It is the country's largest companies that still aren't placing the wager.

The Indian conversation about artificial intelligence has settled into an anxious, deferential, faintly fatalistic mood. India is, after the United States, the second-largest user of ChatGPT in the world, and its largest free user. Every fortnight, a new think-tank paper warns that the country has fallen behind, that we ought to have started ten years ago, that the frontier of artificial intelligence is already a closed shop between Silicon Valley and Hangzhou.

Vivek Raghavan, who built Aadhaar's biometric infrastructure over twelve years and is now co-founder of Sarvam AI, takes a quieter and more interesting view. India, he argues, hasn't missed the bus; Indian capital has refused to board it.

Speaking on a recent episode of the Bharatvaarta podcast, in conversation with host Roshan Cariappa, Raghavan made the technical case bluntly. The frontier of AI, with its GPT-5s, Claudes and Geminis that cost their builders hundreds of billions of dollars, is not where the action for a country like India lies. That money buys the last few per cent of capability.

The first ninety-five per cent is the AI that teaches a child, talks to a kirana-shop owner in Marathi, summarises a contract, or runs a customer-service line. It can be built at a fraction of that cost, by teams that stay one generation behind the absolute frontier and ride the cost curve as it falls.

Raghavan calls this position "frontier minus one". It may be the most underrated strategic argument in the Indian AI debate.

In recent weeks, Sarvam shipped a model that, by Raghavan's own measure, beats the version of DeepSeek that astonished the world a year ago. India, he tells Cariappa, is "behind only in months, not years". For a country told it has missed the bus, that is a remarkable claim, though it does not by itself settle the matter.

The case for frontier minus one

An AI model, at its core, is a very large set of numbers, what engineers call parameters or weights. When a model is described as having a hundred billion parameters, the engineer's job has been to find the right values for those hundred billion knobs such that, given a piece of text, the model produces a sensible next word. Training is the process of finding those values.

Once trained, the model is, as Raghavan puts it, "stateless". It becomes a piece of code that takes input and produces output, and nothing more.

The cost of training scales with model size and the volume of data used. The per-flop cost of compute, meanwhile, keeps falling year after year. A model that costs X hundred million dollars to train this year will cost a meaningful fraction of that to retrain next year. "If I try to train a model that costs me X million dollars or X hundred million dollars to train this year, and I want to train the same model next year," Raghavan tells Cariappa, "you can do that at a cost that is a fraction of what it was."

This is the fast-follower's gift. The frontier labs run at the bleeding edge of what compute can buy, while anyone willing to lag by twelve to eighteen months gets the same capability for an order of magnitude less money.

Raghavan reaches for an analogy from an earlier wave. Forty years ago, a serious computer was the preserve of defence labs and scientific institutions, and the gap between what a researcher used and what a household had was vast. Today, the standard MacBook a student carries to university is good enough for almost everything an ordinary user, or most professionals, actually needs.

The cutting edge has migrated into research labs and high-end engineering, and life on the rest of the curve has become functionally indistinguishable from life at the top.

The same pattern, he argues, will play out in AI. "If the intent is to actually do things that allow most people - both individuals and businesses and governments - the kinds of things that they want to do with AI," he says, "you can actually do like ninety-five to ninety-nine percent of those things using a model that is trained in a fairly cost-efficient way. And that's really the game."

A second insight follows. Frontier models are generalists by construction; they have to be everything to everyone. But many of the most economically valuable applications of AI are narrow. Think of voice agents that talk to bank customers, models that read radiology scans, agents that draft Hindi legal contracts, or tutors that teach Class IV mathematics in Tamil. In such narrow domains, smaller focused models routinely outperform giant generalist ones.

"If you want to go deep into a particular domain," Raghavan says, "even having a small model for that particular domain, you can actually do better than even the largest frontier models, because I'm focused, I've made my domain narrow, and I can do that at a cost."

For India, the calculation is different. The real choice is whether to industrialise frontier-minus-one for a 1.4-billion-person market, shipping cheap, focused, sovereign AI into education, healthcare, governance and small business.

The technical scaffolding is already in place. Sarvam, building on Indian-context data, riding the falling cost curve, and focused on voice and Indian languages, has shown that it can be done.

The stakes are not academic

If frontier-minus-one is so attainable, why is India not full of Sarvams? When Cariappa, at one point, asks Raghavan to put on a policy hat, the reply is among the most sobering formulations of the AI question one is likely to hear from an Indian technologist. AI, Raghavan says, is the new nuclear NPT.

