The Insane Complexity of the Semiconductor Global Supply Chain

ELI5 / TLDR

Every advanced computer chip is the end of an absurdly long assembly line, and at almost every single step there’s exactly one company on Earth that can do that step well. The toilet company makes the ceramic plate that holds the silicon. The MSG company makes the film the chip sits on. A Dutch company makes the only machine that can print the finest patterns, using a stack of mirrors so smooth that a bump scaled up to the size of Germany would be under a millimetre. Because each link is a near-monopoly, you’d think the whole thing is fragile — but every monopoly supplier is also stuck selling to a tiny handful of buyers who can’t go anywhere else either, so they’re all welded together and forced to behave.

The Full Story

The building across from the rice farms

The video opens at a Micron memory plant in Taiwan — 30,000 square metres, billions of dollars of machines running around the clock to make the memory that ends up in data-centre GPUs. And then the gut-punch: the chips that come out of this place aren’t even finished products yet. They’re already the result of over 6,000 suppliers from 40-plus countries.

The framing the whole video hangs on is a paradox:

This is a supply chain that’s simultaneously far more robust than you’ve probably been told, but also full of far more bottlenecks.

Both halves are true, and the second half of the video explains why that isn’t a contradiction.

The toilet plate that holds the universe together

Start at a small factory on Kyushu, southern Japan. A ceramics maker presses ultra-pure aluminium oxide (plus a secret blend of other stuff) into one of the most uniform ceramic plates anyone knows how to make.

Why does this matter? Think about what a chip factory actually does to a wafer — a polished disc of silicon. It throws that wafer back and forth inside sealed chambers, hundreds of times a minute, while spraying chemicals and blasting it with plasma. Something has to hold the wafer still through all of that. And holding it is its own nightmare:

You can’t clamp it — a mechanical clamp would warp something this delicate at atomic scales.
You can’t glue it — the chemicals needed to dissolve the glue later would wreck the chips.
And whatever holds it has to survive wild swings in heat, pressure and force without transmitting a single bit of stress into the silicon.

The solution is the ceramic plate, called a chuck. It grips the wafer with static electricity (the same force that makes a balloon stick to a wall), and because its surface is atomically flat, the force of being flung around spreads perfectly evenly. It even has channels machined into it for flowing liquid helium so the silicon stays cool while it’s being baked in a plasma chamber. These dinner-plate-sized chucks cost six figures each.

The world leader here is Toto — yes, the Japanese company that makes the world’s finest toilets. Their mastery of ultra-pure ceramics came from making toilets, and now data-centre customers are outbidding the hospitality industry for their best ceramicists.

Their priorities have visibly shifted in the last few years as data center customers have started outbidding the hospitality industry for the company’s best ceramicists.

MSG, but for chips

The second oddball: Ajinomoto, the Japanese company that gave the world MSG (the flavour enhancer in your instant ramen). Using the same organic-chemistry expertise, they make ABF — Ajinomoto Build-up Film — a thin insulating film that sits under almost every advanced chip on Earth.

Picture the green square your CPU or GPU is mounted on. That’s the substrate, and it’s essentially layers of ABF film with copper wiring threaded through it. As chips got more complex, the layer count crept from four or five up to well over a dozen on a modern AI accelerator. Other small firms (Ibiden, Unimicron) take the film, sandwich it with copper, drill thousands of microscopic holes through it, and turn it into that green square.

So two companies — a toilet maker and an MSG maker — each hold an effective monopoly on one branch of the chip supply chain. The video’s point is that this isn’t a fluke; it’s an inevitability that the economics keep producing.

The Dutch machine everyone fixates on

Inside the chip factory, the star is the lithography machine — the thing that “prints” the circuit pattern onto silicon. The famous name is ASML, a Dutch company spun out of Philips in 1984. A single ASML machine pulls from over 5,000 suppliers: lasers from Germany (Trumpf), optics from Zeiss polished to single-atom smoothness, vacuum pumps that work by knocking atoms out one at a time. The biggest machines are so large they ship in pieces aboard three Boeing 747s, with engineers travelling alongside to spend weeks reassembling them.

Here the video corrects a common myth. ASML is not the only company that makes lithography machines — Nikon and Canon (the camera companies) make them too. What’s unique is that ASML alone makes the cutting-edge machine capable of single-digit-nanometre resolution. To pull that off, the machine fires a CO2 laser at a falling droplet of molten tin 50,000 times a second, vaporising it into a plasma at 220,000°C, and bounces the resulting flash of extreme-ultraviolet light off Zeiss mirrors:

If you scaled them up to the size of Germany, the largest bump on the surface would still be smaller than a millimeter tall.

But — and this is the under-appreciated point — most chips in the world don’t need the bleeding edge. The chips in your toaster, Wi-Fi router, washing machine, and vaccines are perfectly happy on older Canon/Nikon machines. ASML is also one of the only non-American companies the US government has slapped export controls on, so for buyers who don’t need the very latest, it’s often easier to buy a basic machine elsewhere and skip the diplomatic paperwork.

