Stripe Data Shows AI Is Improving Early Startup Outcomes

Patrick Collison thinks founders spend too much time modeling failure and too little time modeling the result they say they want. Before raising substantial money, he said, they should ask a converse question: what if you succeed?

That question becomes concrete quickly. A company with funding, customers, employees, and momentum may demand a decade or several decades of a founder’s working life. The relevant test is not only whether the business can survive, but whether the founder will want to keep working on it at that scale. Collison pointed to Larry Ellison’s decades at Oracle as an extreme illustration of how long the commitment can become.

For him, Stripe has made that prospect unusually rewarding despite the inevitable administrative and operational work. Nobody starts a company in order to set up payroll, he said, and financial infrastructure contains its share of “arcane and extensive” work. But Stripe’s position gives it a window onto businesses trying to change how the world works. Every new company embodies, in his formulation, an applied theory of some market or aspect of reality—a contrarian thesis about a counterfactual future.

Stripe gets to see those theories early and then work with the companies as they develop. Collison said that 25% of Delaware incorporations are started through Stripe Atlas, and that the company can remain alongside customers from incorporation through their growth into companies such as Shopify and OpenAI. He said he has never encountered a Stripe customer and found the business uninteresting.

Stripe itself was an example of what Paul Graham called “schlep blindness”: the willingness to take on work that looks unglamorous, bureaucratic, or forbidding. Founders need not choose an intellectually tidy problem. They do need to consider whether the totality of the work will remain compelling after the startup has ceased to be an experiment and become an institution.

Harj Taggar asked whether a technically prolific teenager should still write something like a Lisp dialect from scratch when an AI model could plausibly produce one from a prompt. Collison’s immediate answer was uncertainty. There are forms of programming work—hand-writing assembly, optimizing instructions, managing memory layouts—that were once satisfying and important but have largely been delegated to compilers without much collective regret. It is possible, he said, that source code itself will eventually give way to a higher-level practice of giving instructions to models. Emotionally, though, he misses the older mode of making software.

His argument for continued learning rests on an analogy to computer caches. A model or agent can retrieve information, calculate, or generate an answer; but knowledge held in “cognitive L1 cache” is available immediately. The difference matters because a person who knows the relevant constraints and concepts can make many internal iterations without stopping to formulate prompts, inspect outputs, and ask follow-up questions. Even granting powerful model capabilities, Collison expects “neuronal lookups” to remain faster for a long time.

He also pointed to organizational behavior. Companies such as Stripe and AI labs continue to place a large premium on cognitive ability. In his view, it would be premature to renounce the effort to build that ability before there is evidence that its benefits have been exhausted.

Writing is one domain where he still deliberately works without model-generated text. Collison said he dislikes LLM writing both philosophically and substantively, despite models’ demonstrated ability to solve difficult technical or mathematical problems. He has not encountered an LLM-written essay he finds especially compelling. Strong writing may be difficult to reinforce, he suggested, because the relevant utility function is hard to define.

Interpersonal communication, writing, and the ability to reason coherently through the “multidimensional space of reality” remain fundamental, in his view. He sees models as still deficient in some hard-to-specify way on that task. As a matter of personal practice, he said he has never sent one of the prewritten replies now offered by products such as Gmail and WhatsApp.

Patrick Collison has dropped out of college twice to start companies. He left MIT after his freshman semester to work on Auctomatic with Taggar, returned after a few years for another year at MIT, and then left again to start Stripe. That history led him to a relatively unromantic conclusion: leaving is not necessarily irreversible.

As a student, Collison expected to become an academic. He was drawn to physics, had read the Feynman books, and had not grown up in Ireland with startups as an obvious path. At the time, startups were less familiar even on campus, and people considered his decision to leave unusual.

His advice is conditional. Students who enjoy college do not need to leave; he sees little harm in finishing. His own urgency, in retrospect, was somewhat unnecessary. But students who are not captivated by college or its subjects should not assume that departure will permanently damage their reputation. Parents often treat dropping out as an unusually risky act with lifelong consequences, he said; as far as he can tell, nobody has cared.

The urgency that drove him came partly from a habit of speedrunning. Students who have pushed quickly through high school can arrive at college wanting to accelerate there as well. It also came from a belief that startup opportunities in Silicon Valley might be fleeting—that if he did not act within a few years, the window would close.

He now considers that intuition wrong. Silicon Valley has reliably produced an excess of opportunities over many decades. Against the current anxiety that young people must start companies immediately or be consigned to a “permanent underclass,” Collison urged skepticism about narratives of imminent, permanent social transformation. He compared current AI-era thinking to the excitement around aviation, when some observers expected flight to rewrite society in a far more total way than it did. Aviation mattered enormously, but it did not produce the sociological rupture its most excitable advocates imagined.

He would “take the under,” he said, on the claim that the past few years are the final period in which it is possible to start a company.

Patrick Collison described Stripe’s initial problem as unusually direct: moving money on the internet was painful. He and his brother John encountered it while working on Auctomatic. Existing payment systems were unpopular, antiquated, paper-heavy, and often required trips to a bank. The paperwork, he joked, seemed to be in Latin.

