Equipment vs Design: Where Semiconductor Economics Concentrate
Chip-industry economics pool at the two knowledge ends — the toolmakers and the designers — while the capital in the middle absorbs the cycle. Owning the sector means choosing an end.
The instinctive assumption in a chip boom is that the value accrues to whoever makes the chips. It is the one assumption the industry's own history most consistently refutes. Through full cycles, the manufacturers in the middle of the chain have tended to earn the most volatile returns on the most capital, while the durable economics pooled at the two ends — the companies that build the tools and the companies that design the products. An investor who buys the sector without deciding which of those positions they hold has not chosen an exposure; they have chosen an average of three different businesses.
Here is the framing we use: the semiconductor value chain is a barbell. At one end sit the equipment makers — lithography, deposition, etch, metrology — selling the means of production into an oligopoly structure where each process step supports only one or two viable suppliers. At the other end sit the designers, along with the design-software and intellectual-property layers beneath them, selling differentiated products and tolls with almost no capital intensity. In the middle sits manufacturing: fabs and commodity memory, where the capital is heaviest, the product is least differentiated, and the cycle's adjustment is taken in price. The bar connecting the ends is thin for a reason — the middle is where knowledge is most replicable relative to the capital required to exploit it. Wherever that ratio inverts, economics concentrate.
The equipment end: oligopoly by physics
Equipment economics begin with a structural fact: each process step tends toward a one- or two-supplier market. The research cost of keeping a tool competitive scales with the difficulty of each successive node, while the addressable market for any single step is too small to amortise that research across many competitors. The industry's history is a record of this arithmetic working itself out — whole categories consolidating cycle by cycle, until at the frontier of lithography the field narrowed to effectively a single supplier. That endpoint is the extreme case of a general rule: tool markets concentrate because the physics gets harder faster than the market gets bigger.
Concentration alone would invite entry; what locks it in is the switching cost on the customer's side. A fab's process recipe is qualified around specific tools — thousands of interacting parameters tuned until yield is acceptable. Swapping a tool supplier mid-node puts yield at risk, and yield is the most expensive variable in the industry: a point of yield is worth more than almost any procurement saving. So incumbency compounds. The tool that is designed into this node is the default for the next one, and the supplier learns from every installed machine, widening the gap the challenger must close. Layered on top is the annuity: an installed base that needs service, spares, and upgrades for decades, producing a revenue stream that persists through the capex cycle that whipsaws new-system sales.
There is a secular kicker that is easy to miss because it is slow. As nodes shrink, each wafer requires more process steps and more expensive tools per step — equipment intensity per unit of output rises over time. The toolmakers therefore capture a growing share of industry capex regardless of how fast end demand for chips grows. It is one of the few places in the sector where the difficulty of the technology is a tailwind for the supplier rather than a cost.
The design end: product margins on someone else's balance sheet
The design end earns its position differently. The fabless transition — the industry's great restructuring, in which design houses gave up their fabs as manufacturing cost escalated beyond what any single product line could justify — moved the capital intensity of chipmaking out of the design business entirely. What remained is a business that captures end-market differentiation on an operating-expense cost base: the design wins the socket, the foundry carries the depreciation. When a designer's architecture is meaningfully better for a workload, the pricing power that creates flows to margins that are not diluted by a fab's fixed-cost drag, and the downside of a bad cycle is an inventory problem rather than an idle-factory problem — painful, but an order of magnitude smaller and faster to clear.
Beneath the designers sit the purest economics in the chain: design software and licensed intellectual property. These are toll businesses — every chip designed, by anyone, for any end market, passes across their infrastructure, and the toll is paid whether the resulting product succeeds or fails. They are levered to the number of design starts and the complexity of each design, not to unit volumes or pricing, which makes them the lowest-beta way the sector's growth can be held. The general principle across the whole design end: it monetises knowledge with minimal capital, so the cycle reaches it only through demand, never through its own balance sheet.
Three layers of why the middle is thin
That classification is the working output of the framework, and it is the reason Sterling's Embedded Intelligence refuses to treat a chip-sector move as one signal. When the sector rallies, Sterling's first question is which derivative is being repriced: a capex story lifts the equipment end and may say nothing about end demand; a demand story lifts design and may coexist with capex cuts; a pricing story in memory can run against both. The sector label obscures more than it reveals — the barbell position is the exposure.
The strongest case against this framework
The serious objection is that the barbell has grown a third pole. Manufacturing at the leading edge is no longer a commodity: so few producers can execute frontier nodes that the survivors command pricing power, customer prepayments, and returns that rival the ends of the chain. On this view the middle is not structurally thin — it merely used to be crowded, and consolidation has fixed that.
We accept the observation and would refine the framework rather than defend its letter. The barbell was never really about position in the chain; it is about the replicability of knowledge relative to the capital required to use it. Frontier process knowledge has become as hard to replicate as tool physics or architecture — yield learning at the leading edge compounds with cumulative volume in a way a new entrant cannot buy — and so economics have concentrated there too, exactly as the rule predicts. The refined statement: returns pool wherever an irreplaceable knowledge position exists, whether that is a tool, a design, or a frontier process, and they thin wherever manufacturing is replicable, which still describes commodity memory and the trailing edge. The middle did not get thick; a piece of it earned its way into an end. What that concentration does to supply risk is its own subject, treated in the companion piece on foundry concentration.
How to apply this framework
- Classify before you size. For each semiconductor holding, name its derivative: capex (equipment), demand (design), or pricing (commodity manufacturing). A portfolio that holds all three is not diversified within a theme — it holds three different cycles, and should be stress-tested against each separately.
- In equipment, separate the system cycle from the service annuity. New-system revenue rides the capex cycle; the installed base pays through it. The mix between the two tells you how much of an equipment maker's earnings is cyclical exposure and how much is toll collection.
- Track equipment intensity as the slow tailwind. Rising tool cost per wafer means the equipment end captures a growing share of industry capex across cycles — a secular claim that survives any single downturn and compounds quietly underneath the order-book noise.
- Test every moat with the replicability question. Ask what it would cost — in capital, in years, in accumulated yield learning — for a well-funded entrant to replicate the position. Where the honest answer is that capital alone cannot do it, you are at an end of the barbell. Where capital suffices, you are in the middle, whatever the label says.