NVIDIA. AirBnB. Amazon. Uber. Doordash. All of these companies have something clear in common: the world looked visibly different after they succeeded.
NVIDIA made modern PC gaming possible. Airbnb made it normal to host people in your home. Amazon allowed you to buy anything from the internet and have it delivered to your home. Uber let people use their cars to replace taxis. Doordash allowed you to have any food from any restaurant delivered to your home. All of these changed the lives of hundreds of millions of people in visible ways. If you look outside, you can see all of these companies and their effects on society.
People watch movies and play games that were edited on or were run on NVIDIA GPUs. Every house you walk past in a nice neighborhood is potentially an Airbnb. You see Amazon trucks outside every day - you'd be hard pressed to walk outdoors in a city and not see one. You probably have had hundreds, if not thousands amazon packages delivered to your house over your lifetime. Every 20th car you see outside might be an Uber, and you see Uber stickers on cars each time you go outside. Saying "I'll doordash it" has become a verb and you see doordash racks in restaurants everywhere, where people come in, grab the orders on their phone using the app, and leave.
These are small things individually, but they're so woven into daily life that we accept them without thinking about it. The technical complexity and sustained ambition from the collective group of people behind each of them is extraordinary, and has been sustained for decades.
Compare that to companies founded today. Almost all of the largest unicorns started in the last 5–10 years roughly all share one commonality - the world is not changed visibly by their existence. For a smaller but still significant percentage of those unicorns, even if they were maximally successful (in a virtual world where all their financial projections go to 100% and they hit 100% of their SAM) - you still would not be able to tell a difference in a way that is comparable to the companies I mentioned above. Why? There's a massive crop of companies that raised at very high valuations in 2021 - and have not raised more money because it would result in a down round. They're in this state of limbo where they're growing marginally YoY but not enough to live up to the hype of what they raised 5 years ago - because the company never deserved that valuation in the first place and because it was incapable of being world-changing.
The contributing factors
The first and largest factor is the social perception of starting a company. Starting a company was not "cool" 20 years ago. If you told your parents you were leaving college to build a startup, unless they were unusually open-minded, it would be met with skepticism and active dissuasion. This extended beyond parents into the rest of society. Collectively, people would not congratulate you for this. A lot of the time this was (and still is, but much less so for factors I will cover in a later essay), reasonable. The stats on startups failing are the stats - we've all heard them repeated ad infinitum. "1 in every 2 startups fail" or whatever other statements percolate through the rest of society when you tell them you're starting a company.
At some point, this began to change. Starting with Zuckerberg and the mythology of dropout → billionaire changed the social perception. Starting around 2016–2017 and accelerating through the COVID era, the media began treating startup founders differently. The creation of mythology around individual people. The earliest I remember this shift happening was Peter Thiel's talk "competition is for losers" at Stanford and the startup school videos at YC. These made the idea of founding a company legible to the average person. If you didn't live in SF, it was hard to find resources on building a company. What even is a startup? Why do I incorporate as a Delaware c-corp? What's a series A? What is PMF? And so on. This made this small world legible to a much larger group of people. I think that this is generally a very good thing, and the world is better off for it. I myself am a result of this pool. When I was 14, I listened to Peter Thiel's talks. I watched the YC startup school videos. I listened to Elon's talks with Sam Altman. I always wanted to be an entrepreneur, but I would be lying if I said that these had no effect on me. The adverse effect is predictable. Opening the pool to more people necessarily changes the distribution of outcomes.
The additional factor, layered on top of this, is the money. The 2008 financial crisis inadvertently created one of the best times in history to start a company. Interest rates were very low, which means that the cost of starting - and funding - a company was low. It was low-risk for investors to fund ideas, and it was low-risk for founders to start them. But - this era produced some of the greatest companies in the world - namely Uber, Airbnb, and Doordash. So that's not the issue. The issue is the social perception of being a startup founder after 2008 began to shift - and it followed a very observable curve from 2008–2021. Social perception changes who wants to start companies. Capital availability changes who gets to. More money was available, more people wanted to start companies, and the criteria for funding them loosened. These are three semi-independent sources contributing to the change in the distribution of outcomes.
Over time, this creates a feedback loop, which we can see the effects of very visibly today. Once being a founder is high-status and capital is cheap, a new kind of founder appears. Someone whose goal is not to build a world-changing company, or solve any kind of difficult problem, or to improve people's lives. This new type of founder's goal is just…to be a founder. Being a founder itself is the prize. The twitter following. The news interviews. The press releases. The podcasts. None of these are bad in and of themselves. I listen to podcasts and enjoy watching interviews of the world's best founders. People who have differentiated insights are interesting to listen to and watch. The issue comes in when these things themselves are the end goal and the motivation. You optimize for sounding smart. For coming across a certain way. For self-validation. For stacking as many credentials as you can. The company in the mind of this new type of founder takes a backseat and it becomes about them rather than building the largest company possible.
This new type of founder is not a corruption of the startup ecosystem, but it is the natural outcome when you optimize for status and capital for long enough.
How AI accelerates this
AI doesn't change the nature of the dynamic I described above. It compresses and accelerates the selection pressure that produced a generation of companies where being a founder, not building a company, is the prize. Software is now much cheaper to build in terms of time and capital, and it's less differentiated than ever. The result is that a truly world-changing company is much harder to build and what can be defined as a world-changing company is much narrower.
