An industry-defining macroeconomic model for AI · across verticals

    A defensibility map for AI companies.

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

    Is your product a moat, a workflow, or a wrapper a platform will absorb? Supply Chain of Intelligence™ scores every AI product across 10 layers and 50 sublayers, from compute and data to workflows, surfaces, and memory, and tells you where value actually accrues.

    Supply Chain of Intelligence™, the 10 layers of the generative AI stack.

    Prepared by Anand Arivukkarasu, Ex-Meta (Instagram) Product Leader, as a free resource for product leaders.

    Supply Chain of Intelligence™ (SCoI)

    ★ = structurally defensible · 50 sublayers mapped

    See the full 10×5 grid →

    The definition

    The definition is evergreen. The 10 layers and 50 sublayers are the application. Which company sits where is a monthly reading.

    Run the Framework

    Category reframe

    Not another AI stack. A different question entirely.

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

    Axis
    AI Stack
    Supply Chain of Intelligence™
    Category
    AI Stack / AI Value Chain
    Contains both, and adds gatekeeping, currents, flywheels, absorption
    Question
    How is AI built?
    Where does value accrue?
    Lens
    Architecture
    Economics
    Unit
    Components
    Bottlenecks
    Behavior
    Static layers
    Dynamic system
    Discipline
    Technology
    Strategy
    Audience
    Engineering
    Investment & Product
    Output
    Describes
    Predicts

    The AI stack is one input to the Supply Chain of Intelligence — not its competitor.

    Where AI Transformation Happens · 01

    AI is transforming three things. This framework is about one of them.

    Most AI roadmaps confuse the three. They are not the same problem, they do not have the same defensibility, and they should not be scored the same way.

    1. 01

      Internal Operations

      Copilots, RPA, productivity. Cost out.

      Real ROI. Rarely a moat.

    2. 02

      Distribution & GTM

      AI in marketing, sales, support. Reach up.

      Easier wins. Easier to copy.

    3. 03You are here

      Core Product

      AI inside what you sell. The product itself becomes intelligent.

      Hardest. Most defensible. This site is about this.

    Supply Chain of Intelligence™ is a framework for area 03, Core Product. It does not score your internal copilots or your marketing automation. It scores whether the AI inside what you sell is defensible.

    Two Lenses · 02

    On Core Product, you need two lenses, not one.

    Most teams only use the first lens. They ship a real user need, score a viral launch, and then a platform absorbs them in a release cycle. The second lens is what this framework adds.

    LENS 01Necessary

    The User Lens

    JTBD · NMBA · ICP · positioning

    What job is the user hiring this for? What’s the next most valuable action? Who exactly is the buyer? This lens finds demand.

    Tells you if anyone wants it.

    +
    LENS 02The missing one

    The Intelligence Lens

    10 layers · 50 sublayers · defensibility

    Which of the 10 layers does your product actually own? Data? Workflow? Memory? Or are you a thin surface on someone else’s model? This lens proves defensibility.

    Tells you whether you survive the next platform release.

    User Lens without Intelligence Lens → a wrapper with traction. Both lenses together → a moat with users.

    See the 10 layers

    AI is a supply chain. Like gold: ore in the ground, refining, assay, retail, the ring on a finger. Value moves through 10 layers, most products sit on one, usually the wrong one.

    Read the full analogy
    Gold supply chainflow →

    Ground

    Shovels

    Ore

    Refinery

    Assay

    Railroads

    Jeweler

    Collection

    Storefront

    Record Book

    L-1

    Resources

    L0

    Infra

    L1

    Data

    L2

    Models

    L3

    Gates

    L4

    Access

    L5

    Execution

    L6

    Orchestration

    L7

    Surface

    L8

    Memory

    Supply Chain of Intelligencesame chain, different century

    The 30-second aha

    Three companies. Three different layers. Three different ways to be hard to displace.

    L1
    Bloomberg logo
    Bloomberg

    the data

    Owns the data nobody else can buy.

    L5
    Harvey logo
    Harvey

    the work

    Built deep inside the legal workflow, still contested by Claude.

    L8
    Sierra logo
    Sierra

    the memory

    Accumulates per-customer memory that makes leaving costly.

    In the framework: L1 Data · L5 Execution · L8 Memory. Own the three corners and you’ve built the Defensible Triangle.

    Same job. Different fate.

    Two AI-native products. Same wave. Opposite trajectories.

    L7
    Jasper logo
    Jasper

    Sat on one layer

    $1.5B ~$300M

    A thin UX layer over a general model. When the model owners shipped the same surface for free, there was nothing structural left to defend.

