The Real Meaning of Defensibility in the Age of AI

In 2015, defensibility meant switching costs, IP, and network effects. In 2026, those concepts are still real — but AI has reshuffled which ones actually protect you and which ones are illusions.

I've now seen this pattern across 65+ investments. The founders who get defensibility right are not the ones with the best AI model or the most features. They're the ones who built something structurally hard to displace before anyone noticed what they were doing.

What Traditional Moats Look Like Now

Every MBA course covers Porter's five forces and moat theory. The classic list: switching costs, network effects, economies of scale, cost advantages, and intangible assets like brand and IP. These haven't disappeared — they've been re-rated.

AI has brutally compressed the half-life of feature-based advantages. A capability that took 18 months and $3M to build in 2022 can be replicated in 60-90 days in 2026 using foundation models available to everyone. I've watched competitors ship near-identical features to portfolio companies within a single quarter. The product moat — the thing most founders think they're building — is often the weakest protection of all.

Meanwhile, brand and distribution — historically seen as soft advantages — have become harder and more valuable than ever. The reason: AI has made the supply of "good enough" products nearly infinite. In a world where every vertical has five AI-native solutions, the one with the distribution wins.

The New Hierarchy of Defensibility

Here is how I think about defensibility in 2026, ranked from strongest to weakest:

The Data Flywheel Is Real — But Rare

Every founder claims they have a data flywheel. Almost none of them actually do. A real data flywheel requires three things simultaneously: (1) usage generates labeled signal that is actually valuable for model improvement, (2) that signal is proprietary and not replicable from public sources, and (3) the improved model creates a meaningfully better product experience that drives more usage.

The companies that have this are genuinely hard to displace. Glean has processed billions of enterprise search queries across specific company knowledge graphs. Cursor has observed how millions of engineers actually fix bugs in real codebases. These are not datasets you can buy or scrape — they exist only because the product was in production, trusted, and used at scale.

Most "data flywheel" claims I hear in pitches do not meet this bar. "We collect data from users" is not a flywheel. The test: if a competitor launched tomorrow with identical compute and better engineers, could they replicate your data advantage in 24 months? If yes, it is not a moat.

What This Means for How You Build

The strategic implication is straightforward but uncomfortable: if you are spending 80% of your engineering capacity on features and 5% on distribution, your priorities are inverted relative to what actually drives defensibility.

How VCs Actually Evaluate Defensibility

When I look at a company now, I ask a specific question: "If OpenAI or Anthropic released a native version of your product in six months, what percentage of your customers would stay?" Most founders say "most of them." The good founders can actually explain why — with specifics about workflow depth, data exclusivity, and contractual structure.

The data supports why this matters. In CB Insights' post-mortems of failed AI startups, "got outcompeted by a better-resourced player" is the second most common cause of failure after running out of cash. In the majority of those cases, the competitive loss came because the startup's only advantage was product quality — and the larger player replicated it with a team 10x the size.

The companies that survived those competitive battles had at least one structural advantage the attacker could not quickly replicate: a customer community that trusted them, contracts that had 18 months remaining, or training data that had 3 years of customer-specific signal baked in.