PrimetimeGeek independent tech news · since 2008
Understanding the Supply Chain of Intelligence, the defensibility map American AI founders keep reaching for
Anand Arivukkarasu's framework treats intelligence as a supply chain running from power stations to memory, ten layers deep, and argues that value settles at the bottlenecks rather than the screen everyone is looking at. A plain-English guide for US founders, product leaders and investors — what the framework says, where it is genuinely useful, and where it stops.
Every few years American technology gets a new vocabulary problem. Right now it is the word wrapper. A founder pitches an AI product; an investor asks whether it is a wrapper; nobody in the room can define the term precisely enough to settle it, and the meeting ends in vibes. The Supply Chain of Intelligence exists to end that conversation with a coordinate instead of an adjective.
The framework is the work of Anand Arivukkarasu, a former Meta and Instagram product leader now based in San Francisco, published openly at supplychainofai.com with a versioned canonical paper rather than a book deal. Its central claim is one sentence long: intelligence is a supply chain, and value accrues at the bottlenecks, not the most visible node.
The analogy it runs on
Gold is the framework's teaching example, and it is a good one. Between ore in the ground and a ring on a finger sit land and mineral rights, mining equipment, ore, a smelter, an assay office that stamps the hallmark, railroads, a master jeweler, a store, and finally the moment somebody wears it. Most companies own exactly one of those steps. Which step you own decides almost everything about your economics, and it is rarely the glamorous one.
Generative AI, the argument goes, is structurally the same chain. The framework page lays it out as ten layers, fifty sublayers, three tiers, four laws, three currents and a three-dimensional map it calls the Intelligence Cube. That sounds like a lot of scaffolding. In practice most of the working value comes from the layers and the four laws.
The ten layers, in plain English
- L−1 Resources — energy and grid interconnect, cooling water, foundry capacity, rare earths, and the electricians and HVAC technicians who physically build data centers. The layer stack diagrams crop out, and currently the binding constraint on the whole American build-out.
- L0 Infrastructure — silicon, data centers, interconnect, edge compute. The shovels.
- L1 Data — public data, proprietary data, behavioral and sensor data, outcome data, synthetic data. The ore.
- L2 Models — foundation models, fine-tunes, embeddings, routing, reasoning. The smelter.
- L3 Gates — compliance, quality, safety and provenance, editorial and distribution gatekeeping. The assay office that stamps the hallmark.
- L4 Access — APIs, agent protocols, permissions, agent identity. The railroads.
- L5 Execution — the domain work itself: tool use, reasoning scaffolds, operating playbooks. The master jeweler.
- L6 Orchestration — agent loops, human-in-the-loop, routing, state. The workshop.
- L7 Surface — the chat window, the app, the embedded assistant. What the user touches.
- L8 Memory — session memory, entity profiles, aggregated network learning, institutional knowledge. The record book that compounds.
Those ten group into three tiers with very different half-lives: Surface, durable for weeks; Workflow, durable for months; Substrate, durable for years. The single most useful line in the whole framework may be the one attached to that grouping — own the lower layers, or rent them, and rent your future.
The four laws are where the work happens
Law I, intelligence commoditises downward. If your product depends only on generic model capability, the platform underneath eventually absorbs it. The framework's example is Jasper, which fell from a $1.5B valuation to roughly $300M once ChatGPT shipped the same capability inside a surface users already had open. The product did not get worse; the layer beneath it absorbed what the product was charging for.
Law II, value accrues at bottlenecks. Durable value sits at the scarce layer — proprietary data, workflow control, verification, distribution, memory, and right now energy and fabs. The associated exercise is brutally clarifying: name, in one sentence, the bottleneck you own. If the sentence does not write itself, you do not own one.
Law III, the surface captures attention; the chain captures power. A beautiful interface wins the first cohort. Depth decides whether you keep it. Chegg is the framework's cautionary case: a thin content surface with no proprietary data and no memory loop, whose entire layer became free the day ChatGPT arrived.
Law IV, generation and verification must be separate. Wherever 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; the FDA sits above Pfizer. Those positions do not weaken as the models improve — they get more necessary. For US founders this is the most under-exploited idea in the set, because American regulated industries are enormous and the verification seats are mostly unbuilt.
The currents, and why a defensible layer can still starve
Across the vertical layers run three horizontal forces: Demand Gravity, where the budget actually sits; Attention Economics, who owns the on-ramp when generation becomes infinite; and Capital Flows, which the framework asks you to read as a distortion field rather than a value signal. Two currents pointing at a layer is a tailwind. All three is a category. None is a press release. It is a useful corrective to the standard defensibility pitch, because a scarce position nobody has budget for is still worth nothing.
The one piece of vocabulary hygiene worth adopting today
The framework insists that agent is not a layer. It is marketing for a package — minimum viable composition L5 Execution plus L6 Orchestration, usually bundled with a surface and sometimes with memory. Decode it in three steps: name the actual work being done, name the other layers bundled with it, then ask whether any of those are structurally hard for the underlying model provider to replicate. Agent plus proprietary data is a fortress. Agent plus platform access is a railroad. Agent plus memory compounds. Agent plus nothing is a wrapper on a clock.
For an application-layer company the framework offers a target shape 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. Two of the three is a workflow product that gets better. Three of three is something that compounds.
Where it is genuinely useful for American teams
Three concrete uses. For an operator, map your product to the ten layers honestly, mark what you own versus what you rent, then write down in plain language what compresses you and on what timeline — that output survives contact with the next model release in a way that a positioning deck does not. For an investor, the same exercise applied to a target separates the pitch decks that describe an execution product as though it were a full data-execution-memory stack. For anyone writing about the market, it enforces a discipline worth more than the framework itself: every defensibility claim names a layer, every disruption claim names a law, every timing claim names a current.
Where it stops
The framework is descriptive, not predictive, and its author says so — it will not tell you which company wins, only which layers a company owns, which it rents, and which force is about to move the value. That honesty is a strength, but buyers of strategy frameworks should read it as a boundary.
Three further limits worth stating plainly. It is new: published in early 2026 with a versioned paper, which means it has not accumulated the decades of independent practitioner testing behind Wardley Mapping, Jobs-to-be-Done or Christensen's disruption work, and there is no independent empirical study validating that companies which follow it outperform ones that do not. Its taxonomy is explicitly provisional — the author allows that it may be twelve layers tomorrow — so anyone standardising internal reporting on the layer numbers should expect version drift. And like every layer model, it invites false precision: real companies smear across sublayers, and an argument about whether something is L5b or L6a is usually an argument that has stopped being about the business.
There is also a category of claim the framework cannot settle, only sharpen. Whether verification stays separate in a given US industry is a regulatory and insurance question, not a structural one, and the four laws are best read as strong empirical regularities with well-chosen examples rather than as laws in the physical sense.
The bottom line
Judged as what it is — a shared vocabulary for arguing about AI defensibility, given away free by its author rather than sold — the Supply Chain of Intelligence is the most useful new addition to the US technology-strategy toolkit in the agentic era. It does not replace Wardley Mapping, which tells you what is evolving; it answers a question the older frameworks were never built for. When intelligence itself becomes abundant, where in the chain do power and defensibility remain? The canonical paper, the market map and the worked company teardowns are all published openly at supplychainofai.com, and the framework reference is the right place to start if you intend to map your own product this quarter.