Every time a new technology lowers the barrier to entry, the same prediction follows: now that everyone can do this, no one will have an advantage. The camera phone made everyone a photographer. Spotify made everyone a musician. AI makes everyone a software developer.
The prediction is always half right. The floor rises. More people create, more people ship, more people compete. What the prediction always gets wrong is the ceiling. The ceiling rises faster. And the gap between the floor and the ceiling — between the median and the best — doesn't shrink. It widens.
This is the thing about power laws: they don't care about your intentions. Egalitarian technologies produce aristocratic outcomes. Every time.
AI is not going to be different. It's going to be more extreme.
The Shape That Markets Take
When Spotify launched, it did something genuinely radical: it gave any musician on earth access to the same distribution that previously required a record label, a marketing budget, and a lot of luck. The result was an explosion of music. Millions of new artists, billions of new songs. The floor rose exactly as promised.
But here's what also happened. The top 1% of artists now capture a larger share of streams than they did in the CD era. Not smaller — larger. More music, more competition, more ways to find the best, and listeners, unconstrained by geography or shelf space, converged on it. Spotify didn't democratize music. It intensified the tournament.
The same story played out with writing, photography, software. The internet produced more writers than any technology in history and also produced a more ruthless attention economy. More participants, higher stakes at the top, same basic shape: a small number capturing most of the value.
We find this surprising because we think in linear terms — we expect productivity gains to distribute evenly, like pouring water into a flat container. But most complex systems don't work that way. They never did. The power law isn't a market quirk or a failure of the technology's promise. It's nature's default. Technology doesn't create it. It reveals it.
Consider Kleiber's Law. Across every living organism on earth — from bacteria to blue whales, spanning 27 orders of magnitude of body mass — metabolic rate scales as body mass to the power of 0.75. A whale's metabolism isn't proportionally whale-sized. The relationship is a power law, and it holds with extraordinary precision across essentially all of life. Nobody designed this distribution. It's simply the shape that energy takes in complex systems when left to its own logic.
Markets are complex systems. Attention is a resource. When friction disappears — when geography and shelf space and distribution costs stop acting as buffers — markets converge on their natural shape. Which is not a bell curve. It's a power law. The egalitarian story and the aristocratic outcome coexist, which is exactly why each new technology catches us off guard. We see the floor rise and assume the ceiling is following at the same rate. It isn't. It's pulling away.
AI will do this faster and harder than anything before it. The floor is rising in real time — anyone can ship a product, design an interface, write production code. But the ceiling is also rising, and it is rising faster. The question worth asking is: what actually determines where you end up?
When Execution Gets Cheap, Taste Becomes the Signal
In 1981, Steve Jobs insisted that the circuit boards inside the original Macintosh be beautiful. Not the exterior — the interior. The part no customer would ever see. His engineers thought he'd lost his mind. He hadn't. He understood something that's easy to dismiss as perfectionism but is actually closer to a proof: the way you do anything is the way you do everything. A person who makes the invisible parts beautiful isn't performing quality. They're constitutionally incapable of shipping anything less.
This matters because trust is hard to establish and easy to fake — for a little while. We are constantly running heuristics to figure out who is actually excellent versus who is merely performing excellence. Credentials help but can be gamed. Pedigree helps but gets inherited. What's genuinely hard to fake is taste — the sustained, observable commitment to a standard that nobody required you to meet. Jobs didn't have to make those circuit boards beautiful. The fact that he did told you everything about what he'd do when you weren't watching.
For most of the last decade, this signal was somewhat muted. In the peak SaaS era — roughly 2012 to 2022 — execution became so standardized that distribution was the actual scarce resource. If you could acquire customers efficiently, build a sales machine, hit your Rule of 40 — the product almost didn't matter. You could build something mediocre and win if your go-to-market was strong enough. The signal that taste sends got buried under the noise of growth metrics.
AI changes the signal-to-noise ratio completely. When anyone can generate a functional product, a polished interface, a working codebase in an afternoon — the question of whether something works stops being the differentiator. The question becomes: is this actually excellent? Does this person know the difference between good and insanely great, and do they care enough to close that gap even when no one is forcing them to?
This is especially true for business-critical software — the systems companies trust with their payroll, their compliance, their employee data. These aren't products you adopt casually and abandon next quarter. The switching costs are real, the failure modes are serious, and the person deploying it is accountable for what happens. Which means before they sign, they're running every trust heuristic they have. A beautiful product is one of the loudest signals available. It says: the people who built this care. They cared about the parts you can see, which means they probably cared about the parts you can't.
In a world where execution is cheap, taste is the proof of work.
What the New Phase Rewards
This was always true. But for about a decade, the market made it almost impossible to see. The most important skill in software had nothing to do with software.
