A SaaS flywheel is a growth system in which the output of one cohort of users, their content, data, referrals, or revenue, is reinvested to produce the next cohort, and the loop still compounds after you subtract churn, acquisition cost, and the cost of serving each turn. A circular diagram is not a flywheel. A loop that pays for its own next turn is.

The Short Answer: A Loop That Pays for Its Own Next Turn

The word comes from Jim Collins by way of HubSpot, and by the time it reached SaaS marketing it had been flattened into a picture: three words arranged in a circle, usually attract, engage and delight. That version is harmless and useless. It describes an intention, not a mechanism, and it cannot be falsified, which is the first clue that it is not doing any work.

The mechanical definition is older and better. In Growth Loops are the New Funnels (July 2018), Brian Balfour, Casey Winters, Kevin Kwok and Andrew Chen wrote that loops “are closed systems where the inputs through some process generates more of an output that can be reinvested in the input.” That sentence is the whole concept. Everything else is bookkeeping: which output, reinvested into which input, at what cost, and how fast.

So the test of a flywheel is not whether you can draw it. It is whether you can name the output of one turn and point at where it enters the next one.

Where the Word Came From, and What Its Biggest Promoter Sells Now

Jim Collins used the flywheel to describe how good companies become great: a heavy disc that takes enormous effort for the first few turns and then carries its own momentum. HubSpot borrowed it in 2018 and put a circle of attract, engage and delight where the marketing funnel had been. That is the version that spread, and it is the version that lost the mechanism.

Reforge published the loop framing quoted above in the same year. Two ideas from 2018, pointing in opposite directions: one at a diagram of intentions, the other at a closed system you can measure. The diagram is the one at the top of the results.

There is a development none of the five competing pages we read for this article discusses in its text. HubSpot's newest marketing framework is not the flywheel but Loop Marketing, a four-stage cycle of express, tailor, amplify and evolve, built around what HubSpot calls a Taste Profile and scored on four headline metrics: remarkability, resonance, reach and regeneration. It ships with a book, a prompt library and answer-engine tracking. We read that page on 5 November 2026. HubSpot states no launch date on it, so we do not claim one, and its FAQ says Loop Marketing builds on the inbound foundation rather than replacing it. The flywheel page is still live.

Funnel, Loop, Flywheel, Network Effect: Four Words, Four Meanings

These four words are often used as stylistic variants. They are not, and the difference decides what you should build.

Funnel

Best for measuring one conversion path

  • Output returns as input
  • Next turn costs less than the last
  • Each new user adds value for existing users
  • Hard for a competitor to copy

Growth loop

Best for naming the mechanism

  • Output returns as input
  • Next turn costs less than the last
  • Each new user adds value for existing users
  • Hard for a competitor to copy

Flywheel

Best for describing the whole system

  • Output returns as input
  • Next turn costs less than the last
  • Each new user adds value for existing users
  • Hard for a competitor to copy

Network effect

Best for explaining durability

  • Output returns as input
  • Next turn costs less than the last
  • Each new user adds value for existing users
  • Hard for a competitor to copy

Two rows in that grid are our framing rather than anyone's measurement: whether the next turn costs less than the last, and whether a competitor can copy it. The other two rows follow from the standard definitions.

The distinction that gets lost most often is the last one. Andrew Chen, writing in November 2021 while finishing The Cold Start Problem, argues that viral growth is a flavor of network effect, one of three alongside the engagement and economic effects, and that the naive viral factor is the wrong number to chase. His revised model stops treating the viral factor as one static number. People invite across many sessions, so the factor builds up over time, and retention decides how many chances each user gets to invite. Retention is not one term among several. It sets the number of shots on goal.

The Failure Test: Five Conditions That Say Yours Is Not a Flywheel

The five competing pages we read for this article (the ones we could open, out of the nine results on page one on 5 November 2026) are definitions or build guides. Not one of the five gives you a pass-or-fail test with a number in it. Here is that version. If any row below is true of your loop, what you have is a funnel with curved arrows.

ConditionThe number to pullIt fails if
The output has no nameUnits produced per cohort per turn: invitations, indexed pages, labeled data, reviews, integrations, cashYou cannot name the unit, so there is no number to pull
The output is not reinvestedShare of new users traceable to that outputThe share is zero: nothing in the product or the budget routes it back into acquisition, activation or retention
Users leave before the next turnCycle time, against retention measured over that same windowCycle time is longer than the window most users survive, so the loop never closes
Acquisition cost rises as fast as the outputCAC by cohort, against output per cohortCAC grows at least as fast as the output, which is a treadmill rather than compounding
The turn costs more than it returnsGross profit per turn, inference includedGross profit per turn is zero or negative

In AI products the last row is the one that catches people, and it is the subject of a section below. Note what the test does not ask: how well the diagram is drawn, or how many stages it has.

