The Compounding Engine

Whether growth compounds or fades is a number you can compute, not a story you tell afterward.

A company I followed had a month it never stopped talking about. Something it launched caught, the referrals poured in, the numbers tripled, and everyone involved understood that they had finally built the thing every founder dreams of, a product that grew itself. They reorganized around the assumption, hired against it, told investors about it, and waited for the curve to keep bending upward. It did not. Over the following months the surge faded, the numbers settled back toward where they had been before, a little higher but not transformed, and the team was left puzzled and a little bereaved, because they had seen the thing grow itself with their own eyes and now it had stopped, and no one could say why the magic had drained away. The plain answer, which nobody had the framework to give, is that there had never been magic, only a one-time push moving through a loop that was always going to let it fade, and the shape of what happened was as predictable as a dropped ball coming to rest.

I do not tell this as a story about a foolish company, because the company was not foolish, and the mistake it made is one almost everyone makes, including people who have read about viral coefficients and would have said all the right things in a meeting. The pull of a good month is enormous, and the early shape of a fade is identical to the early shape of a compounding, so the misreading is the default that anyone without the number in hand will reach for. The number is the only thing that tells the two shapes apart before the months reveal which one you had.

The reason it was predictable is that growth loops, the mechanisms by which existing demand generates new demand, are a kind of system that has been studied precisely, and their behavior is governed by a single number that decides everything. The number says whether a loop compounds, building on itself without end, or whether it merely amplifies whatever you put into it and then fades when you stop. Most of what gets called viral growth is the second kind misread as the first, a fade mistaken for a compounding because the early part of a fade and the early part of a compounding look identical, and only the framework, or a few painful months, tells them apart. A field that wants to engineer growth rather than pray for it has to be able to compute that number and know which kind of loop it has.

The number that decides everything

Picture the simplest possible growth loop. You have some amount of demand now, a base of customers or users, and over the next period each unit of that demand generates, through whatever mechanism, some amount of new demand. People refer others, content gets shared, usage creates visibility that draws more usage. The question that decides the system’s fate is how much new demand each existing unit produces per cycle, and I will call that number the loop gain, written $g$. If a hundred units of demand this period produce, through the loop, fifty new units next period, the loop gain is one half. If they produce a hundred and twenty, the loop gain is one and a fifth.

Write the demand at cycle $n$ as $Q_n$, and the simplest version of the loop says that the demand next cycle is the loop gain times the demand this cycle,

$$Q_{n+1} = g\,Q_n.$$

The entire future of this system is decided by whether $g$ is above or below one. If $g$ is greater than one, each cycle is larger than the last by a constant factor, and the demand grows exponentially, without limit in this simple model, compounding the way money compounds at a fixed interest rate. If $g$ is less than one, each cycle is smaller than the last, and the demand decays exponentially toward nothing, the loop fading because each generation of demand produces less than a full replacement for itself. The line between these two fates is exactly $g = 1$, where each unit of demand reproduces itself precisely once and the system holds steady. A single number, and on one side of it lies runaway growth and on the other side lies fade, with nothing in between except the knife-edge of exact replacement.

Anyone who has heard of a viral coefficient has met this number under another name. The famous k-factor, the number of new users each user brings, is the loop gain of one particular growth loop, and the famous condition for virality, that the k-factor must exceed one, is exactly the condition that the loop gain exceed one. The framework is not new. What is missing in go-to-market is the habit of taking it seriously, of actually computing the loop gain rigorously and acting on which side of one it falls, rather than watching a number rise for a month and concluding that the loop has crossed a threshold it has not crossed.

Part of why the number matters so much is that you cannot read it off the early data by eye. A loop with a gain of one and one tenth and a loop with a gain of nine tenths look almost identical for the first several cycles, both rising, both feeling like momentum, and they diverge only later, when one keeps climbing and the other rolls over and fades. The eye cannot separate them in the early going, which is exactly when the largest decisions get made, so the field is forever placing its biggest bets at the one moment its instruments are least able to distinguish the two fates. Only the computed gain separates them in time to matter.

What a fading loop actually does

The company that had its famous month had a loop gain comfortably below one, well short of the threshold, and what it experienced was the entirely normal behavior of a fading loop given a sudden push, which is worth working out because it is so commonly misread.

A real growth system is rarely just the loop. There is also external input, the demand that arrives from outside the loop, from paid acquisition, from a launch, from a moment of attention. Add a steady external input $a$ each cycle to the loop, and the system becomes

$$Q_{n+1} = a + g\,Q_n.$$

When the loop gain is below one, this system does something specific and useful. It does not grow without limit, and it does not fade to nothing either, because the external input keeps refilling it. Instead it climbs to a settling point and stays there, a level where the demand lost to the fading loop each cycle is exactly replaced by the external input plus what the loop still produces. That settling point, the fixed point of the system, is

$$Q^\star = \frac{a}{1 - g}.$$

This little expression is one of the most useful things in the whole subject, because it says exactly what a sub-critical loop is worth. A loop with a gain of one half sits at a fixed point of $a$ divided by one half, which is twice the external input, meaning the loop doubles the effect of whatever acquisition you feed it. A loop with a gain of nine tenths sits at ten times the input, a tenfold amplifier. The quantity $1/(1-g)$ is the amplification factor, the multiplier the loop applies to your external acquisition, and it grows large as the gain approaches one, which is why loops with gains near one feel almost magical even though they are not self-sustaining, because a loop that turns every dollar of acquisition into ten dollars of demand is enormously valuable while remaining, in the technical sense, a fade.

