Garry Tan: AI Agents, the Five Kings and the Case for More Billionaires

Y Combinator CEO Garry Tan tells Walter Isaacson why personal AI agents are the next operating system and why five AI “kings” could end the startup era. In this episode of The Builders (taped June 19, 2026, released November 3, 2026), Tan covers how YC picks founders in a 10-minute interview, what he learned from Alex Karp at Palantir, why he thinks big tech is too powerful, how he spends six figures a year on AI tokens, and why he wants more billionaires and less market concentration at the same time.

TLDW

Tan argues that YC exists to break the old access game in startup funding, and that the biggest threat to that mission is now platform power: Google self-preferencing in search, Apple walling off iMessage, and AI labs that may one day gatekeep frontier models. His answer is “little tech”: open app markets, open-source models and a Linux-style personal AI that people control. He is candid that AI may cost jobs in the short term, describes running a personal agent that has read 20 years of his email and notes, and closes by rejecting both “big tech oligarchy” and “mass asset seizure” in favor of a middle American way.

Thoughts

The “five kings” line is the most consequential thing Tan says, and it reframes the rest of the interview. Friends inside AI labs, he says, talk about a “light cone of all future value” collapsing into five companies, meaning almost anything a business does could end up running through one of them. His response is not to break up the labs but to build a Linux for personal AI: an open agent layer that is cheap and easy enough that people pick it over the default. That is a bet drawn from the 1990s rather than the 1910s. Linux never won the desktop, but it won the server, kept the stack inspectable, and put a ceiling on what proprietary vendors could charge. If personal agents settle into a Mac-and-PC duopoly, as Tan predicts, an open option is the main thing that keeps the agent working for its user rather than for whoever hosts it. His side claim that there are “zero” American open-source AI companies, and that rules allowing distillation would help close the gap with Chinese labs, is a sweeping one, but it shows how far he is willing to go on openness.

Isaacson lands his best pushback on the social media idea. Personalized feeds already exist, he points out, and the algorithms still reward outrage because that is the business model. Tan’s reply is that the business model is “hosted by someone else,” and that is the crux of his whole worldview. A personal agent that reads Instagram as a plain API and filters it through your own goals only works if Instagram has to expose that API. That is why his regulatory ask is so narrow: something like the Open App Markets Act, and an end to Apple’s argument that only Apple is safe enough to let software touch your iMessages. Without interoperability rules, “personal AI” is a nicer client for a platform that can switch it off. The social feed idea and the antitrust fight are the same idea viewed from two ends.

The most revealing number in the conversation is Tan’s token bill: $100,000 to $200,000 a year to run his personal agent. He frames it honestly as “a crazy privilege,” and says CEOs who do the same “have wings.” That figure is the real distance between the demo and his prediction that everyone will have a Jiminy Cricket on their shoulder within two to three years, which only holds if costs fall by something like two orders of magnitude. It also exposes the stakes of his openness argument. An agent that has read all his email, his Dropbox and 20 years of therapy notes, and that transcribed and mined Isaacson’s ten best podcasts in 15 minutes before the interview, is the most intimate data object anyone has ever owned. Whether it runs on hardware and models the user controls or on one of the five kings decides whether it is a confidant or a wiretap.

On jobs, Tan is more candid than most AI optimists: in the short term, AI may not create them. His prescription is to pair a displaced domain expert, say a medical biller, with a 22-year-old who can ship 400 times more software than before, and have them start a company in the space the biller knows cold. Then he says the responsibility of CEOs “is not to cost cut,” because that path ends with capital accumulating and nobody sharing it. It is a moral appeal, not a mechanism, and it sits awkwardly beside his own explanation of Google earlier in the interview: a Delaware corporation, he said, is structurally obliged to maximize shareholder value, which is why it self-preferences. The same duty pushes every public company to use agents to cut headcount. If structure explains Google’s behavior, exhortation will not change the CEOs’.

Tan’s closing “middle way” (more billionaires, less concentration) sounds contradictory but is coherent if the billionaires keep coming from new companies, which is YC’s entire thesis. The tension he leaves unaddressed is OpenAI. He points out proudly that it began as YC Research, with Sam Altman in “my seat,” and that “we live on their timeline.” Minutes earlier he called the AI labs “the new big tech.” YC’s greatest success is now one of the candidates for the concentration YC lobbies against. That is not hypocrisy so much as a description of how the cycle works: Brandeis and the trustbusters, then the Microsoft consent decree that Tan credits with letting Google exist, then today. Antitrust in his telling is recurring maintenance on the startup machine, not a one-time fix, and the companies that need it next are the ones his own program produced.

Key Takeaways

Y Combinator now gets 80,000 to 90,000 applications a year for about 800 spots, an acceptance rate under 1%, and the final interview lasts only 10 minutes.
The pitch YC wants is simple: a real user with a real problem, how the technology solves it, and why these founders are the ones to build it.
Tan says 10% to 15% of self-made billionaires aged 50 and under are YC alumni, and YC tests itself by asking whether it would say yes to an unknown Elon Musk, Bill Gates or Steve Jobs in 10 minutes.
YC’s check rose from $18,000 when Tan went through to $500,000 today, so founders without rich parents can afford to take the leap.
Alex Karp put early Palantir engineers in the Mandarin Oriental instead of a Motel 6 because where you stay signals what your software is worth to a general or a Fortune 500 executive.
Tan credits the DOJ’s action against Microsoft with making room for Google, and argues Google now self-preferences in search the same way.
He backs rules like Europe’s Digital Markets Act that force platforms to allow data portability and reverse engineering, and an Open App Markets Act in the US.
YC runs its Washington policy work with one full-time hire, Luther Lowe, on a budget Tan puts at roughly 1/100th of a single big tech firm’s policy spend.
About half of startups that really make it have at least one immigrant co-founder, according to Tan.
YC runs an internal AI agent called Straylight with a “company brain” holding context on every founder, application and email.
Tan spends $100,000 to $200,000 a year on tokens for his personal agent and expects everyone to have something similar in two to three years.
His open-source project GStack packages his Claude Code setup so YC founders can get far more output, and he sees tools like it as part of retraining displaced workers.
He expects “the next browser war” over which AI agent people use, likely ending in two or three dominant players, and wants an open-source “Linux for personal AI” as the alternative.
Tan rejects both big tech concentration and the asset seizure he associates with Bernie Sanders and AOC, arguing for “more billionaires,” more choice and more jobs.

