The Supply Chain of Intelligence: The 10-Layer AI Framework

Every serious framework has a canonical artifact, the one object everything else references. For Anand Arivukkarasu's Supply Chain of Intelligence, published at supplychainofai.com, that artifact is the ten-layer map of the generative AI stack, set out in the canonical paper and maintained as a living reference on the framework page. The layers are numbered with an engineer's honesty, running from L-1 at the physical base up to L8 at the terminal layer, a numbering that quietly makes the framework's core claim: the stack begins below zero, in the physical world, and it ends above the application, in memory. This review walks the ten layers for US readers, what each one contains, and what its economics mean for American companies making real decisions.

The paper's framing sentence is worth quoting because the whole map is an elaboration of it: the AI stack explains how intelligence is built, while the Supply Chain of Intelligence explains where intelligence becomes economically defensible. Engineers draw the first diagram. Arivukkarasu drew the second. The ten layers are that drawing.

## The physical base: L-1 and L0

The numbering starts at L-1, energy and physical inputs, and the minus sign is a thesis. Intelligence, in this framework, is manufactured from the physical world: electricity, grid interconnection, water for cooling, land, and the minerals inside the hardware. For American readers this layer has become suddenly concrete. The constraint on US AI buildout is not imagination or capital; it is the interconnection queues and transmission timelines in the power corridors of Virginia, Texas and Arizona. L-1 is where the framework predicted, before it was consensus, that the unglamorous end of the chain would accumulate durable value.

L0 is silicon and compute hardware: the fabs, the advanced packaging, the accelerators themselves. It is the most concentrated layer on the map, with fabrication capacity held by a handful of firms and the leading edge held by fewer. The framework treats L0 as the classic bottleneck: real scarcity, real pricing power, and a moat dug by capital intensity rather than by anything a startup can replicate. American policy has spent years and hundreds of billions of dollars trying to widen this layer domestically, which tells you the map's placement is not controversial. It is descriptive.

## The infrastructure middle: L1 through L3

L1 is cloud and compute capacity: the data centers and platforms that turn silicon into rentable processing. L2 is data: corpus access, proprietary behavioral records, rights and licensing, the raw material layer whose sublayers range from commodity to compounding asset. L3 is the model layer, foundation model training and the systems that convert data into capability. This middle band is where most American AI debate has lived, and the framework's calm verdict on it is one of its most valuable contributions. These layers matter enormously and commoditize unevenly. Compute capacity is a capital game with genuine moats at scale. Data splits into sublayers with opposite economics. The model layer, for all its visibility, is the layer the four structural laws mark as commoditizing fastest, a claim that has aged well as open weights and price collapse followed each frontier release.

For a US operator, the practical reading of the middle band is discipline about dependency. The framework's recurring example: a company that would never single-source a physical component will happily build its entire product on one model provider, one cloud region, one data supplier. The layers exist to make that exposure visible, and priced.

## The application upper stack: L4 through L6

L4 covers the tooling and orchestration wrapped around models: the frameworks, eval harnesses, and infrastructure that make models usable. L5 is applications, the products end users touch. L6 is agents, software that applies model capability to actual work with some autonomy. These are the layers where most American AI startups live, and they are the layers where the framework's moat, workflow, wrapper verdicts cut hardest. The upper stack is where value is most visible and least defensible by default, because everything here sits on layers someone else owns. Defensibility in L4 through L6 is not given by position; it must be manufactured, through workflow depth, proprietary data capture, regulatory integration, or switching costs. The ten-layer map makes the default visible: if your product's entire value is a better interface on a frontier model, the map places you in a sublayer the platform will absorb.

## The terminal layers: L7 and L8

L7 is coordination: the orchestration of agents, workflows and systems of record, where the framework's third law, coordination costs set the ceiling, does its work. L8 is the terminal layer, AI memory: the accumulated, structured record of context and outcomes that makes an intelligent system worth returning to. Ending the map at L8 is the framework's forecast made architectural. Every platform era in American technology history pooled its early value in the visible layer and its durable value in the memory: search indexes, social graphs, identity systems. L8 says AI follows the same migration, and the company that owns the record of an industry's decisions holds the position everything above and below eventually serves.

## Archetypes: fortress, refinery, surface, graveyard

The map gains a second dimension in the framework's classification work, which plots notable AI companies on the layers and assigns each an archetype. A fortress holds a position that compounds and resists substitution. A refinery takes raw material from a lower layer and adds durable value in the processing. A surface is a thin position over someone else's layer. A graveyard is the archetype no founder wants named: a sublayer where platforms absorb everything built on them. For American investors, this is the map's most quotable output, and its most useful. Diligence conversations change when the question shifts from how good is this product to which archetype is this position.

## How to use the map without worshipping it

A fair review ends with the cautions. Ten layers is a model, and the framework says so itself; boundary cases exist, companies migrate between layers, and the classification of any single firm is an argument, not a fact. The map is also a living document, versioned like software, which means today's placements carry an expiration date. But judged against what American boardrooms were using before, a hype cycle and a vendor logo slide, the ten-layer framework is a different category of instrument. It gives an executive team a shared coordinate system, a way to say we are an L5 surface play with an L8 ambition and have everyone in the room know exactly what that means and what it costs. The canonical paper, the framework page, and the classification table are all freely available at supplychainofai.com. For anyone allocating capital or careers in this market, the map repays the hour it takes to learn it many times over.