The numbers are worth seeing run. Suppose a loop with a gain of one half receives a steady external input of a hundred units a cycle. The fixed point is a hundred divided by one half, which is two hundred, so the loop turns a steady hundred into a standing two hundred, doubling the input. Raise the gain to nine tenths and the same hundred of input supports a standing thousand, because the amplification factor is now ten, which feels spectacular and remains a fade. Now give the half-gain loop a one-time pulse, a launch that drops two hundred extra units in a single cycle on top of the steady input. The next cycle keeps half of that surge, a hundred, then half again, fifty, then twenty-five, and the bump melts back toward the two hundred fixed point over a handful of cycles, each one surrendering half of what remained. Plot those points and you have drawn the famous month and the disappointing quarter that followed, the entire arc, out of a single gain of one half and one pulse.

Now you can see what happened to the company. They fed their referral loop, the mechanism by which each new customer nudged a few others to join, a sudden large external input, a launch, a burst of attention, a one-time pulse rather than a steady stream. A pulse into a sub-critical loop produces a spike that then decays back toward wherever the steady input alone would hold the system, because the loop cannot sustain the elevated level once the pulse stops, and each cycle it gives back a fraction of the surge. The famous month was the spike. The disappointing months after were the decay, the system returning to its fixed point because the loop gain was below one and a pulse through such a loop always fades. They watched a completely ordinary phenomenon, a pulse decaying through a sub-critical loop, and read it as the arrival and then the mysterious departure of self-sustaining growth, when the loop gain had never been near one and the fade was written into the system from the start.

Honest accounting

The loop gain is the number that matters, which means the way it is measured is where the discipline lives or dies, and it is extremely easy to measure in a way that flatters.

The first temptation is to count gross rather than net. A loop might bring in new demand each cycle while also losing demand to churn, and the gain that governs the system’s fate is the net figure, the new demand minus what is lost, per unit of existing demand. A product can look like it has a loop gain above one on gross new users while having a loop gain well below one once churn is subtracted, and the gross figure will promise compounding that the net figure quietly forbids. The second temptation is to ignore cost. A loop sustained only by an incentive that costs more than the demand is worth is really paid acquisition wearing the costume of virality, and counting it as loop gain confuses a thing you are buying with a thing the system generates. The third temptation is to count a one-time pulse as if it were a steady rate, to take the surge from a launch and annualize it into a loop gain it never had, which is exactly the error the famous-month company made.

Each of these flattering measurements has the same effect, which is to report a loop gain higher than the one the system actually runs on, and since the loop gain decides whether you compound or fade, an inflated loop gain becomes a prediction of a future the system will never deliver. A company that believes its loop gain is one and one tenth when it is truly nine tenths has concluded that it compounds when it fades, and it will plan, hire, and raise money against growth that is not coming, and the gap between the believed gain and the real one will surface, eventually, as a painful correction. The discipline of measuring the gain net, after cost, as a sustained rate, is the discipline of refusing to lie to yourself about which side of one you are on, which is the only thing that finally matters.

Under honest accounting, net of churn, net of cost, measured as a sustained rate rather than a spike, a loop gain above one is rare. Truly self-sustaining growth, the kind that compounds without continuous external feeding, is one of the rarest things in business, which is the opposite of the impression the word viral gives, because the word gets applied to every spike, almost all of which are pulses through sub-critical loops. The rarity is worth internalizing, because a field that understood it would stop treating a good month as evidence of a loop gain above one, and would start asking the harder question of what the loop gain actually is, net and sustained, which usually has a sobering answer and always has a more useful one than the story about magic.

Even compounding does not compound forever

Suppose you do have a loop gain above one, the rare and wonderful case. The simple model says you grow exponentially without limit, and that is where the simple model lies, because the loop gain is no fixed quantity. It changes as the system grows, and it changes in one direction, downward, as the addressable market is consumed.

A loop generates new demand by reaching people who are not yet customers, and there are only so many of those. Early on, when penetration is low, nearly everyone the loop reaches is new, and the gain is at its highest. As the system grows and penetration rises, more and more of the people the loop reaches are already customers, so each cycle converts fewer truly new people, and the gain falls. The loop gain is therefore a decreasing function of how much of the market you have already captured, high when you are small and falling toward one and then below it as you saturate. This is the same saturation that governs a single channel, now operating on the growth loop itself, and it has a precise consequence. A system with a loop gain above one grows exponentially at first and then slows as the falling gain approaches one, leveling off as the gain crosses below one, tracing the S-shaped curve that every real instance of compounding growth eventually follows. The exponential is only the early part of an S. The leveling is the loop gain doing what loop gains do as the market fills, not a failure of the system.

