The Supply Chain of Intelligence: Why AI Value Accrues at Bottlenecks

This is the third time this season we have returned to Anand Arivukkarasu's Supply Chain of Intelligence, and readers who followed the first two essays will notice the progression. The first piece gave us the map: resources, compute, data, models, agents and AI memory. The second gave us the rulebook: four structural laws governing how the chain behaves. The new installment, published at supplychainofai.com, asks the question American boards actually ask. Where does the money go?

The answer, in Arivukkarasu's telling, is that value in the intelligence chain does not distribute itself according to effort, brilliance or even capability. It accrues at bottlenecks: the narrow points where demand presses against a constraint that cannot be quickly widened. Everything else in the chain, however impressive, trends toward commodity pricing. This is not a novel observation about AI. It is one of the oldest patterns in American industrial history, and the essay's force comes from applying it with discipline.

Consider the pattern American executives have already lived. The interstate highway system created enormous value, but the durable profits pooled at the toll booths, the fuel stops and the logistics hubs, not in the pavement. The buildout of the internet rewarded those who controlled the constrained points: search intent, the mobile operating system, the cloud region with available capacity. Even the shale boom, America's great resource story, saw margins migrate from drillers to pipeline and export terminal capacity the moment drilling stopped being scarce. Arivukkarasu's claim is that intelligence is now repeating the script, stage by stage.

The essay's most useful move is a simple diagnostic. A bottleneck, in this framework, has three properties. Demand exceeds supply and cannot be quickly balanced. The constraint resists substitution, so customers cannot simply route around it. And the position compounds: holding the bottleneck today makes it easier to hold tomorrow. Any stage of the chain can be tested against these three questions, and the answer changes over time. That last part is the point. Bottlenecks are not addresses. They are weather systems.

Applied to the current American landscape, the analysis is bracing. Compute was the obvious bottleneck of 2023 and 2024, and the market priced it accordingly. But the essay argues compute is already relaxing: supply has scaled, inference costs have fallen by an order of magnitude, and open-weight models have broken the pricing power of closed providers. The binding constraints have moved. Power and grid interconnection now gate new data center capacity in Virginia, Texas and Arizona, with interconnection queues stretching past five years. High-quality, permissibly governed data has become scarce in ways no scraping strategy fixes. And at the far end of the chain, trusted institutional memory, the accumulated context that makes an AI system genuinely useful inside one specific company, cannot be bought at any price. It can only be built.

For US executives, the practical reading comes in three questions, and the essay frames them well. First: where does our strategy assume abundance that is actually scarce? Many American firms are still budgeting as if power, evaluation capacity and integration talent are freely available. They are not. Second: where are we paying bottleneck prices for something already commoditizing? Model access is the standing example; the premium tiers of 2024 are the commodity tiers of 2026. Third: which constraint, if we controlled it, would compound? That question almost always lands on the same answer: the memory and workflow layer closest to the customer, the one stage of the chain that appreciates with use rather than depreciating with each model release.

There is a subtler implication for American investors. If value accrues at bottlenecks, then the correct question about any AI company is not how good its models are but what constraint it sits on. Utilities with generation capacity near data center corridors, firms holding proprietary operational data, and platforms that become the system of record for institutional memory are bottleneck businesses. Application companies reselling model access are not, whatever their growth rate. The essay stops short of naming tickers, but the screening logic is laid out plainly enough for any US portfolio manager to apply.

The caveats deserve their paragraph, as they did in the previous installments. Bottleneck analysis is easy to perform in hindsight and hard in real time; the difference between a durable constraint and a temporary shortage is only obvious years later. Regulation can create bottlenecks overnight and dissolve them just as fast, as US companies learned from export controls. And the framework's confidence about memory as the terminal bottleneck rests on portability and ownership questions that American enterprise contracts have not yet settled. Arivukkarasu flags all of this. A framework that prices its own uncertainty is worth more than one that pretends to certainty.

Read together, the three essays now form a complete argument. The chain is the anatomy. The four laws are the physics. This third piece is the economics: value pools where the flow narrows, the narrow points keep moving, and the discipline that pays is not predicting a single winner but tracking where scarcity migrates next. American readers who want the full treatment, including the bottleneck diagnostic in its original form, will find it at supplychainofai.com. Given how much US capital is currently allocated on the assumption that capability is scarce and coordination is free, the hour it takes is well spent.