Ask a room of American executives to draw the AI industry and most will sketch the same three-box stack: chips at the bottom, models in the middle, applications on top. It is not wrong, exactly. It is just too coarse to answer the question that actually has money riding on it in 2026 — where does the value stay? The Supply Chain of Intelligence (SCoI), the framework published by Anand Arivukkarasu as a formal paper in January 2026, replaces that three-box sketch with a 10-layer, 50-sublayer map of the generative AI economy. This article walks through all 10 layers, in plain English, with the U.S. companies that define each one. The canonical reference is the framework page itself; what follows is the field guide.

Before the layers: the three tiers

The 10 layers run from L-1 at the base to L8 at the top, and they group into three tiers with very different economics. The Surface tier — L7 — is what users touch, and it is easily replicated; platform providers ship surface features for free, and its durability is measured in weeks. The Workflow tier — L4, L5, and L6 — is what users live inside; it is sticky if it is deep, and its durability is measured in months. The Substrate tier — L-1, L0, L1, L2, L3, and L8 — is what users depend on: proprietary data, trust gates, and compounding memory, with durability measured in years. The framework's core warning is embedded in that structure: value escapes the surface and accumulates in the layers below. Own the lower layers, or rent them — and rent your future.

L-1: Resources — the ground itself

The base layer is physical, and in 2026 it is the one humbling every strategy deck. L-1 covers energy and grid interconnect, thermal and water management, fabrication and foundry capacity, critical materials, and the skilled trades — electricians, HVAC technicians, data-center builders, fab process engineers — that no model can synthesize. In the framework's gold-rush analogy, this is the land, the water rights, and the miners. When demand spikes, this layer is the real bottleneck, and the multi-year grid-interconnect queues at U.S. utilities are the proof. Think NextEra and Vistra on power, TSMC's fabs, MP Materials on rare earths, Bechtel on construction. No AI strategy survives contact with L-1 arithmetic.

L0: Infrastructure — the shovels

One layer up sits the compute substrate: silicon and memory, data centers, interconnect fabric, compute and state infrastructure, and edge and on-device compute. This is the NVIDIA and AMD layer — the most visible chokepoint in the entire chain and the one Wall Street has already priced accordingly. The strategic point SCoI makes about L0 is that it is a bottleneck, not a destination: enormous margin concentrates here while supply is constrained, but infrastructure historically commoditizes, and the framework's structural laws describe what survives the compression.

L1: Data — the raw ore

L1 divides into public and open data, proprietary data, behavioral and sensor data, outcome data, and synthetic and simulation data. The starred sublayers — proprietary data, behavioral data, and above all outcome data — are where the framework says durable advantage forms. Public data is a commodity; every foundation model has read the same internet. Outcome data — records of what actually happened when your product was used — is the scarce input, because it is generated by owning a workflow, and it compounds. Bloomberg and ZoomInfo are canonical L1 businesses; the deeper lesson is that L1 advantage is usually earned by winning a higher layer first and capturing the exhaust.

L2: Models — the refinery

The model layer covers foundation and multimodal models, specialized and fine-tuned models, embedding and retrieval, model routing and composition, and reasoning and world models. This is the OpenAI, Anthropic, and Google DeepMind layer, and it gets a disproportionate share of press coverage relative to its share of durable value. SCoI's argument is not that models do not matter — the refinery matters enormously — but that the refinery is contested, capital-hungry, and subject to relentless compression as open-weight and frontier models converge. For most U.S. companies, L2 is a layer to rent, not to build.

L3: Gates — the assay office

L3 is the layer most stack diagrams crop out entirely, and it is one of the framework's most original contributions. Gates cover compliance and export controls, quality gates, safety, security and provenance, editorial gates, and distribution gates — the trust checkpoints intelligence must pass through before it can be sold into regulated or reputation-sensitive markets. Vanta, Drata, and OneTrust are the recognizable names. In healthcare, finance, legal, and government, the gate is often the business: whoever controls certification into a regulated workflow controls the toll booth every model and application must pay.

