Most strategy frameworks die in the gap between the whitepaper and the Monday standup. The Supply Chain of Intelligence — SCoI, the ten-layer defensibility map published by Anand Arivukkarasu, a former Meta (Instagram) product leader and angel investor in San Francisco — is unusually resistant to that fate, because it was built as a working tool rather than a manifesto. But "applicable" is not the same as "applied." This article is the missing middle: how an American AI startup actually runs the framework, what it tends to reveal, and where founders most often flinch when the answer is inconvenient.

The raw materials are all public. The framework page lays out the ten layers, fifty sublayers, four structural laws, three currents, and the Intelligence Cube. The paper, v1.0, published in January 2026, adds the reasoning protocol and the AI Defensibility Audit scorecard. The case studies apply the whole apparatus to real companies — Jasper, Chegg, Tempus, Deere, Waymo, and two dozen others. What follows is how a startup team of five to fifty people can put that machinery to work.

Step one: locate yourself honestly

The first exercise sounds trivial and is not. Place your product on the map: L−1 Resources at the base, then L0 Infrastructure, L1 Data, L2 Models, L3 Gatekeeping, L4 Access, L5 Execution, L6 Orchestration, L7 Surface, and L8 Memory at the top. The common founder error is locating the company where the pitch deck says it lives, rather than where the value actually accrues. A startup that calls itself "an AI platform" but whose entire product is a chat interface over someone else's model is at L7 Surface — the least durable layer on the board, with defensibility measured in weeks.

The corrective question is simple: which sublayers do we control, and which do we rent? Almost every U.S. AI startup rents L0 through L2 — compute, models, and public data are bought, not owned. That is fine. AWS-era startups rented infrastructure too. The problem is only renting, all the way up to L7, with nothing below the surface that a customer would miss if you disappeared. If your honest map shows a single layer with no controlled sublayer beneath it, the framework has already done its first job.

Step two: run the four laws, not the vibes

With the position fixed, apply the four structural laws from the paper. Intelligence commoditizes downward: if your edge is model capability, you are renting an advantage the market reprices quarterly. Value accrues at bottlenecks: is there a constrained stage in your customer's workflow that you control — a proprietary dataset, a compliance approval, an integration no one else holds? The surface captures attention, the chain captures power: demos win meetings; layers below the demo win renewals. And generation and verification must be separate: in every regulated American vertical, the trusted checker is a position in itself, and often the most defensible one available to a startup.

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

— Supply Chain of Intelligence, Framework §01

Step three: name your absorber

Every AI startup has an absorption risk — a platform above or below it that could ship the product as a feature. The exercise the framework forces is to name it specifically. Not "OpenAI might compete with us" but "our product is a thin orchestration of capabilities the model provider is actively building, and our switching costs are a settings page." Jasper's trajectory, covered in the case studies, is the canonical American example: a genuinely beloved surface product, built on rented models, repriced brutally once the model layer shipped the same surface for free. Chegg is the mirror-image case — a company that discovered its position was downstream of a free substitute it could not gate.

Naming the absorber is not fatalism. It is triage. The framework's relocation options are consistent across its case studies: move down into data with outcome linkage — records of what happened after your system recommended something, which no model provider can scrape; move sideways into a gate you can be trusted to hold — regulatory approval, editorial judgment, distribution rights; or move up into memory that compounds — institutional knowledge and learned context that only exists because the customer kept operating inside your product. Tempus and Deere, both mapped on the site, show what the first and second moves look like at scale.

Step four: check the currents before you commit

A defensible position in a layer with no current moving through it is a castle in a dry riverbed. SCoI's three currents — Demand Gravity, Attention Economics, and Capital Flows — run horizontally across all ten layers and decide whether a good position becomes a good business. A startup can hold a genuine L1 data bottleneck in a vertical where procurement budgets are frozen and starve anyway. The live feed exists precisely because layer positions and currents shift as platforms ship; the map is a snapshot, not scripture.

For U.S. founders raising in 2026, this step has a second use: it is the fastest way to pressure-test a VC's thesis. If a fund's enthusiasm for your round is driven by attention flowing to your layer rather than anything structural about your position, that is worth knowing before you sign — because attention currents reverse faster than any other.

Step five: use the Cube to pick your vertical

The Intelligence Cube is the framework's acknowledgment that the same layer behaves differently across industries. An L3 editorial or compliance gate in consumer software is a speed bump; in healthcare or financial services it is a multi-year moat. For a startup choosing between two markets, the Cube reframes the question from "which market is bigger?" to "in which market does my layer position convert into durability?" The case studies make the contrast concrete — the same execution-layer position that evaporates in marketing tooling persists for years in a licensed trade with liability attached.

What the audit usually reveals

Teams that run the full seven-question protocol at the top of the framework page tend to report the same three findings. First, the thing they were calling a moat was a head start — real, but depreciating. Second, they already own one genuinely defensible asset, usually outcome data they never instrumented or a trust relationship they never priced. Third, their roadmap was quietly optimizing the surface while the durable layers went unbuilt. None of these are fatal. All of them are fixable — but only once named.

Where startups should not over-apply it

Two cautions. Pre-product startups can paralyze themselves optimizing a position they have not earned; at zero revenue, speed to the customer's workflow matters more than layer theory, and the framework says as much by rating workflow depth as its own tier of durability. And the fifty-sublayer precision invites false confidence — placing yourself in L5b versus L6c is a judgment call, and treating it as measurement is the way smart teams talk themselves into bad maps. Use the AI Defensibility Audit as a recurring discipline — quarterly, alongside the metrics review — not as a one-time horoscope.

The short answer

Applying the Supply Chain of Intelligence to an AI startup is a five-move exercise: locate your layer honestly, run the four structural laws, name your absorber and your relocation path, check the currents, and use the Intelligence Cube to pick the vertical where your position converts to durability. The framework will not build the moat for you — but as the case studies of Jasper, Chegg, Tempus, Deere, and Waymo show, it will tell you with unusual clarity whether you are building one at all. Start at the framework, work the paper, and check your layer against the market map before your next board deck.