    LAYERS OWNED · L7 only

    L5
    Cursor logo
    Cursor

    Owned four layers

    $9B+ and compounding

    Owns the IDE workflow, the indexing pipeline, the agent loop, and the project memory. Every layer reinforces the others - the model is the only commodity in the stack.

    LAYERS OWNED · L4 · L5 · L6 · L8

    Same job. Different layers. Different fate. The map below shows which layers compound and which collapse.

    Same company. Two lenses.

    Where Cursor sits, through two frameworks.

    Same company, same facts. The AI stack gives you one word. The Supply Chain of Intelligence gives you a map — and an answer to the only question that matters: is this defensible?

    Through the AI Stack

    Application.

    · Done ·

    Categorization. No verdict. No mechanism. No flywheel.

    Through the Supply Chain of Intelligence™

    • L7SurfaceSurface: Owns the IDE — the daily writing surface.
    • L6OrchestrationOrchestration: Agent loop, multi-file edits, human-in-loop.
    • L5ExecutionExecution: Code-aware indexing & retrieval as a domain skill.
    • L8MemoryMemory: Project memory compounds per repo and per team.
    • L1DataData: Outcome data — accepted edits — feeds the loop.

    Flywheel: every accepted edit improves retrieval and the next suggestion.

    Platform risk: model is the only commodity in the stack.

    Defensibility: 4 owned layers reinforce each other.

    One word vs. a map. That's the difference between a category and a strategy.

    The Framework · One Image

    Screenshot this. Paste it anywhere. Cite it where it helps.

    The Framework · At a Glance

    Supply Chain of Intelligence™

    The 10 layers of the generative AI stack, not logistics, not freight.

    L8MemoryRetention, learning, compounding context. What the system remembers.
    L7SurfaceInterface, presentation, experience. How the user meets the intelligence.
    L6OrchestrationWorkflow, routing, coordination. How skills compose into outcomes.
    L5ExecutionApplied skills and capabilities. Doing the actual work.
    L4AccessConnectivity, permissions, integrations, the pipes layer.
    L3GatesTrust, acceptance, approval. Can the system be allowed in?
    L2ModelsIntelligence refinement. Rent early, build custom at scale.
    L1DataThe raw input. What data do you have that nobody else can get?
    L0InfraThe shovels. Chips, data centers, networking, cloud, edge, what is needed to process intelligence.
    L−1ResourcesWhat supports the chain. Energy, water, fabs, materials, skilled trades, the inputs the entire stack consumes.

    The Four Structural Laws

    I

    Intelligence commoditizes downward.

    II

    Value accrues at bottlenecks.

    III

    Surface captures attention; chain captures power.

    IV

    Generation and verification must be separate.

    10 layers · 50 sublayers · 4 laws. The map for every AI strategy conversation.

    SupplyChainOfAI.com

    WORKED EXAMPLE · SALES & MARKETING TECH

    Same category. Different layers. Different fates.

    SWIPE
    COMPANY
    Resources
    L−1
    Infra
    L0
    Data
    L1
    Models
    L2
    Gates
    L3
    Access
    L4
    Execution
    L5
    Orchestration
    L6
    Surface
    L7
    Memory
    L8
    Claude / Anthropic
    L2 giant
    EXPANDING ↑
    NVIDIA
    L0 monopolist
    DOMINANT
    Clay
    $3B · data + workflow
    FORTIFIED
    Sierra
    $15B · agent infra
    FORTIFIED
    Apollo
    GTM data + L2 connector
    L1+L2 SURVIVOR
    Outreach
    Sales Engagement
    COMPRESSES
    Core Significant EmergingEmpty = no presence

    Claude owns L2 and is reaching into L5/L6/L7, gravity at work. Apollo thins toward a data + connector role as Claude becomes the marketer's command center. Much of martech gets compressed unless it deepens into L1 or L8.

    Voices on the framework

    Product leaders, founders, and investors using the 10-layer map.

    Reactions from workshops, 1:1 reviews, and LinkedIn exchanges with people who have applied the framework to their own roadmap, thesis, or category. Venture partners have called it “a macroeconomic, industry-defining model” once the Applications view and vertical maps click into place.

    Read all voices

    Most 'AI frameworks' are taxonomies. This one is absolutely designed to be a macroeconomic, industry-defining model. Once you see the Applications view, the vertical maps, and the Cube together, it stops feeling like a stack diagram and starts feeling like a way to price an entire industry.

