The generative AI gold rush is officially cooling down. For about eighteen months, you could launch a startup by basically wrapping OpenAI’s latest model in a slick interface, slap some marketing copy about “AI-powered” this or that, and investors would line up. Those days are done.
Darren Mowry, who runs Google’s global startup organization, just hit the reset button on everyone’s expectations. And his message is blunt: if your entire company is built on wrapping Claude or GPT with a UI layer, your check engine light is flashing red.
The Wrapper Problem
Here’s the thing about LLM wrappers. They’re easy to build. You take an existing large language model, bolt on some interface that solves a specific problem like helping students study, and boom, you’ve got a startup. The barrier to entry is essentially just API keys and some decent product design.
The problem? Everyone did this. Like, everyone. And now the market’s patience for it has evaporated.
“If you’re really just counting on the back end model to do all the work and you’re almost white-labeling that model, the industry doesn’t have a lot of patience for that anymore,” Mowry said on a recent episode of Equity. “You’ve got to have deep, wide moats that are either horizontally differentiated or something really specific to a vertical market.”
The winners in this space did something different. Cursor, a GPT-powered coding assistant, actually built proprietary technology around how the model gets used. Harvey AI, a legal-focused assistant, went deep into a specific vertical with real legal expertise baked in. They didn’t just resell the model; they added genuine intellectual property on top.
Compare that to the thousand other AI coding assistants or legal tools that launched and subsequently disappeared. There’s no room in the market for thin value propositions anymore.
The Aggregator Trap
Then there’s the aggregator business. This is where startups like Perplexity and OpenRouter fit in. They basically act as middlemen, routing queries across multiple models, giving users access to everything from GPT to Claude to Gemini through one interface or API.
On the surface, this sounds smart. Users get choice. Developers get flexibility. But Mowry’s advice is direct: “Stay out of the aggregator business.”
The fundamental problem is that aggregators aren’t actually adding value in a way that justifies their existence. Users don’t care which model is processing their query behind the scenes. What they care about is getting the right answer. And if an aggregator is just mechanically routing to whichever model might work best based on generic criteria, well, that’s not a service worth paying for.
The real money comes from companies that build proprietary IP to route queries intelligently based on specific use cases. That requires actual domain expertise, not just API orchestration.
A Familiar Echo
Mowry has been in the technology infrastructure game for decades. He spent time at AWS and Microsoft before landing at Google Cloud, so he’s seen this movie before. And it’s playing out the same way again.
In the early days of cloud computing, around 2008 or 2009, startups emerged to resell AWS infrastructure. They marketed themselves as easier entry points, offering billing consolidation, better support, and friendlier tooling. For a hot minute, that worked. But then Amazon built those features themselves. Customers learned to manage AWS directly. And suddenly those middlemen had nothing to offer.
Most of them died. The survivors were the ones who actually added services like security consulting, migration expertise, or DevOps support. They became service companies, not resellers.
History’s repeating with AI aggregators. Model providers are expanding into enterprise features. They’re building better monitoring, governance, and routing logic. The middlemen are getting squeezed.
What Actually Works
So if wrappers and aggregators are heading toward extinction, what survives?
Developer platforms are on fire right now. Replit, Lovable, Cursor all had breakout years in 2025. These companies aren’t just wrapping models; they’re building entire ecosystems around how developers work. There’s real differentiation there.
Vibe coding tools are another winner. AI video generation for filmmakers. Direct-to-consumer applications that put powerful tools in people’s hands. That’s where the growth is.
Beyond AI, Mowry also sees real momentum in biotech and climate tech. These aren’t sexy startup categories in the same way, but both have access to “incredible amounts of data” that can create genuine value. That’s the differentiator.
The Uncomfortable Truth
The uncomfortable truth for thousands of AI startups launched in the last year is that speed to market stopped mattering. Being first to build an AI version of something used to be enough. Now it’s table stakes.
What actually matters now is whether you built something defensible. Whether you solved a real problem in a way that competitors can’t easily replicate. Whether you added your own layer of value instead of just plugging into someone else’s API.
For most of the AI startups that launched by wrapping GPT or aggregating models, the answer is probably no.
So what happens to all of them? Some will get acquired. Some will find actual differentiation and survive. Most will quietly shut down and the founders will move on to the next thing. That’s how markets clear, and that’s what the AI market is doing right now.
The question for founders launching today isn’t whether your idea uses AI. It’s whether your idea would work without it.
