On February 21, Darren Mowry—head of Google’s global startup organization—told TechCrunch that two AI business models have their “check engine light” on: LLM wrappers and AI aggregators. The generative AI boom minted startups by the thousands, but Mowry’s warning cuts straight to the core problem: these companies don’t control the technology they depend on, and as base models improve, the wrapper value evaporates. This isn’t speculation. We’ve watched this exact pattern play out before with AWS resellers in the 2010s, and it ended the same way every time—platform providers squeezed out the middleware, and only companies with genuine intellectual property survived.
The Wrapper Problem: No Moat, No Survival
LLM wrappers are startups that layer a product or UX on top of existing large language models like GPT, Claude, or Gemini. An email-writing tool calling the GPT-4 API is a wrapper if most of its value comes from that API call. AI aggregators take this a step further—they combine multiple LLMs in one interface or route queries across models. Think Perplexity for search or OpenRouter for developer access to multiple models. The pitch sounds reasonable: we make powerful AI accessible and easy to use. The problem is economic and structural.
As base models get smarter, features that differentiated a wrapper product six months ago become standard capabilities baked into GPT-5 or Gemini 2.0. The value proposition shrinks with every model update. Wrappers don’t control their core technology—they’re entirely dependent on external APIs they don’t own. Platform providers can change pricing, restrict access, or simply absorb the wrapper’s key feature into the base product. Mowry put it bluntly: aggregators “aren’t seeing much growth because users want some intellectual property built in”—not just behind-the-scenes compute or multi-model access. Users don’t need middleware when they can go direct to the source.
This creates a fatal dynamic: margin compression. Wrappers can’t control input costs because API pricing is set by platform providers. OpenAI raised GPT-4 API prices three times in 2025. Wrappers must either absorb those costs or pass them to customers, but they’re competing with free or cheap alternatives from the platform itself. Pricing power collapses. Differentiation erodes. The business model implodes.
We’ve Seen This Movie: The AWS Reseller Squeeze
Mowry didn’t just warn about wrapper risk—he cited the historical precedent that proves it’s not just possible but inevitable. Between 2008 and 2012, AWS’s cloud infrastructure boom spawned a wave of reseller startups. These companies offered easier AWS entry points, billing consolidation, tooling, and support. Their value prop was simple: AWS is powerful but complicated, we make it simple. Sound familiar?
Amazon’s response was predictable. They built enterprise tools, improved customer support, and made direct AWS usage accessible to non-experts. Resellers got squeezed out. Customers realized they didn’t need a middleman when AWS could provide the same services directly, often cheaper and more integrated. The resellers who survived weren’t pure resellers—they had proprietary technology, specialized vertical solutions, or unique IP that AWS couldn’t easily replicate. The rest disappeared.
The parallel to AI wrappers is exact. OpenAI, Google, and Anthropic are AWS. Wrapper startups are the resellers. As platform providers improve their interfaces, add multi-model support, or build vertical-specific solutions, the wrapper’s value proposition vanishes. Platform providers have deeper pockets, control the underlying technology, and can move faster because they own the stack. If you’re a wrapper, you’re not just competing with other startups—you’re racing against the platform provider’s roadmap. And they always win that race.
Who Survives: The IP Moat Test
Not all AI startups are doomed. Industry analysis predicts that 80% of wrapper startups will disappear by the end of 2026—ten months from now. But that means 20% survive, and the difference is clear: companies with defensible intellectual property moats make it, those without don’t.
Anthropic, valued near $350 billion, isn’t a wrapper—they build proprietary models with unique safety research and enterprise focus. Perplexity, valued at $20 billion, isn’t just aggregating models—they’ve built search-specific IP with citations, sources, and query optimization that generic LLMs don’t replicate. These companies own their core technology or have proprietary data assets, vertical specialization, or post-training techniques that create genuine differentiation.
Generic wrappers fail the moat test. If your startup’s core value is “we make ChatGPT easier to use,” you don’t have a startup—you have a feature request. And OpenAI is reading your feature requests. The six-month test separates survivors from the doomed: can OpenAI, Google, or Anthropic build your product’s core value in six months? If yes, you’re a wrapper, and your runway is shorter than you think. If no—because you have proprietary data, specialized models, or domain expertise they can’t quickly replicate—you might survive.
Healthcare AI with HIPAA compliance and proprietary medical training data passes the test. Legal AI with exclusive case law analysis and regulatory expertise passes. A prettier ChatGPT interface does not.
Platform Risk Is Not Theoretical
Platform risk—the vulnerability from depending on a third-party platform for core business functions—isn’t abstract. Developers and founders have watched it destroy companies in real time. Twitter’s API shutdown between 2012 and 2023 killed thousands of Twitter client apps and analytics tools overnight. Facebook’s 2018 API restrictions after Cambridge Analytica wiped out social startups. Myspace’s collapse made every app built on its platform worthless instantly.
AI wrappers face the same structural risk. One API pricing change, one terms-of-service update, one platform decision to build competing features, and the wrapper’s business model collapses. Wrappers can’t control costs, can’t dictate access, and can’t prevent the platform from absorbing their differentiators. The old developer wisdom applies: don’t build your castle on someone else’s land. When the landowner’s interests diverge from yours, you’re evicted.
What Developers and Founders Should Know
For developers evaluating job offers, wrapper startups are high-risk bets. Ask the six-month question in interviews: what prevents OpenAI from building this feature themselves? If the answer is vague or relies on “better UX,” that’s a red flag. Talent is already gravitating to platform providers like OpenAI, Anthropic, and Google, or to specialized AI companies with clear IP moats.
For founders building AI startups, the lesson is brutal but clear: build genuine intellectual property or prepare to die. Proprietary data assets, vertical-specific solutions, post-training specialization, and measurable customer outcomes that compound over time are the only defensible moats. UX differentiation alone won’t save you when the platform improves its own interface. Own the connection to your customers—don’t rely entirely on the platform for access, distribution, or value.
For investors, VC write-offs are coming. Billions poured into wrapper and aggregator startups during the 2023-2025 AI boom are now at risk. The smart money is demanding pivots or writing off positions. Future AI investments will require proprietary technology, IP moats, and clear answers to the question: what prevents the platform from copying this in six months?
The Pattern Repeats
The AWS reseller squeeze was predictable. Twitter’s API shutdown was predictable. The AI wrapper extinction is equally predictable. Platform providers always consolidate value when they can, and middleware without defensible IP always gets squeezed out. Mowry’s warning isn’t opinion—it’s pattern recognition backed by industry history. Thousands of AI startups face a reckoning in the next ten months. The survivors won’t be the ones with the prettiest interfaces or the cleverest aggregation logic. They’ll be the ones who built something the platform providers can’t easily replicate.
If you’re building on someone else’s AI models without genuine IP, you’re not racing other startups. You’re racing the platform’s roadmap. And history shows how that race ends.