Between roughly 2012 and 2022, the core architecture of SaaS had been figured out. Cloud infrastructure was cheap and standardized. Developer tools were mature. Building a functional product was hard, but it was a solved kind of hard — you could hire your way through it, follow established patterns, reach an adequate outcome with sufficient resources. What remained genuinely scarce, what actually separated the winners from the rest, was distribution. Could you acquire customers efficiently? Could you build a repeatable sales motion? Did you understand your unit economics well enough to pour fuel on the fire at exactly the right moment?
The founders who thrived in that environment came from sales, consulting, finance. They were fluent in metrics that would have been gibberish a decade earlier — net dollar retention, average contract value, magic number, Rule of 40. They lived in spreadsheets and pipeline reviews and they were, in that context, exactly right to. The conditions of peak SaaS produced peak SaaS founders. It was a rational adaptation.
I found it suffocating.
I grew up in a small town in a state of 250 million people in India. Each year, roughly three students from all of India get into MIT. Every one of them, without exception, comes from an expensive prep school in Delhi, Bangalore, or Mumbai — institutions built specifically to produce this outcome. I was the first person in the history of my state to get in. I say this not to impress but because it's the thesis of this essay lived in miniature: when access is limited, pedigree predicts outcomes. When access opens up, the people who go deepest win anyway. I was the depth bet in a room full of pedigree. It's the only kind of bet I know how to make.
I studied physics, mathematics, and computer science — fields where the deepest insights don't come from process optimization but from seeing something true that others had missed. My master's thesis was on straggler mitigation for distributed machine learning training: what happens when you're running a system at massive scale and parts of it fall behind, and how you optimize around that constraint without losing the integrity of the whole.
When I looked at the startup world in my early twenties, what I saw was a landscape where all of that felt beside the point. The premium was on go-to-market, not on the thing itself. Building something technically extraordinary seemed almost naïve — a distraction from the real game, which was acquisition and retention and sales velocity.
Then, in late 2022, the conditions changed.
What ChatGPT made visible — what it made visceral, in a way that years of research papers hadn't — was that the curve had bent. A new S-curve had started. Phase transitions don't reward the people who were best adapted to the previous phase. They reward the people who can see what the new phase makes possible before anyone else has priced it in.
I quit my job and started Warp.
The bet was specific. The United States has over 800 tax agencies — federal, state, local — each with their own filing requirements, deadlines, and compliance logic. There is no API. There is no programmatic access. For decades, every payroll provider had handled this the same way: with people. Armies of compliance specialists manually navigating a system that was never designed to be navigated at scale. The incumbents — ADP, Paylocity, Paychex — had built entire business models around this complexity, not by solving it but by absorbing it into their headcount and passing the cost to customers.
I could see that agents were brittle in 2022. I could also see the curve of improvement. Someone who had spent years thinking about distributed systems at scale, watching the trajectory of these models up close, could make a calibrated bet that what was brittle then would be capable within a few years. So we made the bet: build an AI-native platform from first principles, starting with the hardest workflow in the category — the one no incumbent could automate because their architecture was never designed for it.
That bet is now paying off. But the broader point is about pattern recognition. Technical founders in the AI era don't just have an engineering advantage — they have an insight advantage. They see different entry points. They make different bets. They can look at a system that everyone else has accepted as permanently complex and ask: what would it take to actually automate this? And then, critically, they can build the answer.
The founders who dominated peak SaaS were rational optimizers working within a set of constraints. AI is removing those constraints and installing different ones. In the new environment, the scarce resource isn't distribution. It's the ability to see what's now possible — and the taste and conviction to build it to the standard it deserves. But there's a third variable that determines everything. And it's the one most founders in the AI era are getting catastrophically wrong.
The Long Game at High Velocity
There is a meme circulating in startup culture right now that goes something like: you have two years to escape the permanent underclass. Build fast, raise fast, exit or die.
I understand where it comes from. AI is moving at a pace that feels existential. The window to catch a wave feels narrow. Young people watching overnight success stories on Twitter reasonably conclude that the game is about speed above all else — that the founders winning are the ones who moved fastest in the shortest window.
This is true about exactly the wrong thing.
Speed of execution matters enormously. I believe this as deeply as anyone — it's baked into the name of the company I'm building. But speed of execution is not the same thing as shortness of horizon. The founders who are going to build the most valuable companies in the AI era are not the ones who sprint for two years and sell. They're the ones who sprint for ten years and compound.
Here's what short-termism gets wrong: the most valuable things in software — proprietary data, deep customer relationships, genuine switching costs, regulatory expertise — take years to accumulate and cannot be replicated quickly regardless of how much capital or AI capability a competitor brings to bear. When Warp processes payroll for a company across multiple states, we're accumulating compliance data across thousands of jurisdictions. Every tax notice resolved, every edge case navigated, every state registration completed trains a system that gets harder to replicate with each passing month. That's not a feature. It's a moat that only exists because we've been building it consistently, at high quality, for long enough that it has mass.
This kind of compounding is invisible in year one. It's faintly visible in year two. By year five it's the entire game.
