By Cornelius George · · 16 min read
I Hate Start-Up Culture
The 90% failure rate isn't a law of nature. It's a curriculum.
Congratulations. You may now stop learning.
Don't graduate. Traction theater, evening showing. I hate start-up culture.
Let me be clear about what I don't hate. The people, for the most part, are amazing. Yes, there are some absolutely disgusting trash humans in the community, every community has them, but the majority are brilliant people trying to solve incredible problems with everything they have. I don't hate the founders. I don't hate the ambition. I love the ambition.
So why do I hate it?
Because I have a kind of mental tick where if something doesn't make sense, I can't let it go. EVER. I am absolutely terrible at accepting "that's just the way it is" as an answer. And start-up culture, for a community that worships innovation, that cries about disruption as the highest form of awesomeness, is the most sheepish culture I have ever been part of. Everyone repeats the same axioms. Nobody asks where they came from.
The one that makes my skin crawl the most: "9 out of 10 start-ups fail."
Say it out loud. Nine out of ten. As if it's gravity. As if it's a law of nature you accept as the cost of doing business, the way sailors accepted scurvy before somebody thought to ask why.
I never accepted it. And when I finally went looking for where the number comes from, it got worse.
The axiom nobody sourced
Try to trace it. The most-linked source is a that opens "Nine out of ten startups will fail." Its citation is a that opens with the same sentence. Fortune's citation for it is nothing. Nothing at all, it's stated as received wisdom. The other big vector is , which asserts "more than 90% of startups fail" in its opening prose with zero citation, then uses the number as background for what its dataset actually studied, which was premature scaling. And found it dead-ends at a 1975 Dun and Bradstreet report... which never contained the number.
Here's what the real data says. . Not great, not 90%. Harvard's Shikhar Ghosh studied 2,000 venture-backed companies and , but read that carefully: it's a claim about investor returns, not survival, and if failure means actually liquidating the company, his number drops to 30 to 40%.
There is exactly one frame where a 90% number is defensible: the fraction of venture-backed startups that fail to produce venture-scale returns. Sit with that. The only version of the axiom that's true is a statement about VC portfolio math. A company can be alive, profitable, employing fifty people, and solving real problems for paying customers, and still count as one of the nine, because it didn't return the fund. The culture took a fact about investor spreadsheets and taught it to founders as a law about businesses.
And it goes deeper, because the 90% isn't just measured by venture math. It's baked into it. The entire portfolio model is built on the assumption that nine of ten will fail: a fund makes its bets expecting one or two to return the whole thing, prices every deal on that expectation, and then prescribes every company the same path to the only outcome that makes the math work. Find your model fast. Lock it in. Scale it. Graduate to the execution machine, because everyone knows execution is how revenue grows.
Except that prescription is the murder weapon. Companies don't fail because they experimented too long. They fail reaching for stability: locking in yesterday's model and scaling it into an environment that has already moved. The failures arrive right on schedule, the number confirms itself for another generation of pitch decks, and everyone nods about the 90% like it's weather. Nine out of ten isn't a failure rate the industry observed. It's a failure rate the industry budgeted for, then manufactured with its own advice. The prophecy funds itself.
So the founding axiom of start-up culture, the one every accelerator repeats on day one, is somewhere between folklore and a misquote, running inside a machine built to keep it true. A culture built on first-principles thinking runs on a statistic nobody can source.
But here's the thing. I don't think the number is wrong because it's folklore. I think something close to it is probably true, for start-ups built the standard way. The failure rate isn't a law of nature. It's an outcome. And outcomes have causes.
To find the cause, you have to ask where the standard way came from.
The man who wrote the playbook
It came from a real place. It has an author, a smart one, and this part I say with respect.
In the mid-2000s, Steve Blank wrote down what became the Customer Development Model, and it conquered the world. Y Combinator, Stanford, Berkeley, every accelerator curriculum, and then, through Eric Ries and The Lean Startup, every founder on earth. If you have ever heard "get out of the building," "validate before you build," or "MVP," you have been taught by Steve Blank whether you know his name or not.
His definition of a start-up is genuinely brilliant: a temporary organization designed to search for a repeatable and scalable business model.
Temporary. Search. Those two words carry the whole model. You search, you find the model, and then you graduate. His process literally ends with a stage called Company Building: departments, specialists, forecasting, process. The organization stops learning and starts executing. Search is the awkward adolescence. The company is the adult.
The four steps run: Customer Discovery, Customer Validation, Customer Creation, Company Building. Search, search, execute, scale. The whole thing points at one destination: become an operating company, because operating companies are what endure.