    Partner, Multi-Stage Venture Fund

    Investment Partner (attribution withheld)

    Macro lens

    JTBD tells you the length of the customer need. Supply Chain of Intelligence tells you the depth of the answer, how many layers you have to own to deliver it durably. 'Trust the output' is one job; you can answer it shallow with a verifier widget, or deep with an L3 gatekeeping layer baked in. The framework finally gave me a vocabulary for that trade-off.

    Bill Leece

    AI Product Leader, ex-Google · Indeed (AI Agents & Evals)

    JTBD × Chain

    I have sat through a hundred 'AI strategy' decks. This is the first one that told me which layer a product was actually on, and which layer it had to move to before the model layer ate it. The diagnostic is brutal in a useful way.

    Ruth Morales Zimmerman

    Investor · Venture & Private Markets Commentator

    Filter

    We were calling ourselves an 'AI platform' and the framework made us see we were a thin L7 surface on top of someone else's L2. We rewrote the roadmap inside a week to compound on L1b proprietary data instead. The language travels, engineering and GTM both speak it.

    Carmen Insignares Newell

    Product Leader · ex-Apple, ex-Amazon Alexa · CEO, Stackforce

    L7 → L1b

    Working on AI and ads inside a platform company, you feel the layer compression in real time, what was an app last quarter is a feature this quarter. The 10-layer map is the first framework that names that dynamic instead of describing it after the fact.

    Mahek Hooda

    Senior Product Manager · AI & Ads · Meta (ex-Microsoft)

    L7 → L4

    The Framework in Action

    Case Studies, Proof Through the Stack

    WORKED EXAMPLE · WRITING TOOLS10 min
    Jasper logoGrammarly logoCopilot in Word logo

    Jasper, Grammarly, Copilot in Word: Same Category, Three Structural Fates

    All three help you write. Jasper owned only the surface (L7c) and dissolved when the model went free. Grammarly owned distribution (L4a + L7c embedded copilot) into every browser and editor, until a bigger L4a owner, Microsoft, integrated the model directly into Word, Outlook, and Teams. Same layer. Bigger railroad. The market is repricing layer ownership, not ARR.

    $1.5B (Oct 2022)~$300M -80%
    L4L7
    Read
    DEEP DIVE · CUSTOMER EXPERIENCE10 min
    Sierra logoSalesforce logo

    Sierra's Memory Moat: Why L8 Beats Salesforce's Agentforce

    Sierra and Salesforce Agentforce look like the same product on stage, an AI agent that resolves customer issues. The Cube projection shows they are structurally opposite. Sierra was architected as L1c behavioral data + L5d operating playbooks + L8c network learning from day one: every resolution compounds into per-customer memory. Agentforce is L5 bolted onto Salesforce's existing L1, with no compounding loop. Same demo, opposite trajectories.

    $0 (2023)$10B+ (2025) Compounding
    L1L4L5L8
    Read

    The Corpus

    All 24 worked examples, in one rail

    Every analysis applies the same 10-layer lens to a real company or category, same framework, different verdicts. Scroll to scan; click any to read.

