Lazard's 91% Consensus: The Valuation Vacuum Replacing Code-First Software — A DeFi Auditor's Reading of the AI Moat Shift

The data shows 91% of private equity secondary investors now describe proprietary data and network effects as the only durable moats left in software. Only 4% say their investment methodology has not changed. Reconstructing the logic chain from block one — this is not a survey. It is a ledger of a paradigm being formally written off. Lazard's investor poll has been circulating since mid-summer, and the percentages are extraordinary by construction. In normal market cycles, institutional opinions on strategic threats cluster around 50% to 70% with a long tail of dissent. Ninety-one percent agreement is not consensus. It is a stampede. What the market is telling us is that the debate over whether AI will disrupt the software industry has closed. The only remaining question is how the disruption happens and which companies survive. A valuation vacuum has opened. The old framework — MRR multiples, growth rates, net revenue retention — is visibly failing. The new framework — AI exposure discounts, data asset quality premiums — has not yet been standardized. Static code does not lie, but it can hide. The same principle applies to portfolios. The institutions that are waiting are not passively idle. They are pricing uncertainty into every software asset they hold, and that repricing is transmitting directly from the PE secondary market into the broader technology economy, including the crypto and Web3 infrastructure stack that I spend my professional life auditing. The first thing to understand about the 91% figure is what it technically implies. When institutional investors align on proprietary data plus network effects as the core moat, they are making a structural bet about large language models. They are betting that general model capability is becoming a public good. Once models can write code, draft documents, summarize meetings, and handle routine support tickets, the functional layer of software is commoditized. What cannot be commoditized is the distribution of data that was never captured in public corpora. Transaction logs from vertical industry platforms. User behavior telemetry from closed ecosystems. Compliance-sensitive records that were never allowed to leak into a training set. Here is the technical boundary condition that most commentary misses. Transformer architectures learn patterns from their training distribution. If a particular dataset — say, insolvency triggers from a regulated lending protocol or audit findings from a private security review — was never in the public corpus, the model cannot directly reproduce that distribution. It may approximate it through analogy, but approximation is not conformance. That structural gap is what makes proprietary data a genuine defense. I have seen this principle hold in smart contract security. The Aave liquidation models I built in 2020 relied on order book and oracle data that no public model understood at the time. The edge was not our code. The edge was our data and the network of nodes feeding us that data. But the 91% consensus embeds a dangerous temporal assumption. The statement "proprietary data is hard to replicate" is true at the current moment of model evolution. It is not true on an infinite horizon. Synthetic data generation is improving. Federated learning is improving. Differential privacy is improving. Model context windows have expanded from 4K tokens to over one million in a matter of years. That means data that was once siloed can be indirectly inferred, distilled through carefully crafted API interactions, or reconstructed from model outputs that were trained on leaked versions of the same information. The investor consensus is effectively saying "today's irreproducibility is permanent irreproducibility." That is a time-horizon error, and it will produce valuation surprises on both sides of the trade. The survey also reveals a hidden technical judgment about model layer homogeneity. If model providers were strongly differentiated from each other, the strategic question would be about which model foundation a software company selects. The fact that 91% of investors bypass that question entirely and go straight to data and network effects tells us that the market has already priced model commoditization. The application layer, the data layer, and the distribution layer are where wins are decided. That is a profound direct analog to the architecture of decentralized finance. In DeFi, the consensus protocol layer has been commoditized — Ethereum is Ethereum, and the real value capture has migrated to oracle networks, data availability layers, and liquidity distribution mechanisms. Let me say that plainly: the same valuation migration happening in software is happening in crypto. Code is becoming a commodity. Data ownership is becoming the new collateral. Based on my audit experience, I can tell you exactly where this framework breaks down. In 2022, I conducted a post-mortem forensic analysis of the Terra USD algorithmic stablecoin. I traced the loop between UST and LUNA and documented 42 specific lines of code that lacked circuit breakers. The technical cause was well known: a feedback loop that amplified small depegs into a death spiral. But the systemic cause was deeper. The market had treated mining and validation as the moat, when the real dependency was oracle integrity and data reliability. Investors did not miss the code. They missed the data chain. The same error is now being repeated in the AI software valuation discourse, only at higher speed and larger scale. The commercial logic of the survey is where the real signal lives. When PE secondary investors say they are moving capital to other opportunities, they are not leaving software. They are rotating into sub-sectors with lower AI exposure risk: infrastructure, security, data services. This is a sector rotation, not an exit. The consequence is a capital drain from undifferentiated application-layer software companies toward the plumbing underneath them. In crypto terms, this is the same rotation we saw when institutional money moved from DeFi applications to Ethereum itself, and then from Ethereum to data availability and oracle layers. The pattern is consistent. Value migrates up the stack when the top of the stack becomes replaceable. This rotation has a direct mechanical effect on valuation levels. The PE secondary market is the price discovery engine for illiquid software assets. When buyers step back, transaction volume drops, sellers who need liquidity accept discounted bids, and the discount rate widens. Based on observable market behavior across the past six months, I estimate software asset secondary market discounts have widened by five