And in 2010, every word of it was right.
Right for a world that's gone
Think about what building anything cost when this playbook was written.
There's an arc that VCs and business-school faculty have converged on, documented by investor Mark Suster and extended by Columbia's Mattan Griffel: launching a tech startup cost about (servers, Oracle licenses, salaried engineers), about $500K by 2005 after open source, about $50K by 2010 after AWS. And this wasn't vibes. A found that after AWS launched in 2006, the number of first-round-funded software startups roughly doubled while industries that couldn't use the cloud grew only 30%. Cheap experiments literally restructured venture capital.
Then AI compressed the part that was left: the humans. Griffel now puts the cost of starting a software business at roughly $500. Y Combinator reported that for , the batch grew 10% per week in aggregate, and companies were hitting $10 million in revenue with fewer than ten people. Garry Tan's words: "That's never happened before in early-stage venture." Sam Altman told Alexis Ohanian his tech-CEO group chat runs . Meanwhile a dev shop will still quote you , which tells you exactly how much of the old cost was ever about the software.
In the world where the playbook was written, search was brutally expensive. A wrong build could kill you, so of course you interviewed customers before writing code. Of course you validated before you built. Of course search was a temporary phase you escaped as fast as possible, because every week in search mode burned money you probably raised from someone else. Blank's model wasn't dogma in 2010. It was arithmetic.
Both constraints are dead.
Search is no longer expensive. With AI, building the thing is often cheaper and faster than scheduling the interviews about the thing. The MVP that took a funded team six months now takes a small team a weekend. You don't have to ask customers to imagine your product anymore. You can hand it to them.
And the other assumption died too, the quiet one nobody talks about: the assumption that once you find your repeatable model, it stays found.
This one was dying before AI. In 1964, a company that made the S&P 500 could expect to stay on it for 33 years. By 2016 that was 24 years, and , with about half the index replaced within a decade. . That was the baseline BEFORE the machines started building.
Then AI compressed the decay cycle from decades to months. ChatGPT reached . And the same force works in reverse: Chegg was a validated, optimized, publicly-traded execution machine with a repeatable model, and within a couple of years of ChatGPT's launch it had . Stack Overflow, the most defensible community moat in software, has seen . These weren't companies that failed to find product-market fit. They found it, graduated, optimized... and the fit left without them. The VC world has started naming the phenomenon: investor Tomasz Tunguz now argues that , and Reforge coined a term for the other outcome: .
The environment doesn't converge anymore. Business models decay faster than organizations can optimize them. The company that is best at executing yesterday's model loses to the company that is best at discovering tomorrow's. Which means "graduate from search and execute" is no longer advice for how to endure. It's instructions for how to build a local maximum and die on it.
The constraints died. The rituals survived. And a culture that worships disruption never noticed that its own operating system shipped in 2010 and never got an update.
The ecosystem doesn't serve investors. It serves the model.
For years I told people the start-up ecosystem exists to serve investors, not founders. I was close, but I had it slightly wrong, and the truth is worse.
The ecosystem serves the model.
Investors internalized Customer Development so completely that it became the lens money sees through. Funding decisions are pattern-matches against Blank's stages: are you discovering, are you validating, do you have PMF, are you ready to scale? Traction theater, pipeline theater, the pitch itself, all of it is a performance of progress along a curriculum written for a world that no longer exists.
I lived this. When I first started, I went through courses that taught you how to build a pitch deck a specific VC liked, formatted the way that specific firm preferred. Not how to find a problem worth solving. Not how to talk to a customer. How to perform the script for one particular audience of one particular firm. That is not an ecosystem teaching company building. That is an ecosystem teaching compliance with a model.
And look at the metric. The measure of a successful founder, the thing that gets headlines and hero-worship, is how much they raised. Not how much they made. Raised. We celebrate the loan, not the business.
Follow the incentive all the way down and you get the ugly part. If money flows to whoever best appears to be progressing through the stages, then founders are paid to look like the model, whether or not the underlying business is real. Most founders respond to that pressure honestly and just burn themselves out performing. Some respond to it the other way. The spectacular frauds of the last decade weren't aberrations of the system. They were the system's incentives, followed to the end of the road. When appearing-to-progress is what gets funded, you have built a machine that manufactures the appearance of progress.
We are optimizing an entire generation of companies toward a shape we can now see is destined to fail, and paying them to hold the pose.
Traction theater, evening showing.
