The Language Tax Same question, up to 179× the energy — and where Persian sits
Ask a language model the same 900 reading questions in English and in Pashto, on the same model and the same GPU. A new EMNLP 2026 study measured the electricity: 17.6 kJ for English, 3,147 kJ for Pashto. 179 times as much, for answers that were right less than half as often. Here is what the paper found across 122 languages, and where Persian sits.
Based on Deng et al. (2026), University of Michigan, University of Aberdeen and University of Cambridge. Figures are the paper's; the Persian angle is ours.
- Pashto vs English
- 179×
- total energy, same 900 questions
- per-token gap
- 8.3×
- cheapest vs costliest language
- languages measured
- 122
- 56 high-resource, 66 low-resource
- Persian vs English
- ≈4×
- energy per answer (our arithmetic on the authors' data)
What was measured, and what was not
The 122-language results come from one open model, Qwen3-8B, on one NVIDIA L40S GPU, answering Belebele reading-comprehension questions with step-by-step reasoning. The authors repeat the key comparison on five models, two GPUs, six batch sizes and three tasks. They did not measure ChatGPT, Gemini or Claude, which would need the companies' cooperation. So these are measured joules for open models, not a bill from any commercial chatbot.
Each interactive figure awards a badge the first time you use it. There are six.
1 · Your language
What does one answer cost in your language?
Pick any of the 122 languages. The numbers are the authors' (Qwen3-8B, Belebele, 900 questions per language); per-answer figures divide their totals by 900.
- Energy per output token
- 0.14 J
- Tokens per answer
- 563
- Energy per answer
- 80.8 J
- 0.022 Wh
- vs English
- 4.1×
- Answers correct
- 67.4%
Energy per answer = the authors' unrounded energy per token × output tokens ÷ 900, from their released data. This reproduces the paper's Table 1 (English: 17,588 J for 900 answers). Accuracy is the paper's lenient score, which also counts answers recovered from output that missed the required format.
2 · Two penalties that multiply
More tokens, and each token costs more
The gap is a product of two factors. Low-resource languages make the model write far more tokens, and each of those tokens costs more energy. Pick a language to see both.
- Tokens per answer
- ×22
- Energy per token
- ×8.0
- Energy per answer
- ×179
22× the tokens × 8.0× the energy per token = 179× the energy per answer
The four low-resource languages in the paper's subset emit 20–25 times as many output tokens as English. That alone would be expensive. But the per-token cost rises too, by up to 8.3 times.
The authors' explanation for why the two travel together: longer answers take more GPU memory per request, so a server can run fewer requests at once, and a less busy GPU spends more energy on each token. They call for ablation studies to test it.
A double penalty
The costliest languages are also where the model does worst. English answers were right 94.6% of the time at the lowest energy; the five costliest languages were all at or below 43%, Shan at 10.6%. The authors stress this is co-occurrence, not cause: both come from weak tokenizer coverage and thin training data.
3 · All 122 languages
One model, 122 bills
Every dot is a language. Tap or hover to read it.
Energy per answer, log scale
Western Persian: 0.14 J/token, 67.4% correct, 80.8 J per answer
Higher energy, lower accuracy
Energy per output token against the share of answers correct
- High-resource
- Low-resource
Western Persian: 0.14 J/token, 67.4% correct, 80.8 J per answer
Most languages cluster between 0.10 and 0.20 J per token. A long tail of low-resource languages in non-Latin scripts, led by Tibetan, Pashto, Tigrinya, Shan and Amharic, sits above 0.4.
Script is not destiny. Simplified Chinese matches English at 0.10 J/token, and Egyptian Arabic, a low-resource variety, is among the five cheapest per answer; the authors attribute this to sharing Modern Standard Arabic's well-covered script.
4 · Where Persian sits
Persian: cheaper than most, about four times English
In the paper's table, Western Persian is a high-resource language and lands in the cheaper half. It still costs about four times as much as English per answer, and the model gets a third of the answers wrong.
- J per output token (English 0.10)
- 0.14
- tokens per answer (English ≈186)
- ≈563
- energy per answer vs English
- ≈4×
- answers correct (English 94.6%)
- 67.4%
Same script, very different bills
Every Arabic-script language in the table, energy per answer
Western Persian: 0.14 J/token, 67.4% correct, 80.8 J per answer
The Arabic script is no excuse on its own: Egyptian, Najdi and Modern Standard Arabic sit among the cheapest languages in the table, while Pashto, in the same script, is the most expensive of all 122. What differs is how much of each language the model and its tokenizer have seen.
For Persian speakers that means two things. The energy gap with English is real but moderate; the accuracy gap is the bigger cost. And it is a property of today's models, not of the language.
