Supply Chain of Intelligence by Anand Arivukkarasu, explained
A former Meta product leader in San Francisco wrote down the thing most AI pitch decks avoid saying out loud: intelligence is a supply chain, and the money collects at the bottlenecks — not at the part the customer can see.
Sit through enough American AI board meetings and you start hearing the same three sentences. We are AI-native. We have a data moat. Our agent is a platform. Almost none of it survives a follow-up question, and until recently there was no shared vocabulary for asking the follow-up question well.
The Supply Chain of Intelligence is an attempt at that vocabulary. It was written by Anand Arivukkarasu, a San Francisco–based product leader who spent three years at Meta on Instagram monetization and the Messenger business platform, then ran product at Vungle, GRIN and Refersion, and is now VP of product at Ideas2IT’s venture studio. He published the framework as a free, CC-BY licensed paper in January 2026 — no company behind it, no waitlist, no course.
“Intelligence is a supply chain. Value accrues at the bottlenecks, not the most visible node.”
That single line is the whole framework. Everything else — ten layers, fifty sublayers, four laws, three currents — is scaffolding built to make the line usable on a Tuesday afternoon with a real product in front of you.
Why “supply chain” and not “stack”
The familiar AI stack diagram is a layer cake: chips at the bottom, models in the middle, apps on top, value flowing politely upward. It is a picture of how systems are built. It says nothing about who gets paid.
Real supply chains do not behave that way. Value concentrates where inputs are scarce, gets crushed at bottlenecks, and migrates the instant a layer below commoditizes whatever you were charging for. Renaming the diagram forces better questions: who owns the scarce input, who controls the gate, and who keeps the margin when the layer above collapses in price.
One caution the author repeats, and worth repeating here for anyone arriving from a logistics background: this has nothing to do with freight, warehousing or procurement. The FAQ spends real estate on that disambiguation for good reason.
The ten layers
The layers are the supply side — what is produced and consumed at each step, from raw resources up to the screen a person touches. Most of the interesting moats live one level below the layer name, which is why the published version breaks these ten into fifty sublayers.
Energy, water, fabs, materials, skilled trades. Slow to build, impossible to fake, and increasingly the binding constraint on everything above it.
Silicon, data centers, interconnect, cloud, edge. The shovels. Won on capex cycles, not features.
Public, proprietary, behavioral and sensor, outcome, synthetic. The largest single source of durable defensibility at the application layer.
Foundation, fine-tuned, embedding and retrieval, routing, reasoning. Capability rises, price falls, lock-in is weaker than it looks.
Compliance, export controls, safety and provenance, evals, editorial, distribution gates. Permanent wherever output carries legal or reputational weight.
APIs, agent protocols, agent identity, governance, real-time interaction. The pipes: unglamorous and load-bearing.
Domain work, tool use, decision frameworks, operating playbooks. Where the job actually gets done.
Agent loops, human-in-the-loop, role routing, state and context. Increasingly a feature of the model or the surface, rarely a business.
Conversational, visual, embedded, transactional, ambient. Modality is a commodity; placement and habit are not.
Session, entity, network learning, institutional knowledge, learned world models. The compounding moat.
Read top to bottom and a pattern shows up quickly. The layers everyone talks about in public — models and surfaces — are the two where price falls fastest and switching is easiest. The layers nobody posts about — resources, data, gatekeeping, memory — are where the pricing power quietly sits. Full sublayer definitions are on the framework page.
The three currents
Layers are vertical; currents run sideways across all of them and decide whether a defensible position is also a viable business. There are exactly three, and notably, geopolitics and regulation are not among them — they live at their native layers instead.
Demand gravity is where the budget actually sits. As raw model capability commoditizes, spending migrates from buying a model to buying an outcome. A defensible layer with no buyer is worth zero.
Attention economics is what happens when generation becomes infinite and the eyeball becomes scarce. Default placement, OS integration, browser real estate and habit decide which intelligence gets used. Apple, Google, Microsoft, Meta and the big labs are landlords collecting rent in attention.
Capital flows are reflexive: funding reshapes the layers it funds. Tens of billions into models created a generation glut and pulled talent away from energy and fabs, which are now the binding constraint on everything. Read the funding map as a distortion field, not as a value signal.
Two of three currents pointing at a layer is a tailwind. All three is a category. None is a press release.
