A single question hung over the room after Apollo’s John Zito spoke: what if enterprise software, the most reliable collateral in modern finance, is less durable in an AI regime than the models assume? The shock was not about quarterly guidance or a new product. It was about a structural mismatch between debt built for stability and a technology that rewrites cost curves, switching costs, and pricing power. If you lever the past, a regime shift is not volatility; it is a breach.
Private equity turned recurring revenue into a bond proxy. The idea was simple: sticky customers, high gross margins, and predictable cash flows fund high leverage. That logic works until the mechanism that creates stickiness changes. Generative AI lowers the practical friction of migration. Copilots write connectors, map data, and automate testing. Support, once a differentiator, is now a prompt away. When switching costs fall, what looked like a moat becomes a drawbridge. In engineering terms, these businesses were bridges rated for static loads; AI introduces dynamic loads and resonance. The same span that handled cars shudders when frequency meets structure. The cash flows were collateral until the operating environment changed the physics.
The last decade trained investors to overpay for small changes in net retention. Many roll-ups were built on revenue-based lending and covenant-lite terms on the assumption that 95 percent gross margins and upsell trajectories cover a lot of debt. In reality, these structures are levered to minor shifts. Drop net retention from 115 to 103, layer in 10 to 15 percent price pressure, add 6 to 12 months of AI integration delay, and coverage flips fast. The variance is not symmetrical. In software LBOs, a point of churn can erase multiple turns of EBITDA. The base rate mindset of the last cycle—steady cohorts, annual price increases, predictable expansions—was a single-climate model. Now the weather changed. Apollo’s 10-K says the risks of AI are not possible to predict. That is not boilerplate. It is an admission of Knightian uncertainty you cannot diversify or model away with a higher discount rate.
In a market with generative features, someone will cut. The first incumbent to bundle AI at no extra cost forces peers into a prisoner’s dilemma: match price and compress margins, or hold price and risk share loss to upstarts with zero legacy burden. Open-source models compress the technology premium. Sales cycles shorten around proofs of concept delivered by bots, not by armies of solution engineers. Winners are those closest to the input costs of compute, data, and distribution; not those with the oldest maintenance base. The textbook moat—switching costs plus proprietary code—faces a different opponent: an ecosystem that reduces the value of proprietary features to table stakes. This is not a normal competitive cycle. It is the kind of phase change that turned paid antivirus into Windows Defender overnight.
Private credit was the silent partner in the software trade. Lenders bought the spreadsheet: recurring revenue, low capex, strong free cash flow. But most of these loans were underwritten with light covenants and rosy recovery assumptions tied to IP value and maintenance contracts. In a world where features commoditize and maintenance pricing weakens, recoveries fall. Apollo probed this fragility when it took bearish positions against loans of Internet Brands, SonicWall, and Perforce late last year—small in size, closed since, but telling. The point was not trade P&L. It was discovering that the capital structure of software is less robust than believed when AI attacks the line items lenders depend on. Expect spreads on enterprise software loans to widen and structures to tighten. The market will start to price obsolescence risk, not just execution risk.
If the cash flows at the application layer wobble, capital migrates downstack. Data centers, power, networking, and memory fabs are the new scarce inputs. Apollo’s co-president says AI infrastructure will need five to seven trillion dollars over five to seven years, only part of it through investment-grade channels. The firm is positioning on that vector too, backing a multibillion-dollar compute deal tied to xAI via Valor. That is a barbell: hedge fragile app-layer exposure while owning the shovels and the right to toll the road. It is not a contradiction. The same team can short overlevered software loans in one pocket and finance substations and chillers in another. In Taleb’s terms, the left tail lives in cash flows that can be competed away; the right tail sits with assets whose value rises with demand for inference and training.
Software moats now look like hard hats and transformers. The strategic variables are megawatts, land near fiber, transformer lead times, interconnect queues, and water rights. Latency becomes a profit center if your racks are where the data lives. The grid is the bottleneck. If you have entitlements, long-term power agreements, and the ability to build at scale, you have pricing power. That is a different kind of defensibility than per-seat software licenses. It is also less sensitive to the specific model winner. Whether open weights or closed models dominate, someone has to power and cool the compute. The locus of value moves toward assets with physical constraints. For investors trained on intangible moats, this is inversion: the fewer atoms you need, the more your economics depend on those who control atoms.
The spreadsheet that underwrote the last wave of software deals was a thin-tail instrument. It extrapolated stable churn, predictable pricing, and modest competition. AI introduces thick tails and sudden correlations. Customer cohorts do not degrade; they regime shift. Margins do not compress steadily; they gap when features get bundled for free. You do not fix this with sensitivity tables. The inputs themselves are wrong. As in probability puzzles, the error is not in computing the odds; it is in choosing the wrong distribution. Apollo’s CEO put it bluntly: fortunes will be made and fortunes may be lost. That is a way of saying variance is widening. In widening variance, leverage is not a tool; it is a hazard. The first job is to survive the path, not to optimize for the mean.
The instinct in private markets is to optimize. Max leverage, max IRR, minimal cash drag. That mindset works in a stable regime. In unstable regimes, redundancy is not waste; it is survival. Balance sheets with room to maneuver will outlast those that chased every basis point. In software, that means funding innovation from equity, not hoping creditors will finance the pivot. It means accepting lower near-term margins to harden retention, reduce customer concentration, and rebuild pricing power with real value, not feature checklists. In credit, it means writing structures that anticipate obsolescence and allow intervention before it is too late. In infrastructure, it means building optionality into sites and interconnects. The first-order question is not who wins AI. It is who can afford to be wrong and still be around when the answer emerges.
The shock in that room was overdue. This industry treated enterprise software cash flows like a law of nature. They were not. They were the product of cost, friction, and complexity that AI is now lowering. Some private equity firms will adapt by moving down the stack, by financing the inputs, and by respecting the asymmetry of uncertainty. Others will defend a spreadsheet that no longer maps to the world. The market will sort them. In the meantime, assume fragility where the consensus still sees stability. When the environment changes, the strongest structures are those designed to flex. The software bet was built to be efficient. AI is testing whether it can be resilient.
