Tracing the invariant where the logic fractures—this is how I read Lazard’s latest private equity secondary market investor survey. The headline is clear: 91% of respondents now cite "proprietary data + network effects" as the primary moat for software companies. Only 4% have not changed their investment approach. This is not a normal distribution. It is a system-level signal that the software industry’s valuation framework has been invalidated, and a new one is being built in real time.
### Context: The Survey as a Stress Test Lazard’s survey, conducted among PE secondary market investors, captures the moment when the market stopped debating whether AI would disrupt software and started asking how. The consensus is near-unanimous. But homogeneity in a complex system is a red flag for a contrarian. I have audited enough DeFi protocols to know that when everyone agrees on the same invariant, the edge cases are where the loss hides. Here, the invariant is "data + network = moat." The fracture? Most software companies do not actually own proprietary data that is both valuable and defensible against AI’s relentless commoditization.
### Core Insight: The Code-to-Data Migration Friction reveals the hidden dependencies. In software, the old dependency was on code complexity and feature velocity. Developers wrote code, users paid for functionality. That model is breaking. AI, specifically large language models, has turned general-purpose code generation into a zero-marginal-cost commodity. The new dependency is on data that is exclusive, unstructured, and difficult to replicate. This is not a theoretical shift—I have seen it in practice. During my 2022 audit of a ZK-rollup, I discovered that the fraud proof window’s race condition was not a bug in the code but a failure to account for the data latency between L1 and L2. The code was correct; the data dependency was the attack vector. The same principle applies here: the software company’s moat is not its codebase but its ability to lock data and network effects that AI cannot easily extract.
But here is the technical nuance: the 91% consensus may be overestimating the durability of data moats. Using my own experience with the 2021 NFT metadata decoupling incident, I found that even “on-chain” data can be hijacked if the backend is centralized. The same risk applies to proprietary data—if the data is stored on a server that can be subpoenaed, or if the data is derived from public sources with a thin layer of processing, the moat is illusionary. The Lazard survey does not distinguish between “true proprietary data” (e.g., user behavior logs from a 10-year-old platform) and “data that happens to be private today” (e.g., a startup’s customer list). The market will eventually punish the latter with a severe valuation discount.
### Contrarian: The Blind Spot of Reliability Precision is the only reliable currency. The survey’s focus on data and network effects overlooks a critical defensive barrier: reliability. In enterprise software, the value of a deterministic system—where a function always returns the same output for the same input—is not replaced by AI’s probabilistic outputs. I have tested this with a prototype AI-oracle synergy in 2026, where I combined Chainlink data feeds with a decentralized ML model. The latency was low, but the accuracy was still bounded by the model’s confidence intervals. Enterprises cannot tolerate a 5% error rate in payroll or compliance software. The “reliability moat” is real, but it is invisible to the 91% because it is not captured in the “data or network” category. This is a pricing inefficiency: software companies that provide deterministic, auditable outcomes will retain pricing power even as AI eats their feature set.
### Takeaway: The Valuation Vacuum Reverting to first principles to find the break: the old valuation framework (EV/Revenue multiples based on growth and NDR) is dead. The new one is not yet standardized. This creates a vacuum where the market is simultaneously pricing in maximum disruption and maximum uncertainty. The only safe bet is to short the narratives and long the code. I will be watching for the first major software company to disclose its “AI exposure score” in its quarterly report. That will be the signal that the new framework has arrived. Until then, the 91% consensus is a crowded trade waiting to be reverted.
Metadata is memory, but code is truth. The survey remembers the past—it captures today’s fear. The code—the actual software, its data architecture, its network topology—will tell us who survives. I am not betting on the consensus. I am tracing the invariant where the logic fractures. And the fracture is here: the 91% are right about the direction, but they are wrong about the velocity. The transition will take longer than they think, and the companies that look like they have no moat today may build one faster than anyone expects.