Darren Mowry runs Google’s global startup organization across Google Cloud, DeepMind, and Alphabet. He went on TechCrunch’s Equity podcast on February 21st and said something that should make a lot of founders nervous. Two types of AI startups, he argued, are essentially walking dead: LLM wrappers and AI aggregators.

And that’s a hell of a lot of companies.

The wrapper problem is not new, but the patience is gone

LLM wrappers build a thin UX layer on top of someone else’s model. Study tools, writing assistants, customer service bots that route everything to GPT or Claude and add a coat of paint. Mowry was blunt about it: “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.”

And he drew a parallel to the early cloud computing era, when startups would resell AWS infrastructure with a nicer dashboard and call it a product. Amazon built its own enterprise tools, customers got sophisticated, those resellers got squeezed out. But the AI version plays out faster and meaner, because at least cloud resellers controlled their own infrastructure. An LLM wrapper does not even have that. And when OpenAI ships a better interface or Anthropic adds the one feature the wrapper offered as its differentiator, the wrapper company’s entire value proposition evaporates overnight. And there is no moat. There was never one.

”Stay out of the aggregator business”

His direct advice. Aggregators, the companies stitching multiple LLMs into one routing interface, are not seeing growth. He named Perplexity and OpenRouter as examples.

So the aggregation layer alone is not a product.

What survives

Mowry listed what he thinks works: proprietary data, deep domain expertise, vertical integration, developer platforms. And he called out Replit, Lovable, and Cursor as companies building compounding value that is hard to clone.

But I keep coming back to one thread running through all those survivors. They do not sit between you and a model. They embed into something you already do. Cursor lives inside your code editor. Replit is the development environment. These tools became the workflow instead of bolting onto it from the outside, and that distinction turns out to be the only thing that keeps you alive when the underlying models get better every quarter.

Because most of us do not write code for a living. But we all live in a browser.

Most AI productivity tools still sit outside your actual work

The average knowledge worker’s relationship with AI right now is a copy-paste loop between a chat window and whatever browser tab actually contains their work, whether that’s Gmail, LinkedIn, a CRM, or an internal portal. McKinsey’s late 2025 survey found the median worker saves less than 30 minutes per week with AI tools. That number is pathetically small given what these models can actually do, and the reason it stays small is exactly the wrapper problem Mowry described, just applied to the user’s daily experience rather than a startup’s cap table. The tool lives outside the work. So most of the effort goes into shuttling context back and forth, not getting anything done.

Browser agents skip that entire loop. They live inside the browser, see the page you’re on, act on it directly. dassi does this as a Chrome extension sitting in your side panel, handling drafting, research, and repetitive tasks right where you’re already working. No separate tab. No copy-paste relay.

BYOK is the anti-wrapper

The other half of what Mowry said stuck with me for a different reason. If wrappers die because they depend on a single model they do not control, then the opposite survival strategy writes itself: let users bring their own model. BYOK browser agents are structurally the inversion of everything he warned about. No dependency on one provider’s API staying cheap. And no scramble to rebuild your workflow when a company pivots its pricing or deprecates an endpoint.

When Gemini 3.1 Pro dropped last week, dassi users could switch to it the same afternoon. Or switch back if it didn’t fit their tasks. Because the tool does not care which model you run. It cares about the work.

Mowry was talking about startups. But the logic applies to every AI tool you pick up — if it locks you into one model and one company’s roadmap, you are betting your productivity on someone else’s survival. And a Google VP just told you that bet keeps getting worse.

I’ve got three different API keys in my browser right now and I am not particularly loyal to any of them. Which, if Mowry is right, might be the only sane way to operate.