Why small, application-specific models are India's path forward in AI

Even as all eyes are trained on the AI Impact Summit underway in New Delhi, the Economic Survey 2025-26 makes a strategic choice that deserves more attention than it's getting. Buried within the usual policy pronouncements about governance frameworks and data sovereignty is a genuine technical bet: small, application-specific models are India's path forward in AI.

The logic is straightforward. Frontier model development is capital-intensive, compute-hungry, and concentrated among a handful of firms with resources India doesn't have. Chasing this is expensive and probably futile. But application-specific models that solve defined problems in Indian contexts - healthcare screening, agricultural advisory, educational assessment, government service delivery - could be developed with fewer resources, run on existing infrastructure, and create genuine economic value.

The recent budget sweetens this with a 21-yr tax holiday for hyperscalers building data centres in India, addressing the compute constraint that might otherwise make even modest ambitions unrealistic.


This is a defensible position. The question is whether India has actually defined what it's betting on.

Read the Survey carefully and you'll find 'small models' describing fundamentally different things: fine-tuned open-weight models (e.g., take Llama, adapt it for Hindi legal queries), purpose-built models trained from scratch on Indian agricultural or medical data, distilled versions of larger models optimised for mobile deployment, and traditional machine-learning systems that aren't language models at all.

These aren't variations on a theme; they're different strategies with different implications. A future built on fine-tuning open-weight models means India's AI capability remains downstream of Meta's licensing decisions. A future built on purpose-built small models requires massive investment in data infrastructure that doesn't exist. The survey doesn't choose. It gestures at all of these as unified, but they aren't.

And, crucially: small relative to what? The frontier keeps moving. Today's small model is larger than GPT-3 was. The survey assumes a stable distinction that doesn't exist.

Startups, hyperscalers, IT services firms and frontier labs all have stakes in how India's small-model strategy unfolds. But the strategic implications are sharpest for two categories: frontier labs that own the base layer India would build on, and IT services giants whose business models depend on what 'building on top' means.

For companies like OpenAI and Anthropic, the survey reads less as a restriction than an invitation. India is signalling where it wants help, and it's not at the API layer.

The implicit message: don't just sell us base-model access and extract value from Indian users. Partner on building India-specific capabilities. Fine-tune for Indian languages and contexts. Co-develop models for Indian healthcare, agriculture and governance. Establish research relationships with Indian institutions. Train Indian engineers in model development, not just prompt engineering.

This is substantively different from global API provision. It requires local investment, genuine capability transfer and a willingness to build more than just customer relationships. The survey's contribution framework - where firms extracting significant value from Indian data must give back through local training, R&D investment or institutional partnerships - makes this explicit.

The smart play is probably to get ahead of these requirements voluntarily. Research collaborations, investment in local talent and partnerships on Indian-context model development will be important in positioning these firms as part of India's AI story rather than a foreign extractive presence.

For TCS, Infosys and Wipro, the survey represents an existential challenge dressed in opportunity language. The explicit warning that AI 'risks hollowing out India's core value proposition if adaptation lags' is directed squarely at them.

The small-model strategy offers a pivot, but only with genuine capability-building. These companies have client relationships, domain knowledge and implementation capacity. If value creation shifts toward application-specific solutions, they could evolve from service providers to solution builders: fine-tuning models for enterprise contexts, building retrieval systems on proprietary data, and integrating AI into business workflows.

But a services company wrapping API calls to foreign models isn't capturing value; it's intermediating it. The survey's emphasis on domestic capability signals that arbitrage won't be strategically favoured. Moving up the value chain means building the capacity to train and adapt models, not just deploy them.

India has correctly identified that it needs a strategy suited to its circumstances rather than one imported wholesale from elsewhere. The small-model bet could work. But a bet you can't define is hard to execute, and 'small models' remains a term that is asked to paper over choices India hasn't yet made.