Part 2 of 2: Why India’s AI future may run through the edge, not the frontier

In Part 1 of this series, we argued that India’s MANAV Vision has a gap at its centre. Data sovereignty without model sovereignty is incomplete. And trying to close that gap by racing to build a single national large language model the size of GPT or Gemini is a path that is expensive, uncertain, and runs against the structural realities of India’s democratic and linguistic architecture.

There is a better path. And its logic begins not with what India lacks, but with what India uniquely has.

The Problem That Was Actually a Misdiagnosis

The conventional framing of India’s AI sovereignty challenge goes roughly like this: India needs a big model trained on Indian data to truly own its intelligence layer. The obstacle is compute, capital, and coordination. Solve those and sovereignty follows.

This framing is not wrong. But it is incomplete in an important way.

The assumption buried inside it is that frontier-scale centralised models are the only architecture that produces sovereign intelligence. That assumption is worth questioning, especially for a country as linguistically plural, administratively federated, and institutionally diverse as India.

The data governance problem that makes a single national LLM so difficult, health records fragmented across states, legal databases inaccessible, agricultural data in departmental silos, language data sparse for dozens of smaller Indian languages, is not a bug in India’s system. It is a reflection of how India actually works. Federal, distributed, contextually specific.

The question worth asking is whether an AI architecture designed around that reality might produce something more genuinely sovereign than one designed around the assumption of centralisation.

What Regional Small Language Models Actually Solve

A small language model, or SLM, is not simply a shrunken version of GPT. It is a model purpose-built for a specific domain, language, or use context. Trained on focused data, optimised for particular tasks, deployable at far lower compute cost, and crucially, ownable by the institution or community closest to the problem.

When you think about India’s actual AI needs at the ground level, this architecture fits remarkably well.

A health SLM trained on ASHA worker interaction data from Tamil Nadu, on district hospital records, on the language patterns of frontline health conversations in Tamil, would outperform a general-purpose foreign model on that task. Not because it is bigger, but because it is right. It would carry the assumptions of the community it serves, not the assumptions of a model trained on Western medical literature and English-language health forums.

The same logic applies to a legal SLM trained on district court judgments in Marathi, a Krishi SLM trained on rainfall patterns, crop advisory history, and mandi price data from Punjab, an educational SLM trained on the National Curriculum Framework and supplementary materials in Odia. Each of these is a tractable, fundable, governable project. None of them require India to out-compute the United States.

The data governance bottleneck also becomes far more manageable at this scale. Getting a state health department to share anonymised records with a regionally accountable SLM initiative is a fundamentally different political and bureaucratic challenge from getting it to contribute to a national data lake where control is opaque and accountability is distant. Smaller scope means clearer consent, faster iteration, and stronger institutional ownership.

The Federation Question Is Real and It Cannot Be Ignored

This is where intellectual honesty requires a pause.

A Tamil health SLM and a Marathi legal SLM are genuinely valuable. But India’s governance challenges are frequently cross-domain and cross-linguistic. A citizen navigating the intersection of land rights, health entitlements, and financial inclusion needs intelligence that can reason across those boundaries. At some point, the edge models need to talk to each other.

How you federate a distributed SLM ecosystem without recreating dependency at the integration layer is the hardest unsolved question in this architecture. If the layer that connects regional SLMs together runs on GPT or Gemini, India has solved the edge while leaving the core unaddressed.

This is not a reason to abandon the SLM path. It is the design challenge that Indian researchers, policymakers, and technologists need to be working on urgently and in public. 

The federation architecture, the protocols by which sovereign edge models communicate without surrendering reasoning to a foreign core, is arguably the most important open problem in India’s AI sovereignty agenda right now. It is also one where India has genuine structural motivation to lead that no Western lab has.

India Does Not Need to Win the Frontier Race to Achieve Meaningful Sovereignty

DeepSeek’s emergence last year reframed something important in global AI discourse. The assumption that the frontier belonged permanently to the biggest compute clusters was wrong. Architectural efficiency can partially substitute for raw scale. A well-designed smaller model trained on the right data can outperform a larger general model on specific tasks.

India can learn from this without copying it. The goal is not a Chinese-style state-directed frontier lab, which requires political and institutional conditions India does not have and arguably should not want. The goal is a public infrastructure layer for AI that mirrors what India built for payments and identity: foundational, open, interoperable, and owned by no single commercial interest.

Bhashini is already close to this for language access. The IndiaAI Mission’s compute initiative creates the substrate. What is missing is an explicit framework that treats regional SLM development as public infrastructure, ring-fences funding for foundational model work rather than application-layer wrappers, and builds the federation architecture that connects sovereign edge intelligence into something with national coherence.

This is, structurally, exactly what ISRO did with space. India did not try to build NASA. It identified the specific capabilities that mattered for Indian needs, built institutional patience around them, and created something that is now genuinely sovereign and increasingly globally relevant.

Why India Is Uniquely Positioned to Export This Model

Here is the argument that should animate India’s AI diplomacy in the coming decade.

No other country is better placed to build and demonstrate a federated, pluralistic, community-rooted AI sovereignty model. The United States and China are both building centralised intelligence architectures that reflect their own political and economic logics. For most of the Global South, neither model is replicable or desirable.

India’s democratic architecture, its federal structure, its demonstrated capacity to build digital public infrastructure at scale, and its extraordinary linguistic diversity make it the natural laboratory for a different paradigm. One where AI sovereignty does not mean a single powerful national model. It means a constellation of contextually sovereign models, governed close to the communities they serve, connected by open protocols, and built on public infrastructure that no single commercial interest controls.

If India builds this and it works, it does not just solve India’s wrapper economy problem. It gives Bangladesh, Nigeria, Indonesia, and Brazil a model they can adapt. It positions India not as a country catching up to the AI superpowers but as the country that invented a different and arguably more democratic architecture for AI sovereignty.

That is a more interesting story than “India’s national LLM.” It is also more honest about what India can realistically build in a timeframe that matters.

For MANAV to Hold, the N Must Be Architected, Not Just Declared

The fire that Modi invoked at the summit did not benefit humanity because one civilisation hoarded it. It spread. It became foundational because it was distributed.

India’s AI sovereignty does not have to be built on the logic of concentrated intelligence. It can be built on the logic of distributed sovereignty, where the intelligence lives closest to the people it serves, governed by the institutions accountable to them, connected by open infrastructure that no foreign actor controls.

That architecture will not emerge from the market alone. It requires the same quality of public imagination and institutional commitment that gave India UPI, Aadhaar, and Bhashini.

The N in MANAV is the hard part. But India has done hard infrastructure before. The question is whether the political will that gathered in New Delhi on February 19 was the beginning of that commitment, or its most eloquent expression.

The difference will show not in the speeches. It will show in the architecture.

END OF ARTICLE