AI Startups in India: Why This Wave Is Global

AI startups in India used to be a bet on the Indian market. Cheap smartphones, cheap data, a billion people arriving online at once, and business models that only worked at Indian prices. That produced real companies. Nearly all of them were domestic by design.

The closing panel at Y Combinator's first Startup School in India made a sharper claim than the usual be-more-ambitious speech. The claim is structural: the thing that used to force Indian software companies to sell into India has been removed, and it was removed by the technology rather than by anyone's strategy. Below is that argument, the parts of it that hold up, and the part founders should handle more carefully than a stage full of good news had time to.

What AI startups in India look like right now

Puneet Kumar co-founded Supr Daily, the Mumbai grocery delivery company that went through Y Combinator in Winter 2017 and was later acquired by Swiggy. He spent time after that exit as a venture partner at Nexus Venture Partners and is now building again. Asked what founders in the room should work on, he gave the thesis in one line: "for the first time we can build global companies out of India."

The companies sitting around him matter more than the line does. Emergent, the AI app builder founded by twin brothers Mukund Jha and Madhav Jha, says it reached $100 million in annual run-rate revenue eight months after launch. Giga, co-founded by Varun Vummadi, sells customer operations automation to DoorDash. Puneet's point about both was that their founders had not lived in San Francisco before starting them, which used to be close to a prerequisite.

Puneet's own history is the control group. He said on stage that Supr Daily reached around $100 million in annual revenue run rate with two engineers, one of whom was him, before AI coding tools existed. That is an extraordinary number. It is also a purely Indian number, because a daily milk and groceries subscription is a business you win one city at a time, and Supr Daily won six of them, all in India.

Why India's last wave was local and this one is not

The distinction Puneet drew is the useful part of the panel, and it is worth restating precisely, because it is a mechanism rather than a mood.

The mobile wave tokenized labor. A phone in every pocket meant anyone could say, in his words, "I have 1 hour free, I can deliver for Swiggy." That created hyper local network effects, which is why the same idea produced separate winners in India, Latin America and the United States. The supply was physically local, so the company had to be local too, and the moat was density you accumulated one city at a time.

His line for what replaced it: "the AI revolution is not local." The supply in this wave is a model endpoint, and it answers the same way from Bengaluru as from San Francisco. There is no local density to accumulate, and none to defend. Puneet's conclusion follows from that: "This is about do you understand this technology 10x better than everyone else," and his claim, which is about people rather than infrastructure, is that "nobody does that better than in India."

Treat that last part as the pitch. The mechanism is solid. The ranking of national talent is a claim nobody can settle, and a founder who builds a plan on it is building on applause. The narrower version is the one worth taking: when the enabling technology is available everywhere on identical terms, the differentiator moves from access to depth. That is a competition Indian engineers can enter on equal footing, which was not true of the previous three waves.

Emergent co-founder Mukund Jha makes the same argument from the operator's side in a separate Y Combinator interview: building a local company for India and building a global company take exactly the same effort, so think global from day one. He puts India at roughly 10 percent of Emergent's revenue.

Selling globally without a warm intro

The old objection to building for the world from India was never ambition. It was access. Selling software into the United States meant a warm introduction, which meant a network, which usually meant living in San Francisco for a few years first.

Puneet's evidence that this has weakened is specific. He described a company in a recent YC batch, run by a third-year IIT student, that cold emailed insurance companies in the United States and sold to them. He called it unimaginable, which is the right reaction. US insurance is one of the harder enterprise buyers on the planet, and one of the slowest.

His explanation for why it works now is that the demand side changed at the same time everywhere: "everyone simultaneously across the world just understands AI is important." Buyers are not doing you a favor by taking the meeting. My read on why: the demand is real, and the supply of vendors who can actually deliver against it is not. So the filter shifts: "This is not about where the budget goes, this is about whose solution is better and can drive better outcomes." Which leads to the instruction: "we have to get rid of our preconceived notions that we need warm connects."

Two caveats are worth holding, because the panel was speaking to a room it wanted to energize.