Tajik, a close relative of Persian written in Cyrillic, is listed as low-resource and costs about 2.7 times Western Persian per answer. Its 58.0% is the lenient score: almost every Tajik answer missed the required format.
5 · Not one model's quirk
The gap holds across models and tasks
The authors re-ran the eight-language subset on five models and three tasks. The ratio between the costliest and cheapest language never fell below 7×.
8.3× costliest ÷ cheapest
- English0.10
- Simplified Chinese0.10
- French0.11
- Russian0.12
- Southern Pashto0.84
- Tigrinya0.73
- Shan0.70
- Tibetan0.85
Bigger models cost more per token for every language (English rises from 0.10 J on Qwen3-8B to 0.51 J on Qwen3-32B), but the divide does not shrink with scale. The high- and low-resource groups never overlap on any model.
6 · Why it happens
Four things the model does in an under-served language
The paper reads the actual answers and finds four recurring patterns.
45×
Token explosion
In Tigrinya the model repeated one sentence 45 times. For Shan, Pashto, Tibetan, Tigrinya, Burmese and Amharic it emitted up to about 4,700 tokens per question, stuck in a loop.
3.4×
Over-generation
In Latvian it understood enough to answer, but at length: 896 tokens, 3.4 times the English answer, at a similar per-token cost.
0.78
Too many tokens per character
Tibetan text averaged 0.78 characters per token, against 4.84 for English: the tokenizer breaks the script into tiny pieces.
EN
Switching language
For Khmer, the model simply answered in English.
None of this is a verdict on the languages. It is how current models, tokenizers and training data treat them.
7 · Your turn
Which answer costs more?
Eight pairs from the full table. Guess which language costs more energy per answer, then see the numbers.
Which costs more per answer?
8 · What the authors recommend
Make the cost visible, then cut it
The paper ends with concrete proposals for researchers, model makers and the companies that serve models.
- 01
Report energy, not just accuracy
Add energy to the Responsible NLP checklist and to model cards. Two systems with the same accuracy at very different energy are not equivalent.
- 02
Report it per language
A model that is cheap and right in English but costly and wrong in Tibetan is two different systems for two communities. One “multilingual” claim hides both.
- 03
Serve each language on its own terms
Detect the language and cap output length, stop early, penalise repetition; tune batch sizes per language rather than for the average.
- 04
Route by language
Send some queries to smaller specialised models, or translate to English, answer, and translate back — except where culture, idiom or register matter.
- 05
Treat it as an equity issue
Equity audits at release, transparent per-language pricing, or subsidised inference for under-served languages.
Who pays, and who is served
The authors flag the risk themselves: a provider could use these numbers to deprioritise expensive languages. Their answer is that leaving the gap unmeasured lets that happen silently. Meanwhile the electricity, water and carbon are paid by everyone, and the worst service falls on speakers of under-served languages.
Read it carefully
Six cautions
- Open models, not chatbots you useQwen, Gemma and Llama on NVIDIA L40S and RTX 6000 Pro GPUs. ChatGPT, Gemini and Claude were not measured.
- The 122-language table is one setupQwen3-8B, one L40S, batch size 256, reading comprehension. Absolute joules change with hardware; the authors say to re-validate as hardware evolves.
- Translated promptsBenchmarks are standardised translations, not text people actually write. The authors read the gap as a lower bound for real, noisier input.
- Co-occurrence, not causeHigh energy does not cause low accuracy or the reverse; both come from tokenizer coverage and training data.
- Modelled totals, lenient accuracyTotals are steady-state energy per token × output tokens, not a meter reading of the whole run. The paper's Table 5 rounds per-token energy first, which inflates ratios; we use the authors' unrounded data. Accuracy is the paper's lenient score.
- Energy, not priceThe study measures joules on the GPU. It does not say what any provider charges.
Sources
- Deng, Shen, Feng, Nwatu, Chung, Chowdhury, Chen & Mihalcea (2026). The Language–Energy Divide: Measuring Energy Costs of Multilingual LLM Inference. EMNLP 2026, pp. 48057–48075 (PDF)
- ACL Anthology record for the paper
- Project page (University of Michigan LIT lab)
- Code: MichiganNLP/language-energy-divide (GitHub)
- Data: MichiganNLP/language-energy-divide (Hugging Face)
All figures are from Deng et al. (2026), quoted exactly; per-language values use the authors' released data (GitHub, Apache-2.0), and ratios marked “ours” are arithmetic on it. Charts redrawn from the paper's published data. The framing, the Persian section and any errors are ours, not the authors'.