The choice of language for a Large Language Model (LLM) has a significant impact on its cost and can create a divide between English speakers and the rest of the world. A recent study reveals that English-language inputs and outputs are much cheaper than those in other languages. For example, Simplified Chinese costs about twice as much, Spanish costs 1. 5 times the price, and Shan language costs 15 times more. Research conducted by the University of Oxford found that processing a Burmese-written sentence with an LLM costs 198 tokens, while the same sentence in English only costs 17 tokens. This means that accessing the service through an API incurs 11 times more cost for the Burmese sentence compared to the English sentence. The tokenization model used by AI companies converts user input into computational cost, making models accessed outside the window of English language more expensive to access and train on. Languages like Chinese, with a different and more complex structure, have a higher rate of tokenization. For instance, a tokenization example by OpenAI's GPT3 tokenizer reveals that the phrase "your affection" would be only two tokens in English but eight tokens in Simplified Chinese. Despite the English phrase being longer (14 characters) compared to the Simplified Chinese one (4 characters), the higher token-to-char ratio makes it more expensive to implement the API for languages other than English.

The cost-effectiveness of English in AI-related expenses is unparalleled, with Chinese costing twice as much as English in terms of required tokens per output. This cost difference stems from the training data available to AI companies. Furthermore, achieving recursive training, or training AI models on their own outputs, is a desire for AI companies. However, research suggests that AI networks become unstable when trained multiple times on their own synthetic data. Different ways of quantifying costs, such as bit or character-counting, still face similar problems as tokenization and cannot surpass the practicality and lower costs of English. This issue is not limited to one model or version but affects multiple language models. The fact that the companies introducing Large Language Models are mostly based in America contributes to the cost difference, as lower usage costs and higher availability of quality data are inherent to the territory. The cost disparity has prompted several countries, including China and India, to develop their own initiatives to train and deploy native-language LLMs to keep up with the pace of innovation brought by English-based AI networks. It is crucial to proceed with caution in the complex realm of AI, considering the far-reaching consequences of each step taken.