Paytm 创始人 Sharma：印度 AI 缺的不是钱是野心

On the morning of October 7, the man who taught India to pay with a QR code stood before a few thousand AI practitioners and told them they were wasting their time. “The problem is we’re unambitious,” Vijay Shekhar Sharma said. Why, the Paytm CEO wanted to know, are Indian builders still making wrappers and shaving minutes off deliveries in 2026? Why was India content to be labelled the “use-case capital” of AI?

“User is good,” he said, “but staying a user is a loser.”

Then he made it personal. A model, he said, costs under $100 million. Intelligence is borderless, and a country must own what it uses. Paytm, he said, trains its own models and runs its own inference. For founders who wanted to do the same, he had a proposition: “Either come over, join us, or build your own and take money from me.” Personal cheques of ₹1–2 crore from his own pocket, potentially more if serious builders showed up, along with Paytm’s distribution network for the founders he chose.

It became the most shared moment of the summit. But the offer deserves to be taken seriously for a reason beyond the headline. Not for funding, which it barely is, but as a diagnosis of what Sharma believes is holding Indian AI back.

What ₹2 Crore Buys

By Sharma’s own arithmetic, ₹2 crore, a little over $200,000, amounts to roughly a quarter of one percent of the cost of a model. It will not fund a full training run. It might buy a small research team’s time for a year. Nobody in Hall 1 misunderstood the difference.

Which means the cheque was never the point. A personal cheque can move quickly. It carries a public endorsement. And, most importantly, it is a dare small enough that a founder cannot easily blame the money for turning it down. Indian AI startups raised $1.56 billion across 206 deals through August, according to market data tracked by Venture Intelligence. Capital, at least in aggregate, has not been entirely absent. Sharma was pricing something harder to measure: nerve.

India’s AI stack · 3 layers · Cheque to scale

Sharma's money points towards the bottom layer of the AI stack. His distribution offer, the one worth taking, points towards the middle one.

The Room Was Already Building

The awkward part of calling Indian AI founders ‘unambitious’ is that some of the people the remark was aimed at were already on the same programme, with receipts.

Ganesh Gopalan of Gnani.ai opened his keynote with the story of an investor who told him Indian companies “only call APIs”. Gnani has since shipped an open-weights 30-billion-parameter model. Gopalan's description of the problem was just as pointed as Sharma’s. Relying on a frontier model, he argued, is “renting your intelligence”. His ambition was larger still: India needs hundreds of frontier-model companies.

Chariot’s Parth Sarthi defined sovereign AI as owning the recipe, “not fine-tuning someone else’s model”. NeuroDx’s Siddharth Panwar presented Sphota, an architecture that refuses to treat tokens as the substrate of intelligence.

These are not founders waiting for ₹2 crore to discover what ambition looks like. What they need is something else: the belief that the work is worth doing from India in the first place.

That was the belief Gopalan’s investor withheld, and the very thing Sharma’s offer, deliberately or otherwise, puts on the table.

The Real Offer Was Distribution

Buried beneath the cheque story was a second offer that could prove substantially more valuable: access to Paytm’s distribution network. Paytm did not win digital payments in India with better cryptography. It won by being on every chai-stall counter first. India discusses foundation models almost entirely as a supply problem: GPUs, data, researchers. Sharma was pointing at demand. A model nobody uses is a paper. A model wired into a few hundred million wallets is an institution.

The rest of Cypher agreed without meaning to. BHASHINI runs 22 million inferences a day on 200 GPUs because the government provides the distribution. ShareChat serves 200 million users using in-house models because the app itself provides the channel. In each case, the model followed the channel. Founders considering Sharma’s invitation do not necessarily need another cheque; they want a channel.

What ‘Unambitious’ Misses

Sharma’s criticism gets the wrapper economy right: most of the 70-plus expo stalls sold agents built on someone else’s weights. But ‘build your own model’ can mean several different things. Treating them all as the same ambition risks confusing scale with strategic value.

Amit Sheth’s case for ‘small AI’ in factories and dams, for instance, offers a different vision of what matters. So does Genpact’s work on sub-$100 distillation, which beats a frontier model on X-rays. Sovereignty, meanwhile, received a more practical definition at the summit. Anuraag Saxena described it as “making sure no one else can unilaterally switch you off”. By that measure, open weights on your own inference, which Gnani sells and Sharma says Paytm runs, is already sovereign.

Who Flinches

A year ago Sharma called India the “AI use-case capital” at the Shiprocket SHIVIR 2025 summit. At Cypher he said “we should be the LLM capital of the world”. The room applauded. Then the audience went back to sessions on context engineering.

That is not necessarily hypocrisy; it is a snapshot of the ecosystem India has built: a handful of labs attempting to develop models that matter, a long tail of companies that excel at the application layer, and a sizeable middle that has yet to say clearly which side it wants to be on.

Sharma’s cheque forces that question into the open. Accept it, and a founder publicly signals a willingness to build. Decline it, and, in Sharma’s framing, the founder risks being labelled a user rather than a builder.

The amount was never meant to fund a model. It was meant to find out who flinches.