Generative AI burns more energy every time it answers in a language other than English, and new research puts a stunning number on the gap: up to 180 times more electricity for a single response. University of Michigan Engineering researchers measured the disparity across three open-source models and 122 languages, finding that around 5.2 billion people must spend more than twice the energy to generate text in their first language that an English speaker spends in theirs.
The penalty compounds with accuracy, because the languages that consume the most energy also tend to produce the least reliable answers of generative AI, turning a physics quirk into a fairness problem. This article breaks down how the study was run, why tokens and scripts drive the disparity, what it means for global inequality and token-priced services, and what companies can actually do about it.
A Study Measures the Divide
A new University of Michigan Engineering study quantifies something users of generative AI have long suspected: the same question costs more when it is asked in a different language. The disparity is greatest in languages that do not use Latin script and have less digital content available for training.
The researchers measured three open-source models: Alibaba’s Qwen 3, Google’s Gemma 3, and Meta’s Llama 3.1. Each model answered multiple-choice questions from Belebele, a reading-comprehension dataset spanning 122 languages, while the ML.Energy framework from Michigan Engineering recorded the electricity consumed by every answer. The models were also tested on grade-school math word problems and open-ended prompts translated into seven languages with Meta’s No Language Left Behind tool.
Across the board, generative AI could use 1.4 to 180 times more energy to generate non-English text. Most Western European languages landed at twice the energy of English or less. Hindi prompts cost 7.9 times more energy than equivalent English ones, and Bengali costs 15.3 times more, despite each language being spoken by hundreds of millions of people.
Closed commercial assistants do not let outsiders plug in energy profilers, so the study leaned on openly weighted models anyone can run and instrument. That choice made the measurements possible and gives the results unusual credibility: three independent development stacks, one consistent finding.
The Michigan team has been building the measuring infrastructure for years, and ML.Energy exists precisely so researchers can audit generative AI power draw without lab access to proprietary systems. Expect this paper to become the template for future language-cost studies, because replication is now a matter of downloading weights and rerunning prompts rather than negotiating access.
The Double Penalty: Energy and Accuracy
The energy gap alone would be an engineering curiosity. What makes it a societal problem is that it arrives paired with a second penalty, and the two reinforce each other in exactly the wrong direction.
“The languages that typically consume more energy also tend to have lower accuracy in the model output,” said Naihao Deng, doctoral student in computer science and engineering at the U-M AI Lab and first author of the study. Users pay more kilowatt-hours and receive worse answers. Generative AI charges a premium at the counter and short-changes the customer at the same time.
Deng’s warning is blunt: most people will want to use AI models by asking questions in their native languages, and they may feel pressure to use them as the tools weave into more parts of society. The double penalty makes it more likely that certain people will be priced out of generative AI, and that companies will be more reluctant to serve models in certain languages.
The inequality loops back on itself. Languages with less online text earn worse accuracy and higher energy costs, which discourages investment, which keeps the training data scarce. Researchers estimate that around 5.2 billion people now face more than double the energy cost per generated word, according to the study’s measurements combined with demographic data from Ethnologue.
English speakers often adopt generative AI as a convenience; much of the rest of the world is being routed toward the same tools by institutional necessity, from government portals to homework platforms. A language surcharge layered onto mandatory adoption is exactly the pattern that turns a technology gap into a class gap, because the surcharge is invisible, metered, and never itemized on the invoice.
Classrooms feel it first. A student in Kolkata drafting an essay in Bengali burns generative AI credits three times faster than a classmate in Leeds writing the same assignment in English, and neither pays nor grade depends on which tokenizer won.
The mechanism behind the divide is deceptively simple, and it starts with the basic currency of every large language model: the token.
When working in other languages, models used more energy because they had to generate more tokens, the basic units of data AI uses internally. A token can be an entire phrase, a single syllable, or one character. Models convert prompts into tokens before processing them, then convert output tokens back into human language, and every step consumes electricity.
The measured multipliers relative to an equivalent English prompt show how steeply the curve rises once a language leaves the Latin-script, data-rich world.
| Language group | Energy vs English | Speakers affected |
|---|---|---|
| English (baseline) | 1.0× | Reference group |
| Western European languages | Up to about 2× | Hundreds of millions |
| Hindi | 7.9× | Hundreds of millions |
| Bengali | 15.3× | Hundreds of millions |
| Measured extremes | Up to 180× | Roughly 5.2 billion over 2× |
The table explains a business decision hiding in plain sight. Companies often price model access by tokens generated, so a language that needs four times the tokens costs roughly four times the money before the energy bill of running the query is even considered.
