India's AI vision: Championing sovereign workflows over models

Context

India is ranked third globally in AI competitiveness. The government has invested ₹10,300 crore (~$1.3 billion) in a five-year AI mission.¹ Private capital is flowing into the sector at scale, with major technology players committing billions of dollars to AI infrastructure and ecosystem development over the coming years.² At the same time, a small but growing group of domestic firms is trying to build large foundation models with ambitions to compete on scale.³ On paper, India presents the profile of a country with strong potential to become an AI leader. However, that isn’t the case yet.

The ambition India is currently projecting to the world is as follows: AI sovereignty equals compute plus foundation models plus data. Build all three domestically, and India joins the elite tier of AI nations. From a policy lens, this sounds lucrative. From an execution lens, it doesn’t hold together, and the gap between these two lenses is where India’s AI strategy needs to be worked upon.

The case made here is simple: India should stop chasing model sovereignty and start optimising for workflow sovereignty. The two paths are not identical. One asks India to outspend the US and China on terrain where they have a decade's headstart (the US leads AI through top research, investment, and policy support, with China tailing behind, driven by major state funding and a layered regulatory framework).^4,5 The alternative path builds on a pattern India has already demonstrated at scale—adapting global technologies into locally relevant systems and deploying them across massive populations. India's digital public infrastructure, including Aadhaar and UPI, has enabled identity, payments, and data systems at population scale,⁶ with UPI alone becoming the world's largest real-time payments system by transaction volume.⁷ These are considerably different paths, with different resource requirements and different odds of success. Pursuing both paths concurrently risks diffusing effort, and a clearer commitment to one could yield faster, more durable results.

What AI sovereignty actually requires

The word 'sovereignty' gets used loosely and often with incomplete context. In practice, full-stack AI sovereignty requires four things simultaneously: domestic compute hardware, frontier model capability, proprietary training data, and the talent to sustain it all. India has made progress in parts of this stack, but it does not yet possess all four at the depth, scale, or maturity needed to claim end-to-end AI sovereignty.

Where India's real advantage lies

India has already built something that most countries in the world have not: a nationwide, interoperable Digital Public Infrastructure (DPI) that serves over a billion people. Aadhaar covers 1.3 billion citizens with biometric identity.²¹ Unified Payments Interface (UPI) processes over 20 billion transactions per month, with more than 400 million active users.^22,23 Jan Dhan accounts have brought formal banking to over half a million of previously unbanked households.²⁴ DigiLocker, Open Network for Digital Commerce (ONDC), AgriStack, and the account aggregator network further complete the picture. This is not a base layer for consumption. It is a base layer for AI. And it is a base layer that no other country with India's scale has assembled.

Here is where things get interesting and where the strategy debate should be happening. The question is not whether India can build a model that competes with GPT-4. It is whether India can build AI-powered workflows on top of this DPI that are impossible to replicate elsewhere, because the underlying data infrastructure, language context, and market conditions are uniquely Indian.

Five domains where India can win

What India should do differently

The policy direction is not wrong. It needs reorientation, not replacement. But reorientation requires transparency about what the current framing obscures.

Start with measurement. India currently measures AI progress in GPUs procured, models announced, and mission budgets allocated. None of these tell you whether AI is changing anything for anyone. The more useful metric is deployment depth—that is, how many government services has AI embedded in the actual workflow, not as a pilot, not as a demo at a summit, but in production, at scale, serving real users in real languages. That number is still uncomfortably small. Fixing it requires a different kind of urgency than announcing a 120B-parameter model.

The data layer also needs honest investment. AIKosh's 9,500 datasets are an impressive start.³⁰ India's strategy should be to leverage its vast and unique datasets (from weather to genomic data to digital payments records) and develop AI-infused workflow IP in key industries. The competitive moat India could build in health AI, agricultural AI, and financial AI depends almost entirely on whether it can curate and interoperate its data assets. Today, models are the base commodity, while data + context and last-mile implementation are the high-value differentiators.

And then there is the DPI–AI interface—the APIs, consent frameworks, and language infrastructure that allow AI to plug into Aadhaar, UPI, Ayushman Bharat. This is the most underfunded and under-discussed part of the entire stack. More GPUs do not help here. Policy clarity, engineering investment, and data governance do.

India's IT industry rose by owning the process layer for global firms. The next version of that story is owning the AI workflow layer—building and operating agentic systems that run financial, healthcare, and governance processes at a fraction of current cost. That is a more durable competitive strategy than building an Indian version of ChatGPT, and it is one that plays to India's actual strengths.

'AI superpower' may mean something different in India's context than it does in Washington or Beijing. In those capitals, it means producing the largest models, attracting the most frontier talent, and winning the R&D race. In India's context—given its infrastructure, its linguistic diversity, its population scale, and its structural constraints—it should mean deploying AI at the last mile, embedded in the workflows of a billion people who currently have no access to the productivity and insight that AI can deliver.

The sovereignty debate is not irrelevant. India should develop multilingual model capability for reasons of cultural preservation and strategic resilience. It should invest in computing infrastructure. It should build the talent pipeline. But these are instruments, not ends.

The end is a country where AI is not a luxury for the English-speaking, smartphone-owning urban professional, but a tool embedded in the farmer's advisory service, the rural health worker's diagnostic kit, the small merchant's credit application, and the first-generation student's tutoring session. That outcome requires workflow thinking, not model thinking.

India has built the DPI rails. Now it needs to run AI on them—and define what AI leadership means for a country that has always found its competitive edge not in invention, but in the intelligent, large-scale application of technology to human problems.

Contributors