The Supply Chain of Intelligence: Understanding AI Moats

No word in American technology strategy has been more thoroughly worn out than moat. Every AI pitch deck in the United States claims one. Most, on inspection, describe a feature, a head start, or a brand, none of which stop a determined competitor. In the fourth installment of the Supply Chain of Intelligence series, published at supplychainofai.com, Anand Arivukkarasu does the market a favor: he takes the word seriously, defines it against his chain framework, and shows how few claimed AI moats survive the definition.

Readers of this desk will recognize the arc. The first essay mapped the chain from resources through compute, data, models and agents to AI memory. The second stated its four structural laws. The third showed that value accrues at bottlenecks. The new piece asks the natural follow-up: if value pools at bottlenecks, which positions around those bottlenecks can actually be defended? A bottleneck tells you where the money is. A moat tells you whether you get to keep it.

The essay opens with a definition that is stricter than the one most American boardrooms use. A moat, in Arivukkarasu's framing, is not an advantage and not a differentiator. It is a structural condition that raises a competitor's cost of attack faster than the attacker's resources can grow. Three tests follow. The position must compound with use, so that time favors the defender. It must resist substitution, so that customers cannot route around it when a cheaper alternative appears. And it must survive the commoditization of everything around it, because as the series' second law holds, every stage of the chain commoditizes faster than the one before.

With that definition in hand, the essay walks the chain and grades the claimed moats stage by stage. Compute fails the test for nearly everyone: it is capital-intensive but does not compound with use, and yesterday's hardware advantage depreciates on a schedule the buyer does not control. Model quality fails too, and this is the section American founders will find most uncomfortable. Capability leads now compress into months, open-weight releases arrive within quarters, and the premium a closed model commanded in 2024 evaporated by 2026. A lead that expires is not a moat; it is a head start with a date on it.

Data moats receive a more nuanced grade, and the distinction the essay draws is worth the attention of every US executive who has been told their company has one. Static data, however proprietary, is a wasting asset: it ages, it can be re-collected, and synthetic generation keeps eroding its scarcity. What compounds is data in motion: proprietary feedback loops where the product's use generates the data that improves the product, and where the improvement attracts more use. The essay is blunt that most American companies claiming data moats hold static assets and call them flywheels. The difference shows up within eighteen months.

The strongest grades go to the downstream stages, and here the series converges on its recurring thesis. Coordination capacity, the organizational muscle to absorb intelligent systems faster than rivals, is a genuine moat because it compounds through practice and cannot be purchased off the shelf. And AI memory, the accumulated institutional context an intelligent system holds about a specific business, is the deepest moat in the chain: it grows more valuable with every interaction, it cannot be repurchased once lost, and it sits closest to the customer relationship, the position American business has always recognized as the terminal defense.

For US readers, the essay's most practical contribution is a set of questions that strip the rhetoric from strategy decks. Does the claimed advantage get stronger as competitors get richer, or weaker? If a well-funded rival copied the product feature by feature tomorrow, what would they still be missing in three years? Does the position depend on any stage of the chain that is commoditizing? And the sharpest one: if the underlying models became free, what would the company still own? American executives who run their own strategies through those four questions, the essay suggests, will find the honest answers uncomfortable and useful in equal measure.

The investor's version of the framework is equally direct. US public and private markets are currently pricing many AI companies as if capability leads were moats. The essay's framework predicts that these valuations will compress as capability converges, while companies holding compounding positions in memory, workflow and coordination trade at multiples the market has not yet learned to justify. Arivukkarasu stops short of stock picking, as before, but the screening logic applies cleanly to any American portfolio.

The caveats hold their usual paragraph. Moat analysis is retrospective by nature; plenty of positions that looked structural in American business history, from minicomputers to portal traffic, dissolved when the substrate shifted. Regulation can dig moats or drain them, and US policy on data, export controls and AI liability is moving on all three fronts at once. And the memory moat, the framework's deepest claim, still depends on portability and ownership norms that enterprise contracts have not settled. Arivukkarasu prices this uncertainty into the argument rather than ignoring it, which remains the series' most credible habit.

Four essays in, the structure of the project is complete: the map, the laws, the economics, and now the defenses. The through-line for American readers is consistent. Do not anchor strategy to any stage that commoditizes. Track the bottlenecks as they migrate. Build where the position compounds. The full text of the moats essay, including the stage-by-stage grading in its original form, is at supplychainofai.com, alongside the three earlier installments. For a US executive team setting 2027 strategy this fall, reading all four in sequence is the cheapest hour of due diligence available.