Frank Slootman, who has built and scaled more software companies than almost anyone alive, puts it simply: get comfortable with being uncomfortable. Not for a sprint. As a permanent condition. The fog of war in an early stage company — the disorientation, the imperfect information, the constant requirement to commit to actions anyway — doesn't resolve after two years. It evolves. New uncertainties replace old ones. The founders who last are not the ones who find certainty. They're the ones who learn to move clearly without it.
Building a company is brutal in ways that are hard to convey to someone who hasn't done it. You live in a state of continuous low-grade terror punctuated by moments of higher-grade terror. You make thousands of decisions with incomplete information, knowing that a long enough string of wrong ones ends everything. The overnight success stories you see on Twitter are not just outliers within an already power-lawed distribution — they're extreme outliers within an already power-lawed distribution. Optimizing your strategy around them is like training for a marathon by studying the finishing times of people who took the wrong turn and accidentally ran a 5K.
So why do it? Not because it's comfortable. Not because the odds are favorable. Because for some people, there is no alternative that feels like actually living. Because the only thing worse than the terror of building something from nothing is the quiet suffocation of not trying.
And because — if you're right about the bet, if you've seen something true that others haven't priced in, if you execute with taste and conviction over a long enough horizon — the outcome isn't just financial. You build something that genuinely changes how people work. You create a product that people love using. You employ and develop people who do their best work inside the thing you built.
That's a ten-year project. AI doesn't change that. It never did.
What AI changes is the ceiling of what's possible in those ten years — for the founders who stay in it long enough to find out.
The Ceiling Nobody's Watching
So what does software actually look like on the other side of this?
The optimists say AI creates abundance — more products, more builders, more value distributed across more people. They're right. The pessimists say AI destroys software moats — that anything can be replicated in an afternoon, that defensibility is dead. They're also partially right. Both camps are looking at the floor. Neither is looking at the ceiling.
There will be thousands of point solutions — small, functional, AI-generated tools that solve narrow problems adequately. Many of them will be built not by companies but by individuals, or by internal teams scratching their own itch. For a certain category of low-stakes, easily replaceable software, the market will look genuinely democratized. The floor will be high and the competition will be fierce and the margins will be thin.
But for business-critical software — the systems that companies trust with money movement, compliance, employee data, legal exposure — something different happens. These are low-fault-tolerance workflows. When payroll fails, people don't get paid. When tax filings are wrong, the IRS notices. When benefits enrollment breaks during open enrollment, real people lose coverage. The person who chose the software is accountable for what it does. That accountability doesn't get outsourced to an AI that vibecoded a solution in an afternoon.
For these workflows, companies will continue to trust vendors. And among those vendors, the winner-take-most dynamic will be more extreme than anything we've seen in the previous generation of software. Not because network effects are stronger — though they are — but because the compounding advantages of an AI-native platform operating at scale, accumulating proprietary data across millions of transactions and thousands of compliance edge cases, become nearly impossible to replicate from a standing start. The moat isn't a feature set. It's what accumulates when you operate at high quality, at scale, in a domain that punishes mistakes, for long enough that no one can quickly replicate what you've built.
This means software markets consolidate harder than they did in the SaaS era. In HR and payroll ten years from now, I don't expect twenty companies each with single-digit market share. I expect two or three platforms capturing the vast majority of value, and a long tail of point solutions capturing almost none. The same pattern will play out across every software category where compliance complexity, data accumulation, and switching costs compound together.
The companies that end up at the top of those distributions will look similar to each other: founded by technical people with genuine product taste, built on AI-native architecture from day one, operating in markets where the incumbents are structurally unable to respond without dismantling their existing business. They will have made a specific insight bet early — seen something true about what AI makes possible that others hadn't priced in — and then stayed in long enough for the compounding to become visible.
I've been describing this founder in the abstract. But I know exactly who he is, because I'm trying to be him.
I started Warp because I believed, in 2022, that the entire stack of employee operations — payroll, tax compliance, benefits, onboarding, devices, HR ops — was sitting on a foundation of manual labor and legacy architecture that AI could replace entirely. Not improve. Replace. The incumbents had built billion-dollar businesses by absorbing complexity into headcount. We would build by eliminating the complexity at the source.
Three years in, that bet is playing out. Since we launched, we have processed over $500M in transactions, are growing quickly, and serving companies that are building some of the most important technology in the world. Every month, the compliance data we've accumulated, the edge cases we've navigated, the integrations we've built make the platform harder to replicate and more valuable to the customers on it. The moat is still early. But it has mass, and it's accelerating.
I'm telling you this not because Warp's success is inevitable — nothing in a power-lawed world is inevitable — but because the logic that led us here is the same logic I've been describing throughout this piece. See something true. Go deeper than anyone else is willing to go. Build it to a standard that doesn't require external pressure to maintain. Stay in long enough to find out if you were right.
The extraordinary companies of the AI era will be built by people who understood that access was never the scarce resource — insight was. That execution was never the moat — taste was. That speed was never the advantage — depth was.
The power law doesn't care about your intentions. But it rewards the right ones.
