From $6 million to $30 million ARR in three months: Garry Tan on YC’s AI boom

When Garry Tan went through Y Combinator in 2008, the batch was 25 companies. Today YC has 14 group partners, each funding 15 to 25 startups, which Tan likes to describe as 14 of those 2008-era batches running at once. The latest one is unlike anything in the program’s nearly twenty years: about 70% of companies are AI-related, and combined ARR went from $6 million at the start of the batch to more than $30 million three months later. Most of those teams are still two or three people shipping a first version of the product.

Tan is president and CEO of YC. Before that he ran the early-stage fund Initialized, founded Posterous (later acquired by Twitter), and worked at Palantir. Sarah Guo and Elad Gil, both investors, host this as a collab episode: part one on No Priors, part two on Tan’s own YouTube channel. He needs little introduction in the Valley, but he still tells the origin story as a subculture, not an institution.

In 2008, he says, YC felt like a punk club in 1990s Seattle, with grunge just getting started. People now treat it as the place you go to start a billion-dollar company. Back then the real door was Paul Graham’s essays. Type “start a startup” into Google and you found PG writing from the Yahoo Stores years. Tan calls himself a case of “perfect mimesis”: he saw someone sit at a computer, write software, and touch a billion people, and wanted the same thing. Walk the YC hall of fame — Stripe’s Patrick Collison, Harj Taggar, Tan himself — and the entry point is those essays.

Gil adds a detail from 2007, when he started his first company. In investor circles the line on YC was adverse selection: who would take that deal? That was exactly when Airbnb and Dropbox got funded. His conclusion: when people start naysaying something, that is the moment you know it is working. Of the 25 companies in Tan’s year, maybe three or four “made it.” Mike Montano and Chris Golda built a comment-analytics startup that sold to Twitter; Montano later spent years as Twitter’s VP of engineering, and Golda still does a lot of YC investing. Scott Phoenix co-founded Vicarious, one of the most watched AI startups of its era, later sold to Google. You meet over pizza and beers, Tan says, and years later you are still in the same game. The hosts agree that Valley longevity is under-discussed: act non-zero-sum early, because those relationships unfold over ten or fifteen years.

Tan maps this AI wave onto two older ones. Social networks had SixDegrees, Friendster, and MySpace, and each failure gave someone a chance to call the whole trend bunk. MySpace died from stuffing the site with ads and from non-technical founders. Then Facebook arrived, dismissed as a college toy; meeting Facebook people in those years, Tan recalls, felt like meeting cult members. Evan Williams built Blogger, got acquired by Google, was mostly forgotten, and never lost the bug for the social space — which eventually became Twitter.

Mobile rhymed the same way. In 2009 and 2010 Gil watched half a dozen companies go from zero to a few hundred thousand users in a week or two on mobile photo uploads, then die. Only Instagram lasted, and its growth was more linear. YC now funds 200 to 250 companies a batch, about 70% of them AI-centric, which Tan treats as that Cambrian explosion running again. The selection thesis underneath has not moved. Extremely technical people — the Zucks, Larrys, and Sergeys — should get extra support. The hard part of chess is not knowing how the pieces move. It is being smart. Find the smart people; the community will teach them the game.

The real break, he says, was GPT-4. Jake Heller, CEO of Casetext, was among the first to get access. Only two co-founders could use it; they shut themselves in a room. They had already tried GPT-3 and 3.5, but those hallucinated too much to bill clients. GPT-4 was the first version that almost never hallucinated, and they decided on the spot that they could charge money. Heller is a lawyer. He broke a lawyer’s day into time-and-motion steps: reading a brief, writing a chronology, going paragraph by paragraph, converting that into a score — highlighting, he says, is scoring — then turning the whole thing into prompts, workflow, tests, and scores. Guo and Gil both backed Harvey, a close cousin of the same pattern. Half the current YC portfolio, Tan jokes, is in the weeds of that same mapping.

He does not buy the “ChatGPT wrapper” sneer that attaches to maybe two-thirds of those AI companies. Calling them wrappers, he says, is like calling all of SaaS and cloud “MySQL wrappers.” A model is a technology you implement, a concentrated form of intelligence. Today it is still an eager intern, maybe 85 IQ, with 100 as the human average; GPT-4o and Claude 3 might just have crossed the average line. You still have to constrain the work. That is why the revenue landed so fast: AI walks in and does the job a person was banging their head against, or the job already paid out to an outsourcing shop. Gil adds the demand side. Enterprises have a CEO mandate to “do AI,” so the CISO and the support team are hunting for a project. The door is being pushed open from above. Of the AI companies, about two-thirds implement frontier models on a concrete job; about one-third build infra — tooling, testing, prompt collaboration, the LangChain and Langfuse layer. Closed-source companies raise large rounds while scrappy open-source “barbarians” commoditize the same category, in parallel.

What YC selects for has not changed: highly technical people who are very clear communicators. Tan can recite the reject list from memory. They have not talked to customers. There is no insight. The product does everything. It is general-purpose. There is no thin edge of the wedge, and no set of users who will absolutely love it. Founders who actually work, he says, work from first principles, which for a startup means a specific set of people — not all lawyers, litigators, and maybe only corporate litigators in one segment.

He ranks information by how many hands it has passed through. Talking to customers is first-hand. Talking to other founders is second-hand. Talking to investors is third-hand. Talking to an investor’s LP is fourth-hand. Building a company off the Wall Street Journal, TechCrunch, or Twitter is a fourth- or fifth-hand retelling of something that happened six or nine months ago. Gil’s picture of early-stage investing is mice running around the feet of an elephant. How do you know which mouse will not get stomped? Tan’s answer is almost too simple: ask them, and if the story is plausible, why not? The power of early-stage is standing next to people in first-hand contact with what is happening.