What I did instead
Before I founded anything, I spent a year downloading over 100 books on entrepreneurship and start-ups into my head. Business model canvases, lean methodology, the whole canon. I read the sacred texts. That's exactly why I stopped believing them, because when you read them all in one year instead of absorbing them over a decade of accelerator sermons, you can see the assumptions, and you can see the assumptions' expiration dates.
What I kept was three questions. I published them in January 2025 as the Tres Preguntas framework (yes, named from the Silicon Valley show, the greatest documentary ever made about my industry, and yes, the goal is tres comas). The long version lives in that post. The short version:
Question 1: Is this problem big enough that solving it creates real economic value?
Question 2: Are we solving it in the best way possible?
Question 3: Are we making it as easy as possible for customers to partner with us to solve it?
Because that's all a business is. Sales. And a sale is nothing more than a partnership between two people agreeing to solve each other's problems. Your problem is money. Theirs is whatever you solve. Every consensual transaction on earth reduces to that, and I have a standing offer to eat my own face on stream if you find a counterexample.
Get to an honest yes on all three and you have product-market fit. Nothing exotic. And here's what I didn't fully understand when I wrote it: the framework quietly maps onto Blank's stages. Question 1 is Customer Discovery. Question 2 is Validation. Question 3 is Creation. For three stages, Steve and I are in almost perfect agreement, which should tell you the disagreement that remains is the one that matters.
Blank's model has a fourth step. Mine doesn't.
His process ends: find the model, then graduate to Company Building, stop asking and start executing. Mine loops: you never stop asking. All three questions stay live for the life of the business, because the answer to every one of them decays. The problem shifts. A better solution appears, usually within months now. The friction moves. A yes is not a diploma. A yes is a reading on an instrument, and the instrument never stops running.
His model is a line. Search, then execute. Mine is a loop. Search and execute, continuously, forever.
That one difference sounds small. It's the whole thing.
The third archetype
Blank gave the world two organizational forms. The start-up, which searches. The company, which executes. His entire lifecycle is the pilgrimage from the first to the second.
I think the AI era has a third form, and it beats both.
A lab. An organization built to run experiments continuously, where the experiments compound: each one produces products, capabilities, and revenue that fund and inform the next. Not a start-up, because it's not temporary and it's not trying to graduate. Not a company, because it never stops searching. The search IS the execution.
And before you tell me labs are what you build after you're rich, look at what's already winning.
. , growing more than 20% month over month, by the company's own account. , the fastest SaaS growth ever recorded, with a team of a few dozen, and . . Around $5 million of revenue per employee, and nobody to perform a pitch deck for.
And if you think this is some AI-bubble anomaly, the precursors were screaming at us the whole time: . . The signal existed before AI. AI just turned it from an outlier into an operating model.
Every one of those is closer to a lab than a company. Tiny teams, permanent search mode, shipping as the discovery mechanism, revenue as the readout. None of them look like Step 4 of the playbook. The org chart with departments and specialists and forecasting, the thing the entire culture teaches you to aspire to, is increasingly the shape of the losers.
The durable organization is no longer the one that graduates from searching. It's the one built to search forever.
Revenue: a readout, not a goal.
Don't graduate
So no, I don't accept 9 out of 10 as a law of nature. I think it's a grade. It's what you get when an entire culture optimizes for a destination, the mature executing company, that the environment stopped rewarding, and teaches its founders fundraising instead of the three questions, and pays them to perform a curriculum instead of build a business.
The advice at the heart of the whole playbook is graduate from search. My advice is the opposite.
Don't graduate.
Stay in the loop. Keep every question live. Build the organization that treats searching not as adolescence but as the permanent job, and let the company be a side effect of the search instead of the goal of it.
And don't mistake this for lowering the bar. I'm not preaching lifestyle businesses and modest expectations, I'm saying the returns are on the other path. Stability is where growth goes to die: the companies that stall are the ones that found a model, built the org chart, and started defending. The ones that compound are the ones that never leave the growth phase, that treat every quarter as another round of experiments. Look again at who's putting up the fastest revenue curves in the history of software: tiny teams in permanent search mode. Staying a start-up isn't the consolation prize. It's the aggressive strategy. Graduation is the conservative play, and the conservative play is what's failing nine times out of ten.
In the future, nobody will care how much you raised. They'll care how much you made, and how much of what you created you kept.
I'm not asking anyone to take this on faith. , the organization Willy and I are building, is the experiment. A lab, in the literal sense, built to search forever, with revenue as one instrument on the panel instead of the destination. Maybe the thesis is wrong. That's what experiments are for.
But I know how I'm betting.
Don't graduate.
If you want the full Tres Preguntas framework, the January 2025 post is . If you want to argue, comments are open. I answer.