Building Where Public AI Labs Won’t Go

One of the AI labs you build on is about to become a public company, and most founders are reading that news exactly backwards.

Anthropic filed a draft registration with the SEC in the middle of 2026 and is lining up a listing that could put it near a trillion dollars, possibly the first AI company to debut at that size. OpenAI filed its own paperwork the same summer, chasing a similar number. The commentary is all about the debut: who gets rich, what the first-day pop looks like, whether the whole thing is a bubble. That is the fun story. It is also the story that matters least to you.

Here is the part that matters. The moment your most important supplier becomes a public company, its incentives change, and they change in a direction you can predict. A private lab can chase research glory and burn cash for a decade. A public one answers to a share price every ninety days. That single shift tells you, with more precision than any roadmap leak, exactly which parts of the market it will fight for and which parts it will quietly abandon. The abandoned parts are where you build.

I run two companies that depend on frontier models for real work, and I have watched this pattern rehearse itself in every platform wave before this one. The winners were never the people standing closest to the platform. They were the people who read the platform owner’s incentives and built in the one place the owner could not afford to go. This post is a map of those places for the age of the trillion-dollar model lab, and a way to decide, on a Monday morning, whether the business in front of you sits inside the safe zone or the kill zone.

The trillion-dollar debut changes the math below it

Start with the numbers, because they set the whole frame. By 2026 the gap between the top models had collapsed. Stanford’s AI Index put the spread between the leading labs at roughly two to three percent, and the spread between the best American and Chinese models fell from more than seventeen points in 2023 to under three. Model quality, the thing everyone raced on for three years, stopped being a place you could win. When six labs are within a rounding error of each other, “we use a slightly better model” is not a business. It is a sentence you delete from your pitch.

So the value moved. It moved into everything wrapped around the model: the workflow, the memory, the interface, the integrations, the unit economics, the trust. The people who study this closely say it plainly now. The model alone is no longer the product. The product is the model plus the harness plus the domain plus the relationship. That is good news for founders, because those are things a solo builder can own and a giant cannot easily copy.

But it is only good news if you know where the giant is heading. A private lab moving into applications is an annoyance. A public lab moving into applications is a freight train with a schedule, because now it has to. When Anthropic or OpenAI carries a valuation near a trillion dollars into the public market, that number is a promise about future growth, and the market collects on that promise every quarter. To keep the promise, the lab has to keep finding enormous, high-margin markets to grow into. That mandate decides where it competes.

Most founders treat “the labs are coming for the app layer” as a single undifferentiated threat. It is not. A public lab is pulled hard toward some markets and pushed hard away from others, by the same force. If you can see the force, you can see the map. I wrote about the general shape of this danger in the note on platform risk for founders, which is about the threat itself. This post is about the inverse: the specific ground the threat structurally cannot reach.

The No-Go Map

Here is the core idea. A public AI lab is not free to enter every market. Its own incentives forbid whole categories of business. I call the set of forbidden categories the no-go map, and I call any single forbidden category a structural no-go zone. Structural is the important word. These are not markets the lab overlooked or has not gotten to yet. They are markets it cannot profitably enter without breaking the promise it made to its shareholders. The barrier is built into what the company is, not into what it has done so far.

There are five of them. A market becomes a no-go zone when it is too small to matter, too deep to model, too dangerous to sign for, too human to scale, or too trusted to buy into. Most durable solo and small-team businesses in the AI era will sit inside one of these five, usually two or three at once. The more of them a market touches, the safer you are.

The rest of this post walks each zone, shows what a real business inside it looks like, and hands you a way to score your own idea against the map. First, the engine that makes the map work, because if you do not believe the giant is pushed away, none of the five zones will feel safe.

Why a public lab migrates up and away

Call it margin migration. A newly public company is under what I think of as the quarter clock: a repeating ninety-day demand to show growth and defend margins, enforced by people who can sell the stock the morning they lose faith. The clock does not care about your mission. It cares about the next print. And it drives product strategy in a very specific direction.

