In an interview on TechCrunch’s Equity podcast, Darren Mowry, who runs Google’s global startup effort across Cloud, DeepMind and Alphabet, delivered one of the starkest warnings yet to founders riding the generative‑AI wave. After two years in which investors funded almost any product with “GPT” or “AI copilot” in the deck, he says the market is now brutally reassessing what actually counts as a business.
Two categories in particular, he argues, are looking less like the future of AI and more like cautionary tales:
- LLM wrappers – apps that put a thin product or UX layer on top of foundation models like GPT, Claude or Gemini.
- AI aggregators – tools that bundle multiple LLMs behind a single interface or API, routing queries between them.
“Startups with these hooks have their check engine light on,” Mowry said, adding that many will struggle to grow or raise fresh capital unless they radically deepen their technology and moats.
What are LLM wrappers – and why are they in trouble?
LLM wrappers were the first big crop of generative‑AI startups: products that call someone else’s large language model in the background and add a bit of UI, prompts and workflow on top. A typical example, Mowry said, might be a study‑helper app that uses GPT or Gemini to generate answers for students, or a sales‑email generator that mostly forwards user input to a foundation model.
The problem is simple: they don’t own the core technology.
“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.
As base models get better, cheaper and more multi‑purpose, the “magic” of many wrappers becomes a commodity feature. What looked like a clever AI product six months ago may now be a checkbox inside GPT‑5, Gemini or Claude or a free plugin inside Microsoft 365 and Google Workspace.
Mowry calls this “very thin intellectual property”: most of the value sits with the model provider, not the startup. When OpenAI or Google roll out a new feature, the wrapper’s differentiation evaporates overnight, and any margin they had is squeezed by falling API prices and rising competition.
‘Wrappers’ that might survive
Not all wrappers are doomed. Mowry points to examples like Cursor, a code‑editing environment built around AI pair‑programming, and Harvey, a legal‑AI assistant, as wrappers that have real advantages.
They work because they don’t stop at calling an LLM:
- They go deep into a specific vertical (software engineering, law).
- They build proprietary tooling, integrations, data and workflows around that domain.
- They embed AI into the core of the product rather than bolting on a chat box.
For these firms, the model is just one part of a larger system that’s hard to copy quickly. Mowry’s message to founders is that survival depends on creating broad or deep moats unique data, domain expertise, distribution, network effects or infrastructure not just clever prompts.
AI aggregators: ‘stay out of the business’
If wrappers are risky, Mowry is even harsher on AI aggregators, calling them one of the least attractive spaces for new founders.
Aggregators are a subset of wrappers: platforms that sit on top of multiple LLMs and offer a unified interface or API so users can send queries to different models from one place. They often layer in monitoring, governance and evaluation tools, promising to help enterprises choose “the right model for the job” without locking into a single provider.
Think AI search and Q&A products, or developer platforms that expose dozens of models behind a single SDK. Some, like Perplexity and OpenRouter, have gained real traction but Mowry cautions that new entrants face brutal odds.
His advice is blunt:
“Stay out of the aggregator business.”
Two forces are working against these startups:
- Limited growth and progression
“Generally speaking, aggregators aren’t seeing much growth or progression these days,” Mowry said, noting that usage and revenue are flattening for many players that don’t have a clear niche or proprietary layer. - Customers want IP, not just routing
Users increasingly expect intelligence and IP baked into the platform, domain‑specific models, proprietary ranking signals, bespoke fine‑tuning, not just a router that shuttles API calls between vendors. “People want to know they’re being routed to the right model for their needs, not because of behind‑the‑scenes compute or access constraints,” he added.
As cloud providers and big platforms roll out their own orchestration tools, the “just aggregate models” pitch looks more and more like a feature, not a company.
The ‘second wave’ of AI: build something only AI can make
Mowry’s warning lands as investors and founders talk increasingly about a “second wave” of AI startups: companies using generative models to create new products and experiences, not just to cut costs or re‑skin existing workflows.
First‑wave startups mostly automated tasks drafting emails, summarising calls, generating boilerplate code. That made incumbents more efficient but rarely produced defensible new businesses. The next wave, argues Inworld CEO Kylan Gibbs and others, will be defined by products that couldn’t exist before LLMs: new types of games, interactive media, companions, tools and services that people will pay for in their own right.
Mowry is effectively telling founders to choose a side:
- Either be the infrastructure (build models, chips, tooling, or deep orchestration with IP), or
- Own the product and the user, with AI as a core ingredient in something genuinely new.
Anything in the mushy middle, a light wrapper on someone else’s model, or a generic aggregator, is likely to be crushed between falling API prices, fast‑moving incumbents and platform features.
What should founders and investors do now?
For startups already in the danger zone, the message is not necessarily “shut down”, but “evolve fast”:
- Deepen your moat: collect proprietary data, build domain‑specific models, or create workflows that are hard to copy.
- Pick a vertical: generic AI helpers are easiest to clone; niche, regulated or complex domains are harder.
- Move up the stack: own more of the experience and technology instead of living as a thin layer on top of someone else’s model.
For investors, Mowry’s comments are a clear signal to be more sceptical of pitch decks that boil down to “we call GPT and add a UI”. The era of funding every wrapper is over; the bar is now real IP, real users and real differentiation.
The generative‑AI boom may have minted a startup a minute, but as Google’s startup chief bluntly puts it, not all of those ideas were built to last.