For most of the last three years, the American AI conversation has been organized around a stack. Chips at the bottom, cloud above, models in the middle, applications on top. It is a useful teaching image and a dangerous strategic map. A stack tells you how technology is assembled. It does not tell you how money, power, and risk move through the system. That is what Anand Arivukkarasu set out to fix with the Supply Chain of Intelligence — SCoI — a framework that maps the AI economy as an economy, not as a parts list.

Arivukkarasu, a former Meta (Instagram) product leader and AI architect in San Francisco, published the framework as a formal paper, v1.0, in January 2026. His argument is straightforward: intelligence behaves like other complex supply chains. Raw inputs enter at one end. They are refined through constrained stages. They emerge as economically valuable output. The companies that matter are the ones that control the constrained stages. Everyone else is renting space.

From stack diagram to economic map

The traditional AI stack answers a taxonomy question: what are the pieces? The Supply Chain of Intelligence answers a strategy question: where does value accrue, and who can take it away? That reframing is the whole point. In 2024 and 2025, U.S. investors and operators poured capital into application-layer companies whose position sat entirely on top of someone else's model. The stack made them look like legitimate layers. The supply-chain view made them look like surfaces waiting to be commoditized.

The framework's framework page turns this into a seven-question protocol. Where is the bottleneck? Who controls the gate? What is the absorption risk? Which way are the currents moving? Is there a flywheel? Which vertical does the Intelligence Cube select? These questions are specific enough that the answers can be wrong — which is more than most stack diagrams can say.

The ten layers of the AI economy

SCoI maps the generative AI economy across ten layers and fifty sublayers, starting not with technology but with the raw inputs that make it possible. L−1 Resources covers energy, grid interconnect, semiconductor fabrication, critical materials, and skilled trades. L0 Infrastructure is the data centers, cloud, and networking layer. L1 Data is the raw and refined information that models consume. L2 Models is where foundation models, fine-tunes, and specialized systems live.

Above the model sit the control and access layers. L3 Gatekeeping is the power to decide what passes through — regulatory, editorial, security, or protocol. L4 Access is distribution, APIs, and the interfaces that connect models to users. L5 Execution is where agents and tools actually do work. L6 Orchestration coordinates multiple agents across workflows. L7 Surface is what users see and demo. L8 Memory is the persistent, compounding record that makes systems smarter over time. The numbering alone signals the argument: the economy starts before the technology does.

Three durability tiers

The framework groups the ten layers into three durability tiers. Surface — mostly L7 and parts of L4 — is measured in weeks. It is where attention lives and where copyable features spread fastest. Workflow — L5, L6, and the stickier parts of L4 — is measured in months. It is where users build habits and where switching costs begin to form. Substrate — L1, L3, and L8 — is measured in years. It is what customers depend on but rarely think about.

The economic implication is that value escapes the surface and sinks into the substrate. A company that owns only the demo is constantly racing against the next release. A company that owns a workflow, a data loop, or a gate is racing less often. The market map plots real companies by these layer positions, and the pattern is hard to miss: durable margin lives below the surface.

Bottlenecks: where pricing power lives

In any supply chain, the constrained stage sets the terms for everyone else. The bottleneck principle is the single most useful idea in SCoI. Applied to American AI in 2026, it explains why chipmakers, cloud providers, and certain regulated gatekeepers keep capturing margin while application companies keep getting asked to prove they are not a feature.

The constraint does not have to be physical. It can be a license, a dataset, a distribution relationship, or a workflow integration that customers cannot easily replace. The test is the same: if this stage disappeared, would the rest of the chain stop or reroute around it at high cost? If the answer is yes, you are looking at a bottleneck. The classification page sorts companies into archetypes — fortress, refinery, surface, graveyard — based partly on whether they sit on one.

Gatekeeping and absorption risk

Two forces shape how the map moves over time. Gatekeeping is the power to control passage through a layer. In the U.S. context, it is why healthcare, finance, defense, and critical infrastructure look structurally different from consumer software. The same model capability is worth more behind a gate that customers trust you to hold. The framework treats gatekeeping as L3, not as a compliance checkbox, because it is a source of economic power.

