Artificial intelligence does not cost the same in every language. New University of Michigan research into AI energy consumption by language found that generating an answer in some languages can require dramatically more energy than producing a comparable English response. The results point to a troubling divide: billions of people may face higher computing costs and weaker performance simply because of the language they use.
The study measured inference energy across 122 languages using three open-source AI models. Depending on the model, language and task, non-English inference consumed between 1.4 and 180 times as much energy as the English baseline. Hindi and Bengali were among the languages with particularly high measured costs. Yet these numbers are benchmark results—not universal multipliers that apply to every model, prompt or deployment.
What the University of Michigan research found
The researchers examined the energy used during inference, the stage when a trained large language model processes a prompt and generates an answer. Inference is distinct from model training: one training run may attract attention because of its scale, but inference occurs every time someone asks a chatbot a question, calls an AI API or uses an AI-powered application.
According to the University of Michigan’s Language-Energy Divide project, energy demand varied substantially among the 122 tested languages and across the three open-source models. The accompanying EMNLP research paper also connects these disparities with differences in model performance.
The headline range—from 1.4x to 180x the English energy use—shows the scale of variation observed in the experiments. It should not be interpreted as meaning that every Hindi, Bengali or other non-English request always uses a fixed multiple of English energy. Results can change with the model architecture, tokenizer, prompt, output length, hardware, inference framework and task.
Why can one language require more AI energy than another?
Language models do not read text as humans do. They break text into tokens: numerical units that may represent a word, part of a word, punctuation mark or even a smaller character sequence. The number and structure of those tokens directly affect how much computation a model performs.
Tokenization can make equivalent text much longer
A tokenizer trained heavily on English may represent a common English word with one token while splitting an equivalent word in another language into several pieces. An idea that takes 20 tokens in English might therefore require many more tokens in a language poorly served by that tokenizer.
This matters because transformer models process tokens through multiple computational layers. More input tokens increase the work needed to understand a prompt, while more output tokens extend the generation process step by step. Poor AI tokenization can consequently raise latency, memory use, API charges and energy demand at the same time.
Writing systems are represented unevenly
Scripts such as Devanagari, used for Hindi, and Bengali-Assamese script can interact inefficiently with tokenizers whose vocabularies and training corpora prioritize Latin-script text. Characters, combining marks, spelling variations and encoded text sequences may be fragmented into more tokens than comparable English text.
The writing system itself is not inherently inefficient. The problem lies in how a particular model represents it. A tokenizer designed with balanced multilingual data may handle the same script much more effectively than one built around predominantly English material.
Models may generate longer sequences
The researchers also identified output length as an important part of LLM inference energy. If a model takes longer to reach an answer, repeats itself, translates internally or produces unnecessary text in one language, it must perform additional generation steps. Because tokens are generated sequentially, longer answers can noticeably increase computational requirements.
Limited training data can reduce efficiency
Many languages have far less high-quality digital training material than English. A model exposed to fewer examples may be less certain about syntax, vocabulary and factual expression. That can contribute to unstable generation, excessive output or difficulty following instructions.
Data scarcity is not simply a question of how many speakers a language has. Hindi and Bengali are spoken by hundreds of millions of people, yet the amount, diversity and quality of machine-readable material available to model developers may still lag far behind English.
The double disadvantage: more energy, lower-quality answers
The language-energy divide is especially consequential because higher energy use does not necessarily buy better results. The Michigan research indicates that languages with high inference costs can also receive lower-quality answers. Users may therefore wait longer or consume more computing resources while receiving a response that is less accurate, less relevant or less fluent.
This is a form of AI language inequality. English speakers benefit from years of data collection, benchmark development and model optimization. Speakers of underrepresented languages can encounter weaker instruction following, mistranslated concepts and reduced factual reliability, alongside inefficient tokenization.
Lower quality can create another layer of energy use. A user may need to rephrase a question, regenerate an answer or switch languages. Each attempt triggers more inference, turning a single unsuccessful interaction into several computationally expensive requests.
How the divide affects users and developers
Consider a healthcare information service that lets users describe symptoms in Hindi. If the model tokenizes those messages inefficiently and produces longer responses, the service may handle fewer requests on the same hardware. At large scale, higher operating costs could limit usage, encourage shorter answers or make the service less affordable in the communities it is meant to support.
