The
Token Tax

Before an AI model reads a single word you type, something happens that most users never see. Your text is sliced into pieces — called tokens — by an algorithm called a tokeniser. These tokens are the actual input the model works with. Not letters, not words. Tokens. And the number of tokens your text produces determines everything: how much the model understands, how long it takes to respond, and exactly how much you pay.

In English, a token is roughly three to four characters, or about 0.75 of a word. The word "business" is one token. "Entrepreneur" might be two. English compresses cleanly because the tokeniser was trained primarily on English text — it has seen these words so many times that it represents them in single, efficient units.

In African languages, the same concept — the same sentence, the same meaning, the same business instruction — can require two, three, five, or even eight times as many tokens to express. Not because the language is more complex. Because the tokeniser has barely seen it before, and treats it as something close to raw character-by-character noise.

Tokenisers are not neutral tools. They are compression algorithms trained on data — specifically, on the same internet text that trained the underlying model. The vocabulary a tokeniser builds reflects the frequencies it observed during training. English words appear millions of times in the training corpus, so the tokeniser learns tight, efficient representations for them. Hausa words, Yoruba morphemes, and Amharic syllables appear so rarely that the tokeniser never developed compact representations for them at all.

The technical term for this is tokenisation fertility — the number of tokens produced per word. A fertility of 1.0 means every word maps to one token: maximally efficient. English averages around 1.3. Morphologically rich African languages — where a single root word can have dozens of grammatically valid forms — routinely achieve fertilities of 3.0, 5.0, or higher on standard LLM tokenisers.

This is the finding of a September 2025 study in arXiv (2509.05486), which evaluated 10 major language models across 16 African languages using the AfriMMLU benchmark — 9,000 standardised questions across five academic subjects. The finding was unambiguous: higher fertility consistently predicted lower accuracy, across all models, all subjects, every time. The tax is not just financial. It is intellectual. The model performs worse in your language because the language costs more to represent.

The token tax is not a single penalty. It is a compounding stack. Each layer amplifies the one below it, and African languages bear all four simultaneously.

Abstract the mathematics for a moment and consider a specific, realistic deployment: a mid-sized African bank building an AI-powered customer service system for its Hausa-speaking customer base in northern Nigeria. It expects to process one million customer interactions per month. Using GPT-4o, with an average prompt-plus-response of 1,000 tokens in English-equivalent meaning:

In English: 1M interactions × 1,000 tokens = 1 billion tokens. At $2.50 per million tokens: $2,500/month.

In Hausa: The same 1 billion units of meaning, after tokenisation at 4× fertility = 4 billion tokens. At $2.50 per million: $10,000/month. The same product. The same output quality — actually lower, because of Tax 4. At four times the cost.

The token tax is not paid once. It scales with every customer interaction, every document processed, every compliance check run. At enterprise volumes — tens of millions of queries — it becomes a line on the P&L that has no English-language equivalent anywhere in the competitive landscape.

The token tax is structural, not immovable. There are five approaches worth understanding — ranked from immediate and tactical to long-term and strategic.

The token tax is not a temporary inconvenience waiting to be patched in the next model release. It is a structural consequence of building AI predominantly on English-language data — and it will persist until the training data changes. For African businesses deploying AI at scale, it is an invisible line item on every AI invoice, a ceiling on AI performance in every African language deployment, and a compounding disadvantage that grows with usage volume.

Understanding it precisely is the first step to managing it. The mitigation strategies above are not hypothetical — they are operational choices available today. The organisations that audit their AI deployments for token efficiency, select models based on language-specific tokenisation performance, and structure their prompts to minimise fertility-inflated token counts will achieve meaningfully better cost and performance outcomes than those that deploy off-the-shelf and absorb the penalty invisibly.

The tax is real. It is quantifiable. And unlike most taxes, understanding it is already half the exemption.