• A cold email that works is a cold email with a working product behind it. Puneet's own framing carries the condition: "if you have a great product, people are open to it." The unlock is not outbound as a tactic. It is that a demo can now travel without you in the room.
• The mandate window closes. Enterprise buyers are experimenting because AI is new and their boards are asking. In two years the same buyer has an incumbent vendor, a renewal cycle and a preferred-supplier list. Cold access is a condition of this moment, not a permanent change to how enterprises buy.

Second mover is a strategy now, not a consolation prize

Ankit Gupta, a general partner at Y Combinator who co-founded Reverie Labs before joining the firm, noted that almost none of the day's success stories were first ideas. When the moderator added that most were not first movers in their space either, Gupta took the point further: "find something that's kind of working and then do it better than them and then beat them."

He hedged it in the same breath, and the hedge is the strategy's whole limit: "unless that first thing had incredible network effects, which few things do, you actually can just beat them from a better product."

It works because coding agents collapsed the distance between product clarity and shipped product, so a team that sees the problem more sharply can close a two-year head start in months. It stops working the moment the incumbent has accumulated something you cannot rebuild by shipping faster: a proprietary data asset, a two-sided marketplace, an exclusive distribution contract. Before you pick a second-mover target, name what the leader has accumulated. If the honest answer is a head start and a marketing budget, go. If it is a data flywheel, you are not competing on product.

The example Gupta cited from earlier in the day was Giga, which won DoorDash with a team of eight against competitors with several hundred employees. It is worth reading that one from the founder's side, because it complicates this section in a useful way: the same deal is the spine of our piece on the forward deployed engineer, where Varun Vummadi describes it as a three-month pilot at a company that happens to be a fellow YC alum. A better product won it. A warm path still opened the door.

He was also blunt about the excuse founders use to avoid this. On the claim that coding agents produce slop, his answer was that the output says more about the operator than the tool: "the people who use them well produce incredibly sophisticated code." The discipline that separates the two is the subject of our piece on moving from vibe coding to agentic engineering.

The one constraint that is genuinely harder for AI startups in India

Here the panel got honest, and here a founder in Bengaluru faces a real problem a founder in Palo Alto does not.

Gupta described spending four months pushing coding agents as hard as he could, and reported a threshold: "unless you are paying for at least that level of usage, you actually are not anywhere close to the frontier right now." The level he meant is the $200 a month tier. He added that "there's a meaningful unlock you get from just letting the tokens rip," and noted that Y Combinator's CEO Garry Tan spends thousands of dollars per day on tokens, which Gupta described plainly as a position of privilege rather than as a benchmark.

$200 a month is a rounding error for a funded US startup. It is a meaningful fraction of a stipend for a student in Chennai. That gap is the real inequality in this wave, and it deserves naming precisely because the rest of the panel's argument depends on it. If the differentiator is depth of technical understanding, and depth is bought with tokens, then token budget has quietly replaced the warm introduction as the access problem.

The panel offered two answers, and both are partial.

1. Open weight models. Gupta said he had been running the MiniMax models and others, and rated them without ceremony: "They're really cheap and they're pretty good." He put the line where it belongs, at the frontier being worth paying for on "a subset of tasks, especially coding" and not on everything else. Arnav Sahu, a partner at Peak XV who spent about four years at Y Combinator, pointed at OpenCode as a YC company built on open models. Where that economics holds and where it stops being free is the subject of our piece on the open source coding agent.
2. Go work somewhere that pays for it. Sahu's advice to the budget constrained was to join a company that gives engineers unlimited token budgets, on the logic that "being early in the game as fast as possible as you can early in your career will just give you such a huge advantage on how the tools work." That is genuine advice, and it is also advice most of a room that came to start companies will ignore.

The direction of the curve is the reason to plan around this rather than despair about it. Both men landed on the same forecast. Sahu put it as "we should safely assume that the cost of tokens will keep coming down." Gupta put it as "cost of compute per level of intelligence is going to go down a lot." The move is not to sit and wait for the price to fall. It is to be precise about which parts of your work genuinely need the best model available and route the rest to something cheap. Every serious team ends up with that discipline eventually. It just arrives earlier when you cannot afford to skip it.