A developer building a Bengali tutoring platform may face a related problem. Two lessons containing the same information could generate different token counts depending on language. If the platform uses a token-priced API, Bengali prompts and answers may cost more even before infrastructure-level energy differences are considered. Lower Bengali AI performance could also require additional validation or repeated calls.
Customer support provides another example. A company may advertise multilingual service but discover that certain languages have higher latency and lower first-response resolution. Without language-level monitoring, the company might blame user behavior or regional infrastructure rather than the model’s tokenizer and generation patterns.
API pricing, accessibility and environmental consequences
Most commercial AI APIs do not bill customers directly for joules of energy. They commonly charge by input and output tokens, model tier or request. Even so, language differences can appear on the invoice because inefficient tokenization turns equivalent messages into unequal token counts.
For a single user, the difference may seem small. For governments, schools, nonprofits and businesses processing millions of requests, it can become a material budget issue. Organizations serving multilingual populations may effectively pay a language premium, particularly when longer outputs combine with greater LLM inference energy.
Accessibility is also at stake. Providers may prioritize the languages that deliver the best quality at the lowest cost, leaving other communities with limited features or smaller usage allowances. That can reinforce the lack of digital content and investment that contributed to weak multilingual AI models in the first place.
The environmental impact depends on total request volume, hardware efficiency, data-center cooling, grid conditions and the source of electricity. A high energy ratio does not translate automatically into the same increase in carbon emissions. Nevertheless, repeated inefficiency across widely spoken languages can expand AI’s electricity demand and environmental footprint.
How developers can reduce the language-energy divide
Build better multilingual tokenizers
Tokenizers should be evaluated by language rather than only by average performance. Developers can compare token counts for semantically equivalent prompts, inspect fragmentation across scripts and expand vocabularies where common words are consistently split. A better tokenizer can lower costs without asking users to change languages.
Use language-specific routing and optimization
Not every request needs the largest general-purpose model. Language detection can route a prompt to a smaller model that performs well in that language, a regionally specialized model or an optimized inference engine. Quantization, caching, batching and limits on unnecessary output can further improve AI model efficiency.
Evaluate translation pipelines carefully
One option is to translate a request into a language the model handles efficiently, run inference and translate the result back. This can improve quality or reduce core-model token usage in some cases. It can also introduce extra model calls, translation errors and loss of cultural meaning. Translation should therefore be tested against direct inference for quality, latency, energy and total cost—not assumed to be automatically greener.
Invest in representative multilingual training data
More balanced data can improve vocabulary coverage, instruction following and answer quality. Useful investment includes professionally created datasets, community-governed language resources and content covering regional knowledge rather than translated English material alone. Consent, copyright and fair compensation remain essential.
Publish language-level measurements
Aggregate efficiency scores can conceal severe disparities. Model providers should report tokenization rates, quality, latency and energy estimates for individual languages and scripts. Developers should run the same checks on their own prompt distribution because benchmark averages may not reflect real users.
FAQ: AI energy consumption across languages
Does AI always use more energy in non-English languages?
No. The study found substantial differences in its tested models and tasks, but there is no universal non-English multiplier. Some languages and prompts may be close to English, while others can be far more demanding. Model design, tokenization and output length all influence the result.
Why were Hindi and Bengali especially expensive in the benchmarks?
The measured costs were associated with factors including token fragmentation, script representation, generated sequence length and uneven training resources. This does not mean Hindi or Bengali are inherently computationally expensive; it reflects how the tested models processed them.
Do more tokens always mean more energy use?
More tokens generally require additional computation, especially during sequential output generation. However, hardware utilization, batching, model size, caching and software optimization also affect energy use, so token count is an important indicator rather than a complete measurement.
What should multilingual application developers measure?
Track input and output tokens, latency, answer quality, retry rates and cost separately for each supported language. Where possible, measure hardware energy as well. Testing realistic user prompts is more informative than translating a small English benchmark and assuming equal behavior.
A model-efficiency problem with human consequences
The Michigan findings show that AI efficiency is also an accessibility issue. If language determines cost and quality, the people most underserved by digital systems may pay more for less. Better tokenizers, representative data and transparent language-level testing can help ensure multilingual AI becomes genuinely multilingual—not merely available in many languages.
