The Supply Chain of Intelligence: What It Means for Founders and Investors

Four essays into Anand Arivukkarasu's Supply Chain of Intelligence series, a pattern had emerged that our research desk kept returning to. The framework worked beautifully as description. It mapped the AI economy from resources through compute, data, models and agents to AI memory. It explained why value pools at bottlenecks and how to grade a moat. What it had not done, explicitly, was tell the two groups who read it most closely what to do on Monday morning. The closing installment, published this month at supplychainofai.com, corrects that. It is addressed directly to founders and to investors, and it is the most prescriptive writing in the series.

The framing device is a fair one. A theory of where value accrues is only useful if it changes behavior. Founders decide where to build. Investors decide where to allocate. Arivukkarasu's argument is that both groups, in the current American AI market, are systematically aiming at the wrong stages, and that the framework's own logic explains why.

## The founder's build order

The advice to founders begins with an uncomfortable observation. The modal AI startup pitch in the United States in 2026 is still a model-layer pitch: a fine-tuned model, a clever orchestration of frontier APIs, a new interface on top of someone else's weights. Under the framework's second structural law, which holds that each stage commoditizes faster than the last, the model layer is the worst place to anchor a company. The substrate improves underneath you, and your differentiation evaporates with each frontier release.

Arivukkarasu's build order runs the other way. Start at the stage where position compounds. For most founders, that means one of two places. The first is coordination capacity: the unglamorous work of integrating AI into a specific industry's workflows, approvals, compliance regimes and handoffs. It is slow to build and slow to copy, and it survives every model upgrade because the upgrade makes the coordination layer more valuable, not less. The second is AI memory: the accumulated, structured record of a customer's context, decisions and outcomes that makes every subsequent interaction better. Memory compounds with use, resists substitution because it cannot be exported cleanly to a rival, and sits at the chain's terminal stage where the framework says value finally rests.

The practical test he offers founders is blunt. Ask which frontier model release would hurt your business. If the answer is any of them, you are building on rented ground. If the answer is none of them, because every better model makes your coordination layer or your memory asset stronger, you are building on the right stage. It is a founder-friendly restatement of the moat essay's commoditization test, and it lands harder in a pitch meeting than in a theory paper.

## The investor's diligence checklist

The investor section is where the essay turns into a working document. Arivukkarasu converts the earlier installments into a sequence of diligence questions, and the sequence is the point. First, locate the company on the chain. Second, ask whether that stage is a current or future bottleneck, because value accrues at bottlenecks and nowhere else. Third, grade the moat with the three tests from the previous essay: does the position compound with use, resist substitution, and survive commoditization of the layer beneath it. Fourth, and this is the new material, ask who captures the memory.

That fourth question does more work than the other three combined. A company can pass the bottleneck test today and still be a pass-through for value that ultimately settles elsewhere. The American market is thick with businesses that sit at a temporary narrow point, a clever routing layer, a proprietary prompt library, an early agent framework, while the durable memory of customer context accumulates in someone else's system. Arivukkarasu argues that the diligence question that separates the next decade's returns from its write-offs is not where the company stands, but where the memory it helps create will live. If the answer is inside the customer's own systems with no portability back to the vendor, the vendor is a contractor. If it compounds inside the vendor's platform, the vendor is a franchise.

He is equally direct about what the checklist rules out. Pure model companies without a proprietary data flywheel fail the commoditization test. Wrapper businesses fail the substitution test. Compute intermediaries fail the compounding test. What remains is a narrow band of company archetypes: vertical coordination platforms, memory-native applications, and the infrastructure that serves the bottlenecks the chain migrates toward, power, interconnection, and the governance layer around agents. For US investors, he notes, this has a portfolio implication. The returns of the AI buildout so far have concentrated in the bottleneck owners, the chip makers and the cloud platforms, and the framework predicts the next concentration one stage further along, in the layers where enterprise memory and coordination consolidate.

## Where the two audiences meet

The essay's most interesting passage is the one where founder advice and investor advice collide. A founder following the build order will deliberately choose a slower, less legible business: coordination work does not demo well, and memory assets take quarters to show up in retention curves. An investor following the checklist will learn to pay for exactly that illegibility, because the businesses that look clean in a demo are usually the ones the next model release will erase. Arivukkarasu frames this as the framework's final law in practice: in a chain where every visible advantage commoditizes, the durable advantages are the ones that are hard to see.

There is a quiet rebuke in that for the American funding market of the past three years, which rewarded velocity of narrative over depth of position. The essay does not gloat. It simply observes that the diligence questions that would have filtered the graveyard of wrapper startups were available in the framework from the first installment, and that the same questions, applied now, point toward a smaller and less fashionable set of companies than the ones currently commanding attention.

## The honest footnote

The caveats are consistent with the rest of the series. Prescriptive frameworks age faster than descriptive ones, and this is the most prescriptive of the five essays. The memory thesis still depends on enterprise data ownership norms that US contracts and regulation have not settled. A breakthrough in model capability that makes coordination trivial would compress the very stages the build order favors, and Arivukkarasu says so plainly. And the checklist, like any checklist, can be gamed by a founder who has read the same essay. The defense he offers is the right one: the tests are about structural position, not presentation, and structural position is the one thing a pitch cannot fake for long.

Read as a whole, the five installments now form a complete arc: the map, the laws, the economics, the defenses, and the playbook. For an American founder choosing what to build in 2027, or an investor deciding what to underwrite, the series is best consumed in order and in full at supplychainofai.com. The final essay rewards the patience. It is rare for a framework to end by telling its readers exactly what to do, and rarer still for the instruction to follow logically from everything before it.