This is also why the next-billion-users question and the token-cost question are the same question. Gupta pointed to Vidit Aatrey of Meesho, who had talked about using voice AI to bring the next billion people online for shopping, and observed that a product like that only works if the models run at a price point those users can reach. Consumer scale in India is a cost engineering problem before it is a product problem, which is the same conclusion that holds for consumer AI generally.

What YC says it actually looks for

Jon Xu, a general partner at Y Combinator and previously co-founder and CTO of FutureAdvisor, was asked to clear up misconceptions about applications. His answer was almost aggressively unglamorous.

On the application itself: "clarity above all else." Founders write to sound impressive. If a partner cannot understand what you are building, nothing else in the form matters.

On what is being evaluated: "we're not investing in ideas." The day's own evidence backed that up, since nearly every founder who spoke had gone through some number of pivots to get to the idea that worked.

On taste, which he said gets misread as visual polish: "it's fundamentally more about intention." Are your design choices backed by insights from customers you actually talked to, and how fast can you get to those insights and put them in the product.

On agency, he recommended the Paul Graham essay Relentlessly Resourceful and framed it as a question about whether you let the conditions of the world happen to you or impose your intent on them. He closed with the one line that is easiest to act on: "your rate of learning matters a lot."

Gupta added the most concrete instruction of the day, a definition narrow enough that you can fail it: "a project is when two people build something that was not assigned to them and get someone to use it."

Three conditions, each of which excludes a common form of resume padding. Two people, so it is not a solo exercise. Not assigned, so it is not coursework or an internship ticket. Someone uses it, so it is not a whiteboard. As he noted, you can complete an entire computer science education and a job without ever having done one.

The reason this matters for finding ideas is mechanical rather than inspirational. Building surfaces ideas that thinking does not: "you can just very quickly find extremely good ideas that would not have been obvious if you were just on a whiteboard thinking." Xu made the same point from the other direction, describing what the strongest young founders do. "They're following their curiosities," and choosing to "work on the things that are barely good enough for the models to do," because that is where "you're going to figure out where the bottlenecks are," and "those bottlenecks are some really good ideas."

That is a searchable definition of a good AI startup idea, and it beats most of them: find the task the current models can just barely do, and build the thing that makes them do it reliably. It also lines up with what founders describe when they are asked directly how to come up with startup ideas, which is almost never a moment of insight and almost always a byproduct of building.

The advice question, handled honestly

Sahu took the part of the conversation most panels skip. His starting point was that "the future of the India AI ecosystem is going to be defined by all the people in this room" rather than by the prior generation, and therefore "The traditional advice of most educational systems and the non-AI native people is irrelevant now."

He extended it to the career calculus: the prestigious safe job may not exist in a decade in its current form, so "historically that safe path might actually now be the risky path."

Then he did something worth copying. He caveated it against the actual conditions of the people listening: "it is very hard for an average young Indian to take the types of risks that someone who grew up in Silicon Valley can." He named the reason, social safety nets, and observed that "the Indian startup ecosystem is booming but maybe not as deep as the US." He went further and said that for someone from a humble background, taking the high-paying stable job is a genuine win, and that he hopes many of them do exactly that.

That caveat is the difference between advice and a sales pitch. The founder path is not strictly better. It is better for people who can absorb the downside. Knowing which one you are is planning, not a character flaw.

His constructive version was about proximity: "in AI actually it's very dangerous to follow the advice of people who are not AI native because they're just not in the game with you," and the remedy is "surrounding yourself with people who are at the cutting edge," which he called a deliberate choice that compounds across a career.

The security review is the first real gate when you sell abroad

Now the part the panel did not have time for, and where the cold email argument meets its first genuine obstacle. This section is mine rather than theirs, and it comes from the other side of the table.

A cold email to a US insurance company can absolutely land the meeting. The second email is from that company's security team, and it arrives with a questionnaire. For an Indian startup selling to regulated buyers in the United States or Europe, that questionnaire is the actual gate, and it is where the everyone-is-open-to-it era quietly stops being frictionless.

Here is what it asks, and what you want ready before the first enterprise call rather than three weeks into procurement.

