The Other Half of the Chip War

Every headline about critical minerals reads the same way: China restricts something, the West scrambles to catch up. Read enough of them and you start to think of rare earths as a supply problem, a shortage to be solved with more mines. That framing misses what is actually happening. This is not a supply story. It is the semiconductor war, seen from the other end, and the two ends are the same war.

The United States controls the top of the compute stack. It gates China's access to the most advanced chips and, more importantly, to the tools that make them. China's answer was to reach for the bottom of the stack, the raw inputs that everything above depends on. And the crucial thing to understand, the thing that turns a metaphor into a mechanism, is that China's controls do not stop at the magnets in motors and drives that get most of the attention. They reach directly into the fab, and they land hardest on the high-power layer.

Start with gallium, because that is where it bites

Modern power electronics increasingly run on gallium nitride. GaN is a wide-bandgap semiconductor that moves high voltage and high frequency far more efficiently than plain silicon can. It shows up in electric-vehicle inverters and onboard chargers, in the radio-frequency front ends of 5G and radar, in fast chargers, and increasingly in the power delivery feeding AI data centers. Adoption has not been slow, either. GaN went from a niche in around 5 percent of new EV platforms in 2020 to something like 30 to 40 percent by 2025 and 2026, across Tesla, Volkswagen, BMW, and General Motors.

The GPU gets the headlines. The power stage that feeds it does not, and the raw material for that power stage is roughly 98 percent Chinese. Gallium is that concentrated. Germanium, which carries the high-frequency and optical layer through silicon-germanium and fiber optics, runs around 90 percent. Antimony, a semiconductor dopant that is also a defense staple, sits near half. These are not the silicon that forms the transistor channel in a mainstream CPU or GPU. That distinction matters and I will come back to it. But they are the materials that make the chip usable at power and frequency, and in the AI buildout, power is the binding constraint.

Why you cannot simply build your way out

Here is the part that makes the gallium chokepoint so much harder to break than the headlines suggest. You cannot open a gallium mine. Gallium is a trace byproduct of refining aluminum from bauxite. Germanium is a byproduct of zinc refining and coal fly ash. Neither exists as a deposit you go dig up. They are captured during the processing of other metals, and that processing is itself concentrated in China. So the chokepoint was never a mineral in the ground. It is the refining stream, and China owns the stream.

That is the same lesson the magnet side teaches, which is why the two halves belong in one story. Rare earths are not rare. You can find them in a dozen countries. What is scarce is the ability to separate them and turn them into high-performance magnets, and there China holds around 90 percent. The deepest lock is not even the separation plants. It is the know-how. That is why, back in 2023, China banned the export of the rare-earth processing technology itself, not just the finished material. It is guarding the recipe, not the dish.

Mine, refine, separate, sinter. The West keeps counting the first step. China controls the middle three.

The flattered numbers

Two numbers get quoted as reasons to relax, and both are traps.

The first is the lifted ban. Late in 2025, as part of a broader de-escalation, China suspended its ban on gallium, germanium, and antimony to the United States. Headlines called it relief. But the licensing requirement stayed in place, the suspension is set to expire in late 2026, and the discretionary gate never left Beijing's hand. A ban that is lifted while the license stays intact is not access. It is a lease on access, revocable on notice, and the notice period is measured in months. Watching the ban is the wrong move. Watch the license and the calendar.

The second is mining capacity. The Western response has been real and well-funded on the magnet side. The Pentagon took an equity stake of roughly 15 percent in the one American company that can mine, separate, and manufacture magnets at scale, set a price floor near $110 a kilogram to blunt Chinese price competition, and backed about a billion dollars of magnet-facility financing. That de-risks capital beautifully. But the binding constraint was never capital. It is earned metallurgical learning and byproduct-refining infrastructure, and both answer to time, not money. The flagship magnet campus does not ship until 2028, and the hardest steps, heavy separation and magnet yield, are exactly the ones a price floor buys slowest. Announced is not the same as delivered.

The honest limit, which makes the point sharper

It would be easy to overstate this, so let me be precise, because the precise version is the stronger one. None of these controls stop a fab from etching a logic die. Leading-edge CPUs and GPUs are built on silicon, and China does not gate silicon. What the controls hit is the periphery that makes the logic usable: the compound-semiconductor power devices, the RF front end, the optical interconnects, and the polishing and coating chemistries inside the fab. The exposure runs through the packaging and RF suppliers and through the materials vendors that feed every fab, not through one marquee chipmaker's die.

That is not a weaker claim. It is a more accurate one, and it points somewhere uncomfortable. The AI buildout is not constrained at the logic die right now. It is constrained at power. Every analysis of the buildout that follows the money ends up at the same rung: transformers, substations, generation, and the power electronics that condition and deliver it. Gallium and GaN sit precisely on that rung. So the mineral chokepoint lands not on the part of the stack with slack, but on the part that is already the tightest.

The tell to watch, and how it lands

The controls also land unevenly, by politics rather than by market. When China eased restrictions, European magnet imports rebounded sharply while United States imports kept falling and never recovered to their earlier levels. For some specific materials the gap is stark: US imports of certain heavy elements collapsed to a small fraction of their pre-restriction volume even as shipments to allied countries resumed. Licensing is a scalpel, not a hammer. It lets the holder decide who gets a stable supply and who gets an unreliable one, and that decision is a lever of statecraft, not commerce.

So the discipline for reading this is simple, and it is the opposite of the headline instinct. Do not count mines, and do not trust a lifted ban. Watch the refining capacity, the separation lines, the magnet-yield milestones, and the licensing gate that is only suspended. Those are the numbers that tell you whether the chokepoint is loosening or just breathing.

Step back and the whole thing resolves into one shape. The chip war and the rare-earth war are not two conflicts. They are one chokepoint logic, aimed from opposite ends of the same stack. The United States holds the top, the tools and the advanced logic. China holds the bottom, the refined inputs and the separation know-how. Each side is discovering that the other's chokepoint is made of something money cannot quickly replace: on one end, decades of tool and process leadership; on the other, decades of refining and metallurgical learning.

You do not remove a bottleneck. You relocate it. This one did not move to a mine. It moved to a refinery and a licensing office, and the country that runs both is the same one holding the other end of the war.