Three Loops That Actually Close

The examples offered for this term are often brands rather than loops. These three are loops, and they come from the same Chen essay, which makes a point about them worth more than the examples: not one of the three was aggressive about inviting anybody.

  • Dropbox, shared folders.The output is a shared folder. It is a useful feature on its own terms and it arrives in front of somebody who does not have an account. Chen calls it both fully consistent with the product's purpose and a source of millions of organically acquired users at no acquisition cost.
  • Slack, inviting colleagues. The output is a teammate who has to be in the workspace before the product can do its job at all. The invitation is not a growth tactic bolted on; it is the product working.
  • Instagram, shared content. Early on the output was a photo posted somewhere else, which carried the product to people who had not installed it. Chen notes it grew without paid marketing.

Run the failure test on all three and they pass for one shared reason: the output is something the user wanted to produce anyway. That is why Chen's conclusion is that retention, not invite mechanics, is what makes a loop turn. The loops built on aggressive prompts are the ones that produced the shark fin above.

What the Data Shows, and What It Does Not

None of the five pages we read addresses this. There is no study, that we can find, measuring flywheel adoption against long-run growth. What exists is adjacent evidence about retention, and it is correlational.

The largest recent sample is SaaS Capital's 2025 B2B SaaS Retention Benchmarks (published September 2025, more than 1,000 private B2B SaaS companies, retention measured from December 2023 to December 2024). Median net revenue retention was 101% and median gross revenue retention 91%. Sorted by NRR band, median growth rates look like this:

Net revenue retentionMedian growth rate
Below 90%15%
90% to 100%16%
100% to 110%21%
110% to 120%30%
120% to 130%38%
Above 130%50%

Median growth across the whole sample was 24%, and moving from the 100-110% band to the 110-120% band is associated with nine percentage points of additional growth. Gross revenue retention, by contrast, shows little correlation with growth in the same data: it behaves like a threshold rather than a lever.

Why the AI Era Breaks the Classic Flywheel Math

The classic SaaS flywheel rested on an assumption that was rarely stated: the next user is nearly free to serve. Seats cost nothing to add, so every extra turn of the loop was close to pure margin, and more usage always meant more operating leverage.

In an AI product that assumption is gone. Every generation, retrieval and agent step carries compute cost, which means a loop can spin faster while gross profit falls. The honest version of the loop is therefore not users into users, but gross profit per turn into the next turn.

The usual rebuttal is that inference keeps getting cheaper, so the problem solves itself. The data says something more specific, and the specific part is not the average. It is the spread.

9x to 900x
yearly price fall at fixed quality, by task

Epoch AI tracked what it cost to reach six benchmark milestones over three years. Matching GPT-4 on PhD-level science questions got about 40x cheaper per year. Across tasks the rate runs from 9x to 900x, and Epoch notes the fastest falls happened in the last year of its window, so they may not hold.

Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, published 12 March 2025, prices from Epoch AI and Artificial Analysis. a16z put the same 2021 starting point at $60 per million tokens in Welcome to LLMflation, 12 November 2024.

A loop running on cheap, good-enough inference sits at the 900x end. A loop that needs the best model available sits at the 9x end, and the gap between those two is the whole design problem. The part the rebuttal leaves out, from a16z's analysis: OpenAI o1 charged the same price per million output tokens that GPT-3 charged at launch, $60. Two years on, that gap has barely closed. On 5 November 2026 we read Anthropic's published price list: its most expensive model costs $50 per million output tokens, its mid-tier model $10, and its cheapest $0.50. Comparing the dearest frontier option of late 2024 with the dearest of late 2026, across two different vendors, the top of the market went from $60 to $50 while the floor kept falling by orders of magnitude.

That asymmetry is the thing to design against. If your loop runs on cheap, good-enough inference, time is on your side and has been all along. If every turn needs the most capable model available, cost per turn is a constraint you engineer around, because waiting has returned about 17 percent in two years.

A worked example. Every figure below is our own arithmetic on inputs we made up to show the shape; none of it is a benchmark or anybody's measurement. The question it answers: how much of its own replacement does one user pay for, and how much does the loop have to supply free?

