The 50 sublayers of the Supply Chain of Intelligence, explained
Ten layers get quoted in pitch meetings. The fifty sublayers underneath them are where the argument actually gets settled, because a moat almost never sits at a layer. It sits at one box inside one.
Say “we own the data layer” in an American board meeting and the room nods. It is a sentence that means almost nothing. Public web scrape and outcome data are both “the data layer,” and one of them is worth zero while the other is close to unbuyable. That gap is why the Supply Chain of Intelligence does not stop at ten layers.
The framework, written by former Meta product leader Anand Arivukkarasu and published free at supplychainofai.com, breaks each of its ten layers into five sublayers, for fifty in total. The ten are the vocabulary. The fifty are the diagnostic.
A layer tells you which neighborhood you are in. A sublayer tells you whether you own the building.
Why five boxes per layer, and why that matters
The choice of exactly five is a discipline, not a discovery. It forces the framework to name the meaningfully distinct positions inside a layer and stop, rather than fanning out into a taxonomy nobody can hold in their head. In practice the five inside any layer are ordered roughly by how hard they are to displace, and the spread between the first and the fifth is usually larger than the spread between two adjacent layers.
That is the practical payoff. Two companies can occupy the same layer and be in entirely different businesses. Both are “memory companies” if you look at L8; one holds session transcripts and the other holds a decade of an institution’s decisions. Only one of them survives the next model release.
All fifty, layer by layer
What follows is a plain-English read of each sublayer, with the part that actually matters to a US operator or investor. Canonical definitions, versioned and dated, live on the framework pages.
Resources
The physical world the whole thing rests on. Nothing here can be shipped in a sprint, which is exactly why it prices like scarcity.
Generation plus the right to plug in. In several US markets the interconnect queue, not the turbine, is the binding constraint, and queue position has become a tradable asset.
Cooling capacity and water rights. It decides which counties can host a campus, and it is where local politics meets a training run.
Leading-edge foundry capacity and advanced packaging. Measured in years and tens of billions, and allocated by relationship as much as by price.
Rare earths, substrates, high-bandwidth memory, transformers, switchgear. Boring components with 18-month lead times that quietly set everyone's ship date.
Electricians, pipefitters, commissioning engineers. The least discussed shortage in American AI, and the one you cannot fix with capital alone.
Infrastructure
The shovels. Won on capital cycles and utilization, not on features.
Accelerators, HBM, custom inference chips. Margin here tracks scarcity, and scarcity is currently upstream in memory rather than in logic.
Shells, power density, siting. A real-estate business wearing a technology costume, and priced accordingly.
Rack-scale fabric, optics, long-haul. Training economics live or die on the network, not the chip count.
Schedulers, storage, caching, checkpointing. Where the difference between a good and bad cloud bill actually gets made.
Phone and laptop silicon, private inference. Cheap, private and always-on, and a genuine threat to per-token pricing.
Data
The single largest source of durable defensibility available to an application company without a fab or a regulator behind it.
The open web and licensed corpora. Everyone has it, litigation is reshaping it, and it is worth roughly zero as a moat.
Contracts, claims, transcripts, clinical records. The classic moat, but only when you hold the rights to train and reuse, which most vendors quietly do not.
What users and machines actually did, captured in the run of work. Cheap to collect if you own the workflow, impossible to buy if you do not.
Did the recommendation work. The rarest and most valuable category, because it is the only one that lets you improve rather than imitate.
Generated coverage for rare and dangerous cases. Fills gaps, and collapses into self-confirmation without real outcome data to anchor it.
Models
Capability rises, price falls, switching costs are lower than the pitch decks claim.
Frontier general models. Extraordinary businesses at the top, brutal for everyone attempting to be the fifth-best one.
Domain models. Defensible only when the fine-tuning data is yours and keeps refreshing, otherwise a six-month lead.
Vectorization, indexing, ranking. Unglamorous, and often the difference between a demo and a system people trust.
Sending each request to the cheapest model that can handle it. A real margin lever, increasingly absorbed by the platforms.
Planning, long horizons, physical prediction. The live research frontier, and where the next capability discontinuity is most likely to land.
Gatekeeping
Permission to operate. Structurally permanent wherever output carries legal, financial or reputational weight.
Sector rules, chip export regimes, data residency. A cost for most companies and an entire business model for a few.
Evals, benchmarks, acceptance testing. Whoever defines the passing grade quietly shapes the market.
