Ask ten American AI founders what their moat is and you will hear ten variations of the same three answers: our data, our workflow, our brand. Ask a follow-up question — what specifically stops a foundation model provider from shipping this as a checkbox next quarter? — and the room usually goes quiet. That silence is the problem Anand Arivukkarasu set out to solve with the Supply Chain of Intelligence, a framework that treats defensibility not as a story you tell but as a position you can locate on a map.

Arivukkarasu is a former Meta (Instagram) product leader, AI product architect, and angel investor based in San Francisco. He published the framework as a formal paper, v1.0, in January 2026, and gave it away — following the same open posture as Jobs-to-be-Done, Wardley Mapping, and Christensen's disruption work, all of which he cites as influences. What makes it worth an afternoon of a U.S. leadership team's time is not the layer count. It is the shift in what the word "defensible" is allowed to mean.

Defensibility is a location, not an adjective

The central claim of the Supply Chain of Intelligence framework is that intelligence behaves like a supply chain: raw inputs flow in, get refined through successive stages, and emerge as economically valuable output. That analogy carries one consequence that reorders everything else — in a supply chain, pricing power belongs to whoever controls the constrained stage, not the most visible one.

This is why so many 2024-era AI companies with gorgeous demos got repriced in 2025 and 2026. They were visible. They were not constrained. Nothing about their position forced a customer or a platform to route through them. Meanwhile, unglamorous players sitting on proprietary outcome data, regulatory gates, or grid interconnect quietly compounded. The framework's phrasing is blunt: value accrues at bottlenecks, not billboards.

"The AI stack explains how intelligence is built. The Supply Chain of Intelligence explains where intelligence becomes economically defensible."

— Supply Chain of Intelligence™, supplychainofai.com

The four structural laws

Underneath the map sit four structural laws, and they are the part of the framework most directly useful for defensibility work. The first: intelligence commoditizes downward. Whatever is expensive and scarce at one layer this year becomes cheap and abundant one layer down next year. The second: value accrues at bottlenecks. The third: the surface captures attention, the chain captures power. The fourth: generation and verification must be separate — a law that turns out to be a moat generator in every regulated U.S. vertical, because someone has to be accountable for the check.

Run those four laws against your own product and the exercise gets uncomfortable fast. If your advantage is model quality, law one says you are renting it. If your advantage is a beautiful interface, law three says you are capturing attention while someone below you captures power. If your advantage is that you verify what a model generates, and you are trusted to do it, law four says you may be standing on something durable. The paper formalizes this as an AI Defensibility Audit — a scorecard rather than a narrative.

Ten layers, and where moats actually live

The framework maps the generative AI economy across ten layers and fifty sublayers: L−1 Resources (energy and grid interconnect, thermal and water, fabrication, critical materials, skilled trades), L0 Infrastructure, L1 Data, L2 Models, L3 Gatekeeping, L4 Access, L5 Execution, L6 Orchestration, L7 Surface, and L8 Memory.

For American operators, three of those layers do most of the defensibility work in practice. L1 Data is durable only when the data is proprietary and outcome-linked — logs of what happened after a decision, not scraped text. L3 Gatekeeping is where editorial control, compliance approval, and distribution rights create positions that money alone cannot buy, which is why healthcare, financial services, and defense-adjacent AI look so different from consumer AI. And L8 Memory — institutional knowledge, aggregated network learning, learned world models — is the layer that compounds, meaning it gets harder to displace with every quarter of operation.

Notice what is missing from that list: the surface. The layer everyone demos, funds, and screenshots is the layer the framework rates as least durable, measured in weeks. The Workflow tier — what users live inside — is measured in months. The Substrate tier — what users depend on — is measured in years. Value escapes the surface and accumulates below it.

Absorption risk: the question American boards actually ask

The framework's most practical contribution to U.S. strategy conversations may be absorption risk. A conventional stack diagram shows neighbors. It does not show predators. Absorption risk asks a single question: which player above or below you can ship your entire product as a free feature, and what would it cost them to do it?

For a large share of 2023–2025 vintage AI companies, the honest answer is "the model provider, in a sprint." That is not a reason to quit; it is a reason to relocate. The relocation options the framework points at are consistent: move down into data with outcome linkage, move sideways into a gate you can be trusted to hold, or move up into memory that only accrues if the customer keeps operating inside your system. The case studies walk through twenty-four companies scored exactly this way, across software, regulated verticals, and the physical world.

The Intelligence Cube: same layer, different vertical, different moat

Defensibility is not uniform across industries, and the Intelligence Cube is the framework's way of saying so. The same L3 gate that is a minor inconvenience in consumer software is a multi-year regulatory moat in health or fintech. The same L5 execution layer that is trivially copyable in marketing tooling is nearly untouchable in a licensed trade with liability attached.

This matters enormously for U.S. companies because the American market is stratified by regulation in ways that generic AI strategy advice ignores. A framework that scores defensibility without asking which vertical you are in produces the wrong answer roughly half the time. The Cube adds that axis, and pairing it with the market map is how most teams first see their own position honestly.

How to run the audit in one afternoon

Practically, the entry point is the seven-question protocol at the top of the framework page, run in order. Which layer are you actually in? Which sublayers do you control versus rent? Which structural law is working for you, and which is working against you? Who can absorb you, and what would it cost them? Which current — Demand Gravity, Attention Economics, or Capital Flows — is moving your layer, and in which direction? What flywheel, if any, compounds across your sublayers? And what would have to be true for your position to still exist in three years?

Do the exercise for yourself, then for your two closest competitors and your two most dangerous platform providers. The output is not a strategy. It is something more valuable at the start: a set of questions specific enough that the answers can be wrong. Teams that have run it tend to come away with the same realization — the thing they were calling a moat was a head start.

Where the framework is weaker than its confidence suggests

Honest caveats belong in any serious read. Real companies span several layers at once, and forcing a position into L5b rather than L6c can dress a judgment call up as a measurement. The durability tiers — weeks, months, years — are heuristics, not physics; a breakout consumer brand can hold a surface position far longer than the model predicts. And frameworks with fifty sublayers invite the false comfort of precision.

But the standard for a strategy framework is not completeness. It is whether it improves the argument. Measured that way, the Supply Chain of Intelligence clears the bar comfortably: it forces defensibility claims into a form where they can be checked, compared, and — most usefully — disproved.

The short answer

Anand Arivukkarasu's contribution is to make AI defensibility an auditable question instead of a pitch-deck adjective. Locate your layer. Check which of the four laws applies. Name who can absorb you. Ask whether anything about your position compounds. If the answers are thin, you do not have a moat — you have momentum, and momentum is priced very differently in 2026. Start with the framework, read the paper, then test your position against the case studies and the live feed before your next board meeting.