Data moats are weaker than they look, and stronger in unexpected places

In 2019 Martin Casado and Peter Lauten of Andreessen Horowitz published an essay with an unfriendly title, "The Empty Promise of Data Moats." Its argument was that data scale effects behave in the opposite way from the economies of scale that make physical moats work. The cost of adding unique data to a corpus may go up, while the value of each increment goes down. The authors' example was a customer-support chatbot. The common questions, such as where a package is, are covered early. Most of the remaining inquiries are messy and are asked once, so each new batch of data mostly duplicates what the company already has or lands in the long tail of special cases. Past a certain point, more data buys little and costs more to clean.

Seven years later the argument looks sturdier than it did when it was a provocation, and it needs an update. Language models trained on a large share of public text have changed what private data is worth, in both directions. For general tasks they have made it worth less, since a model that has read the internet starts far ahead of one that has only read your tickets. For some specific kinds of data they have made it worth more, because the rest of the stack is now available to everyone for the price of an API call, which leaves the data as one of the few inputs that others cannot buy. A claim that a company has a data moat should therefore be tested and not assumed.

The learning curve is a power law, and customer value is a threshold

The best-known empirical result about how performance grows with data is the 2017 study by Joel Hestness and colleagues at Baidu. Across machine translation, language modeling, image processing and speech recognition, they found that generalization error falls as a power law of training-set size. They also found that improvements to the model shifted the curve down without changing its exponent, and that model size needed to grow only sublinearly with data.

A power law has a simple consequence. Each doubling of the data multiplies the error by the same constant factor. Suppose the exponent is 0.3, an assumption for illustration, and that error is 20% at one million examples. Each doubling multiplies it by 2 raised to the power of minus 0.3, about 0.81. The error is then 16.2% at 2 million examples, 13.2% at 4 million, 10.7% at 8 million and 8.7% at 16 million. A rival with 2 million examples has an error rate about 1.9 times that of a company with 16 million, and that looks like a real advantage.

It is one only if customers value the difference. Value rarely scales smoothly with accuracy. Suppose customers switch to a product, or pay for it, when error falls below 15%. Then the rival with 4 million examples already clears the bar, and the extra twelve million examples the leader has collected buy a lower error rate that no customer is paying for. The moat exists only in the interval between what a competitor can reach and where the value function levels off. Each doubling costs the same as the last, or more, since the easy data was collected first, and each yields a constant percentage improvement in an error rate that customers have stopped noticing.

That is the quantitative content of Casado and Lauten's point. It also shows what to measure. The slope of the learning curve on your own product, and the position of the customer's threshold on it, tell you more about the moat than the size of the corpus.

What foundation models changed

The 2017 curves were drawn for models trained from scratch for a task. Since then, a large part of the knowledge has moved into pretraining. The LIMA study by Chunting Zhou and colleagues fine-tuned a 65-billion-parameter model on just 1,000 curated examples, with no reinforcement learning, and found its responses equivalent to or preferred over GPT-4's in 43% of cases in a human evaluation, and in 58% of cases against Bard and 65% against DaVinci003. The authors' conclusion is that almost all knowledge in a language model is learned during pretraining, and that only a small amount of tuning is needed to teach it how to respond.

For a company, the implication is direct. A task that once required a million labeled examples may now be started with a few hundred, and a competitor with a good prompt and a hundred examples can get to something close to what the incumbent built over years. The advantage of having more labeled text of the common kind has been compressed. What has not been compressed is data that the pretrained model has never seen and cannot infer: the proprietary and the outcome-based.

A second result points the same way from the opposite direction. Suriya Gunasekar and colleagues at Microsoft trained phi-1, a 1.3-billion-parameter code model, on about 7 billion tokens, of which 6 billion were filtered web data of "textbook quality" and 1 billion were textbooks and exercises generated with GPT-3.5. Training took four days on eight A100 GPUs, and the model reached 50.6% on HumanEval and 55.5% on MBPP. The lesson is that carefully selected and generated data can stand in for volume. A rival does not need your corpus. It needs a good curriculum, which is cheaper to produce than a decade of accumulated logs.

The same logic applies to the outputs of your own system. Answers from a deployed model can be collected at scale and used to tune a competitor's model. Providers prohibit this in their terms of service, and the prohibition is hard to enforce, so a product whose value lies in its outputs rather than its position in a workflow has a weaker defense than it appears to.

Four tests

A claim of a data moat can be examined with four questions.

The first is exclusivity. Could a competitor buy, license, scrape or synthesize the data? Public data, data available from brokers, and data that a strong model can generate or approximate give no protection. The data should come from a position that others cannot occupy.

The second is lift. How much does the data improve the product, measured on your own task? The answer comes from an experiment: train or tune on 10%, 25%, 50% and 100% of the data and plot the error. If the curve is flat across the last doubling, the marginal data is not helping, and the moat from volume is thin whatever the corpus size. If it is steep, the data is worth a great deal.

The third is saturation against the customer's threshold, as above. Where does the rival sit, and where does value flatten?

