What Actually Makes A Startup Durable
In a Nutshell
Founders must pick harder, more durable problems than pure software, since AI has made building so cheap that replication is trivial. Durability comes from real barriers like regulatory licenses, deep industry sales, or hardware physics. The winning process is relentlessly empirical: launch early, talk to customers constantly, iterate on real feedback, and keep high-agency founders obsessively focused on the current bottleneck rather than endless research.
These notes were generated by AI and may contain inaccuracies.
The cost of intelligence is coming down rapidly at roughly 10x cost reduction per year for the same intelligence. YC provides founders with a million dollars of free tokens plus offers for many millions more. Even if AI appears more expensive than an engineer today, it will solve itself within six to twelve months. An engineer using the absolute best models today is approximately 1,000x better than an engineer without AI. There is no chance that a human is more cost-effective than a model for any programming task when using the best intelligence available.
Founders should join existing communities rather than creating them from scratch. The optimal locations after San Francisco are London and Paris due to their critical mass of startups. Being the only founder in a city makes it difficult because there are no peers to push you forward, and people who don't understand the startup life will hold you back. The London startup ecosystem traces back to a small kernel of fewer than five companies including Go Cardless, which directly connects to TransferWise and Monzo through a family tree spanning dozens of companies.
The direct advice is to find a great startup that embodies startup culture and join it. In the early days of any great startup, it feels like a cult where founders believe something no one else believes, and saying it outside the room would invite laughter. However, the collective set of beliefs, behaviors, and ways of working make the team extremely effective. There will always be one global hub for startups that attracts the most ambitious people and companies with the biggest visions.
Writing correlates directly with thinking, so founders should write most things themselves rather than delegating too much to AI. YC has built tools at remarkable speed to empower partners and share those capabilities with founders. The things AI cannot replace include founder well-being, community, and knowledge sharing among 200 companies obsessing over maximizing AI usage. YC's unique value comes from surrounding founders with people actively experimenting with the latest AI techniques.
Pure software is becoming too easy to replicate. The key question is identifying the hard bit that provides durability. Ten years ago, writing software was itself the hard bit, but this is no longer true. Durability requires something genuinely difficult, whether that's brutally hard B2B sales into specific industries, regulatory barriers like obtaining a banking license, or the physics of hardware like spaceship launches. Most hardware in itself is hard because the atoms are the hard bit.
Founders should avoid the easier end of the spectrum and pick harder, more ambitious problems. Examples of products to avoid today include Calendly or pure software plays, as they lack durability. The spectrum ranges from very easy to very hard, and founders should deliberately choose the harder end.
YC is experimenting with virtual partners for office hours, though early tests like Nikiko doing an office hour with himself felt weird. The irreplaceable elements include being therapists, community builders, and the power of witnessing. Having someone in the room who acknowledges that what you're going through is normal is incredibly powerful. Being a founder is an incredibly lonely job, and robots cannot provide the same witnessing experience.
Partners go deep into understanding each specific business and market rather than giving generic advice from a database of 7,000 companies. All YC partners are themselves successful YC alumni founders who have lived through the founder experience, which creates a fundamental difference from investors who have never been founders.
The biggest regret YC founders have after their batch is not launching soon enough. Founders always tend toward more research and building before selling, but this tendency must be countered. The optimal approach is confronting the market as quickly as possible through user conversations. Optimize constantly for learning, which usually means talking to the market. If people aren't willing to use what you're building, return to building and researching.
The core principle remains building something people want through the cycle of build, talk to customers, build, talk to customers. The pace has changed dramatically, with companies achieving what previously took years now happening in months due to faster iteration cycles. Some projects like deep tech or space have longer timelines, but the fundamental requirement is ensuring what you're building is useful to someone.
Founders are often more comfortable building than talking to customers because market validation carries the terrifying possibility that all current work is wrong. This loss aversion discourages necessary customer conversations. Founders comfortable with research should recognize their bias toward building and deliberately increase customer conversations.
