What Is a Data Flywheel?

By Brian Tr · Updated Oct 7, 2026

A data flywheel is a loop in which using a product generates data, that data makes the product better, and the better product attracts more usage, which generates more data. In AI startups it's the most popular answer to the question "what stops a bigger company from copying us?" It's also one of the most overclaimed answers. Some products have a genuine flywheel. Many have a data pipe with a diagram drawn around it.

This page explains where the idea comes from, how the loop actually works, the specific ways it breaks, and a set of tests for deciding whether a given product has one. It sits inside a wider set of moat questions, covered in the overall AI startup defensibility map, and it stays deliberately narrow: data as the source of advantage, not workflow, distribution or brand.

Where the Idea Comes From

The word "flywheel" is a business metaphor before it's a data one. Jim Collins popularised it in Good to Great, and his own description of the flywheel concept uses the image of pushing a very heavy wheel. Early progress is almost imperceptible, each push builds on the last, and at some point the momentum starts working in your favor. Collins rejects the idea of a single killer move: results come from accumulated pushes in a consistent direction. Our general page on the flywheel effect covers the business-strategy version.

The data version adds a specific mechanism. In the Collins picture, the wheel turns because of sustained effort. In the data picture, the wheel is supposed to turn partly on its own, because every customer interaction produces information that improves the product for the next customer. That claim is what makes it attractive to investors: it promises compounding returns that a competitor can't buy their way past.

It's also the claim that deserves the most scrutiny, because it's really two claims bundled together. The first is that more data makes the product better. The second is that this improvement is large enough, fast enough and exclusive enough to keep rivals out. The first is often true. The second is where most flywheels fail.

How the Loop Actually Works

A working data flywheel has four links, and each link can be the weak one.

1. Usage generates data. Customers do something in the product that leaves a useful trace: a correction, an approval, a rejected suggestion, a completed task.
2. The data is captured and cleaned. The company has permission to use it, stores it in a usable form, and can separate signal from noise.
3. The data improves the product. It feeds model training, fine-tuning, retrieval, prompts, evaluation sets or rules, and the improvement is measurable.
4. The improvement drives more usage. Better output earns more trust, more workflows, more customers or more retention, and the cycle repeats.

The loop only compounds if all four links hold. Plenty of products have link one (everything leaves logs) and none of the others. Collecting data isn't a flywheel. A flywheel is data that's used, that changes the product, and that changes customer behavior.

What kinds of data count

Not all data is equal, and the type decides how defensible the loop is.

The pattern is simple: the closer the data sits to a real decision made by a real expert, and the further it is from anything a competitor could scrape or buy, the more it's worth.

Why Data Became the Default AI Moat

Part of the appeal comes from research. Work on scaling in deep learning by Hestness and colleagues found that generalization error follows a power law as training data grows, across machine translation, language modeling, image processing and speech recognition. Later work on scaling laws for language models reported that loss scales as a power law with model size, dataset size and compute, with some trends spanning more than seven orders of magnitude. The popular reading is "more data, better model, reliably." That reading is the seed of the flywheel story.

A second reason is that learning from people works. OpenAI's InstructGPT paper reported that outputs from a 1.3 billion parameter model trained with labeler demonstrations and ranked outputs were preferred to outputs from the 175 billion parameter GPT-3, despite the model having 100 times fewer parameters. Feedback data, in other words, can matter more than raw scale for how a product behaves.

Both findings are real. Neither proves a startup has a moat. A power law is also a curve of diminishing returns, and a company that pays labelers is describing a process any well-funded rival can copy.

When a Data Flywheel Is Real

A flywheel tends to be genuine when most of the following are true.

The data is proprietary and generated by the product itself. If the best data comes from your own customers' work, a rival can't license it or crawl it. They have to win customers first, which is the harder problem.

The data is hard to get without the workflow. Data that's a by-product of a task people have to do anyway (reviewing a document, resolving a ticket, approving an invoice) accumulates without extra effort from users. Data that requires users to stop and rate things trickles in. For how workflow position creates its own advantage, see AI workflow lock-in.

The improvement is visible to customers. If a better model never changes what a customer sees, the loop doesn't drive usage. The test is whether a customer could tell, without being told, that the product got better.

The domain has a long tail the data can cover. Where inputs are varied and messy, accumulated edge cases are the product. This is the core of the case for vertical AI: narrow domains give data a place to concentrate.

The loop includes outcomes, not just activity. Knowing what customers did is useful. Knowing what happened afterward is what lets the product learn what "good" means.

The closest cousins of the data loop in classic strategy are network effects and switching costs. Data can feed both, but it isn't automatically either. A network effect means each new user makes the product more valuable to every other user. A data loop usually means each new user makes the product a little better for everyone, which is a weaker thing.

When It's Overrated

This is the part founders and investors tend to skip. Martin Casado and Peter Lauten's essay "The Empty Promise of Data Moats", published in 2019, makes the skeptical case directly. Their central argument is that there generally isn't an inherent network effect from merely having more data, and that what looks like one is usually a scale effect, with a crucial difference from classic economies of scale: the cost of adding unique data to your corpus may go up while the value of incremental data goes down.