That concreteness mattered. YC had taught the founders to focus on problems customers feel viscerally and will pay to solve, rather than inventing a hypothetical demand. Payments fit the first description. Yet Stripe also looked implausible because the Collisons were young founders entering financial services without relevant experience. “Fintech” was not even an established sector label at the time, he said. Banks, partners, and others they pitched did not literally throw them out, but their reception often conveyed that two young people starting a payments company did not seem credible.

The idea was therefore both obviously good and obviously hard. Customers wanted it; institutions did not take its founders seriously. Collison believes the reality of the customer problem ultimately overcame the credibility gap.

The decision to start Stripe itself was less grand than the eventual company. In 2009, after attending Startup School in Berkeley, Patrick and John got sushi in Potrero Hill and discussed the payments space as they walked home. Collison remembers precisely where they were when they decided to do it. Their reasoning: they might as well, because it probably would not be that hard.

It became what he called an “ultimate yak shave.” The brothers began working on it the following week, initially alongside college. They went full-time in the summer of 2010 and launched publicly in September 2011—nearly two years after the repository’s first lines of code.

That long public gestation ran against the familiar startup injunction to launch early. Collison does not present it as a general model. In many sectors, waiting that long would be wrong. Stripe needed security, banking partners, money movement, infrastructure, reliability, and the conditions for a genuinely self-serve product. A public launch before those preconditions were in place did not feel responsible.

But Stripe was not built in isolation. Its first production user arrived in January 2010, roughly two months after the first code. Ross Boucher of 280 North initially could only charge a card; his practical requests for a way to view charges, issue refunds, and receive payouts supplied the sequence of features Stripe built next. The company remained in private beta while adding customers every month until the public launch.

For Collison, the important distinction was between delaying publicity and delaying contact with reality. Stripe had a continuous stream of production users, requests, and feedback; its development was “just in time,” driven by use rather than the founders’ imagined requirements. In Stripe’s partner- and reliability-dependent case, that grounding made a later public launch less like building in isolation.

Patrick Collison sees a possible shift in the old lean-startup doctrine of beginning with a narrow wedge, validating it cheaply, and expanding only after learning what customers want. He did not reject the value of customer grounding. Stripe’s own early development depended on it. But he suggested that AI changes the competitive and operational conditions behind the doctrine.

The familiar strategy of identifying a small, overlooked crevice may become more competitive as more people can quickly build software and pursue obvious opportunities. At the same time, AI can make it easier to start organizations with wider capabilities and more potential paths than were practical when capital was scarcer and software production slower.

That could reward founders who begin from more divergent positions—places other companies are not trying to occupy—rather than only finding the most narrowly defined market opening. Collison described this as a need to “more aggressively decorrelate” in the AI era.

He pointed to frontier AI labs and Anduril as successful companies that do not fit the conventional narrow-wedge template. Twenty years ago, he said, lean startup was close to the only viable route for many founders because of the capital available and the difficulty of building broad capabilities up front. That constraint is weaker now. The operating lesson is not to substitute ambition for customer contact; it is that a company can remain grounded in use while starting with a more aggressive initial scope.

Patrick Collison separates two concerns founders often collapse into one. The first is whether frontier labs themselves will expand into every attractive startup market. The second is whether models and agents will eliminate particular tasks or verticals regardless of what the labs choose to do.

On the first concern, he urged caution about treating large organizations as omnipotent. He recalled the earlier startup question—what if Google does this?—and argued that even organizations with talent, capital, and technical resources cannot pursue every priority aggressively at once. Managing many initiatives creates interference, competing demands, and organizational complexity. Google has succeeded in important areas, he said, but it did not do all the things it might have had the material capacity to do. The historical fear that a dominant incumbent will do everything has been overstated in his view.

The second concern is more real and more contingent. Capable models and agents can obviate some kinds of work even without a lab deliberately entering a market, Collison said. That has already happened in certain domains, and the extent will depend on how model capabilities develop.

Patrick Collison said Stripe’s data points toward more entrepreneurial activity and better early outcomes rather than simple displacement. At the time of his remarks, the number of new businesses starting on Stripe was nearly double its level a year earlier—the largest relative year-over-year increase Stripe had seen.

Collison acknowledged an obvious alternative explanation: perhaps AI has simply made it cheap to generate more lightweight, “vibe-coded” products. But he said Stripe’s observed results were not merely an increase in volume. The median business was doing better than a year earlier. The odds of reaching revenue thresholds including $1 million, $5 million, and $10 million appeared to be improving. Companies incorporated through Atlas were also reaching revenue more quickly.

Harj Taggar said YC was seeing a related acceleration in its batches, particularly because enterprises were increasingly willing to buy from startups early. Collison agreed that buyer behavior is central. In more ordinary conditions, an unproven company can struggle to persuade a CIO or CTO to adopt a product that may not survive two years. But businesses now fear the risks of retaining archaic ways of operating. The status quo itself looks dangerous, he said, making buyers more willing to test new vendors and adopt their products at meaningful scale from the outset.

Collison extended that observation to the economy’s structure. Leading AI companies have done extremely well and may continue to do so, he said, but Stripe’s view of new companies adopting AI capabilities and existing businesses retooling makes him less concerned that AI will produce a wholly centralized economy. His expectation, stated with qualification, is of many thousands of winners and a more decentralized world with more broad-based prosperity.