Why Semiconductor Manufacturing Is So Hard to Replicate

Semiconductors are the most economically important manufactured goods whose production is concentrated in the fewest places. Leading-edge logic chips are made at a small number of facilities, and building a new one is not simply a matter of committing capital.

The reasons are worth understanding, because they explain why industrial policy in this area produces results on a timescale of years rather than quarters.

The capital is the easy part

A leading-edge fabrication plant costs enormous sums and takes years to build. That is a real barrier, but capital is the most fungible input. Governments and large firms can raise money.

The harder inputs are equipment, and the knowledge of how to use it.

Lithography

Chipmaking prints circuit patterns onto silicon using light. As features shrank below the wavelength of available light, the industry resorted to increasingly elaborate tricks — immersing the optics in liquid, exposing the same wafer multiple times with offset patterns.

Extreme ultraviolet lithography moved to a much shorter wavelength, which required generating light in a way that has no everyday analogue and mirrors polished to a smoothness that has few precedents in manufacturing. These machines are made by a very small number of suppliers, contain components from a long chain of specialist firms, and are produced in limited quantities.

A country can fund a fab. It cannot conjure a supplier base that took decades to assemble.

Yield is the real product

The metric that determines whether a fab is viable is yield: the proportion of chips on a wafer that work.

A process might begin at low yield and improve over months as engineers identify sources of defects — a contaminant, a slight misalignment, a temperature variance. Since the cost of processing a wafer is roughly fixed regardless of how many good chips come off it, yield largely determines cost per chip.

Why this resists transfer

Much of what raises yield is tacit. It lives in the experience of engineers who have watched a particular process misbehave in particular ways, in accumulated adjustments that were never fully documented because they were discovered empirically.

You can transfer a process recipe. Transferring the judgment that lets a team diagnose an unfamiliar defect pattern quickly is considerably harder, and it is the difference between a fab that runs and one that runs profitably.

The economics push toward concentration

Several forces compound.

Each reinforces the others. The result is an industry that concentrates even without anyone intending it.

Leading edge and everything else

Coverage focuses on the most advanced processes, but most chips in use are not leading edge. Vehicles, appliances, industrial equipment and medical devices largely use mature processes that are decades old.

These are made in many more places and are far easier to add capacity for. The shortages that disrupted vehicle production in recent years were substantially in mature nodes, not the leading edge — which is worth remembering, because policy framed entirely around the most advanced chips does not address the parts most industries actually consume.

Design and manufacturing came apart

For much of the industry's history, companies designed and manufactured their own chips. That model has largely given way to a split: firms that design and sell chips without owning factories, and foundries that manufacture to others' designs.

The separation happened because fabrication costs grew faster than most designers could justify. A company selling a moderate volume of specialised chips cannot fund a facility used at a fraction of capacity, but it can buy capacity from a foundry serving many customers.

This produced enormous efficiency and a specific fragility. A large share of the industry's designers depend on a small number of manufacturers, so a disruption at one facility propagates to products that appear entirely unrelated.

Packaging: the constraint that moved

As shrinking transistors has become harder and more expensive, more of the performance gain has come from how chips are assembled rather than how small their features are.

Advanced packaging stacks or places multiple pieces of silicon in a single component, connecting them with far higher bandwidth than a circuit board allows. It lets a designer combine parts made on different processes — expensive leading-edge logic beside cheaper memory — instead of pushing everything to the most advanced node.

Packaging capacity has consequently become a bottleneck in its own right, and one that receives a fraction of the attention paid to node sizes. Capacity here is also concentrated, for the same accumulated-expertise reasons, which means the shift did not diversify the supply chain so much as move where its narrowest point sits.

What follows for policy

Subsidising construction addresses the most tractable constraint. It does not address equipment lead times, the supplier ecosystem, or the years required to build yield and train engineers.

This is not an argument against the attempt. It is an argument for realistic timelines. A fab announced today produces meaningful volume years from now, and produces it at competitive cost some time after that.

The concentration was not created by a single decision and will not be undone by one.