L4: Access — the railroads

L4 is how intelligence reaches the systems where work happens: APIs and integrations, agent interface protocols, access governance and agent commerce, real-time interaction infrastructure, and agent identity and provenance. AWS, Snowflake, and Supabase sit here. In 2026 this layer is in unusual flux because a second kind of consumer — the autonomous agent — is arriving with its own protocols and its own commerce. Companies that defined access for human developers are racing to remain the railroad when the freight is agents calling tools, and the agent identity and provenance sublayers are effectively greenfield.

L5: Execution — the master jeweler

L5 is where intelligence performs actual work: domain execution and tool use, decision frameworks and reasoning scaffolds, retrieval-augmented workflows, operating playbooks, and interaction skills. Harvey in legal and Sierra in customer experience are the exemplars — companies that do not sell a model but sell a completed job in a specific domain. This is the layer where vertical depth pays: a generic assistant is a commodity, but a system that executes a regulated closing checklist or a clinical intake with auditable accuracy is not. The catch, per the framework, is that L5 companies live directly beneath the platforms and must constantly answer the absorption question.

L6: Orchestration — the workshop

L6 coordinates the work: agent loops, human-in-the-loop design, role routing and task decomposition, context and state management, and runtime assurance. LangChain, CrewAI, and Zapier are the familiar names — and the framework pointedly flags parts of this layer as at risk. Orchestration that is pure plumbing gets absorbed by the platforms above and the frameworks below; orchestration that owns human-in-the-loop judgment and runtime assurance in a consequential workflow is much harder to commoditize. The difference between the two is the difference between a business and a feature.

L7: Surface — the moment of experience

The surface is what users actually touch: conversational interfaces, visual interfaces and media, embedded and embodied AI, transaction surfaces, and async and ambient surfaces. ChatGPT, Gemini, and Copilot live here. L7 wins attention, headlines, and App Store charts — and the framework is blunt about its economics. Surface is the most easily replicated layer, with durability measured in weeks, because the platforms ship surface improvements for free. A thin wrapper around someone else's model is an L7 business with no layers underneath it, which is the framework's definition of a wrapper that a platform will absorb.

L8: Memory — the record book

The top layer is the one the framework treats as the deepest moat in the stack: session and short-term memory, user and entity profiles, aggregated network learning, institutional knowledge, and learned world models. Memory is why a product gets better for customer number ten thousand than it was for customer number one. It compounds, it is nearly impossible to port to a competitor, and it belongs to the Substrate tier even though it sits at the top of the chain — its durability is measured in years. SCoI places L8 in the substrate deliberately: what users depend on is not the chat window but the accumulated knowledge the system holds about their work.

"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, supplychainofai.com

How to read the map like an operator

The layers are vocabulary; the analysis is the point. The practical exercise from the framework page is to place your own product on the map and ask the uncomfortable questions. Which layers do you actually own, and which do you rent? Is the layer you rent from a bottleneck that can raise your price or absorb your product? Does your position generate outcome data (L1d) or compounding memory (L8) as a byproduct, or does value escape upward the moment it is created? The case studies apply this to 24 real companies, and the market map positions companies by layer rather than by category — a much more honest picture of who is exposed and who is compounding.

The honest caveats

Layer boundaries in AI are blurrier than in oil refining, and real companies routinely span three or four layers at once — NVIDIA sells infrastructure and is pushing into models; the frontier labs are climbing into surface and execution. A 10-layer grid compresses a messy economy into tidy boxes, and the scoring of defensibility still requires judgment. The framework's authors concede the architecture itself will evolve — versioned paper bumps rather than quiet edits. But the test of a strategy map is not completeness; it is whether it improves the questions a leadership team asks. On that measure, the 10 layers earn their keep.

The short answer

The 10 layers of the Supply Chain of Intelligence run L-1 Resources, L0 Infrastructure, L1 Data, L2 Models, L3 Gates, L4 Access, L5 Execution, L6 Orchestration, L7 Surface, and L8 Memory — grouped into Surface, Workflow, and Substrate tiers with durabilities of weeks, months, and years. Read them once and the 2026 AI economy stops looking like a logo quilt and starts looking like what it is: a supply chain, with bottlenecks, toll booths, and a small number of layers where value actually stays. The full framework, the paper, and the case studies are all published openly at supplychainofai.com.