Supply Chain of Intelligence™

- § Definition

- § Abstract

Every generative-AI product sits on a 10-layer supply chain: L-1 Resources, L0 Infrastructure, L1 Data, L2 Models, L3 Gatekeeping, L4 Access, L5 Execution, L6 Orchestration, L7 Surface, L8 Memory. The layers are the supply side. Across them flow three Currents, Demand Gravity, Attention Economics, Capital Flows, which decide whether a position at any layer compounds into a business or starves. Four structural laws govern how value migrates through the stack under those Currents. The Defensible Triangle, L1b Proprietary Data + L5 Deep Skills & Playbooks + L8 Compounding Memory, is the most common application-layer fortress. The word "agent" is not a layer; it is marketing for an L5 + L6 (+ L7 / + L8) package and must be decoded before it can be analysed.

The framework is descriptive, not predictive. It does not tell you which company wins. It tells you, for any given company at any given moment, which layers it actually owns, which layers it is renting, and which Current is about to move the value somewhere else. That is usually enough.

- § 1 · Why a supply chain, not a stack

Stacks imply order. Supply chains expose bottlenecks.

"AI stack" diagrams treat the industry as a tidy layer cake, chips on the bottom, models in the middle, apps on top, and imply that value flows cleanly upward. It does not. Value in generative AI behaves like value in any real industrial supply chain: it concentrates at scarcity, gets squeezed at bottlenecks, and migrates the moment a layer below you commoditises what you were charging for. Calling it a supply chain forces the right questions, who owns the scarce input, who controls the gate, who captures the margin when the layer above collapses in price.

The 10 layers below are the supply side. They describe what is being produced and consumed at each step from raw resources up to the surface a user touches. They are deliberately granular, 50 sublayers in total - because most of the interesting moats live one level below the layer name.

- § 2 · The 10 layers

The stack at a glance

1. L-1
   Resources

   What supports the chain. Energy, water, fabs, materials, skilled trades, the inputs the entire stack consumes.

2. L0
   Infrastructure

   The shovels. Chips, data centers, networking, cloud, edge, what is needed to process intelligence.

3. L1
   Data

   The raw input. What data do you have that nobody else can get?

4. L2
   Models

   Intelligence refinement. Rent early, build custom at scale.

5. L3
   Gatekeeping

   Trust, acceptance, approval. Can the system be allowed in?

6. L4
   Access

   Connectivity, permissions, integrations, the pipes layer.

7. L5
   Execution

   Applied skills and capabilities. Doing the actual work.

8. L6
   Orchestration

   Workflow, routing, coordination. How skills compose into outcomes.

9. L7
   Surface

   Interface, presentation, experience. How the user meets the intelligence.

10. L8
    Memory

    Retention, learning, compounding context. What the system remembers.

- § 3 · The 3 Currents

Horizontal forces across the chain

The 10 layers are vertical. The Currents are horizontal, they flow across every layer and decide whether a defensible position is also a viable business. Geopolitics and regulation are not Currents; they live at their native layers (L-1 and L3 respectively). There are exactly three.

Two of three Currents pointing at a layer is a tailwind. All three is a category. None is a press release.

- § 4 · The 4 structural laws

How value moves through the stack

- § 5 · The Defensible Triangle

L1b + L5 + L8, the application-layer fortress

The most common application-layer fortress is a three-layer stack: L1b Proprietary Data nobody else can acquire, L5 Deep Skills & Playbooks organisation-shaped enough that swapping the model out doesn't replicate them, and L8 Compounding Memory (network learning, institutional knowledge, learned world models) that gets better the longer the system runs. Two of three is a workflow product that improves; three of three is an intelligence gate that compounds.

The Triangle is not the only fortress shape. L3 over L2/L5 (Law IV) is a different fortress. L0 + L-1 (NVIDIA, the hyperscalers, the fab operators) is a different fortress. L4 + L5 + L6 + L8 (the platform-native execution stacks like Replit) is a different fortress. But for an application-layer company without a hyperscaler's balance sheet or a regulator's mandate, the Triangle is the shape to aim for, and the shape to test any defensibility claim against.

- § 6 · On the word "agent"

"Agent" is not a layer

"Agent" is marketing for a package whose minimum viable composition is L5 Execution + L6 Orchestration, almost always bundled with L7 Surface, and, if the system remembers across sessions, L8 Memory. L4 Access is the substrate the agent rides on (connectors, protocols, permissions), not the agent itself. Tagging an agent story as "an L4 play" is a category error that recurs in roughly half the analyses written in 2025.

When a company launches an agent, decode it in three steps. First, name the L5 capability, what work does it actually do, and is the work generic or domain-specific. Second, name the other layers it bundles - L1 proprietary data, L4 platform access, L8 cross-customer learning. Third, ask whether any of those bundled layers are structurally hard for the underlying L2 to replicate. If the answer is no, the agent is a wrapper on a clock; if yes, the agent is a Trojan horse for whichever deeper layer the company actually owns.

• Agent + L1 = fortress (Sierra, Harvey).
• Agent + L4 = railroad (Salesforce Agentforce, Microsoft Copilot).
• Agent + L8 = compounding system (the rare ones that get better the more they run).
• Agent + nothing = exposed wrapper that commoditises the moment the underlying L2 ships the same loop.

- § 7 · How to use this framework

A discipline, not a forecast

For an operator, the exercise is to map your own product to the 10 layers honestly. Mark the layers you own, the layers you rent, and the layers you are exposed to. Then run the four laws and three Currents against the map and write down, in plain language, what compresses you and on what timeline. The output is a defensibility statement that survives contact with the next L2 release.

For an investor, the exercise is the same applied to a target. Most pitch decks describe an L5 product as if it were an L1 + L5 + L8 stack. The framework gives you the vocabulary to distinguish the two, and the laws give you the timeline on which the distinction matters.

For an analyst, the framework is a discipline. Every claim about defensibility should name the layer; every claim about disruption should name the law; every claim about timing should name the Current. Loose vocabulary ("AI-native", "moaty", "platform play") is what made the last two years of AI analysis difficult to act on. Precise vocabulary is the contribution this paper is trying to make.