The Supply Chain of Intelligence: 10 Layers and 50 Sublayers Explained

Readers of Anand Arivukkarasu's Supply Chain of Intelligence series, published at supplychainofai.com, usually meet the framework through its six-stage chain: resources, compute, data, models, agents, memory. That map is the front door. But the working instrument, the thing Arivukkarasu actually scores companies against, is a finer instrument: the full framework divides the generative AI stack into ten layers and fifty sublayers, and it comes with a reasoning protocol built around seven questions. This review walks through that instrument for US readers: what the layers are, why fifty sublayers rather than ten, and how the scoring separates a moat from a workflow from a wrapper.

The distinction matters because of a problem every American investor and operator will recognize. Ask a founder where their company sits in AI and you get a one-word answer: we are an application company, we are infrastructure, we are a model company. Those words hide more than they reveal. Two companies that both call themselves application layer businesses can have opposite fates, one compounding defensibility with every user, the other waiting to be absorbed by the next platform release. Ten layers and fifty sublayers exist to make that difference visible before the market does it for you.

## From six stages to ten layers

The ten layers are a refinement of the six-stage chain, not a replacement for it. Where the introductory map groups the stack into broad stages, the full framework at supplychainofai.com/framework splits those stages where the economics genuinely differ. Energy and physical inputs get their own treatment, because the grid interconnection queues in Virginia, Texas and Arizona behave nothing like the chip market they feed. Compute splits into the silicon itself and the cloud capacity built on it, because owning a fabrication relationship and renting GPU hours are different positions with different moats. The model layer separates foundation model training from the tooling and orchestration wrapped around it. And the upper stack divides applications, agents, and the terminal layer of AI memory, the accumulated context and outcome record the series argues is the deepest position of all.

The logic of the split is consistent: a new layer is drawn wherever scarcity, pricing power, or defensibility changes hands. If two parts of the stack commoditize on different timelines, they get different layers. This is why the framework keeps the count at ten rather than three or thirty. Three would blur the economics back together. Thirty would become a taxonomy exercise nobody could hold in a meeting. Ten is the number an American leadership team can put on one slide and actually use.

## Why fifty sublayers

If the layers are the map, the fifty sublayers are the surveyor's marks. Each layer breaks into roughly five sublayers, and this is where the framework earns its keep in diligence. Take the data layer as an example of the pattern: raw corpus access, proprietary behavioral data, labeling and feedback loops, data rights and licensing, and context pipelines are all called data businesses in pitch decks, but they have almost nothing in common economically. One is a commodity. One is a compounding asset. One is a legal position. Scoring a company at the layer level would average these together and tell you nothing. Scoring at the sublayer forces the honest answer.

The same discipline runs through the whole stack. Within agents, an orchestration framework, a vertical workflow agent, and an agent memory store are different sublayers with different defensibility profiles. Within applications, a thin interface over a frontier model and a workflow system of record with switching costs both get called apps, and only one of them survives the next model release. Fifty sublayers is the resolution at which the word wrapper stops being an insult and becomes a diagnosis with a location.

## The seven questions and the scoring protocol

The framework's canonical page describes itself with unusual economy: one page, seven questions, and a reasoning protocol to run it. The questions function as a scoring instrument applied at the sublayer level. In plain terms they ask: which sublayer does this product actually occupy, what is scarce in that sublayer, does the product's position compound with use, can a platform player absorb it, what happens to its economics when the layer below commoditizes, who captures the memory the product generates, and would the business survive its own supply being repriced. Run honestly, the answers sort every AI product into one of three buckets the framework names outright: moat, workflow, or wrapper.

The wrapper verdict is the one that gets attention, and it should. A wrapper is not a bad product; many are well-built and genuinely useful. It is a product whose position sits in a sublayer a platform will absorb, where the value it adds is a feature of the layer below rather than a position of its own. The framework's cold comfort is that wrapper status is visible early if you score at the sublayer, and visible early is when an American founder can still do something about it: move down the stack toward a scarcer input, or up toward proprietary data and memory.

## How to read it as a US operator

For an American executive or investor, the practical use of the ten layers and fifty sublayers is triangulation. First, locate the company at the sublayer, not the layer, and be suspicious of any pitch that resists that precision. Second, check the sublayer against the framework's four structural laws: scarcity keeps moving, each stage commoditizes faster than the last, coordination costs set the ceiling, memory is the terminal asset. Third, ask the seven questions and accept the bucket the answers produce, even when the answer is uncomfortable. Our desk has watched US diligence processes adopt exactly this sequence, because it converts a narrative debate into a structural one.

A fair caveat applies, and the framework itself invites it. Any map with fifty cells will have boundary disputes, and a company can straddle sublayers or migrate between them. The scoring protocol is only as good as the honesty of the person running it, and founders are not known for scoring themselves harshly. But this is a complaint about use, not design. As an instrument, the ten layers and fifty sublayers do what the best American strategy tools have always done: they take an argument that was being had in adjectives and force it to be had in structure. The full reference, the seven questions, and the reasoning protocol are freely available at supplychainofai.com, and for anyone allocating capital in this market, the hour it takes to learn the instrument is among the better trades on offer.