Artificial intelligence spends up to 180 times more energy on non-English queries

Researchers from the University of Michigan have shown that artificial intelligence processing queries in different languages requires significantly varying energy expenditures. Depending on the language and model, energy costs for national languages can exceed those for English queries by 1.4 to 180 times.

• The study covered 122 languages and models including Qwen 3, Gemma 3, and Llama 3.1.
• The main reason for energy disparity is tokenization differences across writing systems.
• Total energy ranged from 17.6 kJ for English to 3,147 kJ for Pashto.
• Around 5.2 billion people speak languages that require at least twice as much energy to process.

Reasons for energy disparity in AI models

A primary factor behind the energy consumption gap is the increase in tokens—the minimal units of text into which models break down input data. Tokenizers primarily trained on the Latin alphabet process other writing systems less efficiently, splitting words into more fragments and increasing memory demands on NVIDIA GPUs.

Additionally, for low-resource languages, models frequently generated much longer outputs, sometimes exceeding English answers by 20 to 25 times in output tokens. The maximum energy difference per output token reached up to 8.3 times.

Impact on accuracy and users

Researchers also found a correlation between energy expenditure and response accuracy, where resource-heavy languages performed worse on reading comprehension and math tests. Because commercial AI services often charge based on processed tokens, this disparity could make AI services more expensive for speakers of non-English languages.

Potential solutions

To bridge the gap, authors suggest developing models specifically optimized for individual languages, expanding training corpora with underrepresented languages, or translating queries into English beforehand if translation quality is high enough.