Red teaming, watermarking, content credentials, audit trails. Rising fast as procurement starts asking for evidence rather than assurances.
Human judgment about what gets published or acted on. The layer this newsroom happens to sit in, and it does not automate away.
App stores, marketplaces, procurement lists, GPO contracts. Someone decides what reaches the buyer, and rents that decision.
Access
The pipes. Nobody markets them and everything depends on them.
Programmatic reach into systems of record. Whoever holds the API relationship holds the account.
MCP and its neighbors. A standards fight in progress, and standards fights decide who is a platform.
Scopes, spend limits, approvals, machine-to-machine payment. The blocker between an impressive pilot and production.
Streaming, events, low-latency transport. Necessary for anything that has to answer while a human is still waiting.
Who is this agent, who authorized it, what is it liable for. Unsolved, and an obvious bottleneck forming in plain sight.
Execution
Where the job actually gets done, and one of the three corners of the Defensible Triangle.
Encoded expertise in underwriting, coding, prior authorization, discovery. Depth is the moat; breadth is a demo.
The rules for choosing under uncertainty. Usually the part a customer cannot articulate and will pay to have built.
Grounding answers in the right documents at the right moment. A plumbing discipline masquerading as a feature.
Operating procedures that survive staff turnover. Boring, and the reason some deployments still work in year three.
Negotiating, de-escalating, and touching real systems and machines. Where software stops describing the world and changes it.
Orchestration
Necessary, rarely a business on its own. Most of this is being absorbed by the model or the surface.
Plan, act, observe, retry. Genuinely hard to make reliable, and increasingly shipped inside the model.
Escalation, approval, override. The design question is who is accountable when the loop closes wrongly.
Assigning work across specialized agents and people. Org design, expressed in software.
What the system knows right now, and what it forgets between steps. Quietly the source of most agent failures.
Guardrails and circuit breakers while the thing is running. Distinct from pre-launch evals, and much less mature.
Surface
Modality is a commodity. Placement and habit are not.
Chat and voice. A default that stopped being differentiating some time in 2024.
Canvases, dashboards, generated interfaces. Strong where the output needs to be inspected rather than read.
Inside the tool you already use, or inside a machine. Where distribution beats quality almost every time.
Surfaces that complete purchases and filings. Being close to money changes both the economics and the liability.
Always-on, low-attention assistance. Small today, and the most plausible next battleground for attention.
Memory
The compounding layer, and the one most product roadmaps treat as a storage detail.
What happened in this conversation. Table stakes, and worth nothing competitively.
What we know about this customer, contract or patient over time. The first version that produces real switching cost.
Learning across customers without leaking between them. Hard, legally sensitive, and enormously valuable when it works.
Why this company decided things the way it did. Currently lives in the heads of people who are about to retire.
A learned, updating model of a domain's dynamics. Rare, early, and the closest thing to a permanent advantage on the list.
What the list changes in a real meeting
The fifty are most useful as a scoring sheet rather than a reading list. Mark each sublayer own, rent or exposed. Owning means you control it and a competitor cannot buy the equivalent this year. Renting means you pay someone and could switch. Exposed means someone else’s roadmap decides your margin.
Most application companies come out of that exercise owning between one and three boxes, which is a perfectly good answer if the boxes are the right ones. The failure pattern is not owning too few. It is owning three boxes that all sit in L6 and L7, the two layers the framework argues are being absorbed fastest by the models and the surfaces beneath them.
The shape worth aiming at, in the framework’s language, is the Defensible Triangle: proprietary data at L1b, real execution depth at L5, and compounding memory at L8. Read as sublayers, that is three specific boxes out of fifty, which is a far more demanding claim than “we play across data, execution and memory.”
Where the list is soft
Boundaries blur. Retrieval appears both as an L2 model concern and an L5 execution discipline, and reasonable people will file the same engineering work in different boxes. The framework treats that as acceptable, since the point is provoking the argument rather than winning a filing dispute.
The market readings also date quickly by design. Which companies sit in which sublayer changes quarterly; the structure is meant to outlast the examples, and the author versions the paper instead of pretending otherwise. Treat the fifty as a map that needs maintenance, not a table of constants.
The short version
Ten layers is a vocabulary you can use in public. Fifty sublayers is the version you use privately, with your own product on the table and nobody selling anything. If your defensibility claim cannot survive being restated as two or three sublayer codes, it was probably an adjective all along.
The full definitions, glossary and market readings are free at supplychainofai.com/framework, and the canonical citation is supplychainofai.com/paper.