To grow into a trillion-dollar valuation, a lab needs markets that are huge, high-margin, and horizontal. Huge, so the win moves the number. High-margin, so the win defends the stock. Horizontal, so one product serves everyone and the sales motion stays cheap. Coding assistants, general agents, enterprise copilots, a services arm aimed at the biggest consulting budgets in the world. Those are the markets that fit the clock, and that is exactly where the labs are marching. The trade press already caught OpenAI standing up a forward-deployed services group to chase the largest enterprise contracts. That is margin migration in the open. The lab is climbing toward the fattest, broadest revenue it can find.

Now flip it. The same clock that pulls the lab toward the fat middle pushes it away from anything thin, narrow, messy, or slow. A market worth twenty million dollars a year is a rounding error to a company that has to add billions to justify its price. Worse than a rounding error, it is a distraction that a shareholder letter would have to explain. So the lab does not enter it. Not because it lacks the technology. Because entering would cost focus and return nothing the market rewards.

That space, between the markets a giant is paid to chase and the markets you are paid to serve, is the incentive gap. Your entire business can live inside it, safely, for years. And you have a set of advantages in that gap that the giant will never match. I call the bundle the founder discount. You have no board demanding a bigger number. You can run a two-million-dollar business happily for a decade. You can sign contracts a public company’s legal team would refuse. You can put a human on a call. You can hold a relationship that predates any model. None of those show up on a lab’s income statement as anything but cost, so the lab treats them as cost and stays out. You treat them as the moat.

One caution before the zones. Margin migration protects you from the giant entering your market. It does not protect you from the giant changing the terms of the tools you rent from it. Those are two different risks, and I will come back to the second one, because it is the real danger of building in this era.

Zone 1: Markets too small to matter

The first no-go zone is the simplest and the most underrated. A market can be too small for a giant and perfect for you at the same time. There is a size of business that is invisible from the top of a trillion-dollar company and life-changing from where a solo founder sits. The gap between those two views is enormous, and it is pure opportunity.

Think about who a public lab can afford to serve. Its sales and support cost structure is built for enterprise and self-serve at scale. A market of four thousand independent inspection firms, or every orthodontist in three states, or the back office of regional freight brokers, does not clear the bar. The total addressable market is real money to you and a rounding error to them. Corporate strategy teams at large companies openly say they only chase adjacent markets when the revenue is large enough to justify the move. Your niche is defined precisely by being below that line.

I have built inside this zone and the feeling is strange at first. You keep waiting for the giant to notice and crush you, and it never comes, because noticing would cost the giant more than winning would pay. A friend runs a tool for a specific kind of medical clinic. The whole market is maybe a few thousand clinics. He has most of them. No frontier lab will ever staff a team to take that away, because the entire market, captured, would not pay for the team. He is safe not because he is fast but because he is small in the exact way that repels giants.

The trap here is confusing small-and-defended with small-and-doomed. A market too small to matter is safe only if it is also hard to serve well. If any developer with an API key can replicate you in a weekend, small does not save you, because ten other small builders show up. Small-TAM is a real moat only when it pairs with one of the next four zones. I made the general version of this argument in the vertical AI playbook. Here the point is narrower: smallness itself is a shield against the one competitor you most fear.

Zone 2: Workflows too deep to model

The second zone is where most of the durable AI businesses of this decade will live. A giant sells horizontal capability. It hands you a model that can, in principle, do almost anything. What it does not do is the last mile: the specific workflow of a specific job, wired into the specific tools that job already runs on, tuned on data that only shows up when you do the work.

The people who fund companies for a living have already voted with their checkbooks. Investor sentiment has moved off thin wrappers and onto workflow ownership, proprietary operational data, and domain depth. The reason is that model quality converged, so the only things left to defend are the things a model cannot supply on its own. Depth of workflow. Data that accrues from usage. Integrations that took two years to build and certify. Domain judgment encoded in a hundred small product decisions.

None of that is a place a public lab wants to go, because all of it is slow, narrow, and low-margin per unit of engineering. The lab optimizes for capability that generalizes. Vertical depth is the opposite of generalizing. It is a thousand ugly special cases that only pay off in one industry. Here is the contrast that decides who wins each layer.

