Open any consulting deck about AI in America today and you will find the same picture: a neat column of boxes labeled chips, cloud, models, and applications. It is tidy, it is familiar, and according to Anand Arivukkarasu — a former Meta (Instagram) product leader and AI product architect in San Francisco — it is structurally incapable of answering the questions that decide who wins. His alternative, the Supply Chain of Intelligence (SCoI), is a full rethinking of the generative AI stack, and it deserves a serious read from anyone allocating capital or building product in this market.

The argument, laid out in the SCoI paper v1.0 published in January 2026, starts with a distinction that sounds academic and turns out to be deeply practical. A stack shows the parts. A value chain shows the flow. Only a supply chain view surfaces the forces that actually decide outcomes: gatekeeping, bottlenecks, currents, flywheels, absorption risk, vertical adjacency, and timing. The framework page calls these “the seven things that go missing” from conventional AI stack diagrams, and the list is hard to argue with once you see it.

What the stack view cannot see

Consider gatekeeping. Every supply chain has chokepoints, and whoever controls one can charge rent for letting traffic through. A stack diagram has no concept for this — it shows neighbors, not predators. The same blindness applies to absorption risk: the constant possibility that a platform above or below your layer ships your product as a free feature. American founders learned this lesson the hard way through 2024 and 2025 as model providers absorbed entire application categories. A stack diagram shows you where you sit. It does not show you who is coming to eat you.

The Supply Chain of Intelligence framework also surfaces bottlenecks that stack diagrams crop out entirely — what SCoI labels L−1 resources (energy, grid interconnect, fabrication, critical materials, skilled trades) at the very bottom, and L3 verification gates and L8 memory at the top. In 2026, with U.S. data-center interconnection queues measured in gigawatts and years, the idea that energy and skilled trades are part of the AI supply chain is not provocative. It is obvious. It just does not fit in the old picture.

The 10 layers, from L−1 to L8

SCoI's structural vocabulary is 10 layers and 50 sublayers, running from L−1 Resources (energy, thermal, fabrication, materials, human capital) up through L0 infrastructure, L1 data, L2 models, L3 gates, L4 access, L5 execution, L6 orchestration, L7 surface, and L8 memory. Several sublayers are starred as especially defensible — proprietary and outcome data at L1, editorial and distribution gates at L3, agent identity and interface protocols at L4, domain execution and operating playbooks at L5, human-in-the-loop at L6, and the compounding layers of aggregated network learning, institutional knowledge, and learned world models at L8.

The layers group into three tiers with radically different durability. The Surface tier — what users touch — is easily replicated, and platforms ship it for free; its durability is measured in weeks. The Workflow tier — what users live inside — is sticky if deep, survivable if owned; durability is months. The Substrate tier — what users depend on, from proprietary data to trust gates to compounding memory — is measured in years. The framework's summary line is worth quoting to any leadership team: value escapes the surface and accumulates in the layers below. Own the lower layers, or rent them, and rent your future.

Currents, flywheels, and the Intelligence Cube

Two more ideas separate SCoI from a static taxonomy. First, three market currents — Demand Gravity, Attention Economics, and Capital Flows — flow horizontally across all 10 layers and decide whether a defensible position actually becomes a business. A company can sit on a genuine bottleneck and still starve if capital is flowing elsewhere. Second, flywheels compound across sublayers: the framework traces loops like L5 execution feeding L1d outcome data feeding L8c aggregated network learning, a cycle that a component list simply cannot represent.

The Intelligence Cube adds the vertical dimension. The same layer behaves differently in legal, health, and fintech — an L3 editorial gate in healthcare is a regulatory moat, while in consumer software it might be an inconvenience. This is why the case studies span software, regulated verticals, and the physical world, with 24 companies mapped against the layers and laws. For U.S. operators in regulated industries, the Cube is arguably the most immediately useful part of the whole model.

Why American teams are rethinking their stack diagrams

"A stack describes parts. A value chain describes flow. A supply chain of intelligence describes the whole system — gatekeeping, bottlenecks, currents, flywheels, absorption — which is the level at which durable AI strategy can actually be reasoned about."

— Supply Chain of Intelligence, Framework §01

The timing of the framework matters. U.S. AI strategy has moved from innovation labs to P&L owners, and P&L owners think in supply chains. When a CFO asks why the AI budget keeps growing while margins do not, “our stack” is not an answer. “We rent layers L0 through L4, we are exposed to absorption at L7, and our only durable position is outcome data at L1d” is. The framework gives American executives a vocabulary for a conversation they were already having badly.

There is also a practical protocol attached: seven questions, meant to be run in order, from definition through the map, the four structural laws, the dynamics, applications, and reasoning. The site even ships a blank 10-by-50 template and invites teams to print it and mark it up. It is an unusually hands-on posture for a strategy document — closer to a Wardley Mapping workshop than a whitepaper, which fits Arivukkarasu's stated influences and his decision to give the framework away.

Where the rethink has limits

Intellectual honesty requires a caveat or two. Real companies span multiple layers at once, and the durability estimates — weeks for surface, years for substrate — are heuristics, not physics. A breakout consumer product can mint a brand at L7 that outlives supposedly deeper positions. And any model with 50 sublayers risks false precision: placing a company in L5b versus L6c can be a judgment call dressed up as measurement. The framework's own authors would likely concede the point; the paper reads as a reasoning protocol, not an oracle.

The short answer

Rethinking the generative AI stack through the Supply Chain of Intelligence means trading a picture of parts for a map of power: who gatekeeps, where value accumulates, which currents are moving, and what gets absorbed next. Whether or not you adopt all 10 layers and 50 sublayers, the core provocation — that American AI strategy has been reasoned about at the wrong level of abstraction — is one of the more useful challenges to conventional wisdom this year. Start with the framework page, test it against your own position on the market map, and see whether your moat survives the questions.