"There are going to be two AI powers in the world," he tells Cariappa. "There are de facto powers, the United States and China. Will any other country join them or not is the question. And if you don't join them, you will have lower rights."

The Non-Proliferation Treaty divides the world into nuclear haves and have-nots, and the difference is not academic. It determines what technology you can buy, what rules you must follow, and whose strategic interests dictate yours. Raghavan's claim is that AI is heading the same way, with the two declared powers setting the terms, and everyone else negotiating from beneath.

The asymmetry is already visible in compute. The most advanced GPUs are American, the most aggressive open models are Chinese, and India sits in the middle, dependent on both and sovereign in neither.

The IndiaAI Mission, which provided Sarvam and ten other model companies the GPU access without which their models could not have been trained, is a useful and important step. Raghavan is clear-eyed about its limits. "It's a great start," he says, "but we need to do more."

Chips and semiconductors, he notes, are a five-to-ten-year game. The mission is table-stakes rather than victory.

The strategic logic that follows is stark. Either India develops a sovereign capacity to build, run and ship AI at scale, within the country, on Indian terms, and audited by Indian institutions, or it accepts a future in which the most consequential infrastructure of the twenty-first century is rented from Washington or Beijing on terms set by them.

Frontier minus one is precisely how a country with a $2,500 per-capita income makes that bet affordable. Why, then, is Indian capital not deploying it?

The ambition deficit

Raghavan's answer, delivered with the diplomatic carefulness of a man who has spent fifteen years inside Indian institutions, is the most quietly damning passage in the conversation. "I think sometimes," he says, "we have lacked ambition to actually go and play in these games."

Indian industry, he tells Cariappa, looks for "guaranteed returns" on investment horizons of one or two quarters, and is uncomfortable funding a three-to-five-year R&D cycle whose payoff cannot be projected on a spreadsheet.

The result is something every honest assessment of Indian capital has noted for two decades. Private-sector R&D as a share of GDP runs at less than half the global benchmark, and the country's largest companies, despite balance sheets that comfortably permit deep-tech bets, place none.

"There are obviously companies with tremendous balance sheets," Raghavan says, "which can afford to do those kinds of things. And they just need to have the confidence that they can win in the world if they make that investment. That confidence is what we need to have."

The diplomatic register cannot quite disguise the message. The familiar Indian businesses, from cement and steel to petroleum, retail and telecom, are paying or have moats. None of them requires anyone to bet five thousand crores on a research agenda that might fail, which is why nobody does. No Indian conglomerate, neither Tata nor Reliance nor Adani nor Birla nor Mahindra, has placed a Sarvam-scale wager on foundation-model AI.

Cariappa, to his credit, names the obvious tension on air. "We're still at $2,500 per capita," he points out. "There are more basic needs and there are easier ways to create wealth and value."

The argument is unanswerable as far as it goes; a country at India's income level has a thousand reasons to extract reliable returns from the businesses it understands. Raghavan's reply, though, goes to the heart of the indictment. That argument, he suggests, used to be safer than it is now.

"As some of these things, the easier ways of very successful business models - some of these things are changing," he says. "So hopefully, as these things change, people will realise that they need to do new things."

The old moats of distribution, capital scale and regulatory familiarity were built in a world where intelligence was a constraint and labour was cheap. AI inverts both inputs, and a company that does not invest in foundational technology now is committing to defending an old moat against an enemy with a flying machine.

The cleanest evidence of the deficit is Raghavan himself. At sixty, he had already spent over a decade as a full-time, unpaid volunteer at UIDAI, helping architect Aadhaar's biometric stack. He had earned, by any reasonable measure, the right to a quiet retirement of open-source projects and government advisory work, and that, by his own account, was the plan.

Then, in late 2022, ChatGPT shipped. Raghavan describes the moment with characteristic understatement. "I said, this is something very important, and we need to figure out how to do this in India. And my initial instinct was, let's go ask the government, let's go and ask philanthropy."

These were the instincts of a man who had spent fifteen years inside the public-system entrepreneurship of UIDAI, where the rule is to find the right institution, build the right alliance, and make the case. "But then I realised this is moving way too fast. And this is too important. And either I can give advice to everybody, gyan to do this, or I can say, okay, let's try to do something." And so he started Sarvam.

