Google VP Says Two Types of AI Startups Are Running Out of Time

Darren Mowry, the vice president who runs Google's global startup organization, told TechCrunch this week that two categories of AI companies have their "check engine light" on. The categories: LLM wrappers and AI aggregators. His advice to founders still building in either lane was blunt - if you cannot point to a moat, the market is about to find out.

This is not a prediction. It is a post-mortem dressed as a warning.

The Numbers Tell the Story

Metric	2024	2026
AI startups raising $100M+ (US, first 49 days of year)	4	17
Cursor valuation	~$400M (est.)	$29.3B
Harvey AI valuation	~$700M	$11B (in talks)
AI wrapper shutdowns/acqui-hires	Rare	Accelerating
OpenAI API pricing (GPT-4 class output)	$60/M tokens	Under $15/M tokens

The paradox is stark. More money is flowing into AI than ever, but the distribution is violently uneven. A handful of vertical specialists are vacuuming up capital at nosebleed valuations while a much larger cohort of undifferentiated wrappers is quietly shutting down or selling for parts.

What Mowry Actually Said

"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."

The Google VP drew a sharp line between two failure modes. First, the thin wrappers - startups that layer a product interface over an existing LLM like GPT-5 or Gemini without adding proprietary value. Second, the aggregators - platforms that route queries across multiple models through a single API or interface, including companies like Perplexity and OpenRouter.

For aggregators, Mowry noted that users now demand "some intellectual property built in" to guide queries toward appropriate models. Simply offering a dropdown menu of LLMs is no longer a product.

His prescription for survival: "deep, wide moats that are either horizontally differentiated or something really specific to a vertical market." Wrapping "very thin intellectual property around Gemini or GPT-5" is, in his words, a sign that a startup is not distinguishing itself.

The Wrapper Tax

The economics are brutal. Consider Cursor, the AI coding assistant now valued at $29.3 billion after raising $2.3 billion. It crossed $1 billion in annualized revenue in November 2025, a five-fold increase in six months. By every growth metric, it is a rocketship.

But Cursor does not build the AI models that power its product. It pays OpenAI, Anthropic, and Google for API access, then resells that access inside a polished editing interface. A single coding session generates hundreds of model calls. At the scale Cursor operates - billions of lines of code generated daily - token usage runs into the hundreds of billions monthly. Even with volume discounts, margins get thin fast.

Worse, Cursor's suppliers are also its competitors. OpenAI, Anthropic, and Google each operate or are building competing coding tools. Model providers increasingly prioritize direct developer relationships over wholesale partnerships. The supplier that sells you tokens today can undercut you tomorrow.

The Vertical Survivors

Not every wrapper is doomed. Harvey, the legal AI startup, is reportedly raising another $200 million at an $11 billion valuation. The difference: Harvey trains custom models on law firms' proprietary documents while maintaining the security and compliance frameworks that legal clients demand. That domain-specific integration creates genuine switching costs.

Mowry pointed to companies like Veeva Systems in life sciences and Palantir in government contracting as the model. Both built moats through deep domain knowledge and entrenched customer relationships, not through the novelty of the underlying technology.

Counter-Argument

There is a case that the wrapper shakeout is healthy, not catastrophic. Seventeen US-based AI startups raised $100 million or more in the first seven weeks of 2026 alone. Cursor's revenue trajectory suggests that execution speed and user experience can create durable advantages even when the underlying technology is commoditized. The company's CEO argues that its rapid innovation cycle, fueled by direct user data, constitutes a moat of its own.

And not every aggregator is stalling. Some are evolving into orchestration platforms with genuine evaluation, monitoring, and governance capabilities - features that enterprises need and that raw model APIs do not provide. The best AI agent frameworks are increasingly built to work across multiple model providers, suggesting that the multi-model future creates real infrastructure opportunities.

The counter-case rests on a bet: that the application layer can move faster than the model layer. If Cursor ships features weekly while OpenAI ships models quarterly, the wrapper stays ahead. If model providers start shipping features weekly too, the wrapper has a problem.

What the Market Is Missing

Mowry's warning carries an irony that nobody in the room wanted to name. Google - the company whose VP is advising startups to build deeper moats - is also the company most aggressively absorbing wrapper functionality into its own products. Google's AI Mode now reaches over 75 million daily users. Every feature a wrapper startup builds on top of Gemini is a feature Google can replicate for free and distribute to its existing user base overnight.

The same dynamic applies across Big Tech. Microsoft has Copilot embedded in every Office product. Amazon is building AI directly into AWS. The open-source vs. proprietary debate matters less when the real question is whether any independent application layer can survive between the model providers above and the platform distributors below.

The AI startup landscape is bifurcating into two tracks. Track one: vertical specialists with deep domain moats, proprietary data flywheels, and switching costs that survive model upgrades. Track two: everything else. The capital is flowing to track one. The obituaries are being written for track two. The question for founders is not whether their wrapper adds value today - it is whether it will add value after the next model release makes last quarter's differentiation table stakes.

Mowry's check engine light is not flashing for the industry. It is flashing for the companies that mistook proximity to AI for possession of it.

Sources: