“No patience for that anymore,”: Google VP flags sustainability concerns for two AI startup categories

The generative AI surge has spawned countless startups, but two once-popular approaches are hitting limits, according to Google VP and global head of startups Darren Mowry. He argues that thin large language model (LLM) wrappers and model aggregators are struggling to differentiate and scale as the market matures. “If you’re really just counting on the back-end model to do all the work…the industry doesn’t have a lot of patience for that anymore,” he said, adding that wrapping “very thin intellectual property around Gemini or GPT-5” is no longer seen as true product differentiation.

Why thin wrappers are stalling

LLM wrappers—apps that primarily layer a user interface or simple workflow on top of foundation models—gained fast traction during the early phase of the boom. Tools that helped students study, draft emails, or automate routine tasks were especially common following the proliferation of model marketplaces and app stores in 2024. But momentum has shifted. Wrappers without deep domain logic, proprietary data, or unique workflows are finding it hard to retain users and defend margins as model providers improve their own features.

Mowry’s view is not that wrappers are doomed; it’s that shallow wrappers are. Products that marry model capabilities with hard-won expertise, specialized datasets, or compliance and workflow depth can still build a moat. He cited examples like Cursor, a coding assistant, and Harvey AI, a legal tool—both “wrapper-like” experiences that embed substantive domain knowledge and developer or professional workflows. In other words, the bar for value creation has risen from “access to the model” to “product plus model plus moat.”

Aggregators face mounting headwinds

AI aggregators—platforms that bundle multiple LLMs behind one interface or API and promise smarter routing, monitoring, or governance—are encountering similar pressure. While these orchestration layers once offered convenience and optionality, customers are increasingly demanding proprietary capabilities, not just a switchboard. Examples in this category include services that dynamically select among models for cost, speed, or accuracy, and provide analytics to oversee usage.

Mowry cautioned founders to “stay out of the aggregator business” unless they can deliver real, defensible IP beyond routing. The reason: as model providers expand enterprise features—security, compliance, monitoring, cost controls, and tooling—the value of independent orchestration can be compressed. That dynamic mirrors an earlier era in cloud computing. In the late 2000s, companies that simply resold cloud infrastructure were squeezed when hyperscalers introduced robust native services. The ones that thrived delivered true services—security, migration, DevOps—not just pass-through infrastructure. Aggregators may face the same margin and differentiation squeeze as models evolve and integrate deeper enterprise controls.

What still looks promising

Despite the cautionary notes, Mowry remains bullish on startups that build deep moats and solve concrete problems. He sees particular strength in:

• Developer platforms that accelerate software creation, testing, and maintenance, where performance and workflow fit matter more than model brand. Companies like Replit, Lovable, and Cursor illustrate momentum in tools that meet developers where they work.
• Direct-to-consumer apps that pair AI with sticky experiences, strong communities, or creator ecosystems—where retention comes from more than raw model output.
• Creator tools and media generation, including video, where new capabilities are unlocking novel formats and enabling solo creators to do studio-grade work. Google’s own advances in video generation, such as Veo, were highlighted as opening new creative frontiers.
• Beyond AI, biotech and climate tech, where massive datasets and simulation capabilities are enabling breakthroughs in discovery, optimization, and deployment.

The new bar for defensibility

Across categories, the message is consistent: sustainable AI startups must own more than an interface. Differentiation is moving toward:

• Proprietary data and feedback loops that improve model performance over time
• Domain-specific logic, compliance, and workflows that are hard to copy
• High-switching-cost integrations and ecosystems (plugins, extensions, marketplaces)
• Operational excellence—reliability, observability, total cost of ownership, and security

As foundation models get faster, cheaper, and more capable, value accrues to companies that transform those capabilities into outcomes users can’t easily replicate. Thin wrappers and generic aggregators won’t clear that bar. The opportunity is still vast—but the market is signaling it expects real IP, real services, and real impact.