Does data really create lasting competitive advantage for startups?

a16z's Martin Casado and Peter Lauten argue that for enterprise AI startups, data by itself is a weak and often illusory competitive moat. They separate the hyped idea of "data network effects" from the more common "data scale effect," and find even the scale effect frequently breaks down: "the cost of adding unique data to your corpus may actually go up, while the value of incremental data goes down." They cite a study (shared with permission) by Arun Chaganty of Eloquent Labs on a customer-support chatbot's training data, where "20% of the effort into the data distribution tends to only get you around 20% coverage of use cases," with coverage approaching "an asymptote of 40% intent coverage" — in their example, no amount of further collection moves the product past that ceiling.

The mechanism is a reversal of ordinary economies of scale. With true network effects, "user acquisition costs go down over time, because the value of joining the network increases," and users have incentive to propagate the network themselves. Data has neither property: new data increasingly duplicates what's already in the corpus ("some of the new data you acquire already overlaps with your existing corpus"), the easy common cases get covered first while what remains is a noisy long tail of one-off queries, data goes stale ("streets change, temperatures change, attitudes change"), and once a pattern is understood, competitors can replicate the "proprietary insight" and erode the edge. Cost of capture rises while marginal benefit falls — the opposite of the flywheel founders assume they're building.

This is a pre-LLM (2019), business-strategy framing of a claim that recent ML work makes more precise: that the value of data is not intrinsic to its volume but depends on what it contributes relative to what a model or corpus already has. What makes synthetic data work across different domains and models? makes the equivalent point inside model training — that the payoff from a dataset's properties depends on domain, model, use case, and scale, not on a universal recipe. And the decomposition in How do quality, diversity, and complexity affect synthetic data differently?, where diversity rather than raw quantity drives out-of-distribution generalization, gives a mechanism for why a16z's "long tail" of one-off queries matters disproportionately: coverage of rare cases, not corpus size, is what a scale effect is actually buying you.

The argument rests on one domain (enterprise chatbot support) and one era of AI products; the authors themselves note a counterexample where more data keeps paying off — "if you have a cancer screen that is 85% accurate, it is far more likely to get used than one which is 80% accurate" — so the claim is a tendency, not a law, and the piece gives no general rule for telling which regime a given product is in ahead of time. It also predates foundation-model-scale pretraining, so it says nothing about whether the same cost/value reversal holds for the giant general-purpose corpora behind today's LLMs, only for bespoke, narrow, enterprise-specific data collection.