Blog

While the generative AI boom rapidly gave rise to numerous startups, as the initial excitement subsides, two once-promising business models—LLM wrappers and AI aggregators—are increasingly viewed as cautionary examples.

Darren Mowry, who oversees Google’s global startup operations across Cloud, DeepMind, and Alphabet, indicates that startups relying on these approaches are exhibiting warning signs.

LLM wrappers are essentially companies that build a product or user experience layer around existing large language models such as Claude, GPT, or Gemini to address a specific problem. An example might be a startup that leverages AI to support student learning.

“If your strategy heavily relies on the backend model to perform most of the work, almost white-labeling it, the market’s tolerance for such an approach is diminishing,” Mowry stated during this week’s Equity podcast episode.

Mowry argues that building on “very thin intellectual property around Gemini or GPT-5” signals a lack of differentiation.

He emphasized that for a startup to “progress and grow,” it must possess “deep, wide moats that are either horizontally differentiated or highly specific to a vertical market.” Examples of robust LLM wrappers include Cursor, a GPT-powered coding assistant, and Harvey AI, a legal AI assistant.

In essence, startups can no longer simply add a user interface on top of a GPT model and expect to gain product traction, as might have been possible in mid-2024 when OpenAI launched its ChatGPT store. The current imperative is to establish enduring product value.

AI aggregators represent a subcategory of wrappers—these startups consolidate multiple LLMs into a single interface or API layer to route queries across models, granting users access to various models. Such companies typically offer an orchestration layer that includes monitoring, governance, or evaluation tools. Consider AI search platform Perplexity or developer platform OpenRouter, which provides access to multiple AI models through a unified API.

While many of these platforms have achieved some success, Mowry’s message to new startups is unambiguous: “Avoid the aggregator business.”

Generally speaking, aggregators are currently experiencing limited growth and advancement. Mowry explains that users now desire “some intellectual property built in” to ensure they are directed to the most suitable model at the right time, based on their needs, rather than due to underlying compute or access limitations.

With decades of experience in the cloud sector, having worked at AWS and Microsoft before joining Google Cloud, Mowry has witnessed similar cycles. He noted that the present situation mirrors the early days of cloud computing in the late 2000s and early 2010s, as Amazon’s cloud service began its ascent.

During that period, numerous startups emerged to resell AWS infrastructure, promoting themselves as easier entry points that offered tooling, consolidated billing, and support. However, once Amazon developed its own enterprise tools and customers became adept at managing cloud services directly, most of these reseller startups were phased out. The only ones that survived were those that incorporated genuine value-added services, such as security, migration, or DevOps consulting.

Today’s AI aggregators face comparable margin pressures as model providers themselves expand into enterprise-grade features, potentially marginalizing intermediary businesses.

For his part, Mowry expresses optimism for