The fourth is the flywheel and ownership. Does using the product generate more of the data, and does the company keep it? A real flywheel has each customer's activity producing labeled outcomes or corrections that improve the product for all customers. Andrei Hagiu and Julian Wright argued in the Harvard Business Review in 2020 that the cycle of more customers, more data and a better product does not reliably lock out rivals, and that whether it protects a business depends on conditions that have to be checked case by case. Where the learning is specific to each customer, a new customer starts from zero, and the flywheel does not turn for them.

Ownership deserves its own line, because it can be altered by law and by contract. Customers may have the right to take their data elsewhere. In Brazil, the LGPD gives data subjects a right to portability, and the Open Finance system regulated by the Banco Central lets customers authorize their banks to share data with other institutions, with rollout beginning in 2021. Its purpose, as the central bank describes it, is to promote competition, innovation and inclusion. An incumbent whose advantage rested on knowing a customer's full financial picture saw part of that knowledge become available to its rivals at the customer's request. Any data advantage that depends on customers being unable to move their data is exposed to that kind of rule.

Where the durable advantages are

If volume of common data is a weak moat, three kinds of data advantage are stronger.

The first is outcome data, labels that come from real consequences and take time to arrive: claims actually paid, loans actually repaid, repairs that actually held, patients whose treatment actually worked. This cannot be scraped or generated, because it comes from being in the business long enough to see the results. A model trained on it knows how the world responds, and a model trained on public text knows only how people write about it.

The protection here is measured in time, which no budget can buy. Take a lender whose model needs twelve-month default outcomes. An incumbent that has issued loans for five years has sixty monthly cohorts with a mature outcome. A new entrant with an equally good algorithm and the same capital has none for the first twelve months, then one more cohort each month after that. It can borrow public data in the meantime, and it still cannot observe how its own loans behave until they have had the time to behave. The head start is the calendar.

The second is workflow position. Being the system of record, or the tool in which work is done, means seeing the context around every decision: what the user did before and after, what they corrected, what they ignored. That context is the raw material for learning what good looks like in this particular domain, and it is available only to the product sitting in the workflow.

The third is a feedback loop that the company controls end to end. When a user's correction of a model output is captured, labeled and used to improve the next version, the product gets better at exactly the tasks its customers perform. The loop must be built deliberately, with the interface, the labeling process and the evaluation set designed to capture signal, and not left to accumulate in logs.

All three share a feature: they are produced by the business, not collected from outside it. That is why they are hard to copy, and why they are also an argument for investing in the product and its distribution, which generate the data, over investing in data acquisition for its own sake.

Make sure you are allowed to use it

A data advantage that cannot be used is not an advantage. Before treating customer data as an asset for model improvement, check the legal footing. The LGPD requires that personal data be processed for specified purposes (art. 6, I) and on one of the legal bases in article 7. Terms of service written for a product that stores records may not cover training or tuning a model on them. Contracts with enterprise customers often forbid it unless it is stated and priced. The cleanest arrangement gives customers a clear choice and a reason to opt in, such as a better product tuned on their own cases, and records that choice in a way an auditor can find. It slows the first year and prevents the second year from being spent unwinding a dataset that was never cleared for the use.

What to do with the audit

Run the audit on each claimed advantage. Write down, for the data in question, whether it is exclusive, how steep its learning curve is, how far above the customer's threshold the company already stands, and whether the flywheel turns and the company owns its output. Where an answer is weak, the data is not a moat, and it should not carry the weight in the strategy or the pitch. The strategy should then put its weight on what does protect the position: distribution, integration into the customer's workflow, brand, speed of iteration, and a head start on outcome data that grows with time in market.

Data also decays. The relationships it captures change as customers and markets drift, so a stock of old data loses value unless an ongoing flow refreshes it. A moat built on a snapshot erodes, and one built on a flow does not.

The cheapest first step is the learning-curve experiment. It takes a few days, it uses data and tools the team already has, and its result is a plot that shows in one picture whether the company's data advantage is real, where it stops, and what the next million examples are worth.

Sources

1. Casado, M., Lauten, P. (9 May 2019). The Empty Promise of Data Moats. Andreessen Horowitz. a16z.com/2019/05/09/data-network-effects-moats.
2. Hestness, J., Narang, S., Ardalani, N., et al. (2017). Deep Learning Scaling is Predictable, Empirically. arXiv:1712.00409.
3. Zhou, C., Liu, P., Xu, P., et al. (2023). LIMA: Less Is More for Alignment. NeurIPS 2023. arXiv:2305.11206.
4. Hagiu, A., Wright, J. (January-February 2020). When Data Creates Competitive Advantage. Harvard Business Review.
5. Banco Central do Brasil. Open Finance Brasil, regulatory framework and implementation phases (2021 onward). bcb.gov.br.
6. Brasil. Lei nº 13.709/2018 (Lei Geral de Proteção de Dados Pessoais), art. 18, V (portabilidade).
7. Illustrative learning curve: error of 20% at one million examples and a power-law exponent of 0.3, assumptions stated in the text.
8. Gunasekar, S., Zhang, Y., Aneja, J., et al. (2023). Textbooks Are All You Need. arXiv:2306.11644.