Chapters

0:09 How Y Combinator Picks Founders in 10 Minutes

Tan explains that Paul Graham found an hour-long interview unnecessary; five minutes would probably do, but felt rude, so YC settled on 10. The constraint works like Twitter’s old 140 characters. Isaacson asks whether YC would have funded Elon Musk fresh out of PayPal, and Tan says that question is the real test of whether YC is working. He thinks his own startup, Posterous, got in on the strength of a “banger demo.”

5:00 Meritocracy, Access and the $500,000 Check

Before YC, Tan says, startup money went to Stanford classmates and golf buddies. He concedes YC is not a true meritocracy but calls the attempt a virtue. Isaacson presses on whether only people with a safety net can play, and Tan points to the jump from $18,000 to $500,000 per company as a deliberate choice so founders do not need rich parents.

7:09 Alex Karp and Mandarin Oriental Software

As an early Palantir employee, Tan went on walks with Karp, whom he describes as a philosopher more than a technologist. Sent to Washington in ripped jeans to demo the finance product, he asked why the team was not at a Motel 6. Karp’s answer was that clients would ask where they were staying, and the hotel would tell them what kind of software Palantir was.

9:02 Antitrust, Self-Preferencing and Little Tech

Tan argues Google could only thrive because the government stopped Microsoft from using Windows to take over the browser, and that Google now shrinks the share of search results going to non-Google properties. His antidote to big tech is “little tech,” backed by open-market rules like the DMA. Isaacson counters that Anthropic and OpenAI rose despite incumbents, and Tan agrees the AI labs are effectively the new big tech.

12:37 Why YC Acts Like a Union in Washington

Tan describes YC as having “union-like characteristics” on policy, from California’s billionaire tax to antitrust in Washington and Europe. Looking into politics, he found astroturf groups claiming to speak for small software companies while pushing rules that helped big tech. YC’s answer is one DC hire and a YouTube channel with more than two million subscribers.

15:59 Immigration and the Childhood Scanner

Born in Canada to a mother whose family fled Burma and a father from Singapore, Tan says roughly half of breakout startups have an immigrant co-founder. Isaacson asks about his father’s unemployment and drinking. Tan describes growing up never knowing which father would come home, which built an emotional “scanner” he now uses with hundreds of founders a year.

17:57 AI Agents as the Operating System, Not a Feature

Tan says the biggest mistake is thinking AI is just chat. He describes YC’s Straylight agent, the working memory of a model compared with a human’s seven-plus-or-minus-three items, and Peter Steinberger’s OpenClaw as a more human-feeling alternative to cautious coding agents. His own agent has read his email, Dropbox and notes, and prepped him for this interview, at a cost of $100,000 to $200,000 a year in tokens.

22:06 Jobs, Retraining and the Ethics of What YC Funds

Tan hopes AI creates more jobs but admits it may not in the short term. He pitches GStack and similar tools as retraining, and asks people to mine their own histories for what they uniquely know. On ethics, he credits partner Gustaf Alströmer with turning away a lucrative startup built on distributing ADHD medication to children, which he says was later shut down after regulatory action.

27:13 Personal AI, the Social Feed and the Five Kings

Isaacson proposes social media that strengthens society, and Tan says personal agents could do it by treating Facebook and Instagram as APIs filtered through your own goals. He compares the moment to Apple’s 1984 ad and predicts a browser war over AI agents. Asked whether consolidation to two or three players would be terrible, he says it is already happening, citing lab insiders’ talk of “five kings.”

31:30 Open Markets, Open Models and the Middle Way

Smart regulation, Tan says, looks like the Open App Markets Act and rules that let American companies distill models the way Chinese labs do. Isaacson draws the line to Louis Brandeis, Teddy Roosevelt and the railroads. Tan rejects both populist asset seizure and big tech oligarchy, notes that OpenAI grew out of YC Research, and ends by calling for more billionaires, more choice and more jobs.

Notable Quotes

“You could probably actually do it in about 5 minutes, but 5 minutes feels rude. So that’s how YC ended up at 10 minutes.”
Garry Tan, on the length of the final YC interview

“Palantir is not Motel 6 software. Palantir is Mandarin Oriental software.”
Garry Tan, recalling Alex Karp’s lesson to early Palantir engineers

“Most tech people are not interested in politics, but politics sure is interested in us.”
Garry Tan, on why YC started lobbying in Washington

“I spend about maybe a hundred to $200,000 a year on tokens, and so I realized that’s a crazy privilege.”
Garry Tan, on the cost of running his personal AI agent

“The responsibility for CEOs and people who have a lot of privilege is not to cost cut. Like, that is the easy way.”
Garry Tan, on how companies should respond to AI

“I think next year will be one of the defining years where there’s the next browser war.”
Garry Tan, predicting a fight over which AI agent people use

“It doesn’t have to be big tech oligarchy, but it also doesn’t have to be mass asset seizure.”
Garry Tan, closing the interview on his middle way

Watch the full conversation here.