Say you launch two AI products in the same month. It’s the same model underneath, and the features. But a year later, one has seen insane amount of growth.

It's noticeably sharper than at launch, users keep coming back, and competitors can't close the gap. And the other is exactly as good as the day it shipped. Maybe worse, because three other teams cloned it over a long weekend.

Both had the same start but ended differently. The instinct is to credit the winner with a better model, a bigger budget, and sharper engineers.

But that's not how it works. The difference is structural. One product is built as a self-feeding loop. The other is a straight line that goes nowhere.

That loop is the data flywheel. And once you can see it, you can't unsee it.

Let me explain.

A data flywheel is four steps.

Step 1: Usage. People use the product.

Step 2: Data. That usage gathers signals, such as what they clicked, what they ignored, what they corrected, what they came back for, and more.

Step 3: A better model. You feed that signal back in, and the thing underneath gets smarter: better recommendations, predictions, answers.

Step 4: More usage. Because it's better, more people use it, and the ones already there use it more. That gives more data, which trains the model. This goes in a loop.

The word flywheel is doing the work here.

The metaphor comes from Jim Collins' Good to Great book.

Picture a giant metal disc on an axle. It's two feet thick, thirty feet across, five thousand pounds. Your job is to get it spinning.

Nothing happens when you push it the first time. When you push it again, leaning your whole weight in, it moves an inch. This one slow movement might take hours.

But you keep going in the same direction. Two turns. Four. Eight. And somewhere in there, the disc's own weight starts working for you.

Momentum kicks in. Now each push lands on top of all the pushes before it, and the flywheel hurls itself forward, "whoosh," as Collins puts it. It's almost impossible to stop.

There's no single heroic shove that did it. It was the accumulation. That's exactly the shape of an AI product. It’s brutal at the start.

There's no data, and every push seems to do nothing. Unstoppable once it's spinning, because every turn of usage sits on top of every turn before it.

If you have been around product for a while, this might smell like a network effect. It's a cousin, not the same. A classic network effect is all about connections between users.

A telephone is worthless if only one person owns it, and it gets valuable with every new line. Same with a social app or a marketplace. Value comes from users being there.

A data network effect is different.

The product gets better because all that usage leaves behind data that makes the underlying model smarter for everyone, including those who never interact at all.

Take Google Search, for example. You and I never connect.

But every time millions search and click a result, or bounce straight back to try another, that behaviour is a signal about which results actually answered the question.

Aggregate enough of it, and the ranking gets better for the next person.

But note that exactly how much click behaviour feeds ranking is something the company describes cautiously. That has been argued over publicly.

Treat the mechanism as real and the precise internals as proprietary.

Many products have two of the most important things: users and data. And yet, they don't compound. "Collect data" isn't enough anymore.

Three things have to be true, all three, or the wheel only sits there.

The gold is feedback that comes free from normal use.

TikTok is the cleanest example.

When the For You feed shows you a video, you don't need to rate it. Whether you watch it to the end, rewatch it, or scroll past in half a second, that is the feedback.

TikTok says outright that a strong signal like finishing a longer video from start to finish carries more weight than a weak one. (TikTok's explanation of the For You feed.)

That's an implicit signal. Captured from behaviour, requiring no effort. The opposite is an explicit signal: a thumbs-up, a star rating, a correction you type in.

Explicit feedback is richer but rarer because asking for it is friction. The best flywheels lean hard on implicit signals, because every interaction becomes a data point.

TikTok gets a verdict on every video... that's a lot of fuel, very fast.

A captured signal makes no sense if it just pools in a warehouse. Tesla is the best example. The cars running on the road double as the data-collection fleet.

In "shadow mode," the self-driving software runs in the background. It predicts what it would do, then compares that to what the human driver actually did.

The mismatches are the cases worth keeping. They get gathered, the models retrained, and updates pushed back to the fleet. (Tesla Autopilot overview)

Did you notice the complete circuit? Deploy, collect, retrain, redeploy.

The product is the data pipeline. That's exactly what most teams are missing, because building the path from raw signal back into a shipped improvement is so hard.

A flywheel with no one pushing is just a heavy disc. This is where the loop calmly demands something from the rest of your product.

Duolingo's model, which it calls Birdbrain, learns each learner's ability from the exercises people complete and tunes difficulty to keep them in the sweet spot.

That's hard enough to learn, not so hard that they quit.

But that only works if people keep showing up. The streaks, the reminders, the gamification aren't separate from the flywheel. They are what keeps it spinning.

Capture the signal. Build the pipe. Give people a reason to return.

If you miss anyone, you will have a wheel that turns once and stops.

All three roads lead to the same place: the dreaded thin wrapper, a product whose entire value comes from the underlying model, with nothing compounding on top.

If a competitor can rebuild it in a weekend by calling the same API, that's the tell. There's no flywheel, so nothing to catch up to.

You are permanently as good as the base model and no better.

This is the real line between a feature and a defensible product.

A feature is a straight line: input, model, output, done. A product with a flywheel is a loop that bends back and feeds itself, pulling away from anyone who started later.

It would be neat to stop there: build the loop, win forever.

Reality is more grudging, and data flywheels can be overstated. "We'll hoover up data and that'll be our moat" is one of the most repeated, least examined claims in tech.

The sceptics, a16z chief among them, point out a few uncomfortable truths.

Data often hits diminishing returns. For many problems, the model improves fast on the first slugs of data, then flattens.

Past a "good enough" point, more barely moves the needle, so a competitor doesn't need all your data, just enough to clear the bar.

Data can also go stale as the world shifts under it, and plenty of it isn't as proprietary as its owner hopes. Rivals can often collect, buy, or generate something close.

So the strongest flywheels share a few traits. The data is hard for others to get. More of it keeps helping rather than plateauing early.

The loop is fast, and the one PMs forget, the improvement is something users actually feel.

What every AI PM should ask is not "do we have data?" (everyone has data). It is "Does our data compound in value, and does each turn produce something users can feel?"

The model you rent is the same one your competitor can rent. The loop you build on top of it is the only part that's yours.

The teams that internalise this stop asking whether their AI feature is impressive and start asking if it's spinning because that decides who's still standing in the long run.