🌍⚡

The Language–Energy Divide

Measuring Energy Costs of Multilingual LLM Inference

Naihao Deng¹, Alissa Shen¹, Yiming Feng¹, Joan Nwatu¹, Jae-Won Chung¹,
Mosharaf Chowdhury¹, Yulong Chen^2,3, Rada Mihalcea¹

¹University of Michigan · ²University of Cambridge · ³University of Aberdeen

📄 Paper 💻 Code 🤗 Dataset

📢 News

Aug 2026 Our paper has been accepted to EMNLP 2026 (Main Conference)! 🎉

📝 Overview

Large language models are increasingly deployed in multilingual settings, yet the energy cost of serving them across languages is poorly understood. We present the first systematic measurement of inference energy across 122 languages, and reveal a stark, persistent language–energy divide: serving low-resource languages costs dramatically more energy — and those same languages are answered least accurately.

8.3×

variation in per-token energy across languages (single model)

179×

total energy gap — English (17.6 kJ) vs. Pashto (3,147 kJ)

📉

Double penalty: costliest languages are also the least accurate

122

languages measured on Belebele; divide holds across 5 models, 2 GPUs, 6 batch sizes, 3 tasks

📊 The divide, visualized

Energy per output token across 122 languages (Qwen3-8B, zero-shot CoT, Belebele). A long right tail of low-resource, non-Latin-script languages exceeds 0.4 J/token.

Energy per response across 122 languages (log scale): the per-token cost compounds with inflated output-token counts to produce up to a 179× end-to-end gap.

🔍 Why it happens

Per-token cost: tokenizers trained on Latin-dominated corpora fragment complex/rare scripts inefficiently, raising energy per token.
Token inflation: low-resource languages elicit far longer (often repetitive) generations — up to ~20–25× more output tokens than English.
Compounding: the two penalties multiply, and larger sequences force smaller serving batch sizes that underutilize the GPU.

The disparity persists across model families and sizes (Qwen3-8B/14B/32B, gemma-3-27B, Llama-3.1-8B), across GPUs (L40S, RTX 6000 Pro Blackwell), across batch sizes, and across tasks (Belebele, GSM8K, LM-Arena).

✅ Recommendations

Energy as a first-class evaluation axis Per-language energy in model cards Language-aware serving & routing Translate-then-process pipelines Equity audits at deployment

📚 Citation

@article{language-energy-divide,
  title  = {The Language--Energy Divide: Measuring Energy Costs of Multilingual LLM Inference},
  author = {Deng, Naihao and Shen, Alissa and Feng, Yiming and Nwatu, Joan and
            Chung, Jae-Won and Chowdhury, Mosharaf and Chen, Yulong and Mihalcea, Rada},
  year   = {2026}
}