Building an AI Investment Thesis: What We Look for in AI Startups

Investing in AI startups in 2026 requires confronting a fundamental tension: the underlying models are improving rapidly, making today's differentiation potentially irrelevant tomorrow, while the application layer is exploding with opportunity. Building a coherent investment thesis requires being precise about where defensible value lives.

The Model vs Application vs Infrastructure Distinction

The AI stack has three layers: foundation models, applications, and infrastructure. Foundation model training requires billions in compute and is dominated by a handful of well-capitalized labs. This is not generally an accessible venture investment area. Infrastructure — inference hardware, model serving, fine-tuning platforms, evaluation tools — is competitive but has real technical differentiation opportunities.

Applications are where most venture opportunities lie, but the moats are different from traditional software. An application built on GPT-4 that doesn't have proprietary data, workflow integration, or switching costs can be replicated when models improve. The question is always: what does this company have that can't be rebuilt on a better foundation model in 12 months?

What We Look for: Data Moats, Workflow Integration, and Network Effects

The most defensible AI companies have proprietary data that improves their models faster than competitors can catch up. This data often comes from being embedded in customer workflows — where every user interaction becomes a training signal. The longer a customer uses the product, the better the product gets for that specific customer, creating genuine switching costs.

Network effects in AI products are rare but powerful. A platform where model outputs are evaluated, corrected, and improved by a community of users creates compounding advantages. Labeling platforms, evaluation networks, and AI marketplaces that aggregate demand can exhibit strong network effects that make them progressively harder to displace.

Red Flags: Model Wrappers and Demo Companies

We're cautious about companies that are primarily API wrappers around foundation models without proprietary data or integration depth. We're also cautious about companies with impressive demos but no clear path to sustainable revenue — AI capability is increasingly commoditized, and the ability to build a working demo says less than it used to. The standard we apply: if the foundation model improves by 50%, does this company get dramatically better, or does it get commoditized?