Data moats: what is real and what is hype

Why the phrase is overused

Almost every AI pitch mentions a data moat. Sometimes it is accurate. More often the claim means that the company has accumulated some data and expects it to matter. Models trained on broad public data are widely available, so the question is what a company has that its competitors, including well-funded ones, cannot easily obtain or recreate.

Four tests

1. Access. Is the data hard for others to get? Exclusive contracts, regulated access or physical presence in a workflow can create real barriers. Data scraped from public sources rarely does.
2. Relevance. Does it improve the specific product in ways customers notice? Data that is large but only loosely related to the task adds cost without advantage.
3. Feedback. Does using the product generate new, labeled data? The strongest position is a loop in which outcomes and corrections flow back into the system.
4. Rights. Can the company legally and contractually use the data as it intends, and will that remain true? A moat without rights can be removed by a single contract term.

A company that passes all four has something. One that passes one has a story.

Real versus hype in practice

Looks like a moat	Often is not, because	Real version
Millions of documents	Generic text is easy to substitute	Documents tied to verified outcomes
A large customer base	Customers can leave with their own data	Data that stays useful only inside the product
A fine-tuned model	Base models improve and erase the gap	Proprietary evaluation data that guides every update
Years of history	Old data may no longer match current behavior	Fresh data from an ongoing workflow

Where durable advantage tends to come from

In our view, the most defensible positions are usually less glamorous than the term suggests. They sit in workflows where the product is part of the customer's daily process, so each use produces a labeled outcome. They sit in domains where data is fragmented or regulated and the company has done the slow work of earning access. And they sit in evaluation: a company with a rich, private set of test cases can choose and tune models faster and more safely than one without. That last point is easy to overlook, and it is not copied by reading a competitor's website.

What decays

Moats erode. Customer behavior changes, regulations move, and competitors with similar access appear. Ask how the company keeps its data fresh, how it detects drift, and what share of its advantage depends on any single source or contract. If the answer is "we have a lot of it", dig further.

Does being early help?

Sometimes. A head start in capturing outcome data can compound if the loop is well designed. But a head start in the wrong data is just a head start in the wrong direction. We would rather see a small, well-labeled dataset tied to a clear task than a large, undifferentiated archive.

A hypothetical contrast

The lesson is not that Company B is safe. Another firm could offer a better clinic product or negotiate similar access. The lesson is that Company B has at least begun a loop that others would have to start from zero.

Rights are part of the asset

Data advantage can be fragile in a quiet way. A customer contract may restrict use to that customer alone. A privacy law may limit retention or require deletion. A provider's terms may change what you can do with outputs. Investors and buyers should read the contracts as carefully as they read the architecture. We would rather see a modest dataset with clean permissions than a large one with unclear ones.

What to ask

• Where did the data come from, and who agreed to what?
• What new data does each customer interaction create?
• If a competitor started tomorrow with the same models, what would they lack?
• How do you test whether more data actually improves results?
• What happens to your advantage if a key contract ends?

These questions fit early companies as readily as established ones. Our notes for startup founders describe how we think about the earliest evidence, and the checklist in evaluating an AI business in diligence shows where this fits in a wider review.

This article is for general information only and is not investment, legal, tax or accounting advice.