The first is supply. If you've ever used Claude Code and it says "This will take 6–8 weeks of work" and you knock it out in an hour, that is a great example of the supply side of software collapsing. Thanks to Claude Code, Windsurf, Cursor, Codex, and so on - more software is being built and shipped than ever before at a rate that is unprecedented. This is great for builders and for the people building things that matter. However - if everyone can build something that works now, what's the moat? Code is a commodity now, and most AI startups haven't internalized this. The argument for the moat is "we'll move faster", which sounds good but doesn't solve the core problem. Shipping fast matters, but for all of the companies I started off describing in this essay, they could cease to ship a single new feature for a year and people would still use them at roughly the same rate. So - shipping fast isn't a moat - or at the very least it's a very weak one.
The second is the capital side. The same VCs who funded the 2021 class are now funding the AI class. High valuations on companies whose differentiation is a function of timing, and whose path to being missed if they disappeared is unclear or simply impossible in their current form. If you ran the disappearance test on the average funded AI startup today, people may be mildly annoyed, but customers would move to an alternative in hours or days.
The conclusion that can be drawn from these two points is your moat needs to come from something AI cannot (currently) replicate. You can still use what always worked, because what makes software companies grow and do well are based on fundamental principles derived from what makes people spend money - which is the goal. The goal is to make people spend money on your company and to make it very difficult for them to stop spending that money. That comes from distribution, unique data, extremely high switching costs, network effects, and so on. The difference now is that these things need to be built very deliberately rather than not being primary focuses.
So, what can be built?
I'll preface this section by saying that everything below is a hypothesis. I can't predict the future, and we live in extraordinary times. I am testing my hypothesis in real life, which I'll expand on later in this section.
The two categories of company I believe can be built AND will endure are companies that interact with physical reality and build data moats from it, and companies that replace or heavily augment the most skilled human cognitive work. The logic comes down to a few things. AI is fundamentally a software phenomenon. That's changing, but hardware is hard and necessitates a much slower rate of change. The cost of collecting data from the real world, from sensors, drones, deployed hardware, from actual people doing actual work outside, is not collapsing in the way building software is. It will eventually, but it will take quite some time. If your company's advantage is a dataset that required real people doing real work in the physical world to generate, then your advantage is structural and that dataset and structural advantage cannot be replicated in a week with Claude Code.
Human replacement in areas and industries that are primarily done by humans - and the most skilled ones - matters for a different reason. The vast majority of AI companies are automating work that was already cheap, like support tickets, email drafting, and so on. These are the companies in the weakest position, for two reasons. First, AI is rapidly getting better at commoditizing this kind of work. Second, the value created is the lowest-order kind, where the only thing the end user benefits from is saved time. If the job I'm automating and saving time for is one anyone could have done anyway, how valuable is that saved time? The answer is it's still valuable, but it's the least valuable type of automation that exists. Companies operating in hard fields - aerospace and mechanical engineering, medical diagnosis, partner-level legal reasoning, senior creative direction - are examples of companies that will endure if built properly. The moat is either something grounded in the physical world or work performed by the most skilled humans in their field. They also require solving problems that don't have a precedent yet for AI automation, or that precedent is being set right now by a select group of companies in each field.
The throughline is that either category of company requires you to build something that AI cannot do alone, either because the data does not exist or the cognitive work is at the frontier of what the most skilled humans can do in their field.
Now, you could argue that Stripe or Figma or another similarly large company is in neither of these cohorts. But - run the disappearance test again. If either disappeared, there would be major issues in society and in the companies that depend on them. Why is this if they don't fit into the categories above? The reason is they changed the entire industry they operated in (Stripe in particular added an entirely new section to an industry) and the world looked different after they succeeded. I touched on this briefly at the beginning of this essay, but the companies that truly matter and become large are ones that make the world - or more specifically, an entire industry - look almost completely different after they existed vs before.
The world is bifurcating rapidly. It will break into companies that fit some or all of the criteria I outlined above, and the hundreds of billions that have been poured into AI companies over the last five years will likely be written down massively. The things that are growing today and will still be here 10–15 years from now will be visible in society OR if they disappeared, it would be a major issue. Today, it's Amazon trucks on every block. Uber drivers are visible everywhere. If you use an app or website, there's a very good chance that the app was designed in Figma. Doordash is a verb. What will the equivalents be in 2035 or 2040? A few companies already embody this and will endure and be a regular part of society by 2035–2040. Zipline is one of them and one of the companies I'm most excited to see continue growing. Speaking more generally, you'll see autonomous vehicles everywhere and it will be normal. Conditions and diseases will be accurately and instantaneously diagnosed by AI systems. Children will largely be taught by AI teachers. Software will design and build and deploy itself. I'm not speculating on any of these. These are logical outcomes of bets being made right now by specific people, and they necessitate that they are built by founders who truly care about solving hard problems and building massive companies.
This really isn't as simple as "some companies will succeed and some will fail", because that's always been the case for as long as companies have existed. The companies that will actually change the reality of daily life will be ones that interact with the physical world, or that replace or augment frontier human talent. Almost all of them will have the common trait of changing an entire industry or changing the world visibly. That's not something you can achieve by shipping faster, although that is important. You need to create things that have unassailable moats. Almost nothing being built in the dominant mode today fits those criteria.
In 10–15 years my hypothesis will be visible from walking down the street the same way it's visible walking down the street today. The loop and throughline I think I've observed in writing this is that really, nothing has changed in terms of principles from the 1990s all the way to today. Whether the moat comes from software alone or physical reality, the principles of building a massive company haven't changed. In 2040, the disappearance test will give the same answer it gives today. The list will just be shorter.