The proprietary data point is the one to sit with. A vertical product that does real work generates data no general model has: what a good outcome looks like in this trade, which edge cases actually occur, how the numbers should reconcile. That data trains features a horizontal competitor cannot match at the same accuracy or safety, and it compounds the longer you run. This is the same engine I described in the data moat playbook. The public lab has more data than you in general and less than you in particular, and particular is where the buyer lives.

Zone 3: Liability nobody at a trillion-dollar company will sign

The third zone is legal, not technical, and it is one of the strongest shields you can stand behind. In many of the most valuable markets, the buyer is not really buying software. They are buying someone to be responsible. A hospital, a bank, a law firm, a factory needs a vendor who will sign a contract, carry insurance, pass an audit, and stand behind the result when a regulator asks who is accountable.

A public AI lab is structured to minimize exactly that exposure. Read the terms on any frontier model and you will find the responsibility pushed firmly onto you, the developer. That is deliberate and correct for a company serving millions of use cases it cannot inspect. But it means the lab cannot walk into a regulated buyer and say the words that buyer needs to hear: we are accountable for this outcome, here is the contract, here is the audit trail, here is the insurance. Saying that at trillion-dollar scale, across every vertical at once, is an unbounded liability. No public company signs unbounded liability across the whole economy.

You can. You can go to one industry, learn its rules cold, build the audit trail it requires, buy the insurance, and sign the contract that makes you accountable. That signature is the product. The model underneath is a component. A founder who will stand behind a specific regulated outcome has something a trillion-dollar lab structurally cannot offer, and the buyer will pay a premium for it precisely because it is scarce. I went deep on operating inside rule-bound markets in the AI-native founder playbook. The compressed version: regulation looks like a barrier to entry, and for the giant it is one. For you it is the moat, because you are willing to carry the risk it refuses.

Zone 4: Work that needs a human in the room

The fourth zone is about margin structure, and it makes purists uncomfortable. Some of the best AI businesses are not pure software. They are sixty percent software and forty percent service, and that mix is a feature, not a failure. A public lab is graded on software gross margins. Every hour of human service in the delivery drags that number down, and the market punishes the drop. So the lab avoids service the way it avoids a leak.

Which means any market where the customer genuinely wants a human involved is a market the giant will not fully take. Onboarding that requires sitting with a client for a week. Implementation that needs someone who understands the customer’s mess. A quarterly review where a real person interprets the output and takes the blame if it is wrong. These are not deficiencies to automate away as fast as possible. In the right market they are the reason the customer chose you and not the self-serve tool.

The economics that repel the giant attract you. A blended software-and-service business can charge more, churns less, and builds relationships that no model upgrade can dislodge. Yes, it scales slower. Slower scaling is a problem for a company that has to double every year to defend a stock price. It is not a problem for a founder who wants a durable, profitable, hard-to-copy business. The quarter clock that forces the lab to shed service is the same clock that hands you the customers who wanted service all along.

Zone 5: Trust that cannot be bought

The fifth zone is the least technical and often the most durable. Some markets are gated by trust, and trust does not respond to model quality. If a buyer already trusts you, your community, or your brand, a better model on the other side of the table does not move them. The relationship is the product, and the software rides on top of it.

This is why an established operator with a small audience often beats a better-funded newcomer with a better model. The audience is not buying the model. They are buying the person or the name they already believe. A public lab can ship a technically superior product into that market and watch it bounce off, because the thing standing in the way is not a capability the lab lacks. It is a relationship the lab does not have and cannot acquire on the quarter clock.

Building in this zone means the moat is you, or the specific trust you have earned, not the code. That has a cost. It is harder to sell the company, harder to step away, harder to hand off. But it is close to impossible for a giant to copy, because the giant cannot manufacture the years of earned belief that make a buyer say yes without shopping around. Trust is the one input on the no-go map that money genuinely cannot buy faster, which is exactly why it repels the player with the most money.

Put the five together and a pattern emerges. Every zone is a place where the thing that makes you money is something a public lab treats as a cost, a risk, or a distraction. That is not a coincidence. It is the whole method. Find where your strength is the giant’s liability, and build there.

