UMich research finds generative AI uses up to 180x more energy on non-English text
radamihalcea · x · 2026-11-06
New University of Michigan research by Rada Mihalcea and Naihao Deng finds that generative AI can use up to 180x more energy processing non-English text than English.
The root cause is tokenization bias: non-English text is split into far more tokens, driving up compute and energy costs. The authors argue fixing tokenization bias is necessary for fairer, greener AI, with direct implications for the cost and carbon footprint of multilingual AI services.
More from Infra
- Alice & Bob study: 23,000-qubit quantum computer could crack RSA-2048 bank encryption in days — ChuckDBrooks · 2026-11-06
- REAX 1.3 claims #1 on HF Decision Index with 74.0, built via Bittensor subnet in four weeks — airesearch12 · 2026-11-06
- Hundreds of protesters storm Microsoft data center site in Amsterdam — Polymarket · 2026-11-06
- Protesters break into Microsoft data center site in Amsterdam — novaRom · 2026-11-06
- Baidu open-sources ERNIE-5.5-Flash-Base: 280B MoE with 14B active params and 1M context — kimmonismus · 2026-11-06
- Data center NIMBYism is a corporate good-neighbor problem, says ex-Kraft vet — BecauseCulture · 2026-11-06