• Where customer data physically lives and gets processed. Not where your company is registered. Which regions your storage and your inference actually run in. If the honest answer is wherever the model provider happens to route it, you do not have an answer yet.
• Who your subprocessors are, in a published list. This is the one that catches teams who optimized hardest on token cost. If you route requests to a cheap hosted open weight model, that host is a subprocessor and belongs on the list. A financial or healthcare buyer will read it, and a provider they have never heard of, in a jurisdiction they have opinions about, turns into a conversation you have to win rather than a line item. Cheap tokens that cost you the deal are not cheap. Make that choice deliberately, with the security review in mind, not purely on price per million tokens.
• What you retain, for how long, and whether it trains anything. Get the no-training and retention terms from every model provider you use in writing, and be able to hand them over on request. Three days of paperwork now, versus a stalled deal later.
• India's own regime, because it applies to you too. The Digital Personal Data Protection Act has moved from statute to schedule: the DPDP Rules were notified in November 2025, with the substantive obligations for data fiduciaries phasing in through 2026 and 2027. Cross-border transfer is permitted rather than banned, subject to restrictions the government may impose. The practical reading: you have real runway to build consent, notice, breach reporting and retention into the product, and it is far cheaper to build them now than to retrofit them into a company that already has customers on two continents.

None of this requires a compliance program in year one. It requires three documents you can write in a week: a data flow diagram, a subprocessor list, and a retention policy. Having them ready turns a four-week procurement stall into a two-day one, and that difference compounds across every deal you will ever run. It is the same principle that runs through the rest of the operating decisions an AI startup has to make: the cheap version done early beats the thorough version done late.

What to do this week

1. Write down which of your assumptions are local. List the three things about your product that only make sense in India: pricing, payment rails, language, distribution channel. For each, decide whether it is a deliberate wedge or an inherited habit. The inherited ones are what keep a global product domestic.
2. Send five cold emails to buyers outside India. Not to test outbound. To test whether your demo travels without you in the room. If nothing lands without a warm introduction, the problem is the demo, not the introduction.
3. Price your frontier usage honestly for one week. Log every task you sent to a model, then mark the ones that genuinely needed the best model available. Route the rest to a cheap or open weight model and measure what breaks. What is left is your real compute floor, and it is almost always smaller than your current bill.
4. Pick the task the models can barely do. One task in your domain that the current model gets right about half the time. That gap is the highest-value place to build, and it is where the bottleneck ideas live.
5. Do a project, by the strict definition. Two people, not assigned by anyone, someone outside the two of you using it. If the last six months contain nothing that meets all three conditions, that is the gap to close before anything else on this list.
6. Write the three security documents. Data flow diagram, subprocessor list including every model provider you route to, retention policy. One week of work that unblocks every enterprise conversation you have for the next two years.

If you want the full operating system for running a company this way, from model and cost decisions to security posture to go to market, that is what we teach in AI Operating System for Startups.

Sources

• India Can Create The Largest AI Companies, the closing panel of Y Combinator's first Startup School in India, the source this article distills.
• Panelists and YC partners on stage: Puneet Kumar, co-founder of Supr Daily (YC Winter 2017, acquired by Swiggy); Arnav Sahu, partner at Peak XV and formerly of Y Combinator; Ankit Gupta, general partner at Y Combinator and co-founder of Reverie Labs; and Jon Xu, general partner at Y Combinator and co-founder of FutureAdvisor.
• Emergent hits $100M ARR eight months after launch, TechCrunch, for the company's revenue claim. Background on its founders: Times of India.
• Varun Vummadi on Giga's Series A and its work with DoorDash.
• Vidit Aatrey, co-founder and CEO of Meesho, whose earlier talk the panel referenced.
• Relentlessly Resourceful by Paul Graham, the essay recommended from the stage.
• DPDP Rules, 2025 notified, Press Information Bureau, and the Digital Personal Data Protection Act, 2023, for the India data protection timeline.
• From the knowledge base: Emergent: How Six Months of Tinkering Led To A $100M ARR Company, Y Combinator, for the argument that building for India and building globally take the same effort.