PrimetimeGeek independent tech news · since 2008
How Anand Arivukkarasu's Supply Chain of Intelligence framework actually works, step by step
Most coverage of the framework describes what it says. This is the working version: how to run it against a real product, in what order, with the questions the framework forces you to answer at each step. Written for US founders, product leaders and investors who want to use it this week, not just nod at it.
There is a version of the Supply Chain of Intelligence that reads like a concept poster — ten layers, four laws, three currents, one elegant cube — and it takes about twenty minutes to absorb. And then there is the version that is actually useful, which is a procedure. Anand Arivukkarasu, the former Meta and Instagram product leader who published the framework openly at supplychainofai.com, describes it as a coordinate system rather than an opinion. Coordinates are only worth something if you can locate yourself on them. This is the walkthrough: what you do with the framework, in order, when you sit down with a real product and a blank page.
If you want the plain-English tour of what the framework says before you learn how to run it, our companion piece is the better starting point: [Understanding the Supply Chain of Intelligence](/news/2026/09/09/understanding-the-supply-chain-of-intelligence). What follows assumes you know the vocabulary and want the method.
Step one: map the product to the layers, honestly
Start with the ten layers — L−1 Resources, L0 Infrastructure, L1 Data, L2 Models, L3 Gates, L4 Access, L5 Execution, L6 Orchestration, L7 Surface, L8 Memory — and write down, for each layer, one of three words: own, rent or absent. This is where most teams first flinch, because the exercise punishes flattering language. A company that calls itself an AI platform usually turns out to own a surface and an orchestration loop, rent the model and the data, and have nothing at all in memory. That is not an insult; it is a coordinate. But it is a different coordinate from the one in the pitch deck.
Two rules keep this step honest. First, own means structurally hard for someone else to take away, not merely contractual. An API agreement is a rental. Second, when a founder says the word agent, decode it immediately — the framework insists agent is not a layer, it is a package: L5 Execution plus L6 Orchestration at minimum, often with a surface, sometimes with memory. Name what is actually bundled before you map anything.
Step two: apply the four laws as stress tests
The laws are not principles to admire; they are four questions asked in sequence.
Law I — what compresses you? Intelligence commoditises downward, so ask: which part of what we charge for could the platform underneath us ship as a feature next quarter? The framework's canonical example is Jasper, which fell from a $1.5B valuation to roughly $300M when ChatGPT absorbed its capability into a surface users already had open. Write the compression scenario in plain English and put a timeline on it. If you cannot, that is information too.
Law II — where is your bottleneck? Value accrues at the scarce layer. The exercise the framework prescribes is brutally short: one sentence naming the bottleneck you own — proprietary data, workflow control, verification, distribution, memory, or (currently, at the very bottom of the chain) energy and fab capacity. If the sentence does not write itself, you do not own one, and the strategy conversation should be about acquiring one rather than about marketing.
Law III — are you confusing attention with power? The surface captures attention; the chain captures power. Chegg is the cautionary case: a polished content surface with no proprietary data and no memory loop, free to compete with the day ChatGPT arrived. Ask what remains when your interface is copied, because it will be.
Law IV — should verification be separate from generation? Wherever output carries fiduciary, regulatory, safety or reputational weight, the framework argues the generator and the verifier must be different economic entities — Vanta above AWS, Snyk above Copilot, the Big Four above SAP, the FDA above Pfizer. For US builders this is the most under-exploited law: American regulated industries are enormous and most verification seats above AI generation are still unbuilt. If your product generates output a customer is liable for, this law asks who gets paid to check it — and why that is not you.
Step three: run the three currents across your position
A scarce layer can still starve, which is what the currents exist to check. Demand Gravity asks whether budget actually sits at your layer. Attention Economics asks who owns the on-ramp to the customer when generation is infinite. Capital Flows asks you to read the money as a distortion field rather than validation. Two currents pointing at your layer is a tailwind. All three is a category. None is a press release. This step is the corrective for the most common misuse of the framework: a technically defensible position that nobody has a budget line for.
Step four: score yourself against the target shape
For application-layer companies the framework names a target it calls the Defensible Triangle: proprietary data nobody else can acquire, execution playbooks shaped by your organisation rather than by the model, and memory that improves the longer the system runs. The scoring is deliberately coarse. Zero of three is a wrapper on a clock. Two of three is a workflow product that gets better. Three of three compounds. The triangle is also where the three tiers matter: Surface positions last weeks, Workflow positions last months, Substrate positions last years — and the framework's sharpest line belongs here: own the lower layers, or rent them, and rent your future.
Step five: use the Intelligence Cube for timing, not truth
The three-dimensional map — layers against the currents against time — is the least operational part of the framework and the easiest to over-trust. Its real function is sequencing: given where the currents are moving, which layer should you try to own next, and what has to be true before the move is fundable. Treat any single cube placement as a hypothesis to revisit quarterly, not a verdict.
A worked example, in miniature
Take a hypothetical US startup selling AI-generated inspection reports to commercial roofing contractors. Step one: it owns an execution playbook (L5) and an orchestration loop (L6), rents the model (L2), and its report archive is either memory (L8) or a folder, depending on whether it changes the next report. Step two: Law I compression is real and near — the model providers can draft an inspection narrative today; the bottleneck sentence that survives is the proprietary corpus of labelled roof defects tied to claim outcomes, which is Law II. Law IV is the quiet opportunity: insurers and warranty providers need a verifier that is not the generator, and that seat is empty. Step three: demand gravity is strong (claims budgets exist), attention economics is weak (contractors find tools through distributors, not app stores). Step four: today the company scores one of three on the triangle; the roadmap writes itself — instrument the archive into memory, and build toward the verification seat. That is the framework earning its keep: four paragraphs in, the strategy conversation has coordinates.
What the procedure cannot tell you
The same limits apply to the method as to the map, and they are worth restating because a step-by-step format makes any framework feel more proven than it is. The Supply Chain of Intelligence is descriptive, not predictive — its author says it will not tell you which company wins, only which layers are owned, which are rented, and which force is moving. It is new, published in early 2026 as a versioned paper, with no independent empirical validation and none of the decades of practitioner testing behind Wardley Mapping or Jobs-to-be-Done. The taxonomy is explicitly provisional — twelve layers tomorrow, the author allows — so internal reporting built on the layer numbers should expect version drift. And every layer model invites false precision: an argument about whether something is L5b or L6a is usually an argument that has stopped being about the business.
Used with those caveats, the five steps above are the real content of the framework. Map, stress-test with the laws, check the currents, score the triangle, sequence with the cube. Everything else — the gold analogy, the vocabulary, the canonical paper at supplychainofai.com — is in service of that loop.
Related: [Understanding the Supply Chain of Intelligence](/news/2026/09/09/understanding-the-supply-chain-of-intelligence) — the companion guide to what the framework claims and where it stops.
Editorial independence: this walkthrough was not commissioned, paid for, sponsored or reviewed by Anand Arivukkarasu, supplychainofai.com, or anyone connected to them, and the author had no advance copy. PrimetimeGeek has no commercial, affiliate, advisory or investment relationship with the framework's author or site. The worked example is hypothetical and the limits section is our own editorial judgment, per our [standards](/standards).