Indian AI startups raised around $1.07 billion in the first half of 2026, up 33% year on year. The largest raise, around $234 million, concerns a company building models covering India's 22 official languages. This is a problem Western labs haven't solved, and it affects far more people than is commonly assumed.
We've covered endangered languages. Here's the neighbouring and far larger topic: living languages, spoken by hundreds of millions of people, that state-of-the-art models handle poorly.
Why models perform poorly in most languages
Three mechanisms compound, and they're technical before they're political.
Corpus imbalance. The web is overwhelmingly English-speaking. A model trained on it sees billions of English sentences and far fewer elsewhere. Quality mechanically follows that distribution.
Tokenisation. This is the most concrete and least known factor. Tokenizers are optimised for English. The same content expressed in a less represented language can consume two to three times more tokens, as we explained in our article on tokens.
The consequences are twofold and rarely stated: it costs more for the same service, and it fills the context window faster, so shorter documents can be processed.
Lack of cultural context. Even when the language is handled correctly, references, administrative practices and local social norms are missing.
This point deserves to be stated plainly. A user writing in a less represented language pays more for a lower-quality service. This isn't the result of any decision: it's a mechanical consequence of tokenisation. But the effect is very real, and it weighs most heavily on populations with the least means. It's the kind of bias that appears in no ethics charter because it's buried in a technical layer.
Why local players are taking it on
The calculation is logical. This market is very large, technically underserved, and Western labs have no strong incentive to address it: the engineering effort is significant and purchasing power per user is lower.
This is exactly the kind of space where a regional player has an advantage: it understands the need, it has access to the data, and it doesn't need to make a global general-purpose model profitable.
This is becoming possible for a reason we've documented: quality open models allow starting from a solid base rather than training from scratch. It's the same mechanism as with Thomson Reuters' legal model: the foundation is public, the specialisation is proprietary.
What this changes in practice
For hundreds of millions of people, the difference is far from theoretical.
An administrative service accessible in one's own language, access to health information, the ability to draft an official request: these are the uses we've described as most useful, in accessibility as in access to legal services.
And this raises a sovereignty question of the same order as the one we follow for compute: a country whose language is not properly handled by any model it controls depends on foreign players for its citizens' digital access to their own administration.
What to take away
There's a lesson that goes beyond the Indian case.
AI is often presented as a universal technology. In reality, it is deeply shaped by the data it was built on, and that imprint sits in technical details that are never discussed — a tokenizer, a corpus distribution.
What's at stake with these regional players is therefore not merely commercial. It's the question of whether this technology will reproduce existing linguistic hierarchies or help correct them. For now it reproduces them, and it takes a deliberate, funded effort for things to be otherwise.