AI Models Burn Up to 180x More Energy on Non-English Text

Generative AI models use between 1.4 and 180 times more energy to process non-English text than equivalent English prompts, according to University of Michigan Engineering researchers who presented their findings at the Empirical Methods in Natural Language Processing conference of the Association for Computational Linguistics.

The disparity is greatest in languages that don't use Latin script and have less digital training data available. Hindi requires 7.9 times more energy than an equivalent English prompt, and Bengali 15.3 times more. Both are spoken by hundreds of millions of people. Most Western European languages come in at twice the energy expenditure or less.

"The languages that typically consume more energy also tend to have lower accuracy in the model output," said Naihao Deng, a doctoral student in computer science and engineering at the U-M AI Lab and first author of the study.

Deng describes a "double penalty" that could lock out both users and providers: "Most people will want to use AI models by asking questions in their native languages, and they may feel pressured to use it as it's incorporated into more parts of society, but the double penalty makes it more likely that certain people will be priced out of AI and that companies will be more reluctant to serve models in certain languages."

How big is the gap in practice?

The researchers estimate that around 5.2 billion people must use more than twice as much energy to generate text in their first language compared with English speakers. That figure combines the team's energy measurements with demographic data from Ethnologue, the language reference database.

The team tested three open-source models — Alibaba's Qwen 3, Google's Gemma 3 and Meta's Llama 3.1 — on multiple-choice reading comprehension questions from Belebele, a benchmark dataset covering 122 languages. They measured energy consumption per question using the ML.Energy framework developed at Michigan Engineering.

Because other benchmarking datasets are not multilingual, the researchers translated grade-school math word problems and open-ended user prompts into seven additional languages using Meta's translation tool, No Language Left Behind. All models used less energy on English text on average. The gap persisted at larger model sizes and across different computer processors.

The National Science Foundation and DARPA partially funded the research; additional funders include Ford, the Laude Institute, OpenAI and the Survival and Flourishing Fund.

Why do some languages cost more?

The energy penalty stems from tokenization. Tokens are the basic unit of data AI models process, representing entire phrases, words, syllables or characters. Models convert prompts into tokens before processing and convert output tokens back into human language.

Scripts such as Devanagari, used for Hindi, Nepali and Marathi, and the Bengali script have primary characters that change form and combine with accents, producing up to millions of unique characters. That inflates token counts.

More tokens consume more memory, which forces models to process tokens in smaller batches, raising energy use per token. The effect compounds the disparity. It also hits direct costs, since companies typically price model access by tokens generated.

The takeaway: the languages with the highest energy bills are also often the ones with the weakest output quality.

Can the gap be closed?

The researchers recommend that companies optimize services for more languages — for example, by routing queries to models tailored to specific languages, or by translating prompts into English before feeding them to a general model, provided the translation is reliable.

"AI developers should be creating more resources for those languages with an energy and cost disparity," said Rada Mihalcea, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering, director of the U-M AI Lab and corresponding author of the study. "Crowdsourcing native speakers to help translate material or create resources to get enough training data could also help improve efficiency."

The study ran on the university's Great Lakes high-performance computing cluster.

For marketers localizing AI-driven customer experiences at scale, the findings set a hard budget constraint: serving Hindi or Bengali audiences may cost up to an order of magnitude more per interaction — until vendors build language-specific routing or better tokenizers for non-Latin scripts.