The English tax — and what AI costs the rest of the world

Post 1 of 20  ·  The Problem

The world's most powerful AI tools were built for English speakers. For the 1.5 billion people who speak African languages, this is not a technical inconvenience — it is a tax.

Try this experiment. Open any major AI assistant and type a question in English. You get a fluent, thoughtful answer almost instantly. Now type the same question in Oromo — the language spoken by more than 50 million people in Ethiopia, the most widely spoken language in the Cushitic family, a language with a rich oral tradition, a formal writing system, and a literature that predates many European languages.

The AI will still answer. But something invisible has already happened before a single word of response appears. Something that costs you money, shrinks your context window, and quietly signals that your language is a second-class citizen in the world of artificial intelligence.

That invisible thing is called tokenisation. And understanding it is the first step to understanding why we built LENA.

The thing that happens before AI reads your words

Every large language model — GPT-4, Claude, Gemini, Llama — begins by breaking your input into small chunks called tokens before it can process them. A token is roughly a word, a syllable, or sometimes just a few characters. The AI never sees raw text; it sees a sequence of numbered tokens drawn from a fixed vocabulary.

Here is the problem. That vocabulary was built by analysing billions of words of text — most of which was in English. So English words map cleanly onto tokens. "Running" is one token. "Government" is one token. "Transformer" is one token.

Oromo does not get the same treatment. Because Oromo was barely represented in the training data for most commercial tokenisers, its words get shredded into fragments. The word mootummaa (government) might become five tokens instead of one. This ratio — tokens per word — is what researchers call fertility.^[1] A fertility of 1.0 means one token per word, perfectly efficient. Most frontier tokenisers achieve around 1.22 tokens per word for English. For Oromo and other African languages, the number is dramatically higher.

A 2026 study measuring 22 African languages across 11 frontier tokenisers found that the median fertility premium for African languages is 1.88× relative to English, rising to over 7× for Amharic when using common tokenisers — meaning the same content costs seven times more to process.^[2] For languages using the N'Ko script, the premium reaches nearly 9×.

Here is what fertility looks like across languages when using the tokenisers built into most commercial AI systems today:

This is not an abstract statistic. If you are a developer in Addis Ababa building a healthcare chatbot for Amharic-speaking patients, you are paying seven times more per API call than your counterpart in London building the same thing in English.^[2] Your effective context window — how much of a conversation the model can hold in mind — is seven times smaller. The tool works for both of you. But it works seven times harder against you.

The fertility tax is only the beginning

Tokenisation inefficiency is visible and measurable. But it is the surface of a deeper problem.

Every large language model trained today was trained primarily on English text — with some French, Spanish, Chinese, and German thrown in. African languages, collectively spoken by over one billion people across 54 countries, account for a fraction of a percent of most training corpora. A recent BCG report found that less than two percent of Africa's roughly 2,000 languages are adequately supported by large language models.^[3] This is not because the languages lack literature, history, or knowledge. It is because that knowledge has not been digitised at scale, and because the economics of AI development have not, until now, made it worth doing.

The consequence is that AI systems are not merely imperfect at African languages — they actively reflect a worldview shaped by the cultures, histories, and contexts that dominate their training data. An AI medical assistant trained on English-language clinical literature will encode assumptions about symptoms, treatments, and healthcare systems that may not match the realities of a rural clinic in Oromia. An AI legal assistant trained on Western common law tradition will misread the norms of customary law that governs daily life for hundreds of millions of Africans.

This is not a failure of the technology. It is a failure of who got to build it, and for whom.

The infrastructure problem underneath the language problem

Even setting aside the quality of AI outputs in African languages, there is a second problem: cost of access.

Training a large language model today requires thousands of high-end GPUs running for weeks or months. A single A100 GPU costs between $10,000 and $15,000 to buy. Cloud rental runs at $2–4 per hour. To train a competitive model from scratch costs hundreds of thousands of dollars at minimum and often millions.

The entire African continent hosts less than one percent of the world's high-performance computing infrastructure.^[4] An analysis of Zindi's network of 11,000 African data scientists found that only five percent have access to meaningful computational power — and the remaining 95 percent rely on their laptops or Google Colab's free tier for AI work.^[5]

This is the context in which we built LENA.

What DeepSeek showed the world

In early 2025, a Chinese AI lab called DeepSeek released a model that matched the performance of much larger, more expensive American competitors — at a fraction of the training cost. The reaction in Western AI circles ranged from surprise to alarm. But for anyone thinking about AI access in the developing world, the reaction should have been something else: relief.

DeepSeek's achievement was not a hardware breakthrough. It was an algorithmic one. By redesigning the way the model processes information — using sparse computation, more efficient attention mechanisms, and carefully engineered training recipes — they proved that the relationship between compute cost and model quality is not fixed. The frontier can be moved without building a bigger GPU cluster.

This matters enormously for Africa. It means the path to competitive AI does not run through billion-dollar data centres. It runs through smarter algorithms.

LENA is our attempt to walk that path — specifically for the languages, contexts, and deployment constraints of the African continent.

What LENA is

LENA stands for Linear Efficient Neural Architecture. It is a new architecture for large language models — not a fine-tuned version of GPT or Llama, but a genuinely different design, built from the ground up with three constraints in mind:

It must run on affordable hardware. The target inference hardware is a consumer GPU — the kind a university computer lab in Africa might actually own. Not an A100. Not a server rack. A single RTX 3060 with 12GB of memory.

It must treat African languages as first-class citizens. LENA's tokeniser was trained specifically to handle Oromo, Amharic, and English. In our tests, it achieves a fertility of 1.61 on Oromo and 1.62 on Amharic — compared to 4.5× and 7.4× for commercial tokenisers respectively. That means an Oromo or Amharic speaker using LENA pays roughly the same per-word cost as an English speaker. The tax is eliminated.

It must be open. The weights, the training data pipeline, the architecture code, and the evaluation results will be released publicly. LENA is not a product. It is a foundation.

Who this series is for

This is the first post in a twenty-part series that will document the LENA project from first principles to published results.

Some posts will be technical — we will walk through the mathematics of state space models, mixture-of-experts routing, and sparse attention, with diagrams and explanations designed for anyone who has studied higher secondary-school or fisrt year university mathematics. Some posts will be practical — we will show the code, the bugs, the wrong turns, and the fixes. Some posts will be about the larger question: what does it mean to build AI that belongs to Africa, not merely AI that works on African languages?

You do not need to be a machine learning engineer to follow this series. You need to be curious about why the world's most powerful technology systematically works better for some people than others — and interested in what it takes to change that.

That is what LENA is about. And that is what this series will show you, step by step.