The restrictions represent a wake-up call that the world was more dependent on US government decisions than realized. A decision in the Oval Office can affect access to the best models globally. This has created demand for AI competitors and sovereignty over AI and inference that didn't exist before.
Theoretically, one-person billion-dollar companies are possible, but adding people provides marginal benefits that outweigh costs at the early stages. The ratio changes as companies grow because coordination problems increase with the square of nodes according to Metcalfe's law. AI will compress companies to stay below Dunbar's number around 150 people where relationships can be maintained successfully.
Solo founders statistically perform worse than teams with co-founders. A solo founder experiment during one YC batch building self-improving autonomous agents led to a crisis of confidence six weeks in. Co-founders serve to pick you up when down or pull you back to earth when hyperactive, modulating emotional extremes.
The complementarity of skills matters less now because AI enables average engineers to become great engineers and non-sellers to become good at selling. The essential co-founder quality is someone you enjoy talking to when sad who can improve your morale. Technical co-founders should find other deeply technical people with high determination, intelligence, and integrity rather than seeking business co-founders. The business aspects are learnable, while finding someone you'd fear competing against is irreplaceable.
YC funds solo founders but sets a higher bar because having co-founders makes the founder experience easier. When selecting co-founders, identify the smartest, most determined, hardest-working person you've worked with, then add an integrity layer of trust and aligned values.
Founders don't need to raise as much capital today as five years ago to achieve the same outcomes. However, AI enables founders to attack harder problems, and more ambitious problems often require more capital. Examples include building small nuclear reactors requiring $800 million raises or obtaining banking licenses. Venture capital will fund solutions to harder problems like curing all disease and producing abundant energy.
YC funds people above all else, and this bet on founders is even more true today than before. Approximately 30% of investment decisions show disagreement between partners, which is considered healthy. A third of companies in recent batches pivoted from their original ideas. The Brex founders applied at 19 with an idea for cardboard virtual reality sunglasses after previously selling a payment network for $50 million as teenagers, demonstrating that strong teams can pursue any idea successfully.
Y Combinator evaluates B2B AI companies by distinguishing between genuine value creation and thin AI wrappers. The term "AI wrapper" is considered somewhat meaningless because everything is powered by AI at this point. The key question is whether companies are adding substantial value on top of core models. If the layer is very thin, YC examines the founding team and assesses whether there's enough depth in the problem space to go really deep. Some of the biggest companies today were once dismissed as AI wrappers, just as earlier companies were dismissed as SQL wrappers or S3 wrappers.
With the cost of implementation having gone down to essentially zero, founders add value in two ways: moving state-of-the-art in one area and building specific technical improvements that provide margin, or selling the product better by leveraging increased model capabilities. YC batches now show a changed ratio between super technical founders working on niche technical issues and founders marketing existing solutions better.
YC has founders in both camps: deep technical experts including an MIT PhD in nuclear physics building nuclear reactors, two former satellite company employees building re-entry vehicles for space manufacturing, and radio engineers who worked on SpaceX launches building radio frequency test equipment. On the other side, there are self-taught "vibe coders." This has prompted YC to re-evaluate how deeply technical founders need to be. The organization was previously biased toward teams with at least one deep software engineer but is now testing whether this remains necessary. The conclusion is that founders still need to be very smart, determined, and systems thinkers, but not all companies necessarily need someone deeply technical. YC funds the full spectrum from MIT nuclear PhDs to bedroom vibe coders.
A European student unable to move to San Francisco due to visa and budget constraints asked about competing with founders who have API budget money. The student received $25,000 in model credits from various sponsors, available to anyone with a valid student email. The advice given was simply to "go build something cool."
Early decision applicants are officially treated the same as standard applicants, but YC partners evaluate why someone is applying for early decision rather than waiting for the next cycle. The question asked is whether they would fund the applicant today. Legitimate reasons might include financial necessity or inability to afford waiting without a job. However, if the reason is reluctance to drop out of school or complete another year and a half, this raises questions about commitment level. The personal factor YC looks for is genuine commitment and drive.