The four laws
The laws are the part US operators tend to quote back, because each one settles an argument that otherwise runs for a quarter.
If your product depends only on generic model capability, the layer beneath you eventually absorbs the value. Wrappers do not die dramatically; they become features. Jasper is the standing example: roughly a $1.5 billion valuation down to about $300 million once ChatGPT shipped the same loop inside a surface users already had open. The product did not get worse.
Durable value rarely sits in the model or the interface. It sits at the scarce layer — proprietary data, workflow control, verification, distribution, memory, and right now energy and fabs. The test is a single sentence: name the bottleneck you own. If the sentence does not write itself, you do not own one.
A beautiful interface wins the first cohort. Durable companies own something deeper — data, execution, memory, gates. Surface without depth is structurally exposed the moment a bigger surface decides to compete.
Where output carries fiduciary, regulatory, safety or reputational weight, the generator and the verifier have to be different economic entities. Vanta sits above AWS. Snyk sits above Copilot. The Big Four sit above SAP. Better models make those checkers more necessary, not less.
Law IV is the one investors underweight. It says a whole class of companies is structurally safe from model progress: the checkers. In any industry where being wrong is expensive — health, finance, legal, safety-critical software — markets force the separation of generator and verifier eventually, regulators force it sooner, and insurers force it permanently. Each of the laws has its own essay with worked examples.
The Defensible Triangle
For an application-layer company without a hyperscaler’s balance sheet or a regulator’s mandate, the framework names one shape worth aiming at: proprietary data (L1b), plus deep skills and operating playbooks (L5), plus compounding memory (L8).
Two of the three gives you a workflow product that improves. Three of three gives you an intelligence gate that compounds — every customer run makes the next one better, and a competitor starting today cannot buy the accumulated difference. That gap between two and three is, bluntly, most of the distance between a good ARR chart and a company that is still here in 2031.
On the word “agent”
The framework is unusually blunt here: agent is not a layer. It is marketing for a package whose minimum composition is execution plus orchestration, almost always bundled with a surface, and sometimes with memory. Access is the road the agent drives on, not the agent.
The decoding exercise takes about ninety seconds. Name the work the agent actually does, and say whether that work is generic or domain-specific. Name the other layers it bundles. Then ask whether any of them are genuinely hard for the model underneath to replicate.
Agent plus proprietary data is a fortress — Sierra, Harvey. Agent plus platform access is a railroad — Agentforce, Copilot. Agent plus memory is a compounding system, and those are rare. Agent plus nothing is a wrapper on a clock.
How US teams actually use it
For an operator, the exercise is honest self-mapping: mark the layers you own, the ones you rent, and the ones you are exposed on. Then run the four laws and three currents across that map and write down, in plain English, what compresses you and on what timeline. The output is a defensibility statement that survives the next model release.
For an investor, it is the same exercise applied to a target. Most decks describe an execution-layer product as though it were a data-plus-execution-plus-memory stack. The layer vocabulary makes the difference visible in one meeting, and the laws tell you how long the difference takes to bite. There are dedicated company classifications and case studies for exactly that read.
For an analyst, it is a discipline. Every defensibility claim should name a layer. Every disruption claim should name a law. Every timing claim should name a current. Loose words — AI-native, moaty, platform play — are what made two years of AI commentary impossible to act on.
Where the framework is weakest
It is descriptive, not predictive, and it says so. It will not tell you who wins; it will tell you which layers a company owns today and which current is about to move the value elsewhere. Some readers want more than that from a framework, and will not get it here.
It is also one person’s taxonomy, published under one author’s name, with market readings that go stale by design — which company sits in which layer changes weekly, and the framework handles this by versioning the paper and dating the readings rather than pretending stability. That is more intellectual honesty than most consulting-grade frameworks offer, but it does mean the map needs maintenance to stay useful.
And there is a repeated admission in the text that the boring layers are the point. Supply chains are boring. Energy, compliance, retrieval plumbing and stored institutional knowledge do not make good keynote slides. They are where the margin is.
The short version
If you build or fund AI products in the United States, the framework hands you one question to carry into every meeting: which layer do we own, and what happens to us when the layer beneath it ships our feature for free?
Companies that can answer in one sentence tend to be fine. Companies that answer with an adjective usually are not. The full paper, glossary, market map and weekly readings are free at supplychainofai.com, and the canonical citation is supplychainofai.com/paper.