Data moat claims: what counts and what does not

A data moat claim counts when four things are true and can be shown: the next unit of data makes the product measurably better; a competitor could not get equivalent data at a similar cost; the company has the rights to use the data as it does; and the loop closes, so that use of the product produces the data that improves it. A claim that passes all four is rare. A claim that passes none is the default, because every AI deck says "proprietary data" and the slide costs nothing.

This page gives the four tests, the evidence for each, and what to write when the claim fails, because a data asset with no moat is still an asset, and the report should say what it is worth.

What does the a16z argument say?

The a16z argument says that what founders call data network effects are usually data scale effects, and that scale effects are a weak defence because the value of additional data falls as the corpus grows while the cost of acquiring it rises. Reveneau's diligence asks for the quality-against-data curve before any discussion of the moat, because the curve is the whole argument in one chart and the company either has it or has never measured the thing it is claiming.

Martin Casado and Peter Lauten published "The Empty Promise of Data Moats" on 9 May 2019. Its central claims, quoted: "There generally isn't an inherent network effect that comes from merely having more data." On diminishing returns, using a chatbot example, "after 40% of queries collected, there is actually no advantage to collecting more data at all." On cost, "getting the next piece of data tends to become more expensive to capture over time," so that "the cost of adding unique data to your corpus may actually go up, while the value of incremental data goes down." And on the starting line: bootstrapping a "minimum viable corpus" is sufficient to start training against, and reaching it "clearly will not be a durable moat" [1].

Seven years on, the argument is stronger, because the models most startups build on were trained on most of the public internet. A corpus that would have been proprietary in 2019 is now a subset of what the provider already has. Foundation Capital's September 2026 analysis of model providers moving into applications makes the updated case: providers have a view across everything built on their platforms, and the data advantage that survives is a feedback loop captured inside a workflow with high exception rates, subjective success criteria and fragmented legacy systems, which is to say inside something the provider does not see and will not build [2].

What are the four tests?

The four tests are the curve, the cost of equivalent data, the rights, and the loop. Each has a specific piece of evidence, and the table gives the claim, the evidence, and what a failure means.

Test	Ask for	Passes when	Fails when
The curve	Product quality (eval score) plotted against data volume, from the company's own runs	Quality rises with volume across the range the company operates in	Quality is flat, or rose early and plateaued at a volume a competitor could reach
Cost of equivalent data	The source of the data and what it cost to acquire, per unit	A competitor would need years, a licence nobody else has, or a customer base to reproduce it	The data is public, purchasable, or generated by a process anyone can run
Rights	The terms under which the data was collected and the provider terms it flows through	Customers agreed to the use, and the provider may not train on it	Consumer terms, silent contracts, or data the company does not own
The loop	A diagram of how product use produces data that changes the product, and a dated example	Use produces labelled outcomes that the evals confirm improved the next version	Data is collected and stored; nothing downstream reads it

The curve is the test that most claims fail, and the reason is simple: measuring it requires an eval suite, because "better" has to be a number. A company without a suite cannot show the curve. It can only show the volume. The eval suite as a diligence artefact is the page on reading a suite, and the curve is the first thing to ask a suite to produce.

Why does the loop matter more than the corpus?

The loop matters more than the corpus because a corpus is a stock that a competitor can match and a loop is a flow that a competitor has to earn customer by customer. Casado and Lauten's point about the minimum viable corpus is that the stock stops mattering once it is large enough to train on [1]. What keeps mattering is whether each new customer's use of the product produces data the last customer benefits from, and whether that data is labelled by something the world would otherwise not label: a decision made, a correction entered, an outcome recorded.

The loop also has to stay the company's own. Data flowing through a model provider's API is subject to that provider's terms. Anthropic's commercial terms state that the customer "retains all rights to its Inputs" and "owns its Outputs", and that "Anthropic may not train models on Customer Content from Services" [3]. Terms differ between consumer and commercial products and between providers, so a company claiming a proprietary loop should be able to show that every hop of the loop runs under terms that keep it proprietary. Code provenance and licence exposure covers the same question for the code; the answer for the data goes in the same table.

Bessemer's August 2025 description of the fastest-growing AI companies, explosive adoption with low switching costs and margin compression [4], is what a data claim without a loop looks like from the outside: growth that a competitor with the same model and the same public data can take back. A working loop shows up in the same place a moat always shows up, which is retention. If churned customers replaced the product easily, the data was not making it better for them.

How do you test a data moat in diligence?

Test a data moat by asking for the curve first and then walking the other three tests with the evidence in hand.

1. Ask for the curve. Eval score against data volume, from the company's own runs, across the volumes it has actually operated at. If there is no eval suite, record "curve unavailable" and move on; the claim is untested by definition.
2. Ask where the data came from and what it cost. Per unit, with the source. Then ask what the same data would cost a well-funded competitor starting today, and whether the answer is money or time.
3. Read the rights. Customer contracts for the collection, provider terms for the processing, account types and dates. A period on consumer terms is a period in which the loop may have leaked.
4. Trace one loop, dated. One example of product use producing data that changed a later version, with the eval run that measured the change. A company with a loop has this in its release history.
5. Ask what the provider sees. Which parts of the loop run through a third-party model, and whether the provider could reproduce the loop from what passes through its API. The claim test on how to verify an AI claim can include an input designed to see whether the company's data changes the answer at all.
6. Read the churn notes. Where did lost customers go, and did they cite quality. A moat that does not show up in retention is a slide.

What to write when the claim fails

Write what the data is worth without the moat. A labelled corpus that a competitor could match in a year is still a year's lead, and a year is worth something at the right price. Training data that was expensive to collect and is now a subset of what the model provider has is worth its licence value and little more. A collection pipeline that produces useful data but feeds nothing is an asset waiting for an engineering project, and the project belongs in the post-close plan.

The one-line report reads: "Data claim: curve unavailable (no eval suite); corpus purchasable at a cost the company could not state; rights clean on commercial terms from March; loop open. Value the data as a lead." Or, when it passes: "Data claim: curve rises across operating range per the suite; source is customer outcomes unavailable elsewhere; rights clean; one dated loop traced in the June release. Value as a moat and verify at renewal data."

AI feature or AI wrapper is the page for the product-side version of the same question, model dependency risk covers what a loop is worth on the day the provider ships the product, and the AI startup due diligence guide puts the data question at the end of the money read because its answer depends on everything before it. Execution speed is the moat is our own view on what defends an AI product when the data does not.