InputSeat-cost productSame product, with inference
Gross revenue per new user, first month$40$40
Variable cost to serve that user for the month$0$18
Gross profit per user per turn$40$22
Cost to buy one user outside the loop$100$100
Share of its own replacement one user funds0.400.22
Invited users the loop must supply free, per user0.600.78

Read the bottom row. Add $18 of inference and the referral rate the loop has to produce climbs from 0.60 to 0.78 per user. A product actually bringing 0.7 invited users per user compounds in the first column and leaks in the second. Same product, same referral behaviour, and nothing in the circular diagram would have told you which column you are in.

Six Numbers That Describe Your Loop

A diagram is not a model. If you want to know whether the thing is real, these are the quantities worth having, and most of them are already in your product analytics:

  • Cycle time. How long one turn takes, from a new user to the output that creates the next user.
  • Loop share. What percentage of new users came from the loop rather than from spend.
  • Survival to the next turn. Retention measured over the cycle time, not over a calendar month.
  • Output per user. Invitations, documents, data points or revenue produced per user per turn.
  • Gross profit per cohort. Revenue minus the cost of serving that cohort, including inference.
  • Payback by loop. How long until the cohort has paid for the acquisition of the one that follows it.

How to Build One, Briefly

Several of the pages we read are build guides with five or so named stages. The short honest version is four questions, and a staged framework will not rescue you from failing to answer them.

  1. What does a user produce that another user wants? Not what you wish they would produce. The three loops above close because the output was already worth making.
  2. Where does that output meet somebody who is not a user? A loop with no exposure point is just a feature.
  3. How long is one turn, and do users survive it? If the cycle outlasts the user, nothing compounds, however good the output is.
  4. What does one turn cost, inference included? That answer decides whether speed compounds for you or against you.

If you want the design side of this rather than the definition, the four loop types and the mapping exercise are in our growth loops playbook, and worked examples from real companies are in the AI-native flywheel teardowns.

FAQ: SaaS Flywheel

What is a SaaS flywheel?

A SaaS flywheel is a growth system in which the output of one cohort of users, their content, data, referrals, or revenue, is reinvested to produce the next cohort, and the loop still compounds after you subtract churn, acquisition cost, and, in AI products, the variable cost of serving each turn. A circular diagram is not a flywheel. A loop that pays for its own next turn is.

What does flywheel mean in business?

In business a flywheel is a system where each cycle makes the next one easier, so effort compounds instead of resetting. Jim Collins used the image for companies that become great slowly and then quickly. In software the term narrowed to growth, where it means a loop whose output, such as referrals, content, data or cash, is reinvested to produce the next cohort of users.

What are some examples of a SaaS flywheel?

Dropbox shared folders, Slack workspace invitations and early Instagram content sharing are the standard examples, and all three work the same way: the output a user creates for their own reasons reaches somebody who is not a user yet. Andrew Chen notes that none of the three relied on aggressive invite prompts, and that many of the products built on aggressive prompts collapsed once the novelty wore off.

What is the difference between a funnel and a flywheel?

A funnel describes one direction of conversion: more in at the top, more out at the bottom, with no mechanism for reinvesting the output. A flywheel describes a closed system whose output returns as input. Reforge made the distinction in 2018, arguing the funnel is too micro a view to answer the question of how a product grows.

Is a growth loop the same thing as a flywheel?

Not quite. A growth loop is the mechanism: one closed cycle where output becomes input. A flywheel is the whole system of loops plus whatever makes it durable, usually a network effect or an accumulating asset. One referral loop is a loop. A product that gets more valuable for existing users every time a new one joins, and therefore keeps generating invitations, is closer to a flywheel.

Does a SaaS flywheel still work when every action costs inference?

It works when you measure the loop in gross profit rather than in users. The classic SaaS flywheel assumed the next user was nearly free to serve. In an AI product every turn carries compute cost, so a loop can accelerate while gross profit falls. Cheap models did get dramatically cheaper, but frontier output prices have not followed, so cost per turn is a design input, not a problem that time solves on its own.

Is there research proving that flywheels cause SaaS growth?

No study we can find measures flywheel adoption against long-run growth. What exists is adjacent: retention benchmarks showing that higher net revenue retention correlates with faster growth, and cost curves showing what has happened to inference prices. Correlation is not the same as a mechanism, and anyone citing a benchmark as proof that flywheels work is overstating what the data says.

Where to Go Next

If you came here to find out whether the circle in your board deck is real, run the five-condition test above and then put numbers on the six quantities. Check the second condition first, because it is the cheapest to test: does anything route the output back into acquisition?

And if you are earlier than that, before any loop exists, the question is not which flywheel to build but which move belongs to your stage. That sequencing is the subject of the stage-by-stage growth playbook: one rule, four stage gates, and the single job that matters inside each one.

We publish one teardown or playbook like this every week. If that is useful, the SaasFlywheel newsletter sends them as they land.