Sovereign by Design: How India Is Building Its Own AI Stack, One GPU at a Time

Something has shifted in Indian technology. For years the story was one of adoption — India as the world's back office, its engineers building and running the artificial intelligence dreamed up elsewhere. In 2026 the story is increasingly about ownership. A cluster of homegrown companies is building foundation models in Indian languages, the government is subsidising the scarce compute they need, and the phrase on every policymaker's lips is "sovereign AI." The ambition is no longer merely to use the technology, but to own the stack that produces it.

The clearest signal came from the funding rounds. Bengaluru-based Sarvam AI, which builds Indic-language foundation models, crossed into unicorn territory in June 2026 after raising 234 million dollars at a 1.5-billion-dollar valuation, in a round led by HCLTech with a 150-million-dollar strategic investment. Ola's Krutrim reached unicorn status around the same window. India minted multiple AI unicorns within the space of a year, an acceleration that would have seemed implausible when the last major model boom began abroad.

What "Sovereign AI" Actually Means

Sovereign AI is a phrase that can sound like a slogan, so it is worth being precise about what it entails. At its core it means a country controlling the full chain that produces artificial intelligence, rather than depending on foreign platforms for critical capability. In the Indian framing, that breaks down into a few concrete pillars:

• Indian data. Training models on Indian data, curated and governed under Indian rules, rather than on datasets that under-represent the country's languages and contexts.
• Indian compute. Access to GPUs and data centres located in India, so that the physical infrastructure of AI is not entirely offshore.
• Indic-language models. Foundation models that natively understand Hindi, Tamil, Bengali, Marathi and the dozens of other languages that global models handle poorly, if at all.
• Auditable government deployments. AI systems used in public services that can be inspected, governed and held accountable domestically.

The motivation is part economic, part strategic. A country that rents its intelligence from foreign providers is exposed to their pricing, their policies and their priorities. Building the stack at home is a hedge against that dependence — and a bet that Indic-language AI is a market global players will never serve as well as local ones.

The IndiaAI Mission: Compute As Public Infrastructure

None of this is possible without compute, and compute is exactly where the state has intervened. The IndiaAI Mission, approved by the Union Cabinet in March 2024 with an outlay of around Rs 10,372 crore over five years, was designed to bridge the gaps in India's AI ecosystem — and its centrepiece is subsidised access to graphics processing units, the specialised chips on which modern AI is trained.

The Mission set out with a target of roughly 10,000 GPUs and has blown past it. By 2026, reports indicated that more than 18,000 GPUs had been empanelled in the early phases, with the broader compute ecosystem making tens of thousands more available at subsidised rates — figures ranging up to 38,000 and even 45,000 GPUs cited across the shared-compute facility as procurement expanded. The idea is to treat AI compute as a form of public infrastructure: rather than leaving startups and researchers to bid against deep-pocketed global firms for scarce chips, the state aggregates demand and offers subsidised access through a national compute portal, alongside aggregated datasets and support for indigenous foundation models.

Compute is the oil of the AI economy, and the IndiaAI Mission is an attempt to build a public refinery — pooling the scarce, expensive resource that would otherwise be the preserve of a handful of giants.

The Companies Leading The Charge

Sarvam AI has become the poster child of the movement. Founded in 2023 by Vivek Raghavan and Pratyush Kumar — both alumni of the AI4Bharat Indic-language research initiative at IIT Madras — the company spans model development, inference infrastructure and enterprise applications tailored to Indian languages. Its trajectory has been steep: from a 41-million-dollar raise in late 2023 to a 1.5-billion-dollar valuation in under two and a half years, with reports that the government was weighing a small equity stake through compute support extended under the IndiaAI Mission, and that a further large round was being assembled with backers including global infrastructure investors.

Krutrim, backed by Ola, has pursued its own path into the unicorn club, building models and AI infrastructure. Together with a broader crop of startups, these firms represent the emergence of a genuine domestic layer of model-builders, not just application developers wrapping foreign APIs.

A Workforce That Has Already Adopted AI

The demand side of the equation is striking. A 2025 BCG survey found that around 92% of Indian workers reported using AI tools weekly — among the highest rates anywhere in the world, well ahead of many advanced economies. India's startup funding for AI reportedly surged sharply, with one report citing growth of around 277% in 2025. The picture that emerges is of a country whose workforce and entrepreneurs have embraced AI faster than almost anyone, creating a large, hungry home market for the very models the sovereign-AI push aims to build.

