The protocol remembers what the regulators forget. But the market remembers what the consensus forgets.

Lazard’s latest survey of private equity secondary market investors isn’t a data point—it’s a seismic shift in how capital prices software. 91% of respondents now identify “proprietary data + network effects” as the core moat. Only 4% haven’t changed their investment approach. This is not a slow drift. This is a stampede. And in a bull market, stampedes are the most dangerous time to look for exits.

Context: The survey that broke the old frame

Lazard’s 2025 survey (conducted in June, published August 15) captures the moment when institutional investors stopped debating whether AI will disrupt software and started pricing the how and who survives. The 91% figure is statistically anomalous—typical survey consensus ranges from 50-70%. When 9 out of 10 investors agree on a single moat thesis, the market has already priced in that belief. The 4% who haven’t changed are noise. The real signal is the 96% who have shifted their lens.

But what exactly are they shifting to? The old valuation framework—MRR multiples, growth rates, NRR—is being discarded. A new framework is forming: base multiple × AI exposure discount × moat quality premium. The problem? That framework isn’t standardized yet. We’re in a valuation vacuum. Old rules don’t work, and the new rules haven’t been written. This is where alpha is born—and where most capital will get trapped.

Core: The moat consensus is a map, not the territory

Let’s dissect what the 91% actually means. They’re saying: “Generic LLM capabilities are a public good. The only defensible advantage is data that the model can’t replicate, wrapped in a network that compounds the data asset.”

This is technically correct—but only at a snapshot. Based on my experience auditing DeFi protocols during the Terra collapse, I’ve seen how quickly perceived moats can evaporate. The same applies here. The “proprietary data” moat depends on three unstated assumptions:

1. Model intelligence has a ceiling. If LLMs approach AGI, synthetic data and reasoning alone may approximate the value of proprietary datasets. The 91% are betting against that possibility.

1. Network effects are AI-complementary, not AI-substitutable. This is true for platforms like collaboration tools or marketplaces, where user behavior data feeds into personalized AI. But many “network effects” are just user lock-in, which AI can erode by offering a better experience.

1. The cost of deploying AI doesn’t destroy unit economics. Traditional SaaS has 70-85% gross margins. Adding inference costs (API fees or GPU clusters) compresses margins. The 91% are implicitly assuming that revenue growth from AI features will offset margin compression. History shows otherwise: every feature commoditization wave in software has compressed margins first, growth later.

Contrarian: The consensus is a trap

When 91% of investors agree on something, the market has already priced it. The contrarian angle isn’t to disagree—it’s to ask: What is this consensus missing?

First, the 91% are ignoring the “reliability moat.” In enterprise B2B software, deterministic behavior (the software does exactly what it’s told) is still a competitive advantage over stochastic AI models. The survey didn’t ask about “hallucination risk” or “compliance certainty.” This suggests investors are over-rotating toward data assets and underestimating the value of trust in mission-critical workflows.

Second, the consensus assumes that AI-native companies will replace incumbents. But the reality is more nuanced: incumbents with data can buy AI startups. Lazard’s survey is itself a signal of impending M&A—when secondary market investors become cautious, primary market exits shift to acquisition. The “moat” narrative is a self-fulfilling prophecy: it justifies why large players will pay premiums for data-rich targets, while small players without data get starved of capital.

Third, the 4% who didn’t change their methods might be the smartest in the room. They’re not ignoring AI—they’re using a longer time horizon. They know that the current “proprietary data” advantage could be legislated away (data portability laws, privacy regulations) or technology-arbitraged away (federated learning, synthetic data). The 91% are pricing a moat that may not exist in 3 years.

Takeaway: The next crisis is already coded in this consensus

Speed without direction is just volatility. The 91% consensus is a directional bet—but it’s also a crowded trade. The real opportunity lies in the gaps: software companies that own neither data nor network effects but have deep domain expertise and deterministic reliability. These are the hidden value plays that the market is systematically undershooting.

Crisis is just code with a high gas fee. The valuation vacuum won’t last forever. When the new framework solidifies, the first movers who built it will capture the spread. Open source is a promise, not a product. But the promise of a new valuation paradigm is the only product that matters now.

The protocol remembers what the regulators forget. The market remembers what the consensus forgets. And the savvy investor remembers that 91% agreement is usually the last signal before the crash.