Scale AI is his favorite “the elephant missed” story. You would have assumed Google would do it — Google already paid huge internal labeling teams and had both the ops and the need. Amazon already had Mechanical Turk. Nobody built the company. That, Tan thinks, is the good news for startups. Megatech moats are so unassailable, and the profits so large, that the giants are somewhat incapable of stomping anymore. “It’s a weak elephant. It didn’t do its push-ups every week.” Necessity is the mother of invention.

The other expensive lesson is fundraising. In 2012 and 2013, writing his first checks, Tan heard “don’t raise too much money” and filed it as a negotiation tactic. Years later he watched some of the best technical teams raise too much, then freeze when it was time to let go of 50 to 80% of staff — people they had hired, managed, and fought beside for years. Too much money equals potentially not doing the right things, and the worst version is not even being able to cut back to the point where the company can get on track. Gil adds that companies are path-dependent. A set of firms from the last five years still sit on huge war chests and painful valuations. The only question left is whether they can make the hard decision.

YC will fund foundation-model work, but it still wants a demo you can ship in three or four months on half a million dollars. Cruise is the template. Kyle Vogt was staring at a research project that needed $100 million. He worked backwards to a $10,000 automated-driving attachment on an Audi A4. In 2013 you could drive it up and down Highway 101 — no city streets, highway only — a demo that, Tan says, resembles Tesla Full Self-Driving today. That was enough to attract talent and capital and become a self-fulfilling prophecy. Current YC companies include Diffuse Bio on the biology side and a team on robotics foundation models.

The machine that pushes near-term milestones is weekly accountability. What did you do? Did you follow up? Are you talking to customers? Tan has joked about a rule he has not shipped yet: miss your goal that week, do push-ups on the spot. He is only half kidding. “Our realities are basically socially constructed.” Learn to run fast for a long time, and compounding takes care of the rest. Half of that culture comes from Paul Buchheit, the creator of Gmail, who invented group office hours. Tan compares it to Alcoholics Anonymous for founders: mild coopetition. You want your friends to succeed, but if you show up without having come correct, you will feel bad that week. Paul Graham’s gift was the opposite end of the telescope — helping someone see that a thin edge of the wedge, done well, could be a $10 billion or $100 billion company.

The founder picture is getting younger again. Tan was 27 when he did YC, which is roughly the average age now; this summer’s batch may land closer to 25 or 26. For a stretch YC skewed older, because the fashionable work was vertical marketplaces. If you had never been in a shipyard in Africa, you could not even participate. AI reversed that. The world is shifting week to week, a strong software engineer is almost the whole ticket, and far more 19-, 20-, and 21-year-olds are dropping out so they do not miss the boom. Sitting in the room, Tan thanks Sam Altman and Greg Brockman for the megabet on transformers and scaling. Without that, he is not sure where startups would be.

The check size moved the same way. In 2011 YC jumped from about $15,000–$20,000 to $170,000; the next year produced Instacart and Coinbase, the year after DoorDash. In 2022 the deal went from $120,000 to $500,000, and another wave of young technical teams showed up. The median YC company in 2014–2015 raised under $1 million after the batch; now it is closer to $1.5 to $2 million. Brand compounds. Founder quality compounds. Everyone, Tan says, underestimates that.

Tan wants builders in the room, not on Zoom. Going from zero to one, he says, you need the smartest people around you, and San Francisco dinners are where toolmakers, foundation-model builders, and researchers break bread with people deploying the same models into every kind of knowledge work. Those are, in his view, some of the highest-value conversations on the planet right now. AGI might arrive soon, or it might take longer, in which case the work is squeezing RAG, embeddings, workflow, and tests. The gap between a good AI app and an outrageously stellar one is in the details, and the people writing the concepts and the open source live within about five miles of where they are sitting.

Asked what he wants YC to be in five years, he reaches for ideology as much as tactics. There is no substitute for human intent, ingenuity, and belief. YC should be the defining magnet for people who are ambitious, optimistic about technology, and smart — “the people who we really desperately need to save from working at Goldman Sachs.” Save them from investment banking, consulting, middle management, and big tech. “We’re not numbers, we’re free people.” If you have a problem that capitalism and technology can solve, fill out the 12-question form. People who came before you, “pure of heart,” will try to help you manifest it, and they will put you in a room with hundreds or thousands of people who believe the same thing. “If not us, then who? It has to be us.”

The last stretch of the episode leaves YC for San Francisco civic life. Tan is a product of public schools. If he could not have studied algebra in middle school or calculus by senior year, he would not have gotten into Stanford, would not have studied computer science, and would never have built anything. The city brain-drains the most ambitious people on earth, then tells them their kids cannot learn algebra. That, he says, does not make sense, and the reason it persists is that people are afraid to say basic things. The fight is on in the hallways of Stanford’s School of Education, where ideologically motivated researchers run the apparatus of “science,” while mathematicians and computer scientists at Berkeley and Stanford ask why children that age should not be taught math. Removing math hits hardest the families who most need public school. Gil, also a public-school product, calls it an ideological removal of ladders.

The actionable layer is unglamorous. Know the issues, talk about them at dinner, vote, and use the voter guide from GrowSF, the civic group Tan helped start with friends, which walks through initiatives and budget items. He does not think the work is that complicated. The hard part is saying the thing people got too scared to say. Guo closes on a less partisan note: reigniting grassroots civic engagement, and making it possible to talk about these issues with nuance, is a good thing even before they start part two on Tan’s channel.