The Gap With OpenAI And Google

Enthusiasm should not obscure the distance still to be covered. The frontier of AI is defined by a handful of American and Chinese labs — OpenAI, Google DeepMind, Anthropic and their peers — whose largest models are trained on compute budgets measured in billions of dollars and on GPU fleets far larger than India's national pool. India's total subsidised compute, impressive against its own past, is still modest set against the clusters those firms command.

The challenge is not only hardware. Frontier model-building requires deep research talent, vast high-quality datasets, and the capital to absorb the enormous cost of training and re-training at scale. India's advantage lies less in matching the frontier head-on than in owning the layers where local knowledge is decisive: Indic languages, Indian data, cost-efficient models tuned for domestic use cases, and government deployments that demand auditability and data residency. The realistic goal is strategic autonomy in the layers that matter most to India, not a like-for-like race with the trillion-parameter giants.

The Ecosystem Challenges

Several hurdles remain. Under-utilisation of allocated funds has been a recurring criticism, with reports that a significant share of the IndiaAI Mission's outlay went unspent in a given financial year — a reminder that appropriating money is easier than deploying it well. India also depends on imported GPUs, leaving it exposed to global supply constraints and export controls, even as it builds domestic semiconductor capacity through separate projects. And sustaining a model-building ecosystem requires patient capital and research depth that take years to mature.

Talent is both a strength and a vulnerability. India produces a vast pool of engineering graduates and has one of the world's largest developer communities, but the specialists who can lead frontier research — the small cadre able to design and train large foundation models — are globally scarce and heavily courted. Many of the country's best researchers have historically gravitated to labs abroad. The emergence of well-funded domestic players like Sarvam, several of them founded by returnees and by veterans of homegrown research initiatives such as AI4Bharat, is beginning to offer a reason to stay or return. Whether that reverse brain drain becomes a trickle or a flood will shape how far India's ambitions can travel.

Data, Language And The Home Advantage

India's most defensible edge lies in its languages and its data. The country is home to hundreds of languages and a linguistic diversity that global models, trained overwhelmingly on English and a handful of high-resource languages, serve poorly. A model that genuinely understands Hindi, Tamil, Telugu, Bengali, Marathi and their dialects — including the code-switched, transliterated way Indians actually write and speak — is not something a Silicon Valley lab is well placed to build. This is the market that companies like Sarvam are targeting, and it is one where local knowledge, local data and local partnerships are decisive advantages rather than afterthoughts.

The same logic applies to government and enterprise deployment. Public services delivered in citizens' own languages, chatbots that handle welfare queries in regional tongues, and document processing across India's administrative languages all require Indic-first models. Add the demand for data residency and auditability in sensitive government use cases, and the case for sovereign, domestically built AI becomes not just a matter of national pride but of practical fit. It is in these layers — language, local context, governed deployment — rather than in a raw race to the largest model, that India's sovereign-AI bet is most likely to pay off.

By The Numbers

• Rs 10,372 crore — the IndiaAI Mission's five-year outlay, approved March 2024.
• 10,000 → 18,000+ → tens of thousands — the GPU target and its expansion across the subsidised compute ecosystem.
• 234 million dollars — Sarvam AI's unicorn-making round at a 1.5-billion-dollar valuation.
• ~92% — the share of Indian workers reporting weekly AI use in a 2025 BCG survey.
• ~277% — the reported growth in India's AI startup funding in 2025.

What Comes Next

The next phase will test whether India can convert inputs into outputs — whether subsidised GPUs and well-funded unicorns translate into models that are genuinely useful, widely deployed and trusted in government and enterprise. Watch for the maturation of Indic-language foundation models into products, the pace at which the Mission's compute is actually consumed, and whether domestic chip fabrication begins to reduce the dependence on imports.

The strategic wager is clear. India has decided that in the age of AI, it would rather build than rent — and that owning its data, its compute and its languages is worth the cost and the risk. The unicorns and the GPUs are the down payment. Whether they buy genuine sovereignty, or merely a well-funded catch-up, is the question the rest of the decade will answer.