Token pricing quietly taxes languages at different rates, and most users never see the exchange rate. Understanding tokens is the first step to auditing whether generative AI bills you fairly across your languages, and several cloud vendors now publish per-token rates that make the arithmetic painfully literal.
Scripts That Burn Brighter Than Others
Not all writing systems are punished equally. The problem escalates for scripts whose characters require more code to display, and the deeper reason lies in how those scripts combine.
The Devanagari script used in Hindi, Nepali, and Marathi, and the Bengali script, have primary characters that change form and combine with accents to produce up to millions of unique characters. Tokenizers trained mostly on English slice these scripts into long ribbons of tiny fragments, and each fragment is another step of memory and energy.
Bar lengths below encode the measured energy multiplier per language relative to English, using the upper bound for Western European languages.
Memory is the quiet amplifier. More tokens occupy more computer memory, leaving less available for processing information, so models handle tokens in smaller batches, which burns more energy per token. That effect compounds the disparity compared with English text rather than merely adding to it.
The compounding matters for climate accounting too. Datacenter dashboards track total gigawatts, but the per-language breakdown reveals that generative AI’s carbon cost is distributed along linguistic lines, a dimension almost no sustainability report currently mentions.
Even punctuation is not neutral. Scripts that attach marks to base characters force generative AI to choose between storing giant character tables or emitting extra token fragments, and both options cost energy that Latin typing simply never incurs.
What Generative AI's Numbers Reveal
Zoom out from the laboratory and the measurements read like a risk dossier for anyone deploying generative AI across borders. The same prompt carries three different bills depending on the language it is written in.
The first bill is electrical and lands on the operator. The second lands on the customer through per-token pricing. The third, an accuracy penalty, lands on everyone, because errors in high-stakes translated queries carry their own cost.
| Bill | Who pays | Why it grows for non-English |
|---|---|---|
| Energy per response | Operators and grids | More tokens, smaller batches |
| Invoice per request | Businesses and users | Per-token pricing multiplies output |
| Accuracy shortfall | Everyone downstream | Scarcer training data per language |
Companies watching these numbers draw an uncomfortable conclusion: serving some languages simply earns less margin per query. Left alone, the market drifts toward under-serving the very languages whose speakers stand to gain most from AI assistance, the inequality the Michigan researchers warn about.
Data scarcity is the root cause behind both halves of the penalty. Languages with less digitized content give tokenizers fewer whole words to learn and models fewer examples to reason from, so generative AI pays twice at the same counter, once in electricity and once in errors.
Energy audits already exist for utilities, and fairness rules already exist for essential services. A public, per-language energy disclosure for widely deployed generative AI would combine both traditions, giving regulators a metric as legible as fuel-economy stickers and giving researchers a trend line to watch.
Nothing stops enterprises from demanding that disclosure today. Procurement teams routinely insist on service-level reports for uptime and latency; language-level energy and accuracy reports are the same conversation with one more column.
The double penalty also skews research itself, because most published benchmarks are run in English and reported as universal. When generative AI results are finally published per language, the gap between showcase demos and deployed reality narrows, and procurement teams gain the comparison tables they currently have to build by hand.
That reporting shift is already beginning. Conference review boards are nudging authors to disclose multilingual evaluation, and energy figures like the Michigan numbers give the conversation a hard unit of measure. Expect generative AI procurement documents next year to quote watts-per-answer the way cloud contracts once began quoting cost-per-gigabyte.
Why Model Size Does Not Save You
For buyers in multilingual markets, the practical takeaway is budget arithmetic rather than ethics lecture. If Bengali queries cost fifteen times an English one, a support ticket flow in Dhaka is a different capital project than one in Dublin, and quotes built on English pilot numbers will miss badly.
The study’s own spreadsheet of energy use by language gives every organization a starting calibration before generative AI workloads ever touch production. Re-running it against your own prompts costs an afternoon and predicts the year.
A natural hope is that the divide is a growing-pain artifact, destined to vanish as models scale. The study’s answer is unwelcome: the disparity persists at larger model sizes and across different types of computer processors.
The energy gap between English and non-English prompts held whether the researchers measured small variants or larger ones of the same families, and it survived swaps of processor class. Scale changes the absolute wattage of generative AI; it does not change the relative tax imposed by language.
The cause is baked into tokenization itself, a design layer that every major architecture shares. Until tokenizers allocate symbols more generously to dense scripts, a trillion-parameter model fragments Bengali into the same expensive confetti as a seven-billion one.
