The Wrapper Mirage

There is a quiet panic in the application layer, and the founders building there can feel it even when they won't say it out loud.

For two years, the easiest pitch in technology has been "ChatGPT, but for X." Wrap an API. Add a clean interface. Charge a markup. Raise a seed round on the slide that says "AI-native." It worked because the demand was real and the execution was cheap. But cheap execution is exactly the problem, because anything you can build in a weekend, someone else can build in a weekend too – including the company whose model you are renting.

My claim is simple, and I'll defend it the whole way down: most of today's celebrated application-layer AI startups are a short-term trade, not a durable business. In a year or two, the majority of these thin-wrapper companies will either die or get commoditized into irrelevance. The value that lasts is moving in the opposite direction – upstream, toward the people building the foundation models, the infrastructure beneath them, and the deep-tech frontier where progress is gated by physics and science rather than by a clever prompt.

We have already watched the dress rehearsal. When OpenAI shipped its agent natively inside ChatGPT, an entire cohort of standalone "agentic AI" startups simply deflated. Not because they were badly run. Because their whole reason to exist was a feature the platform decided to absorb.

Start with the thing nobody wants to put on a pitch deck: the economics don't survive contact with reality for most of these companies.

Classic software is a beautiful business. You build it once, the next customer costs you almost nothing, and you keep around eighty cents of every dollar. Wrapper economics invert that. Every single query runs the meter. You are paying the lab for inference at retail – they bake in their own margin before you see a cent – so your gross margins land at roughly half of what real software earns. The typical AI-native company today is clearing somewhere around fifty cents on the dollar, with model costs alone eating a quarter of revenue. The worst-optimized ones are down near twenty-five.

That is a structural twenty-to-thirty point penalty for the privilege of building on borrowed intelligence. And here is the detail that should end the debate: even Cursor, the single most celebrated app-layer company of this cycle, ran at negative gross margins until recently. It cost them more to operate the product than they could charge for it. They only clawed their way to positive margins after they started building their own model. Sit with that. The crown jewel of the application layer became viable the moment it stopped being purely an application.

The lab doesn't even have to try to kill you. The math does it for them.

The deeper problem is dependency. Your entire company sits one API key, one price change, one feature announcement away from extinction – and the entity holding all three is also your most capable potential competitor.

This is not a hypothetical fear. Last year Anthropic cut off a fast-growing AI coding tool's access to Claude with less than a week's notice (Windsurf), the moment it looked like OpenAI might buy it. The reasoning was almost charmingly blunt: why would we keep selling our best model to a company about to belong to our rival? One decision, and a startup's core engine was gone.

We have a word for this from the last era – Sherlocking, when Apple quietly ships the feature your app was selling and your business evaporates overnight. The foundation labs now do this at the speed of a model release. They watch what's working in the ecosystem they host, and they build it in. The wrapper is, by design, the cheapest thing to copy.

Jasper is the cautionary ghost here. An AI writing darling that raised at a $1.5B valuation right before ChatGPT went mainstream. Within a year, it had slashed its revenue forecast, done layoffs, quietly marked down its own valuation, and watched its CEO walk. Revenue would later fall by more than half. Jasper didn't fumble the execution – it built a toll booth on a road the platform decided to pave for free.

Now, the smartest people in venture disagree with me, and I'm not going to pretend otherwise. Their argument runs like this: foundation-model intelligence is getting cheaper by roughly an order of magnitude every year, the models themselves are commoditizing into interchangeable pipes, and so the durable value must flow to whoever owns the customer, the data, and the workflow. The model is the dumb plumbing. The app is the product.

And they have proof. Cursor scaled from a hundred million in annual revenue to over two billion in about a year. Harvey put AI into the hands of more than a hundred thousand lawyers and raised at an eleven-billion-dollar valuation. Perplexity is worth north of twenty billion. They have real workflow depth, real switching costs, real data flywheels. The fair version of my critic's case is that these companies have become, functionally, too big to fail – too embedded in how their users work to be dislodged by a model upgrade.

I take that seriously. I just think it proves my point instead of breaking it.

Look closely at what actually makes Cursor or Harvey defensible, and none of it is the wrapper. It's the proprietary data loops, the domain integration measured in years of accumulated workflow, the switching costs that take quarters to unwind – and, tellingly, the fact that the strongest of them started building their own models. Cursor's defensibility didn't come from sitting prettily on top of someone else's intelligence. It came from becoming, in part, a lab itself.

In other words, these companies escaped the application layer by acquiring foundational characteristics. They didn't win as wrappers. They won by ceasing to be wrappers. They are the handful that climbed out of the pit, and we are using their survival to argue the pit is safe.

It isn't. AI tools built on commoditized models without a real moat were the single sharpest correcting category of startup shutdowns this past year. For every Harvey, there is a graveyard you've never heard of, full of companies that had a great demo and a great quarter and no answer for the day the platform shipped the feature for free. Building your career around being the 1 in a 100 isn't a strategy. It's a lottery ticket with a founder's salary attached.

If you don't believe me, believe the capital. In the first quarter of this year, global venture funding hit a record, the overwhelming majority of it pouring into AI – and four upstream companies building models and autonomous systems swallowed roughly two-thirds of all of it. Funding into foundational AI startups in that single quarter doubled the entire prior year. Meanwhile Nvidia posted a record eighty-one billion dollars in a single quarter at a seventy-five percent margin.

This is the gold-rush pattern, and history is merciless about it. The miners mostly went broke. Levi Strauss, selling them denim, got rich. In the dot-com boom, Cisco sold the internet's plumbing and briefly became the most valuable company on earth while the consumer apps became punchlines. The durable value of every platform shift accrues to two kinds of company: the one selling the irreplaceable input, and the one solving something genuinely, physically hard.

Which is why the most interesting place to build right now isn't on top of a model. It's where the moat is made of physics – fusion, novel materials, drug discovery, defense hardware, the chips themselves. You cannot reverse-engineer a fusion magnet or an FDA approval over a weekend the way you can reverse-engineer a prompt. That difficulty is not a bug. That is the entire point. Difficulty is the only moat that doesn't evaporate when the next model drops.