← News

ResearchSources: arXiv · 18 Nov 2026

Study finds multilingual LLM energy costs vary up to 179 times across languages

An arXiv paper reports that serving large language models in non-English languages can cost far more energy, with low-resource languages paying the highest bills and scoring lowest on accuracy.

In plain words

Researchers measured how much electricity large language models use when answering in different languages. They found that some languages, especially low-resource ones like Pashto, need far more energy than English for the same requests, and they also produce less accurate answers.

What was announced

The paper, posted on arXiv, reports a systematic study of inference energy consumption across languages using the ML.Energy framework. The authors found that energy per output token varies by up to 8.3 times across languages, and that total energy for a fixed set of requests varies by up to 179 times, from 17.6 kJ for English to 3,147 kJ for Pashto, the cheapest and most expensive cases they measured.

According to the authors, the gap comes from two compounding factors: higher per-token energy for languages written in complex or rare scripts, and more tokens being generated for low-resource languages. They also describe a combined cost and performance penalty, since the languages with the largest energy footprints tend to score lowest on task accuracy. The divide persisted across the models, hardware and tasks tested, which the authors interpret as a systemic energy inequity in multilingual deployment. They recommend treating energy as a first-class evaluation axis, adding it to reporting checklists and model cards, and adopting deployment-side mitigations.

What it does not show

  • The evidence is the paper's abstract only; the full article's methods, model list and hardware details are not available for verification.
  • The measurements come from the authors' own study and have not been independently confirmed.
  • The abstract does not specify which models, hardware configurations or tasks were tested, so the breadth of the claimed persistence is unclear.

For security and IT teams

Scope
The study addresses inference energy consumption in multilingual LLM serving and recommends deployment-side mitigations, but the abstract does not describe specific operational measures for production systems.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.