If you work anywhere near AI in the United States right now, you have seen the stack diagram. Four boxes, sometimes five. Compute at the bottom. Foundation models above it. A tooling or orchestration band. Applications on top, usually with a little chat bubble. It is on venture decks, cloud provider blog posts, conference keynotes and roughly every internal strategy memo written since 2023.

The diagram is not wrong. It is just answering a question nobody is losing money on. Engineers already know that an application calls a model and a model runs on GPUs. What boards, buyers and investors keep getting wrong is something the stack diagram is structurally incapable of showing: which of those boxes actually keeps the margin, and which one gets absorbed by the layer beneath it the next time OpenAI ships on a Tuesday.

That gap is the reason the Supply Chain of Intelligence framework exists, and it is the cleanest way to explain what it does differently. So let us put the two side by side properly.

What the traditional AI stack actually describes

The conventional stack is an inheritance from cloud computing, and it inherited cloud's assumptions along with its shape. In the cloud era, layers were reasonably stable: IaaS, PaaS, SaaS. Each layer had recognisable owners, and boundaries between them moved slowly enough that a diagram stayed accurate for a few years.

Applied to AI, the layers usually come out as: infrastructure (chips, data centers, networking), models (foundation, fine-tuned, embeddings), tooling or middleware (vector databases, frameworks, evaluation, orchestration), and applications (the product a human uses). Some versions add a data layer, some add agents, some split inference out of infrastructure.

As a build map, it works. If you are designing a system, this is the right vocabulary: it tells you what to buy, what to host, what to abstract and where the dependencies run. As a description of technical reality, it is fine.

What it cannot express is any of the following. Which layer is scarce this quarter. Which layer is a gate somebody must pass through. Which layer the layer below wants to eat. How long any given layer stays a business. Whether the sector — legal, health, defense, fintech — changes the answer. Or the single most consequential fact in the market: that a layer can be technically essential and economically worthless at the same time.

What the supply chain view changes

The Supply Chain of Intelligence takes the same territory and re-reads it as a chain of production rather than a pile of components. Ten layers, numbered L-1 through L8: resources, infrastructure, data, models, gates, access, execution, orchestration, surface, memory. The framework's own comparison is blunt — the AI stack explains how intelligence is built; the supply chain explains where intelligence becomes economically defensible.

Three things fall out of that change of frame, and each one is invisible on a standard stack diagram.

First, the chain extends below the chips. L-1 is energy, grid interconnection, cooling water, foundry capacity, critical minerals and the electricians who physically build data centers. American readers watching interconnection queues in Northern Virginia and central Texas stretch past five years understand this instinctively: no software roadmap moves faster than the megawatts arrive. Traditional stack diagrams start at compute, so they simply cannot show the constraint that currently governs the whole industry.

Second, it adds a layer for gatekeeping. L3 — compliance, export controls, quality gates, safety and provenance — has no equivalent in the standard stack, because architecture diagrams do not model who is allowed to say yes. Vanta and Drata are not middleware. They are the assay office that stamps the gold.

Third, it treats memory as its own layer at the top rather than as application state. Session memory, user profiles, aggregated network learning and institutional knowledge compound over time, which makes them behave economically unlike anything else on the chart.

The comparison, laid out plainly

Purpose. The stack answers a build question: what components does this system need? The supply chain answers a defensibility question: is this a moat, a workflow, or a wrapper a platform will absorb?

Shape. The stack is a pile — layers sit on layers. The supply chain is a flow with narrow points, which is what lets it talk about bottlenecks at all. Value does not accrue evenly across a chain; it collects wherever supply is constrained.

Bottom boundary. The stack starts at silicon or cloud. The chain starts at electricity, water, minerals and human trades.

Top boundary. The stack usually ends at the application or the user. The chain continues past the interface into memory, because what accumulates after the interaction is frequently worth more than the interaction.

Time. This is the sharpest difference. A stack diagram is timeless; it implies every layer is equally permanent. The supply chain assigns each tier a shelf life. Surface layers are durable in weeks, because platforms ship interfaces for free. Workflow layers are durable in months, and only if you genuinely own the workflow. Substrate layers — resources, infrastructure, data, models, gates, memory — are durable in years, because proprietary data, trust roles and compounding memory are slow to build and awkward to copy.