Raghavan, who had built one civilisational system for the country, knew every Indian institution that ought to be funding the next one. He looked at the available options of government, philanthropy, and the conglomerate balance sheets that Indian newspapers describe as "patient capital", and concluded that none of them would move fast enough or boldly enough. The only way to make the bet, in the end, was to make it himself.

When the most senior public-system technologist of his generation has to start a startup at sixty because no Indian institution will, the question stops being whether India has missed the AI bus, and becomes what the conductors are doing.

Raghavan is too gracious, and too dependent on those same institutions, to put it that bluntly. He frames the failure as a mindset that has not yet shifted, but hopes the shift is coming. "It has to shift," he says.

He believes the conglomerates will eventually realise that the old moats are collapsing and that defending them is more expensive than building new ones, and reports seeing "early green shoots" of intent. He has the patience of a man who has been inside Indian institutions long enough to know that bullying them does not work.

Green shoots, however, still have to grow, and Sarvam on its own cannot carry the country.

The Aadhaar lesson, told properly

There is a precedent, and it is one Indians often cite without telling the whole story. Aadhaar and UPI are routinely held up as proof that India can build civilisational technology, the country's standing reply to anyone who argues it cannot do hard things. That much is true; the rest of the lesson is uncomfortable.

Raghavan, who knows that history better than almost anyone, draws it out. India Stack, he says, was conceived as digital public infrastructure rather than as a private platform, a third path between the American model, in which private platforms own the rails, and the Chinese model, in which state platforms own the rails and the private sector serves them. It was, deliberately, sovereign by design.

"We designed it to be sovereign," he says, "because we said that it is so critical that at the wrong point, you know, these things can't be." The architects neither chased the latest foreign technology nor waited for international best practice. They picked a principle, sovereignty, and built around it.

The other half of the lesson is more uncomfortable. India's most successful piece of contemporary digital infrastructure was built by a network of public-system entrepreneurs, philanthropic capital, and a cohort of volunteers like Raghavan, working under the political cover of a government willing to back the bet.

The country's largest companies were beneficiaries of this infrastructure rather than its architects. Reliance's payments business runs on a UPI it did not build, and every Indian bank now relies on an Aadhaar layer that no bank built.

"We could have asked Google to make an identity system," Raghavan tells Cariappa. "Why didn't we do that? I'm sure they can." The rhetorical question carries its own answer. Some technologies are too foundational to outsource, and the institutions willing to take that view, in the Indian experience, have not been the country's largest balance sheets.

Frontier-minus-one AI, Raghavan implies without quite saying it, is the same kind of bet. It is foundational, too important to outsource, and, as with Aadhaar in 2009, it is being made by a small group of people who have decided not to wait for permission from the institutions that ought to be leading.

The pattern is real, but it cannot be the country's only setting. Aadhaar took twelve years and a generation of volunteer technologists to reach maturity, with UPI riding on top of it. AI, moving as Raghavan keeps reminding Cariappa in a feedback loop that compounds faster than any technology before, will not wait twelve years for Indian capital to find its courage. The conglomerates may have, at most, two.

The choice

There is a sentence Raghavan returns to, almost as a refrain, when Cariappa presses him on what Indian industry needs to do. "They just need to have the confidence that they can win in the world if they make that investment."

The sentence is a deceptively gentle one. In the standard reading, confidence is something corporate India will eventually acquire when conditions are right, precedents are sufficient, and consultants have written the reports. Raghavan, having lived through the building of Aadhaar and now of Sarvam, treats it differently. Confidence of this kind, in his telling, has to be decided rather than waited for.

The frontier-minus-one window is open now, and will not be open indefinitely. The cost curve that makes a $200-million model affordable today will, in the other direction, also raise the floor on what counts as sovereign capability tomorrow.

The geopolitical window, in which the United States is competing with China hard enough to make a third pole strategically attractive, is similarly open and similarly perishable.

The IndiaAI Mission, the GPU access, the early model companies, and the goodwill of a government that has, for once, moved faster than the technology, together amount to the most favourable environment Indian deep tech has ever seen. It will not last either.

India has not missed the AI bus, and Raghavan, with characteristic understatement, has already shown why. The harder question is what the people with the country's capital, credibility and largest balance sheets intend to do with the brief, miraculous window in which all three could still matter.

They will either decide to win this, or they will explain to their children why they didn't.