91% of investors agree: proprietary data and network effects are the moat. That’s not a consensus. That’s a red flag.
When a statistic hits 91% in a room of sophisticated institutional investors, it’s not a signal of clarity—it’s a stampede of cognitive convergence. Lazard’s survey of private equity secondaries investors captures the moment perfectly: 91% cite “proprietary data + network effects” as the core moat for software companies in the age of AI. Only 4% haven’t changed their investment approach. The rest are either pivoting or waiting.
But here’s the problem. Consensus in financial markets rarely predicts the future. It prices the past. The architecture of trust, engineered for failure, is being rebuilt on a foundation of sand.
Context
Lazard’s survey, released in mid-2025, polls a broad set of PE secondaries investors—LPs, GPs, and intermediaries—about how AI is reshaping software investment. The headline numbers are stark: 91% believe the moat is data and network effects; only 4% say their approach hasn’t changed. The rest are either reallocating capital to other sectors or waiting for clarity. The survey is a snapshot of a market in the middle of a paradigm shift—from valuing software by growth metrics (ARR, NDR) to discounting it by AI exposure.
In a bear market where survival matters more than gains, this data is exactly what readers need: a signal that the old valuation framework is dead. But the new framework is not yet born. The result is a valuation vacuum—risk for some, alpha for others.
Core: Systematic Teardown of the Data Moat Thesis
Let’s dissect the 91% consensus. The argument sounds logical: AI models are becoming commoditized; the last differentiator is proprietary data. Fine. But the devil is in the technical details.
First, the “data moat” is a time-bound illusion. As someone who spent six weeks auditing the 0x Protocol v2 exchange contract in 2017, I learned that what looks like a structural advantage often hides a critical vulnerability. Today’s proprietary data can be tomorrow’s open dataset. Synthetic data generation, federated learning, and model distillation are rapidly eroding the uniqueness of many “private” datasets. The architecture of trust, engineered for failure, is exactly this: investors assume data superiority is permanent, but it’s a function of the current model’s training data ceiling. Once models reach near-AGI capabilities, or even just 10x longer context windows, inference alone can approximate many proprietary distributions.
Second, network effects in the AI era are not what they were. Traditional network effects (more users → more value) are now being challenged by AI-driven platforms that can generate synthetic network effects. A single LLM-powered chatbot can simulate thousands of user interactions, creating a data flywheel without real users. The 91% consensus implicitly assumes that network effects are human-driven and hard to replicate. But the 0x audit taught me that automated scanners miss what humans do—and vice versa. Today, AI can mimic—and even accelerate—network effects. The moat is not durable.
Third, the 4% who didn’t change their method are the contrarian signal. In my experience tracing 185,000 BTC through Alameda’s wallets after FTX’s collapse, the minority often holds the truth. The 4% likely understand that software companies with real defensibility—like those with regulatory moats, hardware lock-in, or deep workflow integration—are not threatened by AI. They are still investing in the same framework because the framework still works for a subset. The 91% are overcorrecting.
Let’s talk about the “waiting” behavior. In my Celsius on-chain forensic analysis, I saw the same pattern: investors waiting for clarity while the ship sinks. The Dencun upgrade stress test simulation I ran in 2024 revealed a similar dynamic: everyone was excited about proto-danksharding, but the gas fee volatility for small L2 users was ignored. The PE secondaries market is now ignoring the fact that the data moat comes with a massive cost: data collection, storage, compliance, and governance. The architecture of trust, engineered for failure, is built on the assumption that data is a free asset. It’s not. It’s a liability.
Contrarian: What the Bulls Got Right
The bulls are not entirely wrong. Some software companies do have genuine data moats—especially in regulated industries like healthcare, legal, and finance. The network effects of platforms like Salesforce or Microsoft Teams are real and reinforced by integration complexity. The 91% consensus is not a hallucination; it’s a simplification. The bulls correctly identify that AI will not destroy all software, only the undifferentiated layers. The code generation, basic CRUD, and simple automation will be commoditized, but the companies that own the workflow and the data have a window.
However, the size of that window is overestimated. The AI-agent smart contract vulnerability I exposed in 2026 showed that even “autonomous” systems can be compromised by simple prompt injection. If AI agents can bypass multi-sig wallets, they can also learn to mimic network effects. The moat is shrinking, not widening.
Takeaway
The 91% consensus is a lagging indicator. It tells you where the market’s head is, not where it’s going. The real opportunity lies in the 4% who haven’t changed—and in the companies that have regulatory or hardware moats, not just data. The architecture of trust, engineered for failure, is the current valuation framework. The next architecture will be built by those who understand that data is not a moat. It’s a responsibility.
Bear market lesson: survival matters more than consensus. The 91% are right about the trend, but wrong about the timeline. The window for data moats is closing faster than anyone expects. The next sell signal will come when the first major software company’s “proprietary data” is replicated by open-source models. That day is not hypothetical. It’s already happening.