Every argument about chips, whether it concerns export controls, shortages, or the concentration of manufacturing in a few places, rests on a handful of physical and economic facts. The physics stopped cooperating two decades ago, the response was to make manufacturing dramatically harder rather than to stop, and the resulting supply chain has a small number of points through which nearly everything must pass. Those three facts explain most of the politics.
A switch with no moving parts
Silicon in its pure form conducts poorly. Introduce a small quantity of an element carrying a spare valence electron and the crystal gains mobile negative carriers, and introduce one short of an electron and it gains mobile positive carriers instead. The useful behaviour appears where the two meet, because the junction conducts readily in one direction and resists in the other.
A field effect transistor is a conductive channel between two doped regions, with a gate above it separated by a thin insulating layer. Voltage on the gate creates an electric field that attracts carriers into the channel or repels them out. The gate draws essentially no steady current because it is insulated from what it controls, commanding the channel through a field rather than a current, which is why a chip can hold tens of billions of these without the control circuitry consuming the entire power budget.
Digital logic is built from complementary pairs arranged so that in either stable state one device is off. A gate holding a value draws almost nothing, and power is consumed only when the state changes, charging and discharging the capacitance of the wires being driven. This is the single most important property of the technology, because power consumption scales with activity rather than with transistor count, and it is what made large scale integration possible at all.
When shrinking stopped being free
For roughly thirty years, making transistors smaller improved everything at once. Reduce dimensions and supply voltage together and devices switch faster, occupy less area, and use less power per operation, so power density stays roughly constant as density rises. Higher clock speeds arrived as a side effect rather than a goal.
It broke in the mid two thousands, and the reason was voltage. A transistor is not a perfect switch, and the sharpness with which it turns off is limited by carrier statistics rather than manufacturing quality. As supply voltage fell, threshold voltage had to fall with it, and leakage current in the supposedly off state rose exponentially.
Below a certain point the leakage cost more than the switching saved. Voltage scaling stalled, power density began climbing each generation, and the industry hit a thermal wall.
What followed the thermal wall
The consequences shaped everything after. Clock frequencies stopped rising and have not meaningfully risen since, so performance had to come from parallelism, which is why processors grew cores rather than gigahertz and why a generation of software had to be rewritten. Chips began to contain more circuitry than could be powered simultaneously, so designs increasingly dedicate area to specialised blocks that sit idle until needed, because idle silicon is cheaper than hot silicon.
It is worth being clear that the well known observation about transistor counts was never a law of physics. It was an economic observation that held while the physics cooperated, and when it stopped, the trend continued only because manufacturers were willing to spend far more to sustain it.
Wrapping the gate around the channel
As channels shortened, the gate lost authority over them. In a flat device the gate touches the channel on one side only, and once source and drain are close enough their own fields influence the channel directly, so the transistor leaks when it should be off and its threshold varies with geometry. That is fatal for a design containing billions of nominally identical devices.
The fix was geometric. Stand the channel up as a narrow fin and drape the gate over it and the gate controls three sides instead of one. The next step encircles the channel entirely with stacked horizontal sheets, recovering control again and allowing drive strength to be tuned by sheet width rather than in discrete fin steps.
This is also where process node names became marketing. The numbers no longer correspond to any physical dimension, and comparing them across manufacturers tells you very little without actual density, performance, and power figures.
Printing something smaller than the light
Patterns are put onto silicon photographically. The wafer is coated in a light sensitive polymer, an image is projected through a mask, and exposed regions are chemically removed to leave a stencil for etching or deposition, repeated dozens of times to build the layers of a device. Resolution is set by the wavelength and the optics.
Deep ultraviolet at 193 nanometres served for many generations, extended by immersing the final lens in water and then by splitting one layer across multiple exposures with separate masks. Multiple patterning works, and every additional mask adds cost, cycle time, and alignment error.
Extreme ultraviolet
Extreme ultraviolet at 13.5 nanometres was the way out, and it is difficult in a way that is hard to overstate. Light at that wavelength is absorbed by everything including air and glass, so the entire optical path is under vacuum and every element is a mirror rather than a lens.
No material reflects appreciably there, so the mirrors are built from dozens of alternating layers engineered to reflect by interference, and each still absorbs a meaningful fraction of the light. Generating the light is equally awkward, with molten tin droplets fired through a vacuum chamber and struck twice each by a high power laser, vaporising them into a plasma hot enough to radiate at the required wavelength.
The whole assembly is among the most complex machines ever built, it is produced by a single company, and no substitute exists. That is the geopolitical story compressed into one line, and it is why export controls target manufacturing equipment rather than finished chips. The equipment has no alternative source. Chips have many.
Yield and the economics of concentration
Defects occur at some rate per unit area of wafer, and a die containing a defect in a critical location is scrapped. The proportion of good dies is therefore not linear in area, it falls sharply as the die grows, so doubling the area of a die can more than double its effective cost.
This is why designs are increasingly split into smaller pieces fabricated separately and joined in the package, with pieces that gain nothing from the newest process built on a cheaper mature one. The cost is that connecting them requires advanced packaging with its own difficult tolerances, which has moved packaging from an afterthought to one of the areas where the most important engineering now happens.
Why only a few can play
A leading edge plant costs tens of billions and its equipment depreciates over a handful of years, so that capital must be recovered at very high utilisation before the process is obsolete. There is no viable small scale version of this business, and an operator who falls a generation behind loses the customers who need the newest process and cannot fund the next plant.
The result is severe concentration at the leading edge, and the concentration is not confined there. Wafers, photoresists, ultra pure gases, packaging substrates, design automation software, and licensed circuit designs each have their own short list of suppliers, and a shortage can originate at any of those layers.
The disruption to vehicle manufacturing earlier this decade illustrated it precisely, because it was not a shortage of advanced processors but of unglamorous parts made on mature processes that nobody had reason to expand.
Why this became politics
Two conditions together make a technology strategic. It must be required by nearly everything, and it must be producible by nearly no one. Semiconductors satisfy both, embedded in vehicles, medical devices, industrial control, communications, financial systems, and weapons, while the capability to make the most advanced of them exists in very few places and the equipment to build that capability in fewer still.
A universally required input with a highly concentrated supply is leverage, and leverage attracts policy. The mechanism does not depend on any particular set of national interests, and any technology with the same structure would attract the same attention.
Note: the gate insulator in a modern transistor is a few atomic layers thick, and the industry moved to alternative materials because silicon dioxide at that thickness leaks by quantum tunnelling. Manufacturing at this scale is negotiating with quantum mechanics, not merely engineering.