Moats in AI: The Test Each Claimed Advantage Has to Pass

Most lists of AI moats are lists of nice things to have. A moat is something narrower: a cost asymmetry that a competitor cannot close by spending a normal amount of money, and that does not close on its own while you sit still.

What a moat has to be

Three properties, and all three are required. An advantage is a moat when a competitor reaching parity must spend materially more than you did; when the gap does not close simply because time passes; and when the advantage grows as a by-product of ordinary operation rather than needing to be re-bought.

Almost everything called a moat fails the second property. A lead is not a moat. Being first is not a moat. Having spent a lot is not a moat — it may be the opposite, if the cost of doing the same thing is falling for everyone who comes after you.

The condition: accumulation beats decay

Write the advantage as a stock that both accumulates and decays:

dA/dt = g - d * A

  g   accumulation rate: how much advantage ordinary
      operation adds per unit time
  d   decay rate: how fast a competitor's cost of
      replication falls, or your advantage erodes

Steady state:  A -> g / d

An advantage is durable when g is meaningful and
d is small.
An advantage is a lead, not a moat, when g = 0 —
whatever you built stops growing and only decays.

The useful move is to estimate d honestly for the thing you are claiming. d is not about your competitor’s effort; it is about how quickly the world makes your achievement cheaper to reproduce. In a field where the cost of reaching a given capability has been falling quickly, d is large for anything defined by capability, and that is the crux of the whole argument.

Three questions

• Does it accumulate from use? Something that grows every day you operate has g > 0. Something bought once has g = 0 and only decays.
• Can a competitor buy it? If the advantage is available for money on an open market — hardware, an open-weight model, an advertising channel, a dataset that is licensed to anyone — it is a cost, not a moat. Costs are matched by anyone with capital.
• Is it getting cheaper for the next entrant? If reproducing what you did costs less each year, you are on a treadmill. This is the question that disposes of most technical advantages in this field.

The candidates, scored

Model weights

Fails the third question decisively. The cost of reaching any given capability level has fallen repeatedly through better data, better training methods, distillation from stronger models, and cheaper compute. A checkpoint is therefore a depreciating asset: it does not get better while you hold it, and the cost of matching it goes down. Weights can be a lead of some months, and treating that lead as a moat is the most common error in the category.

Data

Conditional, and the conditions are stricter than the claim usually allows. A data advantage is real when it is generated by operation rather than purchased, when it cannot be reconstructed by someone who did not have it, and — the one that is almost never checked — when more of it demonstrably improves the product. Many proprietary corpora fail the third: the model is already good at that distribution, and additional examples move nothing you can measure. If you claim a data moat, the evidence is an ablation on your own evaluation set showing quality as a function of data volume. If that curve is flat, there is no moat, whatever the volume.

Distribution

The strongest candidate, and the least discussed in technical circles. Being the surface where the work already happens — the inbox, the editor, the record system, the device — is not getting cheaper to acquire; if anything the reverse. It accumulates through ordinary operation, and it cannot be bought outright. This is why a capability advantage often loses to a distribution advantage: the capability arrives at the incumbent eventually, and the users do not arrive at the challenger.

Switching costs

Real, but only where they are real. An API call has a switching cost near zero. A system holding years of records, permissions, audit trails and integrations has one that is very large, and it accumulates without anyone working on it. Everything covered under where switching costs actually live applies here, viewed from the other side of the table.

Scale in serving

Partly real and bounded. Higher utilisation and better throughput per accelerator lower unit cost, and both improve with volume. But the inputs are rentable, so a well-capitalised entrant can buy the same hardware, and much of the software that raises throughput is published or open. It qualifies as an advantage in the way that operational excellence in a logistics business does — real, defensible with effort, never permanent.

Regulatory and compliance posture

Underrated and growing. Certifications, audit history, data residency commitments and an approved-vendor status take calendar time that money cannot compress, and the requirements ratchet upward. This is a moat with a positive g and a small d, and it is boring, which is why it is undervalued in conversation and priced correctly by enterprise buyers.

Brand and default status

Real where the buyer cannot easily judge quality, which describes most consumer and much enterprise purchasing of AI, since evaluating a model properly is expensive. Being the safe default is worth a real premium and accumulates through use.

What this means if you are building

The pattern across those rows is consistent: advantages defined by capability decay quickly, and advantages defined by relationships, accumulated state and institutional position decay slowly. That is not a claim about this technology being special. It is what happens in any field where the technical frontier moves faster than the institutional one.

• Ask what your product accumulates. If nothing gets better on its own as customers use it, g = 0, and you are defending a lead.
• Prefer advantages that are boring to describe. An approved-vendor status and five years of a customer’s records are worth more than a prompt nobody else has.
• Assume the model layer is contested. Building on the assumption that your model advantage persists is the assumption most likely to be falsified, per the drift argument.
• Do not confuse a moat with a business. Plenty of good businesses have no moat and compete on execution. The mistake is pricing and planning as though a moat exists when it does not.