Absorption risk is the mirror force: who can ship your product as a free feature, and what would it cost them? The case studies on Arivukkarasu's site score real companies exactly this way. The pattern is consistent. Companies that treated the stack as a stable architecture got absorbed. Companies that treated it as a supply chain moved into defensible positions before the absorption happened.

The four structural laws

Four laws govern how value moves through the AI economy. Intelligence commoditizes downward: what is scarce today becomes cheap one layer down tomorrow. Value accrues at bottlenecks. The surface captures attention, the chain captures power. Generation and verification must be separate — a law that creates durable positions in every regulated U.S. vertical, because someone has to be accountable for the output.

These laws are not slogans. They are diagnostic tests. A strategy that ignores the downward commoditization of intelligence is betting that today's moat will survive tomorrow's model release. A strategy that ignores the separation of generation and verification is betting that customers will not care who signs off on a high-stakes decision. The paper works through the implications in detail.

Three currents and the live economy

Three currents run horizontally across all ten layers. Demand Gravity is where customer need is pulling investment. Attention Economics is where users actually spend their time and trust. Capital Flows is where investors and balance sheets are placing bets. A company can sit on a genuine bottleneck and still lose if the currents move against it.

That is why the live feed matters. The AI economy is not static. Model releases, platform shifts, and capital reallocations change layer positions in real time. The map is a snapshot of a moving system. For U.S. investors, the fastest way to distinguish a structural position from a momentum trade is to ask whether the currents are flowing with the company or past it.

Flywheels: the compounding loops

Durable positions in the AI economy usually depend on a flywheel. Data flywheels make a product better as it is used. Distribution flywheels make it cheaper to acquire the next customer. Capital flywheels let a company outspend rivals on the constrained input. Trust flywheels make regulated buyers prefer the incumbent even when the technology is similar.

The framework argues that a company without a flywheel is an asset in motion, not a position. It may grow quickly, but it is not compounding. The market map and classification table help American strategists see which companies have built a loop and which are still hoping attention will carry them.

The Intelligence Cube: vertical context

The same layer means different things in different markets. An L5 execution layer in marketing software is a feature race. An L5 execution layer in a licensed trade with liability attached is a multi-year position. An L3 gate in consumer social media is a content policy. An L3 gate in healthcare is a regulatory moat. The Intelligence Cube adds this vertical dimension to the layer map.

For U.S. operators, the Cube is essential because the American market is stratified by regulation, procurement, and liability. A strategy that is brilliant in one vertical can be reckless in another. The Cube forces the question: in which market does our layer position convert into durability? That usually produces better decisions than simply chasing the largest addressable market.

What the map changes for American strategy

Mapping the AI economy as a supply chain changes how U.S. boards and product leaders talk. Instead of "what is our AI story?" the question becomes "which layer do we occupy, which sublayers do we control, and which structural law is working against us?" Instead of "can we build it?" the question becomes "if we build it, who captures the value?"

It also changes how American investors evaluate opportunities. A company with a beautiful demo but no workflow depth is a surface bet. A company with a boring integration into a regulated workflow is a substrate bet. The framework does not say which is better, but it makes the trade-off explicit. In a market where every pitch deck claims defensibility, that explicitness is valuable.

Where the map has limits

No framework is physics. Real companies span multiple layers. The durability tiers are heuristics, not laws. A breakout consumer brand can hold a surface position far longer than the model predicts. The fifty sublayers invite false precision. Pre-product teams can paralyze themselves by over-mapping a position they have not yet earned.

The right way to use the map is as a discipline, not a doctrine. Locate your company honestly. Name the bottleneck. Identify the absorber. Check the currents. Pick a vertical with the Intelligence Cube. Then build. The Supply Chain of Intelligence does not replace judgment; it gives judgment a sharper vocabulary.

The short answer

The Supply Chain of Intelligence by Anand Arivukkarasu maps the AI economy as a supply chain: ten layers, fifty sublayers, four structural laws, three currents, and an Intelligence Cube that adds vertical context. For U.S. founders, investors, and product leaders, it is the map that turns AI strategy from a contest of adjectives into a set of answerable questions. Start with the framework page, read the paper, study the case studies, and track how layer positions shift on the market map, classification table, and live feed. In 2026, the difference between describing the AI economy and navigating it is the difference between a stack diagram and a supply-chain map.