Your Next Million AI Users May Not Use English

Many AI startups launch first in English due to easier access to models, benchmarks, tools, and enterprise buyers, but this approach risks overlooking a vastly larger multilingual opportunity. According to the International Telecommunication Union, 2.2 billion people remained offline in 2025, mostly in low- and middle-income countries, where users expect digital products in their native languages, not just translated English interfaces.

Simply adding translation APIs to English-based models is insufficient: general-purpose models show uneven performance across languages, and token requirements can vary significantly. A language needing more tokens per concept raises inference costs, even if the model itself is priced similarly.

Founders should audit token economics, compare general-purpose and language-focused models using real regional inputs, and leverage sovereign or institutional language resources like India’s BHASHINI and BhashaDaan. Legal and technical vetting of such datasets remains essential, even with government backing.

Finally, interfaces should match local communication habits, voice-first or visual design may outperform text boxes where typing or literacy is challenging. Scaling multilingual AI demands starting with focused markets, measuring task success and support costs, then expanding incrementally rather than launching globally all at once.