Wrappers are dying
The survivor story Is more interesting.
In 2023, if you had a laptop, an API key, and a Figma account, you could raise a seed round. The pitch wrote itself. By 2024, Carta was tracking 966 US startup shutdowns, up 25.6% from the year before, and a significant chunk were exactly what you’d expect: polished frontends stapled to an OpenAI API call, charging subscription fees for something ChatGPT would eventually make free.
The AI wrapper class is collapsing. This is not a controversial statement anymore. A Google VP said it on TechCrunch last week. A Forbes Council post said it the week before. What both pieces got wrong, and what actually matters, is what happens to the one or two companies in this category that won’t collapse. Because those companies are going to be worth more than anyone currently thinks.
The MySQL Rebuttal + Mowry Pivot
There is a popular defense of the wrapper startup that goes like this: calling a company an AI wrapper is like calling Salesforce a MySQL wrapper. Technically true. Completely meaningless. The YC partners made this argument on their Lightcone podcast, and they are not wrong in the abstract.
But the analogy only holds if the company actually built something on top of the infrastructure. Salesforce did not wrap MySQL. It built an entirely new category of organizational software that happened to use a relational database underneath. Most AI wrappers did not do that equivalent. They built a UI and called it distribution. They built a prompt and called it IP. They built a landing page with a gradient and called it a company.
Google’s VP of global startups made a sharper analogy last week. He compared today’s AI aggregators to the AWS resellers of 2010, middlemen who marketed themselves as easier entry points until Amazon built enterprise tools directly and squeezed them out. The only survivors were the ones that had added real services before Amazon noticed them: security, migration, DevOps consulting.
He is right about aggregators. The more interesting question is which wrappers are the Rackspace of this cycle, the ones that built real services before the platform noticed them, and which ones are just reselling compute with a gradient. That distinction is worth a few billion dollars and almost nobody is drawing it cleanly.
The test is simple: if OpenAI ships a feature and your product evaporates, you were never a company. You were a gap in someone else’s roadmap. And OpenAI closes gaps for sport.
Dozens of “chat with your PDF” startups learned this when ChatGPT added document uploads natively. Dozens of AI meeting summary products learned it when Zoom, Teams, and Google Meet built the feature themselves. The wrapper model assumed the gap was permanent. It was not. It was just early.
The Survivor Numbers
Here is something the “wrappers are all dead” crowd is not adequately pricing: Cursor went from $1 million to $1 billion in annualized revenue in under two years. It is now valued at $29.3 billion. Half the Fortune 500 uses it. Replit grew 1,556% year over year, going from $16 million ARR at the end of 2024 to an estimated $265 million through 2025. Lovable hit $70 million ARR by June 2025, having launched in late 2024, making it among the fastest-growing startups in European history.
These companies were all dismissed at some point as AI wrappers. Cursor was “just a GitHub Copilot clone.” Replit was “just a browser IDE.” And yet OpenAI reportedly tried to acquire Cursor before the company declined. It then bid $3 billion for Windsurf, another player in the space, only to have that deal fall apart when Microsoft blocked it over IP concerns. At which point Google stepped in and paid $2.4 billion for Windsurf’s talent and licensing rights. Two of the biggest AI companies on the planet engaged in a bidding war over companies that a certain strain of VC conventional wisdom had written off as undifferentiated wrappers.
It also helped that Cursor launched into a category tailwind where every developer on earth simultaneously decided AI coding tools mattered, timing that cannot be engineered backward into a clean framework. But timing explains the speed, not the destination. The question is what they built that made the destination worth $29 billion.
The Behavior Loop
The thing that separates Cursor from a dead AI coding assistant is not its model. Its model is, by design, someone else’s model. What Cursor built is a proprietary understanding of how software engineers actually work: the keystrokes, the context-switching patterns, the way developers structure files, navigate repos, write comments, recover from errors. Every session is a training signal. Every user interaction compounds into something increasingly hard to replicate from scratch, even with a better underlying model.
