Chip Demand Signals Need Inventory by Node
Semiconductor demand headlines average across the industry. Actual conditions differ violently by node, from trailing-edge logic to leading edge and memory.
Semiconductor demand headlines average across the industry. Actual conditions differ violently by node, from trailing-edge logic to leading edge and memory.
The short answer: read demand as inventory weeks by node and end market, because one industry number describes nothing.
Evidence note: Industry bodies such as the SIA aggregate wafer and revenue data, while public datasets from the ITU and BIS frame the digital and financial context. The aggregation is exactly the problem: shortages and gluts coexist in the same quarter.
| Segment | Demand signal that matters | Common headline error |
|---|---|---|
| Trailing-edge logic | Capacity bookings, auto orders | Treated as commodity cycle |
| Leading-edge logic | Design wins, tape-out counts | Read as a demand boom |
| Memory | Bit growth vs bit demand | Priced as a single market |
| Analog and power | Distribution inventory weeks | Ignored until shortage returns |
Related reading: how supply-chain clocks reprice delivered markets.
One industry, many markets
The semiconductor industry reports a single revenue number, and it is almost meaningless for planning. Trailing-edge capacity serving automotive and industrial demand, leading-edge capacity serving compute, and memory serving everything behave on different cycles with different capital intensity.
**A quarter can simultaneously carry a memory glut, a trailing-edge shortage and rational leading-edge pricing.** The average of those states describes none of them.
Any demand read must start by naming the node and the end market. Everything else is commentary.
Inventory weeks are the honest gauge
Revenue says what was sold. Inventory weeks say what is sitting unsold in distribution and at customers, and that is the number that predicts the next quarter's orders.
Track weeks of inventory by product family at distributors and at major customers. Rising weeks with rising revenue is a pull-forward warning. Falling weeks with flat revenue is the early shortage signal.
Compare against the segment's healthy band. Five weeks may be lean for one family and bloated for another; the absolute number has no meaning without its history.
Order patterns lie during transitions
When buyers fear shortage, they double-order. When they fear glut, they cancel and let inventory run down. Both behaviours distort visible demand in the same direction as panic, not as end consumption.
Correct for this with consumption proxies: billings at equipment makers, assembly and test utilisation, and end-market sell-through where available. Orders are intentions; consumption is fact.
The classic error is forecasting from bookings during a double-ordering phase. The correction always arrives as a sudden, violent cancellation cycle.
Capacity decisions outlive the cycle
Leading-edge fabs take years and billions to build, so capacity decisions made in a shortage arrive into whatever market exists then. This structural lag is why the industry oscillates.
Trailing-edge capacity is cheaper but not instant, and mature-node expansion carries its own qualification timelines with automotive and industrial customers.
Analysts should map announced capacity against the node-specific demand view, not the aggregate. Aggregate capacity headlines have justified some of the worst capital allocation in the industry's history.
A working read for each quarter
For each node and end market, record: revenue, weeks of inventory, order-to-consumption gap, and capacity additions inside 24 months. Four numbers, one table, no adjectives.
Flag divergences explicitly. If memory inventory falls while logic inventory rises, the headline will say something flat and wrong.
Date everything and keep the history. Semiconductor cycles reward the analysts with long memory and punish everyone else.
Teams that need a consistent cross-market view, rather than one clip of data at a time, often pair this kind of desk check with independent market intelligence so every conclusion carries its source and date. The point is not another report. It is a method that survives the next quarter.
What the data cannot tell you
Distribution inventory covers only the part of the market that flows through distribution. Large direct customers hold their own stock, and their build-to-order positions are visible only in their own disclosures, partially and quarterly.
Utilisation and capacity data also arrive with a lag and a smoothing. Fab operators report on their own cadence, and the most informative utilisation numbers are often estimates long before they are confirmations.
And no dataset sees design pipelines in real time. The tape-outs and design wins that decide leading-edge demand two years out are private until a company chooses to mention them, which is precisely when the market has already moved.
Who this analysis does not help
It will not help anyone trading the equities on a daily horizon. These signals move over quarters and the market prices information faster than the information is published.
It is also not a device engineering guide. Component selection turns on specs, lifecycles and supply agreements at the part level, which no node-level view resolves.
A quarterly desk routine that works
Fix the matrix: four segments, and for each, four numbers this quarter: revenue, weeks of inventory, order-to-consumption gap, and capacity arriving within 24 months. The table fits one page and explains most headlines.
Refresh consumption proxies quarterly: equipment billings, assembly and test utilisation, and end-market sell-through where disclosed. Orders belong in a separate column labelled intentions.
Review the order book for concentration. A node whose demand is five customers deep behaves differently from one spread across thousands, and the correction cycles differ accordingly.
Write one paragraph per segment naming the state: shortage, balance or glut, and what would change the call. Naming the falsifier is what keeps a desk honest through a cycle.
Rule of thumb: never forecast from bookings alone. Double-ordering and cancellation waves are the industry's normal behaviour, and consumption proxies are the only correction.
Frequently asked questions
Why does aggregate chip revenue mislead?
Because nodes and end markets run on different cycles. A memory glut and a trailing-edge shortage can coexist inside the same headline number.
What is the best early warning indicator?
Weeks of inventory by product family. It moves before revenue and orders reveal the change.
How do double orders distort forecasts?
They inflate visible demand during shortage fear, then return as cancellations. Consumption proxies correct for it.
Is trailing-edge capacity less strategic?
No. Automotive and industrial demand depends on it, and its qualification timelines make expansions slow and sticky.
Where can weeks-of-inventory data be found?
Distributor disclosures, customer balance-sheet commentary, and industry channel surveys. No single source covers it, which is why the estimate should be shown with its method.
Is leading-edge capacity a bubble?
It is a bet on specific compute demand with a multi-year lag. Node-level analysis, not industry sentiment, is the only fair way to judge it.
How early do inventory signals appear before revenue turns?
Typically one to two quarters at distribution, earlier for large direct customers if their disclosures are read carefully. Early, but not instantly, and never uniformly across nodes.
Should memory be analysed separately from logic?
Always. Memory is a commodity with its own supply discipline, and averaging it into logic reads has misled more semiconductor commentary than any other habit.
Sources and method
This article uses the following public sources. Figures retain the source definition and date. It is market analysis, not investment, legal or medical advice.