    Open the analysis index
    JasperGrammarlyCopilot in WordL4 · L7

    WORKED EXAMPLE · WRITING TOOLS

    Jasper, Grammarly, Copilot in Word: Same Category, Three Structural Fates

    Verdict: L7c surface vs L4a railroad

    CheggChatGPTL7

    L7 EXPOSURE

    Chegg: From $12B to 99% Collapse, The Fastest Value Destruction in EdTech

    Verdict: L7b only, no L1b/L3a/L8b

    GammaCopilotGeminiL2 · L4 · L7

    ARCHETYPE ANALYSIS

    Gamma at $2.1B: The Thin-Layer Graveyard in Real Time

    Verdict: L7b on rented L2a

    Stack OverflowChatGPTGitHub CopilotL1 · L2 · L7

    L1 MIS-PACKAGED AS L7

    Stack Overflow: When Your Community Becomes Training Data

    Verdict: L1b mis-packaged as L7b

    Apollo.ioZoomInfoL1 · L7

    STRUCTURAL DIVERGENCE

    Apollo vs ZoomInfo: Same Layer, Opposite Strategies, Different Fates

    Verdict: L1b headless vs L1b + L7b tax

    SierraSalesforceL1 · L4 · L5 · L8

    DEEP DIVE · CUSTOMER EXPERIENCE

    Sierra's Memory Moat: Why L8 Beats Salesforce's Agentforce

    Verdict: L1c + L5d + L8c stack

    Stability AIMidjourneyL2 · L7 · L8

    MODEL LAYER TRAP

    Stability AI vs Midjourney: Why Open-Source L2 Couldn't Monetize

    Verdict: L2a without L1b/L4a/L8c

    SalesforceNotionChatGPTL5 · L6 · L7 · L8

    THE FIVE ERAS · STRUCTURAL THESIS

    From Dashboard to Skill Hire: The Death of Per-Seat Software

    Verdict: Era 3 → Era 5 transition

    Harvey AIL1 · L3 · L5 · L8

    VERTICAL STACK

    Harvey AI Through the Layers

    Verdict: L1b + L3a + L5b + L8d

    McKinseyOpenAIL1 · L2 · L6 · L8

    CONSULTING × MODEL LAYER

    McKinsey + OpenAI (Lilli): When the Consulting Firm Owns the Memory, Not the Model

    Verdict: L1 + L8 OVER L2

    BloombergL1 · L2 · L3 · L4

    VERTICAL STACK

    BloombergGPT: Why a 50B-Parameter Model Beats GPT-4 in Finance

    Verdict: L1b + L2b + L3a + L4a stack

    KlarnaOpenAIL1 · L5 · L8

    L5 + L8 IN PRODUCTION

    Klarna: 700 Agents Replaced, $40M Saved, The First Honest Number on Agent Economics

    Verdict: L1c + L5a + L8c stack

    Cognition (Devin)CursorL2 · L7

    L7 ON RENTED L2

    Devin at $2B: The Autonomous Coder With No Layer Beneath It

    Verdict: L7c agent on rented L2a

    PerplexityGoogleL4 · L7

    L4 DISTRIBUTION

    Perplexity vs Google: The Answer Engine vs The Default

    Verdict: L4a absorbs L7a

    CursorGitHub CopilotL4 · L6 · L8

    L4 + L6 STACK

    Cursor at $9B: The IDE That Quietly Became the Most Important L4 in AI

    Verdict: L4a + L6c + L8d stack

    AnthropicL2 · L3

    L3 TRUST PLAY

    Anthropic's Enterprise Wedge: Selling L3 When Everyone Else Sells L2

    Verdict: L2a + L3a + L3c wedge

    AdobeL1 · L3 · L4

    L1 + L4 STACK

    Adobe Firefly: The Only Image Model an Enterprise Can Legally Use

    Verdict: L1b + L3a + L4a stack

    Character.AIGoogleL2 · L7 · L8

    L8 WITHOUT L2

    Character.AI: The L8 Memory Moat That Couldn't Stand Without L2

    Verdict: L8b without owned L2a

    GleanL1 · L6 · L8

    L1 + L6 ENTERPRISE STACK

    Glean at $7.2B: The Enterprise Memory Layer Microsoft Was Supposed to Own

    Verdict: L1c + L6d + L8d stack

    Tempus AIL1 · L3 · L8

    VERTICAL & REGULATED · MEDTECH

    Tempus AI: When the Data Layer Sits Inside the Clinic

    Verdict: L1b + L3a + L8d stack

    John DeereL-1 · L1 · L8

    PHYSICAL & INDUSTRIAL · AGROTECH

    John Deere: Why the Tractor Is the L-1 Moat

    Verdict: L0e + L1d + L8d stack

    TeslaWaymoL-1 · L1 · L8

    PHYSICAL & INDUSTRIAL · AUTONOMY

    Tesla vs Waymo: Two Bets on Which Layer Wins Autonomy

    Verdict: L0e + L1c vs L1b + L8d

    Apollo.ioClaude / AnthropicL1 · L2 · L7

    SAASPOCALYPSE · SURVIVOR PATTERN

    Apollo: How Giving Up the SaaS Stack Became the Smartest Bet in B2B Data

    Verdict: L1b moat + L2 connector, the thin-stack survivor

    DripifyLinkedInL1 · L3 · L7

    WORKED EXAMPLE · GATEKEEPER ARBITRAGE

    Dripify vs LinkedIn: The L7 Arbitrageur Living Off an L1 Gatekeeper

    Verdict: L7 surface arbitrage on an L1+L3 bottleneck

    The Diagnostic

    Where Do You Actually Sit in the Stack?

    1.

    What layer do you think you own?

    2.

    What sublayer is actually defensible?

    3.

    What happens when L7 becomes free?

    4.

    Are you rising by gravity, or climbing down too late?

    5.

    Do you own any part of the Defensible Triangle?

    Explore the Framework

    One worked example per week

    One company. Scored on the 10 layers. Verdict in plain English. No filler, no upsell.