Their reasoning gives a useful checklist of failure modes.

Diminishing returns

The essay's chatbot example is the cleanest illustration. Collecting the first slice of queries covers the common questions and delivers big gains. By the time a company has seen 40 percent of queries in the example, there's no advantage left in collecting more. The curve flattens, and a rival who reaches the flat part of the curve is on equal footing. A power law in a research paper and a flat curve in a business plan are the same shape seen from different distances.

The cold start is bounded

The same essay argues that bootstrapping a minimum viable corpus is achievable, and so can't be a durable moat on its own. If a new entrant can reach "good enough" with purchased data, synthetic data, a few design partners or a strong base model, the incumbent's head start shrinks to a head start on the long tail only.

The long tail is expensive

In a later piece, Taming the Tail, Casado and Matt Bornstein argue that long-tailed distributions are endemic to ML problems and that this can mean diseconomies of scale: a growing share of the work lands where value per instance is lowest. They also report that development costs for AI applications can run three to five times higher than in typical software. A flywheel that has to feed on rare cases is a flywheel with rising friction, not falling friction.

Public data erodes advantages from the other side

Foundation models are trained largely on public data, and that stock is finite. Epoch's research on limits of LLM scaling based on human-generated data projects that models will be trained on datasets roughly the size of the available stock of public human text between 2026 and 2032. The relevant lesson for startups isn't about the end of data. It's that as general models get better at what public data can teach, the part of a product that depends on public data stops being a differentiator. What's left is whatever the public can't see.

Better base models can absorb your edge

If a startup's data advantage is "we fine-tuned on examples that improve a common task," a stronger foundation model released next quarter may close that gap without any new data at all. This is the practical risk behind the question in AI wrapper vs AI native: an advantage that lives in the model layer you don't control is rented, not owned.

Feedback quality, not quantity

Volume is easy to measure and easy to brag about. Quality is harder. Thumbs-up buttons get clicked by people in a hurry, labels drift when different reviewers disagree, and logs record what users did, not whether it was right. A flywheel fed by noisy feedback can spin without improving anything. The cleanest signal is usually an expert's correction or a verified outcome, and these come from evaluation and trust practices that most teams build late.

A Practical Test: Do You Have One?

Questions that separate a working flywheel from a slogan:

The curve question matters most. A startup that can plot quality against volume of proprietary data knows which of three situations it's in: still climbing, where more data is worth a lot; flattening, where additional data is worth little; or already flat, where the moat has to come from somewhere else. Most founders haven't drawn this plot.

A second habit helps: keep a fixed evaluation set that never enters training. Without one, any improvement claim is anecdote. Related measurement ideas appear in AI product adoption metrics, and the infrastructure side of capturing usage is covered in telemetry loops for in-product AI and the product telemetry advantage.

How a Flywheel Fits Into a Larger Moat

Even a strong data flywheel rarely stands alone. In classic strategy language, an economic moat is a durable advantage, and durable usually means several reinforcing sources. Data tends to pair with:

• Workflow position. Being where the work happens is how the data gets generated at all.
• Distribution. More customers means more data, and the question of how you win those first customers is separate, see AI distribution moat.
• Switching costs. Accumulated customer-specific configuration makes leaving painful, which keeps the data coming.

Treat the flywheel as an amplifier of other advantages rather than a replacement for them. Casado and Lauten's warning is that treating data as a magical moat can misdirect founders from what they actually need to win. The useful version of the idea is narrower and more testable: for this product, in this domain, does proprietary usage data measurably improve results in a way a rival can't quickly match?

Common Mistakes

• Counting logs as data advantage. Volume without labels or outcomes teaches little.
• Assuming the curve never flattens. Diminishing returns apply to your data as much as anyone's.
• Ignoring data rights. If customer contracts don't allow learning from their data, the loop is legally blocked.
• Building the loop around a rented model. If the improvement can be replicated by upgrading a base model, it's not yours.
• Measuring training volume instead of customer-visible quality. The loop closes when customers notice.
• Skipping the control group. Without a held-out evaluation set, you can't tell improvement from drift.

Key Facts

• Origin of the metaphor: Jim Collins describes the flywheel as accumulated pushes in a consistent direction, with no single killer innovation.
• Scaling research: Hestness et al. found power-law generalization error scaling across four ML domains, and Kaplan et al. reported loss scaling as a power law with model size, dataset size and compute.
• Feedback beats size in one case: InstructGPT outputs from a 1.3B model were preferred to those from the 175B GPT-3, despite 100x fewer parameters.
• The skeptical case: Casado and Lauten (2019) argue that having more data does not by itself create a network effect, and that the cost of unique data may rise while its incremental value falls.
• Long-tail economics: Casado and Bornstein (2020) say AI application development costs can run 3-5x higher than typical software.
• Public data limit: Epoch researchers project training datasets will roughly match the stock of public human text between 2026 and 2032.