The most successful founders launch early and often. The biggest issue YC observes is founders being too theoretical, too in their heads, and too afraid to face rejection. They convince themselves that more weeks of research or building is the smart approach, when it's not. A founder of Cursor showed a timeline of Hacker News launches with zero upvotes, then two upvotes, then two upvotes, eventually leading to acquisition by xAI for $60 billion. The key lesson is to launch early, make contact with the real world, get data, and update plans accordingly.
Startups are an extremely empirical exercise rather than a theoretical one. Smart founders selected from thousands of applicants often come from academic environments where sitting in a library thinking hard produces top grades, but this translates poorly to startup success. Humans are bad at predicting the future. The key activity is forming hypotheses quickly, running tests with the real world, and updating approaches based on data received. YC's catchphrase is "launch early and talk to users."
Top founders set ambitious goals and reassess them every two weeks, holding themselves accountable. They focus on overwhelming the biggest actual problem facing their startup. Every two weeks during YC meetings, these founders present a new problem because they've overwhelmed the previous bottleneck. They throw everything at the current bottleneck, make progress, then return with a new challenge. This creates a tight timeline of constantly reassessing the most important work and focusing intensely on it rather than being distracted by smaller fires.
The biggest mistake when using AI is looking at results and moving on without iteration. The self-learning cycle is critical: building feedback loops where AI memorizes feedback and updates based on corrections. Once founders invest in building these loops, results become game-changing within weeks.
To build an AI-native company, start with a single narrow loop rather than trying to implement everything at once. First, make all information legible and queryable to agents by organizing emails, Google Drive files, and customer support tickets. Use tools like Obsidian or GBrain for this. Then pick one narrow process to automate, such as sending follow-up emails after sales calls with key discussion points, or having an agent transcribe sales calls and build prototypes of product ideas mentioned so demos can be shown by call end. Implement the narrowest possible version where value is visible, then improve iteratively.
When restarting a startup from scratch as an AI-native company, the last thing to automate would be talking to customers. Direct customer conversations keep founders focused on what matters and provide complete context on what to build. Even with self-serve PLG models, some direct customer exposure is necessary. Co-founder conversations should also never be delegated, as co-founder issues are the main reason companies fail during YC batches.
While writing software is cheaper, pivoting still means throwing away substantial built work, including hard components, so it's not necessarily cheaper overall. Pivoting should be based on a hypothesis framework: testing whether the fundamental business direction is correct versus specific product choices. Real evidence that the core hypothesis is wrong justifies pivoting after exhausting all options. Most pivot decisions stem from running out of enthusiasm rather than genuine exhaustion of possibilities. When founders get rejected repeatedly and feel sad, they often consider pivoting, but the grass is rarely greener elsewhere, especially when moving from deep expertise in one area to shallow knowledge in another.
The better version of pivoting occurs when working on something difficult exposes founders to an even better idea through customer feedback. An example is a company collecting payments for sports teams that discovered users wanted the same solution for their businesses instead. Customer-pulled pivots tend to be easier decisions. YC distinguishes between founders needing encouragement versus those who have genuinely exhausted themselves or all options.
High agency is fundamentally a belief that actions will produce output and create impact. It can be developed by starting with projects that are difficult enough to push limits but tractable enough to see results, then progressively making them harder. Side projects work well for this development. Finding a co-founder who is smart and determined helps sustain momentum. The progression builds from small projects to eventually building significant companies. No one realistically starts by building rocket ships to Mars; they build smaller things first like Zip2, then PayPal, then Tesla.
With building becoming easier, more companies are being built at early stages, making differentiation critical. When every billboard in San Francisco advertises AI companies, the obvious answers no longer work. The founder of Brex previously bought all San Francisco billboards when no one else was doing so, but this strategy no longer differentiates. Distribution matters more than ever, and the meta-game for reaching customers evolves constantly. At the early stage, founders are playing the game of reaching product-market fit through white-glove service, in-person meetings, and building exactly what individual customers need rather than competing with scaled distribution tactics used by companies that already have product-market fit.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