This is the flywheel the dead wrappers never built. They captured user sessions as revenue events. The survivors captured them as data assets. One produces a subscription. The other produces a moat.
Replit took a different path to the same destination. It did not just build AI coding assistance on top of a browser IDE. It built the full development environment: deploying, hosting, managing infrastructure, running agents, and made that environment the home base for a new class of builder who had never opened a terminal before. The switching cost is not “I would miss the autocomplete.” It is “my entire deployment stack lives here and my team has three months of project history on this platform.” That is not a feature. That is a workspace. Workspaces are sticky in ways features never are.
Lovable went further still, targeting the non-developer entirely and building toward an outcome: a shipped product, not just written code. When your product is defined by what gets created rather than what gets typed, you are playing a different game than the AI coding assistant commodity market.
Foundation Models Coming
The foundation models will eventually ship a version of your product. This is not a theory. Jasper AI saw its valuation cut by 20% after ChatGPT launched competing content features. OpenAI eliminated four major wrapper categories within 18 months of identifying market demand. There is a credible argument that platform providers deliberately encourage wrapper proliferation as market research: letting thousands of companies experiment, then integrating the successful patterns natively and capturing the value while eliminating the intermediaries.
OpenAI acquiring Windsurf and trying to buy Cursor were not the actions of a company that believes application-layer startups are irrelevant. They were the actions of a company that realized the application layer creates distribution, behavioral data, and user habituation that pure model capability cannot replicate from the API layer. Google paying $2.4 billion for Windsurf’s talent and licensing rights confirmed the same fear from the other side.
The window for wrappers to build something defensible before the foundation model companies arrive is short. Short does not mean zero. It means you need to know what you are racing toward before the platform notices you are winning.
The Three Rules + The Named Category
The pattern in the survivors is consistent enough to be a framework.
They own the environment, not just a feature within it. Cursor owns the editor. Replit owns the deployment stack and the project. Lovable owns the creation context from idea to shipped app. None of them are plug-ins to a larger workflow. They are the workflow. Owning the environment means accumulating context that is genuinely costly to migrate, not just annoying to migrate.
They built a data asset the foundation models do not have. Cursor knows how professional software engineers work in production codebases. Replit knows how non-technical builders describe and iterate on products in natural language. This behavioral data makes their products better in ways that cannot be closed by OpenAI dropping a new model weight. It is not the intelligence that is proprietary. It is the understanding of the human on the other end.
They expanded the addressable market rather than fighting for existing users. Cursor did not try to take enterprise developers away from GitHub Copilot. It created a new kind of power user. Replit did not fight for the professional developer. It went after the twenty million people who had an app idea and no computer science degree.
The coding vertical got to these answers first because the feedback loop is fast. The category that has not sorted itself out yet is legal.
Harvey AI is the strongest candidate for the Cursor of legal: deep workflow integration, behavioral data on how lawyers actually research and draft, enterprise contracts with switching costs baked into the product. But the category is not settled. There are still dozens of “GPT for lawyers” wrappers that will die exactly the way the PDF chat companies died, the moment a foundation model ships a legal reasoning feature natively and their entire value proposition lives in a single JSON payload.
The wrapper that survives in legal will be the one that owns the workflow environment, not the one with the best document summarizer. It will know how partners at a specific firm actually structure arguments, where deals stall in diligence, which clause language their clients reject. That knowledge does not come from a better model. It comes from thousands of sessions inside the actual work. The markdown file is a snapshot. The environment is the asset.
The companies that escape the graveyard will not look different from the ones that do not, not at first. Same API calls. Same gradient backgrounds. Same pitch about transforming workflows.
The difference is invisible until it is not. One of them is building a behavioral dataset their users cannot migrate. One of them owns an environment their users live inside. One of them is expanding a market that did not exist before they arrived.
By the time it is obvious which is which, the acquisition offers are already on the table and the window is closed. The question is not whether to build a wrapper. It is whether you are building the right thing inside it.