The AI Chip Supply Chain

Before anything can be "made," you need the raw stuff: ultra-pure silicon (from quartz sand), plus a long list of niche metals and gases such as gallium, germanium, rare earths, neon, and cobalt. Most of these are mined in one set of countries but refined (purified into usable form) in another. It works like crude oil: lots of places have the rock, but only a few have the refineries. Why it matters: refining is the real chokepoint, and China owns it. "For 19 of 20 important strategic minerals, China is the leading refiner, with an average market share of 70%." Beijing has turned this into a weapon, banning exports of several of these materials to the US.

Raw silicon isn't ready to use. This step makes the precise inputs a chip factory consumes: polished silicon wafers (the round dinner-plate-sized discs chips are printed on), photoresist (a light-sensitive coating that acts like photographic film, onto which the chip pattern is "exposed"), mask blanks (the master stencils), and ultra-pure gases and chemicals. Why it matters: this is the most chokepoint-dense layer of the whole chain, and Japan owns it, with over 50% share of 14 of the most critical chipmaking materials. In 2019 Japan briefly cut these supplies to South Korea, proving materials can be weaponised faster than machines.

Lithography is how the circuit pattern gets onto the chip: a machine shines light through a stencil to "print" features thousands of times thinner than a human hair. The newest method, EUV (extreme ultraviolet), uses light so short it must be generated by blasting molten tin droplets with a laser 50,000 times a second. Alongside lithography sit other tool-makers for etching, coating, and inspecting wafers. Why it matters: this is the most concentrated chokepoint in the entire chain. Exactly one company on Earth, ASML, makes EUV machines. And ASML itself depends on a chain of sole-source suppliers, so even ASML can't be copied without copying all of them.

A fab (fabrication plant) is the factory where all the previous steps come together: wafers, machines, and chemicals print and build billions of transistors onto silicon. The "leading edge" means the newest and smallest generations, today's 3nm and 2nm nodes. ("nm" once meant transistor size; now it's mostly a marketing label for a generation.) A single leading-edge fab costs $15–48 billion and takes 3–5 years to build. Why it matters: ~99% of the chips used to train frontier AI models are made by one company, TSMC, in one earthquake-prone, conflict-exposed corner of Taiwan. There is no near-term substitute. A blockade of Taiwan is the single biggest risk to the whole AI economy.

Before a chip is built, it's designed. EDA (Electronic Design Automation) is the specialised software used to lay out billions of transistors, and you cannot design a modern chip without it. IP means reusable design blueprints (like Arm's processor designs) that companies license instead of building from scratch. And at the top sit the chip designers themselves, like Nvidia. Why it matters: this is where US control is broadest. Three US-controlled firms own the design software the entire world uses, which doubles as a legal hook (the "FDPR") letting Washington claim jurisdiction over any chip designed with their tools, anywhere.

An AI chip is useless without fast memory feeding it data. HBM (High Bandwidth Memory) is memory stacked vertically right next to the processor, like a skyscraper instead of a sprawling suburb, for huge speed. Advanced packaging (TSMC's "CoWoS") is the precision assembly that fuses the processor and its HBM stacks onto one module. Beneath it all is a special film, ABF, that insulates the chip's wiring. Why it matters: this, not the chip itself, is often the real bottleneck on how many AI GPUs can ship. HBM is Korea-dominated, packaging is sold out at TSMC through 2026, and the ABF film is made by one company in Japan, arguably the deepest single point of failure in the whole chain.

Loose chips have to become a working machine. This step assembles them into rack-scale systems (a single Nvidia rack can cost ~$3.9M), wires thousands together with networking and fiber-optic transceivers so they act as one giant computer, and, increasingly the hardest part, supplies the electricity to run them. A big AI data centre now draws as much power as a small city. Why it matters: assembly is ~90% Taiwanese; the fiber-optics are mostly Chinese (an awkward US dependency); and the newest bottleneck is power itself. The US grid, transformers (2–3 year waits) and turbines now limit AI build-out from the inside.

The final step: actually training the AI. This needs software (CUDA and frameworks like PyTorch that turn math into instructions the chips run), enormous cloud computing capacity from the giant hyperscalers, and the frontier labs that run the training. Unlike everything below it, this layer is the most open and substitutable: software can be copied or routed around, and model weights spread freely. Why it matters: the real lever here is compute (who has the chips and clusters), not software. ~71% of the world's AI compute sits in five US hyperscalers. China has nearly closed the model-quality gap, proof that software, not silicon, is the part the US can't truly control.