You can watch this confusion play out whenever a fast-growing company’s growth rate dips for the first time. The dip gets treated as an emergency, a signal that something has broken, and a scramble follows to find the cause and fix it, when the dip may be nothing more than the loop gain sliding down the saturation curve exactly as it always would once penetration crossed a certain level. The growth rate of a healthy compounding system declines from its first day, slowly at first and then unmistakably, and a team that does not expect this will misread its own success as the onset of failure and intervene against a process that needed no intervention.

This matters because it changes what you watch. A compounding system that is slowing is often read as a system that is breaking, prompting panic and intervention, when it may simply be a healthy loop moving along the inevitable S as penetration rises and the gain falls. The engineering question is not whether growth is slowing, because all compounding growth slows, it is whether the slowing matches the saturation, whether the gain is falling because the market is filling as expected or because the loop itself is deteriorating for some other reason. Telling those two apart requires knowing the loop gain and how it should fall with penetration, which is a quantitative question with a quantitative answer, and which the field mostly addresses with vibes and alarm.

The hypothesis

Let me state the claims as a hypothesis sharp enough to test and to be wrong.

The hypothesis has three parts. The first is that the fate of a growth loop is governed by its loop gain $g$, with sustained self-sustaining growth requiring $g > 1$ under honest accounting, net of churn and cost and measured as a steady rate. The second is that most growth described as viral has a loop gain below one, and behaves as an amplifier of external input with a fixed point at $a/(1-g)$, so that the characteristic spike-and-fade pattern after a launch is a pulse decaying through a sub-critical loop rather than the arrival and loss of true compounding. The third is that the loop gain falls as penetration rises, so that even super-critical loops are self-limiting and trace an S-curve, and a slowing of compounding growth is the expected behavior of a filling market rather than a sign of failure.

These are testable. Measure the net new demand a loop generates per cycle per unit of existing demand, rigorously, and check whether it exceeds one. Watch what happens after a pulse, and see whether the elevated level sustains, which would indicate a gain near or above one, or decays back toward a fixed point, which would indicate a gain below one. Track the gain as penetration rises and see whether it falls as the saturation account predicts. If you found loops sustaining elevated levels after pulses with no continued input, or found gains that did not fall with penetration, the model would be wrong in instructive ways. I expect the model to hold, because the famous-month company is the single most common growth story there is, a pulse through a sub-critical loop, celebrated as virality and then mourned as its loss.

What it changes

A field that computed loop gain rigorously would talk about growth completely differently. It would stop celebrating spikes, because it would know that a spike is consistent with any loop gain and tells you almost nothing about whether the system compounds. It would distinguish, sharply, between loops that amplify and loops that compound, valuing the amplifiers correctly as the multipliers they are while reserving the word self-sustaining for the rare loop that truly earns it. It would read a slowing as a measurement to be checked against the saturation the gain should produce, rather than as an automatic crisis. And it would put its engineering effort where the number actually moves, on raising the loop gain through the design of the loop, rather than on manufacturing pulses that feel like growth and fade like the pulses they are.

This would also change how growth is funded and promised. A great deal of capital is raised on the strength of a spike, on a chart that points steeply upward for a quarter, and a discipline that read loop gain rigorously would know that such a chart is consistent with a loop that fades, and would discount it accordingly, asking for the sustained net gain before believing in compounding. The same discipline would protect founders from their own best months, from the seductive quarter that feels like escape velocity and is a pulse through a sub-critical loop, by handing them a number to consult that is harder to fool than the feeling of a good run. Knowing the gain is, among other things, a defense against the most expensive optimism in the business.

The whole framework rests on one quantity measured rigorously, and that is also where it becomes fragile, because measuring the loop gain rigorously means knowing how much new demand the loop actually caused, net of everything, and causation is the hardest thing to know in any system made of people. The number that decides everything is a number about cause, about how much of next cycle’s demand this cycle’s demand is truly responsible for, and separating that from everything else that moves demand, the season, the market, the other things you were doing, reaches past anything better dashboards could fix and down into the question of whether cause and effect can be known at all from the data a go-to-market system produces, which turns out to be a question with a difficult and clarifying answer.

---

References

• On iterated maps, fixed points, and stability in discrete dynamical systems, see standard treatments of difference equations and nonlinear dynamics.
• On the viral coefficient (k-factor) and the condition k > 1 for self-sustaining growth, and on amplification of acquisition by sub-critical loops, see the growth and network-effects literature; the analysis here treats that literature critically, distinguishing amplification from true compounding.
• On saturation and the logistic (S-curve) form of bounded growth, see the standard models of constrained growth in population dynamics and diffusion of innovations.