When capital floods into a buildout, it does not pay for products. It pays for whatever the products cannot be made without. A railroad boom bids up land and rail and the men who lay it; a fiber boom bids up conduit and right-of-way and lasers; the current compute boom bids up chips and the machines that package them and the power that runs them. The market draws a ring around the whole supply chain and stamps every link inside it with the same word: scarce. Then it pays a premium for the word.
That premium is where the mistake lives, because scarcity is not one thing. It has a half-life, and the half-life runs in two opposite directions.
Some bottlenecks get harder to clear as the technology they serve advances. This is the rare and valuable kind. The constraint is set by physics or by a process so capital-intensive and so far down a learning curve that no amount of money buys past it on any horizon a portfolio cares about.
The compute buildout has a handful of these. There is one firm that volume-manufactures both leading-edge logic and the advanced packaging that stitches accelerators together, and the packaging step gets worse with every generation: more dies, tighter pitches, hybrid bonding measured in microns where the tolerances tighten as the stacks grow. The machine that bonds those dies has effectively one credible maker. The equipment that tests a finished AI chip faces test times that rise faster than the chips themselves, because complexity compounds inside the package. And the cleanest case sits outside silicon entirely: a fleet of always-on, carbon-free nuclear generation that took fifty years and a closed regulatory door to build and that no buyer can replicate inside a decade, no matter the check.
What unites them is direction. The moat widens as the field progresses. Each node, each stack, each gigawatt of demand makes the chokepoint scarcer, not less. You are buying a position that the future strengthens.
The other kind looks identical from a distance and behaves as its mirror image up close. The margin is real today, but it exists because supply is briefly tight or because one customer signed one contract. It is a peak, and peaks mean-revert.
Commodity memory is the textbook case. In a shortage its makers post gross margins north of eighty percent, numbers that belong to software, on a product that is physically a commodity, fungible by design and built by three rivals who have spent forty years teaching each other that discipline breaks the moment one of them adds a fab. The margin is not a moat. It is the shortage, wearing a moat's clothes. The same shape recurs wherever a single design win or a single tight quarter is being read as a permanent franchise: the rental economics of a graphics-card landlord, the connectivity socket that gets re-bid every silicon generation, the accelerator that has exactly one anchor customer.
The tell is in the multiple, and it is counterintuitive. A cyclical at peak margins often trades at a low forward price-to-earnings, and the untrained eye reads cheap. It is the opposite. The low multiple is the market's discount mechanism doing its job, marking earnings it knows are about to roll. To buy it you have to believe two things at once: that the peak earnings persist, and that the multiple re-rates upward from here. You are asking to be paid twice on the same bet. That is the signature of a value trap, not a value.
If the market told the two apart, the prices would sort themselves and there would be nothing to write about. It does not. The most reliable mispricing in a buildout is that the deepening scarcities and the decaying ones carry the same premium, and frequently the wrong way round.
In mid-2026 the double-monopoly fab, the one genuine picks-and-shovels position that captures value no matter which chip designer or which model lab wins, trades at a lower forward multiple than the cyclical designers it manufactures for. The market pays up for the customer and discounts the supplier the customer cannot live without. Meanwhile the memory makers, riding the most violent margin cycle in their history, command valuations that price the peak as a plateau. The premium is inverted relative to the half-lives. The thing that decays is priced to compound; the thing that compounds is priced to decay.
There is a corollary that the crowd resists more than any single name, and it points at the layer everyone is watching. The frontier model itself is the least defended position in the entire stack.
Capability converges. The gap between the best model and the third-best closes every cycle, open weights drag the floor up underneath all of them, and the price of a token falls toward the cost of the electricity to emit it. A business whose product can be distilled by a competitor or approximated by a free download does not own a chokepoint. It owns a brand and a head start, both depreciating. The labs are not bad businesses. They are bimodal ones, lottery tickets with real expected value and almost no defensibility, and the market is pricing several of them at forty times a revenue line that loses more than a dollar for every dollar it books.
Value migrates down, away from the model, toward the things a model cannot be without and cannot route around: the fab, the package, the test, the power. Those are the positions that survive the convergence, because you cannot open-source a foundry or torrent a gigawatt.
None of this is about a particular ticker, and the specific prices will be stale before the ink dries. The frame is the durable part, and it reduces to a single question you can carry into any buildout, the next one as much as this one.
Ask of every scarce thing whether the scarcity deepens or decays with the thing it is scarce in. If demand growing makes the bottleneck worse — tighter, rarer, more physically constrained — you are looking at a position the future pays you to hold. If demand growing invites new supply that competes the margin away, you are looking at a rental, fairly priced only if you treat it as one. Own the deepening scarcities through the cycle. Rent the decaying ones, and only at a price that assumes the peak ends. And refuse, on principle, to pay a deepening-scarcity multiple for a decaying-scarcity margin, which is the trade the crowd is most eager to sell you at the top.
The buildout will keep stamping the whole chain with the same word. The work is reading the half-life it forgot to print.