How language LLMs will lead the India's AI leap

Hindustan Times

14-06-2025

The next great power struggle in technology won't be about speed or scale, it'll be about whose language AI speaks. Because trust in technology begins with something deeply human: being understood.
You trust a doctor who speaks your language. You trust a banker who understands your context. So why would you trust an algorithm that doesn't know who you are, where you're from, or what your words mean?
This question is being asked by governments, developers, and communities across the Global South who have seen how powerful large language models (LLMs) can be—and how irrelevant they often are to people who don't speak English or live in Silicon Valley.
In India, the response until now has been BharatGPT. This is a collaboration between startups like CoRover.ai, government-backed platforms like Bhashini, and academic institutions such as the IITs. Its aim is not to chase ChatGPT on global benchmarks. Instead, it hopes to solve problems at home—helping citizens navigate government forms in Hindi, automating railway queries in Tamil, or enabling voice assistants in other regional languages. CoRover has already deployed multilingual chatbots in sectors like railways, insurance, and banking. The value here isn't just in automation. It's in comprehension.
This isn't unique to India. In South Africa, Lelapa AI is working on InkubaLM, a small language model trained in African languages. In Latin America, a consortium is building LatAm GPT, rooted in Spanish, Portuguese, and indigenous dialects. Each of these projects is a rebellion: against invisibility, against standardization, against a worldview where the technology speaks only in one accent.
What's driving this shift? 'Current large language models do not adequately represent the linguistic, cultural, or civic realities of many regions,' says Shrinath V, a Bengaluru-based product coach and Google for Startups mentor. 'As governments begin exploring AI-powered delivery of public services, from education and legal aid to citizen support, they recognize the need for models that reflect local languages, data, and social context. Regional LLMs are being positioned to fill that gap,' he explains.
Manoj Menon, founder of the Singapore-based research firm Twimbit, is on the same page as Shrinath: 'With AI there are several nuances that come into play—how we train them to be contextually relevant for our local, national needs.'
At the heart of it lies something more political: digital sovereignty. Shrinath breaks it down and says, 'Data sovereignty is no longer an abstract idea. Countries don't want to depend on models trained on data they don't control. Indigenous models are a way to retain that control.'
It boils down to geopolitical leverage. Nations that build their own models won't just protect cultural identity—they'll shape trade, diplomacy, and security doctrines in the AI era. 'This is a reasonable argument,' says Menon. 'How we interpret a particular subject or issue depends completely on the context. Hence geo-politics is a significant input. Also the ability to train based on local issues and context.'
Viewed through this lens, the shift underway towards frugal AI is more radical than most people realise. These are models that don't need massive GPUs or high-speed internet. They're lean, nimble, and context-rich. Think of it like this: if ChatGPT is a Tesla on a six-lane highway, BharatGPT is a motorbike designed for rough, narrow roads. Not as flashy. But it gets where it needs to go.
'Most countries will want a say in shaping how AI is adopted, governed, and deployed within a sovereign context,' points out Shrinath. This matters because AI is starting to mediate access to public services—healthcare, legal advice, welfare. And in that context, a model that doesn't understand a citizen's language isn't just ineffective. It's dangerous. It can mislead, it can exclude and it can fail silently.
So yes, Silicon Valley still leads the headlines. But away from the noise, something deeper is unfolding. A shift in who gets to define intelligence, in whose language it speaks and in whose image it is built. Regional AI, says Menon, 'won't go head-on with what is built in Silicon Valley. They will complement it and their opportunity will help AI be more relevant locally.'
These regional AI efforts don't seek applause, they seek agency. They aren't chasing scale, they're chasing significance instead. This revolution is not being televised, it's being trained.