Mapping the AI Supply Chain: Where the Spending Actually Lands
A dollar of AI capex fans out across a fixed chain of income statements — and the economics at each node are set by market structure there, not by the size of the flow passing through it.
Sophisticated investors already know to distinguish AI narrative from AI revenue. The subtler error survives that filter: treating AI capex as a single pool of demand that lifts every company it touches, when it is actually a flow that decomposes through a fixed chain of income statements — and the share each link keeps has almost nothing to do with the size of the flow passing through it. Two companies can sit in the same buildout, book revenue from the same data center, and have opposite economics. The map of where the spending lands is not the map of where the value stays.
Start with what a dollar of data-center capex actually buys. The largest share goes to compute hardware — the accelerators themselves, and around them the memory that feeds them, the networking that stitches thousands of them into one machine, and the systems integration that assembles the racks. A second tranche goes to the electrical chain: generation or grid connection, transformers, switchgear, backup power, and the cooling plant that carries the heat away. A third goes to the physical shell: land, structure, construction labor, and the long tail of fittings. Upstream of all of it sit the foundries, the advanced-packaging capacity, and the equipment makers who supply them. Each of these is a node, each node is a market with its own structure, and the structure — not the flow — sets the economics.
Toll booths and pass-throughs
The mechanism that splits a node's revenue into margin or mere throughput is the same everywhere in industrial economics: what does the buyer do if this supplier raises price? At a toll-booth node, the answer is "pay it" — because the input is sole-sourced or nearly so, because it is designed into the system at a level where substitution means requalifying an entire architecture, or because capacity is reserved years ahead and the queue itself is the product. Frontier accelerators, the software ecosystems that lock developers to them, high-bandwidth memory during periods of qualification scarcity, and advanced packaging capacity have all functioned as toll booths for stretches of the cycle. At a pass-through node, the answer is "take the next bid" — construction, standard electrical gear in normal supply, generic shell development, commodity components. These nodes can report spectacular revenue growth during a buildout while competing away nearly all of the pricing, because the buyer faces many interchangeable sellers and procures by auction.
The distinction is invisible in a revenue chart and decisive in an income statement. During the expansion phase, both node types grow, correlations across the chain run high, and the market often rewards them with similar multiples — which is precisely the mispricing. A pass-through node's growth is real but rented; it holds margin only while demand outruns industry capacity, which is a statement about the cycle, not about the business. A toll booth's margin is structural; it holds until the market structure itself changes. Paying a structural multiple for cyclical margin is the characteristic error of buildout investing, and it is committed most often by investors who correctly identified the buildout.
The bottleneck migrates
The second layer of the framework is dynamic: pricing power lives at the bottleneck, and the bottleneck migrates. A buildout of this scale is a convoy moving at the speed of its scarcest input. Whichever node is the binding constraint collects economics wildly out of proportion to its cost share — expedite fees, take-or-pay reservations, multi-year prepayments — because at the margin it is not selling a component, it is selling the ability of the entire project to exist. But abnormal margin at a bottleneck is the industry's signal to add capacity there, and capacity additions at a node do not end the buildout; they move the constraint to the next node. Compute scarcity gives way to packaging scarcity, packaging to memory, memory to power delivery, power equipment to grid interconnection and energization timelines. The overall flow of spending persists, but the identity of the toll booth changes — and with it, which income statements are collecting rents.
Paid once, paid per refresh, paid per operating year
The third layer is duration, and it is the one most often skipped. Nodes differ not just in what share of the flow they keep but in how many times the flow visits them. Some nodes are paid once per buildout: land, shell construction, core electrical infrastructure. Their revenue is a function of new facility starts, and it stops when the footprint stops growing — even if the installed base thrives. Some are paid once per refresh cycle: accelerators, memory, much of the networking layer. Because the compute fleet ages out on a short competitive clock, this revenue recurs as long as the installed base is maintained at the frontier — a materially more durable claim than the construction node holds, though hostage to the refresh cadence. And some are paid per operating year: power, cooling operations, maintenance, connectivity, and the landlord's lease itself. These nodes convert a one-time buildout into an annuity, which is why capital markets price them on an entirely different basis. Two suppliers with identical revenue this year can face opposite futures purely because one is monetizing facility starts and the other is monetizing the installed base. Sterling's Embedded Intelligence classifies supply-chain exposure along exactly this axis first, because duration of claim is the difference between a cyclical windfall and a business.
The illustration: picks and shovels that stopped selling
The picks-and-shovels metaphor is usually deployed as if it settled the question — sell to the miners, win regardless. History is less kind to the metaphor than its users are. In the telecom buildout of the late 1990s, the purest picks-and-shovels businesses were the equipment vendors, and they briefly earned extraordinary economics as the bottleneck node of their cycle. But their claim on the flow was of the once-per-buildout kind: when carrier capex stopped, their revenue did not decelerate — it collapsed, because there was no installed-base annuity underneath it and no refresh clock forcing repurchase. Meanwhile the durable winners of that era's aftermath were operating-year nodes: the businesses that ran services on top of the overbuilt capacity once it became cheap.
The memory industry offers the complementary lesson from the other direction. For most of its history, memory was the canonical pass-through node — a commodity procured by auction, with margins whipsawed by the capacity cycle. What periods of qualification-constrained, application-specific demand have shown is that the same physical product category can temporarily become a toll booth when the market structure at the node changes: fewer qualified suppliers, capacity reserved in advance, pricing negotiated rather than bid. The node did not move; its structure did. Both lessons compress to the same rule: the metaphor tells you where the spending goes, the market structure tells you what stays, and the duration class tells you for how long.
The strongest case against this framework
The serious objection is that in a demand environment this strong, the map is pedantry: when the tide is rising fast enough, every node earns above its structural entitlement, correlations approach one, and the investor who spent the cycle grading market structures underperformed the one who simply owned the flow. There is a version of this that is empirically fair — in the steep phase of any buildout, node selection has tended to matter less than participation, and an over-refined framework can become a reason to own too little of a generational capex wave.
We accept the observation and reject the conclusion, because the objection is really a statement about when the framework pays, not whether it does. Undifferentiated exposure and structured exposure look identical while spending accelerates; they separate violently at the inflection. When the flow decelerates — not ends, merely decelerates — pass-through nodes give back margin first, once-per-buildout nodes see revenue vanish rather than slow, and toll-booth and operating-year nodes keep collecting. The map is not a tool for predicting the turn; it is a tool for ensuring that whenever the turn comes, you are holding claims on the flow that survive it. That is a portfolio-construction discipline, and its cost during the boom — owning somewhat less of the most cyclical winners — is the premium on the insurance.
How to apply this framework
- Classify every supply-chain holding by node structure. Ask what the buyer does if this supplier raises price: pay it, or take the next bid. Sole-source design-ins, reserved capacity, and ecosystem lock-in mark toll booths; auction procurement marks pass-throughs whose margin is rented from the cycle.
- Then classify by duration of claim. Paid once per buildout, once per refresh cycle, or per operating year. Facility-start revenue ends when the footprint stops growing; refresh revenue recurs while the fleet stays at the frontier; operating-year revenue annuitizes the installed base. Identical revenue, different futures.
- Track the bottleneck, because rents rotate toward slow supply responses. Capacity announcements at the current constraint are the leading indicator of margin reversion there — and of pricing power arriving at the next-slowest node, historically the power and grid chain, where supply takes years to answer.
- Refuse structural multiples on cyclical margin. When a pass-through node trades at a toll booth's valuation, the market is capitalizing scarcity that its own capacity cycle is already working to remove. The reverse — a toll booth priced as a commodity during a lull — is where the map earns its keep.