3 Minutes
The frenzy around generative AI has cooled, and with it comes a harsh truth for many fledgling companies: not every business model built on large language models will survive. The question now is simple and brutal — who has a real moat?
Darren Mowry, a senior Google executive, says two specific startup archetypes are under severe pressure: LLM wrappers and multi-model aggregators. Both rode the wave of easy access to foundation models, but that tailwind is fading fast.
LLM wrappers are the neat product layers and user interfaces built on top of third-party models. Think of an app that helps students study or a coding assistant that leans on a major provider's model but adds a customized workflow. These products can feel valuable at first. They ship quickly. They attract users. Yet, if the only thing you own is a prettier UI, you don’t own much.
Some exceptions exist. Startups like Cursor in developer tools or Harvey in legal tech have invested in deep, defensible intellectual property — data, vertical expertise, fine-tuned models, or unique integrations — and that’s what gives them staying power. But those cases required building inward: collecting domain-specific knowledge, engineering proprietary data pipelines, and negotiating real enterprise contracts. Not every founder is willing or able to do that heavy lifting.
Then there are the aggregators: platforms that offer a single API or interface to multiple underlying models. On paper, aggregation solves fragmentation, and some players such as Perplexity and OpenRouter have added genuine value through monitoring, routing logic, and observability. But as the big model providers expand enterprise-grade features and pricing options, the margin for middlemen shrinks. Why would a company pay an intermediary when a provider can offer a turnkey enterprise stack?
The parallels with early cloud computing are striking. In the early cloud era, many startups tried to be generic infrastructure brokers. Over time, only companies that focused on security, migration, or a finely tuned vertical service thrived. Mowry points out that the AI landscape is moving the same way: specialization and defensible assets matter more than thin wrappers or simple stitching layers.
For founders, this is a strategic crossroads. Build a product with real intellectual property and deep customer integration, or don’t assume the market will tolerate yet another surface-level UX bolted onto someone else’s model. Investors are already asking the hard questions about moats, margins, and long-term value.
Products without defensible moats and enterprise-grade differentiation won't survive the next phase of AI maturation.
The era of easy arbitrage is ending; the era of engineering and domain expertise is beginning. What will your startup choose?
Leave a Comment