Post Thread

University of Michigan researchers found generative AI uses 1.4 to 180 times more energy for non-English text, with Hindi needing 7.9× and Bengali 15.3× more than English, a disparity affecting about 5.2 billion people whose usage data is tracked via Ethnologue and visualized in the energy-by-language spreadsheet.

The team measured Qwen 3, Gemma 3 and Llama 3.1 on the Belebele reading comprehension benchmark using the ML.Energy framework, with additional math and chat prompts translated by No Language Left Behind, and noted lower accuracy alongside higher cost.

The extra cost comes from tokenization: scripts like Devanagari that change form and combine with accents can produce up to millions of unique characters, inflating token counts and memory use, so authors Naihao Deng and Rada Mihalcea from the U-M AI Lab, using the Great Lakes cluster, suggest language-specific models or translation pipelines to reduce the penalty.


https://news.engin.umich.edu/...-energy-on-non-english-text/