At equal recurrence, what a customer can do without its supplier predicts how the revenue behaves under shock. Enterprise software is the current instance.
Two revenues can carry identical recurrence and rest on different mechanisms of persistence, and the dashboard reads only the recurrence. A company can show the same annual recurring revenue as its competitor, the same net retention, the same growth curve, and lose most of that revenue within two quarters if one team walks out, not its own but the supplier’s. Stated once and directly: at equal observed recurrence, supplier-support dependence predicts a revenue’s response to shock, because it reveals whether persistence rests on a customer capability that was transferred, on an external operation still running, or on other mechanisms of retention. The domain of validity is narrow on purpose. The claim is comparative and conditional, strongest where the recurring line is presented as evidence of an installed capability, weakest where no one ever mistook operation for autonomy.
Persistence runs through several mechanisms that recurrence cannot tell apart: reproduced value, supplied operation, exit friction, coordination effects, institutional constraint, passive continuation. Only the first is the customer producing value on its own. Call it autonomy, and note at once that it is not what a dashboard reads. A strategy firm whose recommendation is internalized holds an autonomous revenue; the same firm on a permanent retainer holds a support revenue, and only the second falls when the retainer is cut. A surgeon who trains the unit reproduces the capability; the theater that stops the day the surgeon leaves never did. Recurrence is agnostic to which of these holds it up, which is why it is the wrong primary instrument for revenue quality.
The move here is to change the unit of analysis. Standard practice observes revenue and its retention; this note observes the capability that generates the revenue, and treats the recurring line as an imperfect commercial projection of that capability rather than as the thing itself. The lineage is deliberate and old, from Penrose on the firm as a bundle of productive services, through Sen on capability, Nelson and Winter on routines, to Teece on dynamic capabilities. What is new is applying it to the one line accounting treats as primary. The recurring number is not falsified; it is demoted from cause to symptom.
Model the deployment as a chain, not a single link. A capability stock A rises with transfer support and the client’s own internal investment, at some absorption efficiency, and falls with depreciation, turnover and forgetting, plus an external term set by the supplier’s pace of innovation that can deprecate the stock, raise it by abstraction, or substitute for it. Capability then produces reproduced value under operational support and context; value produces billable usage, filtered by inertia, network and institutional terms; usage produces supplier revenue through contract structure and bargaining power. Revenue does not depend on capability directly. It sits three transformations downstream, which is exactly why a rise in capability can lower revenue: a capable client consumes more efficiently, negotiates harder, or substitutes. An open-source tool that keeps running without its maintainer is the pure case, high autonomy with almost no revenue. Autonomy that is good for the customer can be negatively correlated with the supplier’s revenue, and any metric that assumes the two move together is measuring the wrong object.
Support is not a scalar. It is a vector of expertise, configuration, training, operations, maintenance and governance, and it splits by function. Operational support sustains current output; transfer support raises the customer’s future production function. The first rents output, the second sells capability, and the invoice distinguishes neither. Withdraw operational support and revenue drops to what the stock carries alone. Withdraw it from a client whose stock was genuinely built, and revenue dips, then climbs back as the client keeps investing on its own. That rebound is the observable signature of a capability that was transferred rather than rented, and it is legible before any cancellation appears in the retention numbers.
The chain forces two distinct measures, never one. A Capability Autonomy Index concerns the client’s stock: its ability to sustain value without the supplier. A Support-Adjusted Revenue Persistence concerns the supplier’s flow: the behavior of the revenue once support is removed. Collapsing them into a single recurring number is the original error, and naming them apart is half the contribution. What survives a withdrawal, it should be said plainly, is not autonomy alone but autonomy plus lock-in, inertia, network and institutional persistence. The quantity of interest is therefore a counterfactual, a revenue that would have persisted with no support, no captivity, no inertia, all else equal, which is a world never directly observed. That is why the autonomy index is closer to a causal estimand than to an operational metric, a limit rather than a defect.
“Withdrawal” names three experiments, not one. Removing this supplier tests dependence on it. Removing every equivalent external resource tests autonomy proper. Replacing the supplier with a peer tests substitutability. Only the second measures autonomy, and conflating the three is the fastest way to misread a shock. The strongest objection is that net retention already captures all this: a revenue held up by support will eventually churn, and retention will record it. It will, eventually. The claim is narrower and earlier, that the support mechanism is legible before the cancellation, in the elasticity, not after it, in the lost logo. So the honest status of the argument is a hypothesis with a specified test.
The first and modest test is predictive: does an ex-ante measure of support improve the prediction of churn or consumption decline beyond ARR and net retention, ΔAUC above zero. If it does, recurrence was hiding information, and that is enough to justify measuring it. Causal identification comes second and harder, through natural experiments, service-contract endings, supplier reorganizations, geographic team withdrawals, read in difference-in-differences, because a deliberate withdrawal is endogenous: support is pulled precisely when an account matures, or slides, or gets cost-cut. Predictive lift first, causal estimand second, is the order that keeps the argument falsifiable rather than merely plausible.
If part of the asset is the customer’s capability stock, the useful contract term is capability built, not volume consumed, and capability is observable enough to write down: the share of operations run without assistance, the mean time to internal resolution, the share of changes the client makes itself, the pass rate of a recovery test performed with the supplier absent. Index the engagement on those, not on usage. This does not require the aligned supplier to disappear. It may keep mutualizing scale, security, regulatory and innovation load, and that dependence is productive. The criterion is not maximal withdrawal but the removal of unnecessary dependence, a presence justified by continuing marginal value rather than by a maintained incapacity. The accounts will not adjudicate this, because support cost and revenue quality are recorded in different rooms, with no ledger connecting the spend to the capability it did or did not transfer. At more than seven hundred billion dollars of hyperscaler capacity built for 2026 on the four companies’ own guidance, and built ahead of the revenue meant to fill it, the question that decides which revenue survives is not who consumes, but who could carry on without the vendor in the room. A capability that disappears when the supplier leaves was never transferred. The revenue may keep recurring, but its persistence is operationally rented, at a price no line in the accounts is built to show.
Full argument, with the five propositions, the complete chain model, the three-withdrawal protocol and the discussion of limits, in the PDF below (5 pages).
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.