topic: Green IT and Sustainability
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203572236
