Patrick Collison: "What If You Succeed?"
Patrick Collison, CEO of Stripe, discusses his journey from dropping out of MIT to founding Stripe, emphasizing the importance of grounding decisions in real customer problems and the value of deep knowledge even in an AI era. He shares data showing unprecedented growth in new business formation and argues that this is the best time ever to start a company, with AI creating more decentralization rather than concentration.
Summary
Patrick Collison opens by reflecting on his 20-year relationship with Harj Taggar and his early work on Chroma, a Lisp dialect. When asked whether a modern teenager should still write code dialects given AI capabilities, he argues that despite AI's power, retaining knowledge in what he calls 'cognitive L1 cache' remains valuable because internal knowledge retrieval is faster than querying external agents. He notes that companies still place enormous premiums on cognitive ability and that he personally avoids using AI-generated writing suggestions, finding LLM output deficient at reasoning about multidimensional aspects of reality.
Collison discusses his decision to drop out of college twice—once after freshman year to start a company with Taggar, and again to found Stripe. He explains that while he initially felt urgency to pursue opportunities before they disappeared, in hindsight this concern was unfounded. He advises college students not to fear missing out on a supposed 'last window' to start companies, comparing contemporary doomism to historical millenarian predictions like those surrounding aviation's invention.
Regarding Stripe's founding, Collison explains that while the idea seemed simultaneously obvious and ridiculous—two young founders starting a financial services company when 'fintech' didn't even exist as a sector. Banks were skeptical, but the venture succeeded because it was grounded in a concrete, viscerally-felt customer problem: the genuine difficulty of processing payments online.
On Stripe's lengthy two-year private beta before public launch (September 2011), Collison explains this wasn't recklessness but necessity. Financial services required extensive security, partner relationships, and infrastructure work before scaling. Critically, they had production users almost immediately—their first customer in January 2010, just two months after the first lines of code. This provided constant grounding in reality rather than hypothetical assumptions, allowing just-in-time development driven by actual customer requests.
When asked about whether AI-enabled cheap development should encourage more ambitious product launches, Collison suggests the traditional lean startup doctrine may need updating. He notes that the internet is far more competitive than 20 years ago, so aggressive de-correlation through ambitious starting points might now be necessary. Many recent successful companies (labs, Anduril) took this anti-lean-startup approach, suggesting that cheaper development capabilities make this more feasible today.
Collison addresses the intellectual rewards and losses in building Stripe. While financial services involves tedious tasks like payroll setup, he's fortunate because Stripe works with the world's most innovative companies—25% of new Delaware corporations start with Stripe's Atlas service. He argues that asking 'what if you succeed?' before raising money is crucial; he needed to genuinely want to work on payments for decades, which he does because every business represents an applied theory worth understanding.
On fears that big AI labs will crush startups, Collison argues the historical record doesn't support this concern. Google appeared omnipotent 20 years ago but didn't accomplish everything despite material capabilities to do so, because organizations struggle to prosecute hundreds of priorities simultaneously. While specific tasks may be obviated by AI, the broader trend Stripe data reveals is unprecedented: more businesses are starting than ever before, with 2x year-over-year growth—the largest relative jump they've seen. Crucially, median businesses are performing better year-over-year, and probability of reaching revenue milestones ($1M, $5M, $10M) is increasing across the board.
Collison attributes this growth to enterprises being 'spring-loaded' to adopt new approaches because they fear being left behind by archaic operations more than they fear startup risk. The risk of status quo now appears as dangerous as the risk of the new, creating an unprecedented window for startup adoption. Combined with Stripe data showing more AI adoption than ever, Collison argues the future appears to be heading toward decentralization with thousands of winners rather than concentration among a few companies.
About this episode
<p>In 2009, Patrick and John Collison went to Startup School in Berkeley, got sushi in Potrero Hill afterward, and decided on the walk home to start Stripe. The reasoning, as Patrick remembers it, was that “we might as well because it probably won't be that hard.”</p><p>It took two years to launch.</p><p>Seventeen years later, at Startup School 2026, he talks with YC's Harj Taggar about dropping out of MIT twice, why founders should ask what happens if they succeed, and what Stripe's own data says about the best time to start a company.</p>
Key Insights
- Collison argues that retaining knowledge in 'cognitive L1 cache' remains faster and more valuable than querying AI agents, even granting the models' full capabilities, because internal knowledge retrieval requires fewer round trips than external queries.
