What happens when the safest cash flows in private markets suddenly carry tail risk of obsolescence? Private equity spent a decade treating subscription software like regulated utilities with better margins. Then AI showed up with a demolition kit. The result is a Darwinian moment: not a tech story, a capital structure story. Leverage, which flatters stability, punishes ambiguity. And ambiguity just became the base case.
The dominant thesis was simple: buy sticky software with 90 percent gross margins, raise prices a little, cross-sell, lever it 6-8 times, and refinance later at tighter spreads. This worked because switching costs were high, features were scarce, and predictability made lenders generous. AI collapses that triangle. Features commoditize faster, switching costs fall as systems become interoperable, and hyperscalers bundle intelligence across suites for free or close to it. The market is telling you as much. Hedge funds have cut net exposure to software to the lowest level in over five years, a vote that the moat has narrowed and the variance of outcomes has widened. That is not a rotation. It is a repricing of durability.
Software LBOs were engineered for average-case stability, not for fat-tailed downside. In probability terms, a small chance of product displacement can dominate expected value once you add 500-700 basis points of interest expense and covenants. Think Kelly criterion misapplied: gains are linear, losses are nonlinear. If net revenue retention slips from 120 to 105 and churn inches up, the equity story goes from compounding to coverage math overnight. With rates still elevated and refinancing windows narrower, the distribution of outcomes skews down. Fragility shows up first as silent drift—missed upsells, lengthening sales cycles—and then all at once in the refinancing meeting.
AI also changes where the rent is collected. If your product depends on inference from a foundation model, your gross margin now pays a compute tax to the cloud provider. That turns software’s classic 80-plus gross margins into something structurally lower or more volatile. Meanwhile, platform vendors weaponize distribution. A feature that once supported a standalone $20-per-seat SKU is now bundled into an enterprise suite for zero marginal price. Game theory applies: in a bundling war, niche vendors face a prisoner’s dilemma—invest heavily in AI just to keep up, or hold back and get buried by default settings. Either way, capital intensity rises and pricing power falls. Kodak did not die because photography vanished; it died because the profit pool moved. The same migration is under way from mid-stack software to compute, data, and distribution.
Public markets often see fragility first. Net hedge fund exposure to software has fallen to a multi-year low as managers split tech into AI asset owners versus AI-pressured distributors. This is a rational update, not panic. The signal is that investors no longer grant blanket durability to subscription revenue. They are discriminating between workflows embedded in regulated or physical processes and point tools that can be automated by a general model or bundled away. In Bayesian terms, the arrival of competent AI is new information that raises the posterior probability of margin compression and churn for a wide swath of vendors. When fast money de-risks, private markets should listen. Liquidity gives you prices. Illiquidity gives you stories.
The private credit boom extended the software trade with covenant-lite structures and aggressive add-backs that assumed linear growth and benign refinancing. Now a cohort of LBOs faces maturities in the next few years with weaker unit economics and higher base rates. Mark-to-model masks the crack until the day it does not. Interest coverage that once came from 90 percent gross margins and steady net dollar retention now has to contend with higher compute costs, more competitive pricing, and delayed procurement as buyers test AI-native alternatives. Lenders can tighten terms, but that only shifts risk back to equity. The structural error was confusing high recurring revenue with immunity to technological substitution. Utilities face regulated rate cases. Software faces unregulated algorithms.
AI has also created collateral damage: zombie companies hollowed out as top talent is absorbed by platforms. When a startup’s core developers and researchers are acqui-hired into giants, what remains is a brand, some contracts, and technical debt. That shell can limp along under private ownership, but its innovation engine has been removed. This is not just a venture problem. PE roll-ups that relied on accretive M&A and internal build will find a shallower pool of targets with real moats and a deeper pool of assets whose value sat in humans who have left. In biology, ecosystems that lose keystone species become brittle. In software, losing the few people who know why the system works creates the same brittleness, only masked by ARR until renewal season.
Dealmakers now face a selection problem. If every investment committee can plausibly argue that AI will disintermediate a target, the assets that do come to market may be the ones where that risk is hardest to price. Adverse selection shows up as clean data rooms and optimistic cohort analyses that hide the simple question: what happens when the buyer can replace your core feature with an agent or a native AI add-on? Some banks are even creating vehicles so multiple investors can pool into the small set of AI winners at nosebleed valuations, concentrating exposure further. Meanwhile, private market marks inflate on the back of a few mega-rounds at the same time lenders grow stricter on ordinary software. That divergence is not a signal of safety. It is a classic late-cycle pattern where capital chases narrative while cash flows face friction.
There is still software that benefits from disorder. Systems entangled with the physical world, regulated workflows with audit trails, and products where the data exhaust improves the product faster than a general model can imitate—these can absorb AI and get stronger. But the inversion is key: do not ask whether AI is an opportunity. Ask whether the business gains convexity when the environment gets more volatile. Does competition force you to cut price, or does volatility drive mission-critical usage? Does AI reduce your cost to serve faster than it pressures your price? Does your vendor concentration drop or rise as you adopt foundation models? In engineering, sound structures carry dynamic loads with a margin of safety. The LBO era built too many software bridges for static loads. AI turned on the wind.
The private equity software boom thrived on a mistaken equivalence: recurring meant resilient, and resilient meant leverable. AI has broken that chain. The market is now sorting businesses into those that own distribution, differentiated data, or rights to the compute tax—and those that do not. The latter are not doomed, but their cost of capital should be higher and their leverage lower. The most dangerous assumption in finance is that tomorrow will look like yesterday with a small error term. AI makes the error term the story. In that world, survival favors those who subtract debt, add optionality, and treat stability as something to be earned, not assumed.