The supplier IPO risk test

Now the danger I promised to return to. The no-go map keeps the giant from entering your market. It does nothing about the giant changing the deal on the tools you rent from it. When your supplier goes public, that second risk gets sharper, and you have to manage it on purpose. I call it supplier IPO risk, and it is a specific flavor of the concentration risk every builder on a platform carries.

A newly public lab, chasing the quarter clock, has three moves that can hurt you even while it stays out of your market. It can raise prices to defend margin, which hits your unit economics overnight. It can deprecate the model or endpoint you built on, forcing a migration on its schedule, not yours. And it can ship a first-party feature that eats a piece of your product, not your whole market, but the piece that was your wedge. None of these require the lab to enter your niche. They only require it to optimize its own numbers, which after an IPO it is legally obligated to do.

The defense is not to avoid the platform. You cannot build in this era without renting frontier capability from someone. The defense is to build so that any one of those three moves is survivable. That means keeping your prompts and orchestration portable across at least two providers, so a price hike or a deprecation is a config change and not a rewrite. It means owning the parts of your product that do not depend on the model at all, the workflow, the data, the relationship, so a first-party feature dents you instead of killing you. I laid out the switching mechanics in the note on vendor lock-in and switching cost, and the broader “will my business survive the platform” question runs through the commoditization clock. Run this quick test against any AI business, yours or one you are considering.

A business that passes all three is one where the supplier’s IPO is a headline you read with mild interest, not a threat to your survival. That is the goal: to be a customer of the frontier, never a hostage to it.

What most founders get wrong

The common advice, repeated at every demo day and in every thread, is to build on the frontier. Use the newest model, ride the capability curve, stay as close to the cutting edge as you can. It sounds obviously right. It is quietly the most dangerous advice in the field.

Here is the flip. The frontier is the single worst place to build a durable business, because the frontier is exactly where the public labs are forced to compete. When your product’s whole value is that it wraps the latest model a little more nicely, you have set up shop in the one market a trillion-dollar company is legally obligated to win. You are not riding the wave. You are standing in front of the freight train, admiring how fast it is coming. The thin-wrapper failure mode is common enough that I gave it its own writeup in the wrapper trap, and the deeper point is this: proximity to the frontier feels like safety and is the opposite.

The safe place is the boring place. A market too small to matter, wrapped in a workflow too deep to model, protected by liability nobody at a giant will sign, delivered with a human in the room, sold on trust that predates every model. That business will never trend on tech Twitter. It will also never be taken from you by a company worth a trillion dollars, because taking it would violate the promise that company made to keep its valuation. Founders chase the frontier because it feels like the center of the action. The durable money is at the edges the center cannot reach.

I want to give the opposing case its due, because it is not empty. Building on the frontier can be right for a very specific kind of company: one that intends to move fast, get acquired, and be gone before the platform catches up. That is a real strategy and some people run it well. But it is a trade, not a business, and it depends on timing the exit before the giant arrives. If your goal is to own something that lasts, the frontier is a place to visit for capability and a terrible place to plant your flag. The honest counterweight actually sharpens the rule: the frontier is for renting power, the edges are for building wealth.

What to do Monday morning

Turn the map into a decision. Take the business you run now, or the one you are thinking about starting, and put it through four steps before lunch.

First, score it on the no-go map. Give it one point for each of the five zones it genuinely sits in: too small to matter, too deep to model, too dangerous to sign, too human to scale, too trusted to buy. Be honest, not generous. A score of zero or one means you are exposed, a business whose only defense is speed, and speed does not beat a company with a hundred billion dollars of cash. A score of three or more means you are standing on ground the giant is structurally kept off. If you score low, your job this quarter is to move the business into more zones, not to ship faster inside the kill zone.

Second, run the supplier IPO risk test from earlier. Take each of the three moves in turn, a doubled price, a killed endpoint, a first-party feature that eats your wedge, and write one sentence on what you would actually do in the next week if it happened. If any answer is “rewrite the product” or “shut down,” you have found the thing to fix before you scale, not after.