to fifteen percentage points. That is not a slow drift. That is a repricing event. The same dynamic appears in crypto private sales and token unlocks. When secondary buyers of locked tokens step back, the implied discount for the same token in the primary market widens mechanically. This is not speculation. It is arithmetic. The new valuation framework essentially becomes: base revenue multiple times an AI exposure discount factor times a moat quality premium. The problem is that no one has agreed on what those second and third terms should equal. A software company with ten million dollars of ARR and a proprietary dataset of insolvency events could be worth either four times revenue or twelve times revenue depending on how the market weights the data asset. That divergence is where alpha is created. The investors who can distinguish real data moats from marketing fiction will generate excess returns for the next three to five years. The infrastructure dimension of this shift is the most underappreciated. Traditional software companies had a cost structure dominated by engineer salaries, producing gross margins of seventy to eighty-five percent. Add AI to the stack and you add inference costs. Every API call to a language model is a line item. Every self-hosted GPU cluster brings depreciation and electricity. This cost structure change alone would justify investor hesitation even without the existential disruption narrative. In my audits of DeFi protocols, I have seen the same phenomenon. Gas costs and oracle subscription fees are eroding the margin story of yield-generating protocols. The protocols that hide these costs are the ones that blow up later. The protocols that price them honestly are the ones that survive. The contrarian angle deserves equal time. The 91% consensus is itself the risk. When consensus is this uniform, the market is not discovering a truth; it is constructing one. The self-referential nature of this consensus will cause capital to flee from software companies that lack a compelling data story, even when those companies are actually protected by reliability and trust. Enterprise software vendors that have spent a decade building deterministic payroll processes, regulatory reporting, and audit trails — software where hallucination is not an acceptable failure mode — are being penalized by a framework that celebrates data moats but ignores the value of certainty. In B2B and in DeFi, deterministic correctness has economic value. The survey suggests investors are not pricing that correctly. There is also an agency problem embedded in the source of the survey itself. Lazard is an M&A advisory and asset management firm. Publishing a survey that says "the market is in wait-and-see mode" serves two functions simultaneously. It manages seller expectations downward in a market where Lazard represents sellers. And it signals to buyers that the moment to enter is approaching — which brings buyers to the table where Lazard also represents them. I am not suggesting the survey is fabricated. I am suggesting that the platform of the survey is part of the phenomenon. The same is true in crypto research reports that are released in advance of protocol token sales. The report itself is a market-moving instrument. Listening to the silence where the errors sleep — what the survey does not say matters as much as what it does. It does not quantify how much capex a software company needs annually to maintain a data moat. It does not address whether the legal and regulatory environment will allow companies to lock up data as property. The EU AI Act, data portability regulations, and the aggressive privacy enforcement in Singapore and the broader APAC region directly threaten the legality of "proprietary data as moat." My 2025 work on Standard Chartered's DeFi gateway taught me that KYC/AML data is both the most valuable asset and the most regulated one. A data moat that violates compliance frameworks is not a moat. It is a liability waiting to be liquidated. The industry impact will play out on a compressed timeline. In the next six months, software asset discounts on secondary markets will widen further. In twelve to eighteen months, primary market software investment will decline while AI infrastructure investment rises. In twenty-four to thirty-six months, the M&A wave will hit with full force. Large-cap software companies with cash and strong data positions will acquire undervalued vertical SaaS players at prices that look absurdly low from a 2024 perspective. The same playbook is visible in crypto. Well-capitalized L1s are already acquiring wallet infrastructure, oracle startups, and data providers at distressed valuations. The AI disruption narrative is accelerating a consolidation cycle that was already underway. For the crypto and Web3 investor reading this analysis, the application is direct. The chains and protocols that own proprietary data — transaction graphs, credit histories, compliance attestations, MEV patterns — are the ones that will accrue value as AI commoditizes the application layer. The protocols that offer generic smart contract execution will face the same compression that generic SaaS is facing. The valuation framework for protocols will shift from "total value locked" to "data asset uniqueness and regulatory conformance." Protocols that secure data access and prove data provenance will command premiums. Protocols that merely move tokens will trade at discounts. Security is not a feature, it is the foundation. The same discipline that forces me to verify every line of bytecode before signing off on a protocol audit forces me to recognize that the Lazard survey is not a report on reality. It is a report on the perception of reality. But perception drives capital, and capital drives structural change. Whether the underlying data moats are real is, for the next two years, almost irrelevant. What matters is that ninety-one percent of institutional investors believe they are. That belief is now baked into how software assets are priced, how portfolio construction decisions are made, and how private equity capital is allocated. The question that should keep every software founder and every protocol team awake is not whether the data moat story is true. It is whether the market's belief in that story survives the first wave of model distillation attacks and regulatory constraints. If the moats are real, the current discount creates one of the largest buying opportunities in a decade. If the moats are theater, then the 91% consensus is a crowded exit at the top of a fictional cliff. The next twenty-four months will reveal which reality is encoded in the ledger. As an auditor, I know the answer. The ledger does not care about beliefs. It only records the transaction. Trade accordingly.