The generative AI boom has created a wave of new startups almost overnight. But as the market matures, not every business model built on large language models is proving sustainable.

According to Darren Mowry, who leads Google’s global startup organization across Cloud, DeepMind, and Alphabet, two categories of AI startups are flashing early warning signs: LLM wrappers and AI aggregators.

His message is clear. If your startup relies too heavily on someone else’s model, you may not have a business that lasts.

The Problem With Thin LLM Wrappers

LLM wrappers are startups that build a user interface or product layer on top of existing large language models such as OpenAI’s GPT, Anthropic’s Claude, or Google DeepMind’s Gemini.

In practice, this might look like:

An AI study assistant for students
A content-generation dashboard for marketers
A lightweight productivity tool powered entirely by a third-party model

The issue? If the backend model does all the heavy lifting and the startup adds only a thin UX layer, there is little defensible value.

As Mowry puts it, simply “white-labeling” a model is no longer enough. The market has moved past the phase where adding a clean interface on top of GPT could generate traction by itself.

In 2024, when ChatGPT launched its GPT Store, opportunities for lightweight wrappers seemed endless. In 2026, investors and enterprise buyers are asking a different question:

What is your moat?

Startups need:

Deep domain expertise
Proprietary data
Workflow integration
Vertical specialization
Or strong horizontal differentiation

Examples of more defensible AI-powered products include:

Cursor, which deeply integrates AI into developer workflows
Harvey AI, built specifically for legal professionals

The takeaway for founders: if your value disappears the moment a model provider ships the same feature, your differentiation is too thin.

Why AI Aggregators Are Under Pressure

AI aggregators are a subset of wrappers. These startups combine multiple LLMs into a single interface or API layer, routing queries between models and often offering orchestration tools such as monitoring, governance, or evaluation layers.

Examples include:

Perplexity AI, an AI-powered search engine
OpenRouter, which provides access to multiple AI models through a single API

While many of these platforms gained rapid adoption, Mowry’s advice to new founders is blunt:

“Stay out of the aggregator business.”

Why?

Because model providers themselves are increasingly building:

Enterprise tooling
Governance layers
Routing intelligence
Optimization features

As foundation model companies expand vertically into enterprise features, they squeeze the margins of middle-layer aggregators.

Customers also expect more than simple routing logic. They want embedded intelligence that ensures the right model is used at the right time based on business context, not just cost or compute constraints.

Without proprietary IP or workflow ownership, aggregators risk becoming replaceable.

A Familiar Pattern From Cloud’s Early Days

Mowry sees a clear historical parallel.

In the early days of cloud computing, startups emerged to resell Amazon Web Services infrastructure. They positioned themselves as easier entry points with consolidated billing, support, and tooling.

But once AWS built robust enterprise features and customers became more comfortable managing cloud services directly, many of those resellers disappeared.

The only survivors were companies that added real value:

Security services
Migration expertise
DevOps consulting
Managed operations

The same consolidation pressure is now playing out in AI.

Where Mowry Sees Real Opportunity

Despite his warnings, Mowry remains optimistic about several AI-driven sectors.

1. Developer Platforms and “Vibe Coding”

Startups focused on developer productivity had a record-breaking 2025. Companies like:

Replit
Lovable
Cursor

have attracted major investment and enterprise adoption.

These platforms embed AI directly into core workflows instead of simply wrapping a model with a UI.

2. Direct-to-Consumer AI Tools

Mowry also expects strong growth in consumer-facing applications that put powerful AI tools directly into users’ hands.

For example, film and TV students can use Veo, Google’s AI video generator, to bring creative ideas to life without traditional production budgets.

The opportunity lies in enabling entirely new capabilities, not just automating existing tasks.

3. Biotech and Climate Tech

Beyond pure AI infrastructure, Mowry highlights biotech and climate tech as sectors benefiting from unprecedented access to data and AI-driven analysis.

With massive datasets and improving model performance, startups in these verticals can create value in ways that were not technically feasible a decade ago.

Control F5 Perspective

For founders and CTOs, the message is strategic:

AI is no longer a novelty layer.
Foundation models are infrastructure.
Sustainable value comes from integration, ownership, and domain depth.

The next wave of successful AI startups will not be defined by access to GPT-level models. They will be defined by how intelligently those models are embedded into real workflows, proprietary systems, and industry-specific problems.

In other words, the era of “AI on top” is fading.

The era of “AI built in” has already begun.

Source