The earliest wave of generative AI startups could raise significant capital on the strength of a compelling demo and a credible founding team, with relatively little scrutiny of long-term defensibility. That bar has moved. Investors who have now sat through several years of AI pitches are asking sharper questions: what happens to this business if the underlying foundation model provider ships a similar feature natively? What's the actual gross margin once inference costs are accounted for at scale? Is the data or workflow advantage here defensible, or is it a thin wrapper around someone else's API?
The 'wrapper' problem
A significant share of AI startups built in the last few years are, structurally, an interface layer on top of a small number of foundation model providers. That's not inherently a weak position-plenty of successful software companies are built on top of infrastructure they don't own. But it does mean the startup's survival depends heavily on decisions made by a company it doesn't control, and foundation model providers have shown a consistent pattern of absorbing popular use cases into their own core products over time. Investors have gotten more disciplined about pricing that risk into valuations.
Where defensibility actually comes from
The AI startups attracting the most durable investor confidence now tend to share a common trait: they've built an advantage that doesn't depend on having the best model. That might be proprietary data that improves outputs in ways a generic model can't replicate, deep integration into a specific industry's workflow and compliance requirements, or a distribution advantage that would be expensive for a larger competitor to replicate. In each case, the model is a component, not the product.
What this means for founders
The practical implication for founders raising capital now is that the pitch has to answer the defensibility question directly rather than let strong usage growth stand in for it. Rapid adoption in an AI startup's early months is a weaker signal than it used to be, because much of that growth reflects general market curiosity about AI tools rather than durable product-market fit. The diligence questions that matter now are about retention after the novelty wears off, and about what specifically happens to the business's margins and moat as the underlying models improve.
A healthier market
None of this is bad news for the sector as a whole. A funding environment that rewards defensibility over demo polish is a more durable one, and it's likely to produce startups better positioned to survive the next model generation rather than be made obsolete by it.