Predictive claims. A stack makes none. The supply chain makes four, stated as falsifiable laws: intelligence commoditizes downward; value accrues at bottlenecks; the surface captures attention while the chain captures power; and generation must stay separate from verification.

Why that last law is the one to remember

Separation of generation and verification is the part of the framework with the most precedent behind it, and it is the part that most cleanly contradicts the stack view. On a stack diagram, verification is a feature — a box you could add to any layer. In the chain, it is a structural boundary that platforms cannot cross.

The reasoning is economic, not technical. Where output carries fiduciary, regulatory, safety or reputational weight, the generator and the verifier have to be different economic entities. The model cannot audit itself. The code generator cannot certify its own security. The drafter cannot approve its own filing. This is why the Big Four still audit companies running SAP, and why no American CISO accepts the same vendor writing the code and signing off on it.

Which produces a useful prediction. Foundation model labs will keep expanding upward into execution, orchestration and surface, because buyers accept it. They will not expand across the trust boundary into auditing their own output, because buyers do not. OpenAI will ship agents. OpenAI will not issue its own SOC 2 report.

That is a claim a stack diagram cannot make, disprove, or even phrase.

The same company, read both ways

Take a coding assistant. On the traditional stack it is an application: it sits on a model, which sits on GPUs. Clean, correct, uninformative.

On the chain, the questions get uncomfortable. Which layer does its revenue actually depend on? If the answer is L7 — a nice editor window over a general model — the first law applies and the platform below absorbs it. If the product also owns the agent loop, the hosting, the auth, the database and the deployment target, it holds L5 through L8 and the fate changes entirely. Two companies in the same category on the same stack diagram, structurally different outcomes.

Or take the cautionary case everyone in American AI already knows. A writing tool valued at $1.5 billion sat, in stack terms, in exactly the same place as any other application. In chain terms it held nothing scarce, and the layer below shipped its core capability directly. Wrappers do not survive contact with the layer they wrap; wrappers become features.

When to use which

This is not a replacement argument, and the framework does not make one. The two diagrams answer different questions and both questions are real.

Use the traditional stack when you are designing a system, choosing vendors on technical grounds, drawing an architecture review, or onboarding an engineer. It is the correct map for build decisions.

Use the supply chain view when you are writing a strategy memo, pricing a round, deciding whether to build or buy, defending a roadmap to a board, or assessing whether a supplier will still exist in three years. It is the correct map for money decisions.

A practical way to run both: draw the stack to see what you depend on, then place your revenue on exactly one chain layer — the one it actually comes from, not the one you would prefer — and ask the three hard questions. If the platform below you shipped your core capability free tomorrow, what remains? What scarce thing do you own that a funded competitor cannot buy in a quarter? Does anything in your product compound across customers, or does every new customer start you at zero?

Where the supply chain view is weaker

Two honest limits, and the framework's author names both.

The taxonomy is a claim, not a measurement. It is published as a versioned paper precisely because ten layers may become twelve as the field moves. Any single company placement is a snapshot, and reasonable analysts will argue about whether a given product lives at L5 or L6. The stack diagram, being simpler, is harder to get wrong — it is also harder to get anything useful out of.

And timing remains the soft spot. Knowing a layer will commoditize is much easier than knowing when. Plenty of companies have been structurally correct and eighteen months early, which in this market is indistinguishable from being wrong. The framework tells you the direction of the compression; it does not date it.

The short version

The traditional AI stack is a wiring diagram. The Supply Chain of Intelligence is a map of where the tolls are collected. If your job is to make the system work, keep the stack. If your job is to explain why your company will still be here after the next model release, the stack will not help you and the chain might.

The framework, the versioned paper, a sortable classification of notable AI companies by layer and archetype, and the case studies are published free at supplychainofai.com. Our earlier explainer covers the ten layers and the four laws in full if you are meeting the framework for the first time.