Third, pick one zone and go deeper instead of wider. The instinct when a giant looms is to broaden, to add features, to chase a bigger market for safety. That runs you straight toward the giant. The correct move is the opposite: take the zone you are already strongest in and drive it deeper. More workflow, more proprietary data, more regulatory coverage, more service, more trust. Depth in a no-go zone compounds. Breadth toward the frontier evaporates. The way to think about what only you can do, and to protect it, runs through the incompressible core, and the wider question of what is even worth building when building is nearly free is in this companion piece.

Fourth, when you evaluate the next idea, start from the map, not from the model. The failing question is “what can I build with the newest model.” The winning question is “where can I build that a public trillion-dollar lab is structurally forbidden to follow.” Screen every idea through the second question first. If you want a wider survey of where those openings sit across the market, I keep an updated view in the AI opportunity map. The IPO wave everyone is watching is not a threat to you. Read correctly, it is the clearest signal you will ever get about where the giants must go, which tells you exactly where they cannot.

Frequently asked questions

Does a public AI lab really avoid small markets, or does it just get to them later?

It avoids them structurally, not temporarily. A public company chasing a trillion-dollar valuation has to add revenue large enough to move its number and defend its margin every quarter. A market worth tens of millions a year cannot do that no matter how easy it is to enter, so entering it costs focus and returns nothing the stock market rewards. The barrier is the lab’s own incentive structure, and that does not ease over time. It tightens as the valuation grows.

What is a structural no-go zone in one sentence?

It is a market a public AI lab cannot profitably enter without breaking the promise it made to shareholders, because the market is too small to matter, too deep to model, too dangerous to sign for, too human to scale, or too trusted to buy into. The word structural means the barrier is built into what the company is, not into what it happens to have done so far.

Isn’t building on the frontier model the safest bet since it’s the best technology?

The best technology and the safest business are different questions. The frontier is where the public labs are obligated to compete, so a product whose main value is wrapping the newest model sits in the one market a trillion-dollar company must win. Use frontier capability as a rented component, but build your durable value in a no-go zone the labs are pushed away from, not on the frontier they are pulled toward.

How do I manage the risk that my AI supplier goes public and changes its terms?

Treat it as concentration risk. Assume the supplier can double your price, deprecate the model you built on, or ship a first-party feature that eats your wedge, all of which a public company may be obligated to do to defend its numbers. Keep your orchestration portable across at least two providers so a change is a config edit, not a rewrite, and own the parts of your product, the workflow, data, and relationship, that do not depend on the model at all.

What if a giant lab decides to enter my niche anyway?

If your niche truly sits in three or more no-go zones, the lab entering would cost it more than winning would pay, so it stays out for the same reason it stays out of every other small, messy, liability-heavy corner of the economy. If the lab does show interest, that is a signal your market is larger, cleaner, or more horizontal than you thought, which means you were not as protected as the map suggested. The fix is to move deeper into the zones, not to try to out-run a company with unlimited cash.

Does this mean I should never build anything horizontal?

Horizontal is fine as a delivery model and dangerous as your whole moat. Plenty of no-go-zone businesses sell a broadly useful tool, but the thing that protects them is vertical: the specific workflow, data, liability, service, or trust underneath the general surface. If your product is horizontal all the way down, with nothing a public lab is structurally kept away from, you are exposed regardless of how useful it is.

How is this different from just saying “go vertical”?

Going vertical is one of the five zones, the workflow-depth one, and it is a good instinct. The no-go map adds four more shields most vertical advice skips: market size that repels giants, liability they will not sign, service that drags their margin, and trust they cannot buy. It also keys the whole choice to a new input, the supplier’s public-company incentives, which tells you not just to go vertical but exactly which verticals and features a trillion-dollar competitor is forbidden to chase.

Is the IPO wave good or bad for small AI founders?

Read correctly, it is good, because it is the clearest map you will ever get of where the giants must go. The quarter clock forces public labs toward huge, horizontal, high-margin markets, which tells you precisely which small, deep, messy, human, trusted markets they are forced to leave alone. The event everyone is treating as a threat is actually a floodlight on the safe ground, if you build at the edges it points away from.