Headline & intro
The party for “thin” AI startups is winding down. A senior Google executive is effectively telling founders that two of the most popular generative AI business models of the last two years – LLM wrappers and AI aggregators – probably won’t make it to Series C. That matters far beyond Sand Hill Road: it’s a signal that the easy money phase of the AI boom is ending and that infrastructure giants are tightening their grip. In this piece, we’ll unpack what Darren Mowry is really saying, who should be nervous, and where the next defensible opportunities in AI are likely to be.
The news in brief
According to reporting by TechCrunch, Darren Mowry, who leads Google’s global startup organization across Cloud, DeepMind and Alphabet, believes two types of AI startups are in trouble: basic LLM wrappers and AI aggregators.
LLM wrappers are products that mostly reskin existing models like GPT, Claude or Gemini with a user interface or light workflow layer. AI aggregators sit above multiple models, routing queries between them and adding monitoring, governance or evaluation tools.
Mowry argues that if a startup offers only a thin UX and relies entirely on third‑party models, investors and customers are losing patience. He points to examples like coding assistant Cursor or legal tool Harvey as wrappers that do have deeper moats. But as a category, he thinks both simple wrappers and aggregators face weak growth and margin pressure as model providers build more enterprise features themselves. Instead, he’s bullish on AI-native developer tools, consumer applications and data-heavy sectors like biotech and climate tech.
Why this matters
This is more than one executive giving portfolio advice. It’s a public statement from a hyperscaler that the “there’s an AI app for that” phase is over. A lot of founders have built startups that are, bluntly, feature ideas that probably belong inside Notion, Figma, Google Workspace or Microsoft 365 – not standalone companies.
Who benefits?
- Big model providers and cloud platforms: If wrappers and aggregators struggle, power consolidates with those who own models, GPUs and distribution. Google, OpenAI, Anthropic, Meta and a handful of cloud players become even harder to dislodge.
- Deep vertical players: Startups that combine models with proprietary data, domain expertise and integration into messy real‑world workflows suddenly look much better. Think AI in radiology, underwriting, industrial maintenance, compliance or logistics.
Who loses?
- “UI on top of GPT” startups that raised at 2023/24 valuations without meaningful IP, data or network effects. Many will quietly pivot to agency work or get acqui‑hired.
- Pure aggregators whose main value is API convenience. As Mowry notes, customers increasingly expect the platform itself to know when to route to which model; they don’t want to pay a premium for a middle layer that offers limited intelligence.
The immediate implication: generative AI is normalising into software, with all the usual expectations around unit economics, defensibility and switching costs. Hype alone no longer clears an investment committee.
The bigger picture: from gold rush to consolidation
We’ve seen this movie before. In the late 2000s, when AWS took off, a wave of startups sprang up to “simplify the cloud” – reselling compute, offering unified billing or dashboards. Most vanished once Amazon, Microsoft and Google built those capabilities into their own platforms. The survivors moved up the stack into security, compliance, DevOps automation or managed services.
The same pattern is now playing out in AI:
- APIs are commoditising. Model quality is converging for many generic tasks. Prices trend down, while open-source models get better. Owning the API gateway, by itself, is a shaky business.
- Platforms are integrating up the stack. OpenAI, Google, Anthropic and others keep shipping agents, orchestration tools, evaluation suites and governance capabilities. Every time they add a feature, a slice of the aggregator/wrapper market disappears.
- Distribution beats clever prompts. It’s easier for Microsoft to roll an AI feature into Office and reach hundreds of millions of users overnight than for a startup to convince those users to add yet another tool.
Historically, platform shifts create a brief window where point solutions thrive before being absorbed or outcompeted. Mobile apps, browser extensions, SaaS add‑ons – most categories saw a handful of durable independent winners, surrounded by a graveyard of one‑feature clones.
Mowry’s comments are essentially saying: in generative AI, we’re leaving that experimental phase. The opportunity now is not to be “the ChatGPT for X”, but to own a hard problem – with data, integrations and trust that hyperscalers can’t easily replicate.
The European angle: thin wrappers meet thick regulation
For European founders, this warning intersects with something even more structural: regulation and market structure. The EU AI Act, GDPR and the Digital Services Act all push companies toward strong governance, data protection and explainability. That is expensive to build – and nearly impossible to do well if your startup is just a skin on top of someone else’s model.
European buyers – especially in Germany, France, the Nordics and the DACH region – are already privacy‑sensitive and procurement‑heavy. They will ask hard questions about where data goes, how models are evaluated and what happens if the upstream provider changes terms. A simple wrapper has very little room to manoeuvre here.
On the flip side, Europe has genuine advantages in vertical AI:
- Deep industrial bases (automotive, manufacturing, energy, logistics)
- Strong healthcare and public sector institutions
- A growing open‑source AI ecosystem (from Stability AI to Aleph Alpha and many university labs)
Founders in Berlin, Paris, Barcelona, Ljubljana, Zagreb or Tallinn who combine proprietary European datasets with compliance‑by‑design can build moats that a US‑centric GPT skin simply cannot. The same applies to startups serving regulated sectors across the continent or exporting trustworthy AI into Latin America and Africa.
The message for European ecosystems: don’t chase Silicon Valley’s 2023 wrapper playbook. Use regulation and domain depth as a weapon, not a constraint.
Looking ahead: what to watch in the next 24 months
A few things are likely over the next two years:
- Silent consolidation of wrapper startups. Many “AI copilot for X” products will become features inside incumbents. Expect acqui‑hires, asset sales and founders quietly rebranding into consulting.
- Aggregators morph into platforms or die. The ones that survive will not sell “access to many models”, but rather decision engines, evaluation frameworks and governance layers tuned to specific industries (finance, healthcare, defence, critical infrastructure).
- Model providers will keep climbing the stack. Expect more end‑to‑end vertical offerings from OpenAI, Google, Anthropic and others – from call‑centre solutions to document processing pipelines.
- Data moats become the new GPU flex. Startups that control unique, hard‑to‑recreate datasets – labelled medical images, industrial telemetry, legal corpora, proprietary support logs – will be able to negotiate better terms with model providers or fine‑tune their own.
For founders and investors, key questions now are:
- Can we articulate our moat without mentioning “we use GPT‑X/Gemini/Claude” in the first sentence?
- If our upstream provider doubled prices or launched a direct competitor, would we still have a business?
- Do we own a workflow, a community, a dataset or a trust relationship that is painful to dislodge?
Those who answer “yes” have a fighting chance. The rest are playing on borrowed time.
The bottom line
Mowry’s warning is not the death of AI startups; it’s the end of a phase where merely wiring a chat box to a large language model counted as innovation. Power is shifting back to those who own models, data, distribution or deep domain integration. For everyone else, this is the moment to decide: evolve from wrapper to real product company, or accept that you’re building a feature for someone else’s platform. The only wrong move now is pretending the check‑engine light isn’t blinking.