- Collison contends that the premise of a 'last window' to start companies before AI takes over is historically unfounded, analogizing contemporary urgency to past predictions about aviation that overestimated sociological transformation.
- Collison explains that Stripe's early success was grounded in a genuinely felt customer problem rather than hypothetical assumptions, which distinguished it from ideas that seem obviously good in theory but lack real demand.
- Collison describes Stripe's two-year private beta as a necessity of the financial services domain, not recklessness—it allowed just-in-time development driven by actual production users starting in January 2010, just two months after initial code.
- Collison suggests that in an AI era with cheaper development, the traditional lean startup doctrine of identifying niches through incremental iteration may need replacement by more aggressive de-correlation and ambitious initial product scopes.
- Collison argues that Google's failure to dominate despite material capabilities to do so suggests that organizational complexity and priority management limitations prevent companies from capturing all opportunities, making lab dominance fears overstated.
- Stripe data shows 2x year-over-year growth in new business formation (the largest relative jump ever recorded), with median businesses performing better and higher probabilities of reaching revenue milestones, indicating unprecedented conditions for startup success.
- Collison observes that enterprises now fear the risk of maintaining status quo operations more than the risk of adopting new startup solutions, fundamentally shifting the adoption calculus for new ventures in their favor.
Topics
Transcript
. Okay, Patrick, thanks so much for being here. Welcome to Startup School. Great to be here. Harj and I first met 20 years ago and we started a company together. Sorry, am I giving away the introduction? Yeah, I thought this was my interview, but sure, keep going. Well, we started a company together many years ago, and I learned a huge amount from Harsh. So it's really fun to do this. All right. Well, actually, I mean, speaking of that, so when I think when I first met you 20 something-ish years ago, at the time, your most impressive achievement, I would argue, was Chroma, your dialect of Lisp. At the time, your most impressive achievement, I would…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator Startup Podcast
Paul Graham On Startups, Ambition, and Great Founders
Paul Graham discusses YC's evolution over 21 years, arguing that startups funded today are more ambitious than in the past, and explores how founder ambition, shipping speed, and formidability drive success. He shares perspectives on AI's surprising development as a 'bullshitting' system rather than a perfect-to-complex progression, and explains how YC's batch model has remained fundamentally unchanged despite scaling.
The World’s Largest Electric Aircraft Just Flew
HART Aerospace has successfully flown the world's largest electric aircraft, a 100-foot wingspan hybrid-electric plane designed to reduce regional air travel costs. The company, founded by Anders Forslund, evolved from a 3D-printed model seven years ago to a 40-person team building a commercially viable airliner that swaps jet engines for electric motors.
Max Junestrand: You Need The Willingness To Learn Faster Than Anyone Else
Max Junestrand of Legora describes the company's rapid growth from zero to $100M ARR in 18 months by building an AI operating system for lawyers. He discusses the importance of learning the market deeply, maintaining competitive culture, hiring for growth potential rather than prestigious credentials, and betting on improving AI models without fine-tuning.
Susan Kare: Designing Icons & Graphics For the Original Mac
Susan Kare, the iconographer for the original Macintosh, discusses her journey designing system fonts, icons, and graphics under severe technical constraints (16x16 black and white pixels). She shares design principles learned from mentors like Paul Rand, her experiences across multiple tech companies, and the philosophy that simplicity, metaphor, and meaningful design create universal, memorable user interfaces.
Chelsea Finn: This is the State of the Art in Robotics
Chelsea Finn from Physical Intelligence discusses advancing robotics toward general-purpose models that achieve high reliability through reinforcement learning, memory systems, and diverse training data. She demonstrates how robots can perform complex real-world tasks autonomously and argues that physical AI has reached a ChatGPT-like era comparable to language models, with models now being deployed in real-world applications.