Why Semiconductor Inventories Lead the Cycle

Here is the framing we use, and it is an accounting identity before it is a theory: chipmaker shipments equal end consumption plus the change in channel inventory. In stable stretches the second term hovers near zero and everyone forgets it exists — shipments track consumption and demand models work. At turning points the second term takes over: the channel is either stockpiling, so shipments run above what the world is actually consuming, or destocking, so shipments run below it. The violence of the semiconductor cycle — the part that moves share prices — lives almost entirely in that second term. End demand for computing has tended to grind; the channel swings. The cycle you trade is an inventory cycle layered on a much gentler demand cycle, and the inventory cycle turns first.

The channel nobody models

Between a wafer and a user sits a longer chain than most demand models acknowledge: the chipmaker's own finished goods, distributors, contract manufacturers, the component shelves of device makers, finished-device stock, and — increasingly, for the largest buyers of compute — strategic stockpiles held by the end customers themselves. Every stage holds a buffer, and the size of each buffer is set not by end demand but by lead time: how long that stage believes it must wait for resupply, and how badly it would be hurt by running out. A chip that costs little can idle a production line worth vastly more than the part — so when availability is in doubt, the rational buffer is generous. The channel, summed across stages, can hold a large multiple of any single month's true consumption, which is exactly what gives the second term its capacity to dominate.

Three layers of why the channel amplifies

This is why Sterling's Embedded Intelligence reads chip-sector results with the identity open on the desk: for any reported quarter, the question is not whether revenue grew but how much of the growth is the second term. Sterling's working habit is to line up producer shipment growth against the best available read of end sell-through; a wide positive spread is the channel filling, whatever the press release says about demand, and a wide negative spread is the channel draining, whatever the headlines say about collapse. The spread mean-reverts by arithmetic — inventory cannot accumulate or drain forever — and knowing which side of that arithmetic a company sits on is worth more than another demand forecast.

The illustration: phantom demand and the digestion after

The strongest case against this framework

The serious objection comes in two forms. The first is structural: when a genuine platform shift is underway — a new class of compute demand growing off a small base — inventory dynamics are second-order noise against a demand trend that is doing the real work, and a framework obsessed with channel mechanics will keep calling tops in a secular ramp. The second is practical: inventory data is lagging and dirty. It is reported quarterly, restated, distorted by write-downs and by strategic stockpiling that never intends to be consumed, so the variable the framework depends on cannot be observed cleanly in real time.

Both objections deserve a real answer. On the first: structural growth raises the trend; the channel governs the deviation around the trend — and turning-point losses are taken on the deviation. A secular ramp does not repeal the identity; it means consumption is growing fast, which makes it easier for the channel to hide accumulation inside impressive shipment numbers, and the eventual digestion is a pause in a rising line rather than a collapse. The framework does not forbid owning a platform shift; it tells you which part of the reported growth will need to be given back. On the second objection we largely concede the data problem and route around it: the working signals are directional, not level-based. Lead times are observable near real time and turn before inventory is reported; the shipment-versus-sell-through spread can be estimated from public reporting; and inventory days are most informative at their extremes, where measurement noise cannot explain the reading. The framework asks less of the data than a point forecast would — direction and extremity, not precision.

Reading a turn while it happens

The sequence of a downturn is repeatable enough to write down. Lead times stop stretching and roll over. Order momentum fades while shipments are still strong — the backlog is being consumed, not replenished. Inventory days peak at the customers before the producers. Producer guidance breaks, usually framed as sudden, though the channel data had been pointing there for quarters. Shipments then run below consumption for as long as digestion takes, and the trough in reported revenue arrives while end demand is already stabilising. The upturn mirrors it: orders resume before end demand accelerates, because restocking is itself demand — the second term flips positive and briefly flatters every producer's growth rate.

For a portfolio owner, the most useful implication is about the worst-looking moment. When shipments are running below consumption, the situation is arithmetically self-correcting: the channel cannot drain below zero, so undershipment must end, and reported growth at that point snaps back toward true consumption without any improvement in end demand at all. Symmetrically, when shipments have run above consumption for an extended stretch, reported strength is borrowing from future quarters, and the giveback is equally arithmetic. Neither observation times the week of the turn — but both tell you which direction the surprise is loaded, and at cycle extremes that is the tradable information.