Hardware vendors are attacking the memory and batching sides of the ledger, work that lowers the floor for everyone, as the industry has been documenting with full-stack co-design efforts. But cheaper chips shrink all bills proportionally; they do not erase the ratio between an English prompt and a Hindi one.
Three concrete changes would move the ratio: tokenizer vocabologies rebuilt for script coverage, models trained with language-balanced data budgets, and inference stacks that batch by script rather than mixing everything. None requires new physics, and all are engineering choices waiting on incentives.
The incentives exist in miniature already. Regional startups compete on multilingual quality, open communities maintain language-specific variants of popular models, and academic labs keep publishing tokenizer efficiencies. Scaling those efforts is coordination, not research.
So the burden shifts to deliberate engineering, the routing, translation, and language-specific optimization the Michigan team recommends. Nothing about the trajectory is inevitable, and nothing fixes itself by adding parameters alone.
Fixing the Language–Energy Gap
The researchers do not stop at diagnosis; they hand companies a practical playbook for shrinking the disparity without waiting for the next hardware generation.
The first lever is specialization. Companies could route user queries to models tailored to specific languages, sending a Bengali question to a model that handles Bengali tokens efficiently instead of forcing a generalist to chew through fragments. A second shortcut already exists in the wild: translating user prompts into English before feeding them to a more generalized 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, Janice M. Jenkins Collegiate Professor of Computer Science and Engineering at Michigan and director of the U-M AI Lab. Crowdsourcing native speakers to translate material or create resources could build the training data that makes both accuracy and efficiency possible.
Translation-first architectures carry a caveat worth stating plainly: machine translation of legal, medical, or idiomatic content can fail silently, trading an energy penalty for a comprehension risk. The recommendation works best when reliability is measured, not assumed.
Hybrid designs point the way. Keep a small, sharp language-specific front door for common requests, escalate the ambiguous cases to the giant generalist, and the average cost per query falls while the tail stays safe. Generative AI economics improve most when architecture stops pretending every question needs the same hammer, and routing tables become the cheapest fairness tool available.
Smaller companies can act alone today. A startup serving one language region can fine-tune an open model, shrink its token waste, and undercut global per-token pricing while winning accuracy in its own market.
Specialization plus crowdsourcing remains the most direct route to fair generative AI economics, because both attacks shrink the token waste that drives the divide at its source.
What Buyers Should Ask Next
If your organization deploys generative AI in more than one language, the Michigan study converts a vague unease into a procurement checklist with hard numbers behind it.
Start by measuring which languages your users actually speak and what each language costs you per thousand requests in tokens and watts. A support desk serving English, Hindi, and Bengali is running three different businesses cost-wise, and per-token invoices hide the difference. Benchmark a handful of real prompts per language with an energy profiler and compare, exactly as the study did with ML.Energy.
The industry conversation is already moving toward making every layer cheaper, from Nvidia’s full-stack memory codesign work that shrinks batching waste to open ecosystems where Hugging Face tracks who profits from widely reused models. The same open model ecosystems that let Michigan test Qwen, Gemma, and Llama are the fastest route to language-specialized variants, because tokenizer and vocabulary fixes can propagate through them without waiting for a single lab.
Buyers can also ask vendors two blunt questions: what is your energy cost per response in our top five languages, and what is your accuracy in those same languages? The research says the answers correlate, and vendors who refuse to measure both are telling you which penalty they plan to keep.
The study’s authors see the path clearly: better tokenization, smarter routing, and crowdsourced data can close most of the gap within a few model generations. Whether they do depends on whether the 5.2 billion people affected ever become a demand signal powerful enough to matter. Your next multilingual invoice is one way to send it.
Track one number quarterly: average energy per resolved answer, split by language. It is a single metric, it fits on a slide, and every layer of your generative AI stack will react the moment leadership starts asking why the Hindi line is higher than the English one.
Every fix in this story runs through the same decision: whether the people building generative AI treat language efficiency as a feature. Generative AI that charges 5.2 billion users a hidden surcharge is not doomed, but generative AI that audits the surcharge, publishes it, and then shrinks it is the version that deserves the grid it burns.
References
- TechXplore: Generative AI uses more energy on non-English text
- University of Michigan: AI energy use by language (data spreadsheet)
- Deng et al., The Language–Energy Divide: Measuring Energy Costs of Multilingual LLM Inference, EMNLP 2026
- Ethnologue: demographics of world language speakers
- Belebele: reading comprehension across 122 languages
- ML.Energy: energy measurement framework from Michigan Engineering
- Meta: No Language Left Behind translation tool
- Character inventory growth in Indic scripts (LREC 2024)
- Orthographic depth and character combination in scripts
- Science X editorial team: story credits
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
