How Semiconductor Capex Turns Into Supply — and Why the Lag Is the Cycle

Here is the framing we use: capex is dated money. Every capacity commitment in this industry carries two dates. The decision date is set by present conditions — current margins, current backlog, current board confidence. The delivery date is set by physics and logistics — how long it takes to pour a cleanroom, receive and install the tools, qualify the process, and climb the yield curve. The semiconductor cycle lives in the gap between those two dates, because the demand information embedded in the decision is stale by the time the supply arrives. Read a capex announcement without dating the money and you will systematically mistake tomorrow's supply for evidence of tomorrow's demand.

From board approval to good wafers

The pipeline between a capex decision and sellable output runs through distinct stages, each with its own clock. First comes site work and the shell — the building and its extraordinarily demanding cleanroom environment. Then the tools: lithography, deposition, etch, metrology equipment ordered from a small set of specialised suppliers whose own lead times stretch precisely when every chipmaker is ordering at once. Then installation and qualification, in which the process is tuned until it produces working devices at acceptable yield. Then the ramp itself, because a fab does not switch on — yield and throughput climb over quarters. The industry's own shorthand for the finish line is telling: not capacity, but good wafers out. Capacity that exists on a press release and capacity that ships revenue are separated by years.

Two properties of this pipeline matter more than its length. The first is lumpiness: a modern fab is a step function, not a dial. Capacity cannot be added in small increments matched to demand; it arrives in large discrete blocks, which means even perfectly forecast demand would produce alternating stretches of tightness and slack. The second is irreversibility past a certain point. A half-built shell gets finished, because the remaining cost is small relative to the sunk cost. Delivered tools get installed. Cancelling late-stage spending saves little and forfeits a hard-won position in the equipment queue. Supply in transit therefore tends to arrive whether or not the demand that justified it still exists — the pipeline can be slowed, but it cannot usefully be recalled.

Three layers of why the lag is procyclical

A lag alone would make the industry merely slow to adjust. What makes the lag the cycle is that the timing of capex decisions is itself systematically procyclical — the industry commits the most supply exactly when it should commit the least. That is not a behavioural accident; it follows from three stacked mechanisms.

Sterling's Embedded Intelligence applies exactly this inversion when it reads capacity news: a capex disclosure is filed as a supply-side document with a delivery date attached, not as demand corroboration. Sterling's discipline is unglamorous — date the money, then ask what the demand assumption for the delivery year has to be for the spending to earn its cost of capital. That second question is where most cycle errors become visible, because at the peak the implied assumption is usually that the best conditions in memory persist for years.

The illustration: a shortage that scheduled a glut

The strongest case against this framework

The serious objection is that the cobweb description belongs to an earlier, more fragmented industry. Decades of consolidation have left a handful of producers controlling most capacity in each major segment, and concentrated industries can practise supply discipline: fewer actors, better information, a shared memory of what overbuilding costs. On this view, the lag still exists but no longer bites, because the decisions at the top of the cycle have become more restrained.

We accept half of this and reject the conclusion. Consolidation changes the amplitude of the cycle, not its structure — the lag between commitment and delivery is set by construction and tool lead times, which no market structure shortens. And discipline is a regime, not a law: it holds while demand is ordinary and defects when a platform shift makes share-grabbing rational again, which is exactly when the largest commitments get made. A second development cuts the other way entirely: state-subsidised capacity. Governments now fund fabs for resilience rather than returns, which means a growing slice of supply is committed acyclically — indifferent to price signals on the way in, and therefore indifferent to them on the way out. Subsidised capacity weakens the one force that historically ended gluts, namely the refusal of private capital to fund the next round. The honest synthesis is that concentration argues for gentler cycles at the leading edge, subsidy argues for harsher trough economics wherever subsidised capacity lands, and neither repeals the lag.

Reading capex as a supply schedule

The practical output of the framework is a change in what you do with a capex headline. First, benchmark it against the industry's depreciation charge rather than against last year's spending — capex near depreciation is maintenance of the existing supply curve, capex far above it is expansion, and the ratio is the cleanest single expression of how much future supply is in the pipe. Second, decompose it: spending on shells is optionality with a long fuse, while spending on tools is near-term supply, because tools are what turn a building into wafers. A period in which announced projects are heavy on shells and light on tool orders is committing far less supply than the headline number suggests — and the reverse is more dangerous than it looks.

For a portfolio owner the consequence is a timing asymmetry that runs against instinct. The moment of maximum reported strength — record capex, stretched lead times, customers prepaying for allocation — is the moment the largest future supply is being scheduled, and the moment of maximum reported distress is when future supply is being cancelled. Neither observation is a trading signal by itself; demand can grow into capacity, and platform shifts sometimes justify the spending. But the burden of proof moves. When capex is running far above depreciation, the bull case must explain why delivery-year demand will absorb a supply curve that is already committed; when capex has collapsed, the bear case must explain how weak pricing survives a pipeline that has been emptied. Dating the money tells you which side owes the explanation.