The big five nobody talks about

Lithography hogs the spotlight, but it’s only one tool on the factory floor. The video introduces the rest of the “big five” of chip equipment:

ASML — lithography (printing the pattern).
Applied Materials — deposition (laying down metal and insulation one atom at a time), ion implanters, polishers, measurement tools. The 40th most valuable company on Earth, and most people have never heard of it.
Lam Research — plasma etching (carving the printed pattern into the silicon).
KLA — inspection and metrology (the very expensive microscopes that check whether the chips are any good before too many bad ones pile up).
Tokyo Electron (TEL) — the developer-track tools that handle every wafer before and after lithography, plus cleaning and thermal steps.

Each is the undisputed world leader in its slot. Three of the five are American, which the video notes is partly why they fly under the radar — fewer geopolitical headlines, and regulators who’ve often never heard of them. To become fully independent of this chain, a country would have to recreate all five companies plus their thousands of suppliers — a multi-decade project, if it’s possible at all. Many of these firms have been refining their processes since before modern silicon chips even existed.

The four foundries (really, the one foundry)

All these machines flow into just four companies that actually run the factories: Samsung, GlobalFoundries, Intel, and above all TSMC. Except for Intel, these foundries don’t design chips — they just manufacture other companies’ designs. The video’s analogy:

Like Ford designing the new Focus and then subcontracting out the entire manufacturing to a completely separate business that also happens to be building the new Toyota Corolla on the same factory floor for a direct competitor.

It’s strange, but the equipment is so expensive and the skill so concentrated that no chip designer can justify doing it in-house. TSMC alone makes north of 90% of the world’s leading-edge chips — and the T stands for Taiwan, which sits in China’s crosshairs. The video’s vivid framing:

It would be a bit like the Strait of Hormuz, but if 80% of the world’s oil flowed through there instead of just 20%. And that oil was made of glass.

TSMC is building fabs in Arizona and Japan to hedge, but admits its overseas plants run a generation or two behind — not for lack of machines, but because actually using them requires decades of “soft skills” that are notoriously hard to transfer.

Down to the sand

Before any of this, you need the raw wafer — a 300mm mirror-polished silicon disc. Half the global wafer market is two Japanese firms, Shin-Etsu and Sumco, with smaller shares for GlobalWafers (Taiwan) and Siltronic (Germany). It starts as high-purity quartz sand, refined to electronic-grade polysilicon (impurities measured in parts per billion), melted, pulled into a single giant crystal, then sliced and polished. A single disruption ripples through every fab on Earth — as in 2021, when a fire at a Shin-Etsu supplier helped trigger a global chip shortage that left half-finished cars sitting on dealership lots for nearly two years.

Memory, packaging, and the single line in Taiwan

Logic chips (the brains) are only half the story; memory is now just as critical. Without DRAM and the stacks of high-bandwidth memory (HBM) bolted next to a GPU, even the fastest logic chip is useless — it would have nothing to read from fast enough to matter. Memory comes from Samsung, SK Hynix, and Micron. SK Hynix is the sole qualified supplier of the most advanced HBM3e stacks Nvidia uses, packaged with bonding-wire techniques almost nobody outside South Korea can do at volume.

Then there’s the join. To attach those memory stacks to a GPU, you use a packaging process called CoWoS (chip-on-wafer-on-substrate). In 2025, nearly all CoWoS capacity sat inside a single TSMC facility — meaning for about 18 months the entire global supply of frontier AI chips was rate-limited by how fast one packaging line in Taiwan could move. Nvidia could design all the chips it wanted and SK Hynix could make all the memory it wanted, but they couldn’t be combined any faster than that one line allowed.

The software you can’t escape

One last bottleneck: Nvidia, Apple, AMD, Broadcom and Qualcomm don’t make chips — but they also don’t make the tools they design with. Drawing a modern chip means modelling tens of billions of transistors while accounting for quantum uncertainty as a real engineering input. Only two companies make this software competitively: Cadence and Synopsys. A single seat can cost over half a million dollars a year, and designers pay it because there’s no alternative — switching would mean rebuilding their entire pipeline on untrusted tools. Both are American, so their software is on the US export-control list too.

Why it holds together

The video closes by resolving its own paradox. Yes, nearly every link is a monopoly — a single point of failure. But each monopoly supplier sells to only a tiny handful of buyers who also can’t go anywhere else. Monopoly power on one side is offset by monopsony power (one buyer, or very few) on the other.

Their monopoly power is somewhat offset by the monopsony power on the other side.

So everyone is forced to cooperate, because there’s genuinely nowhere else for any of them to go. The whole tower is rigid and fragile and stable all at once. And as every new AI data centre comes online, it keeps bidding up the price of ultra-pure aluminium oxide, MSG-grade chemistry, Zeiss glass, Korean packaging expertise — and, by extension, the world’s finest toilets.

Key Takeaways

A single advanced chip is the output of 6,000+ suppliers across 40+ countries, and at most steps there is exactly one company in the world that can do it well.
Electrostatic chucks: wafers can’t be clamped (warping) or glued (chemicals) during manufacturing, so they’re held by atomically-flat ceramic plates using static electricity, with helium channels for cooling. Dinner-plate-sized, six figures each.
Toto, the toilet maker, is the world leader in those chucks — ultra-pure ceramics expertise transferred straight from toilet-making.
Ajinomoto (MSG inventor) makes ABF, the insulating build-up film under nearly every advanced chip; layer counts have risen from ~4-5 to over a dozen on modern AI accelerators.
ASML is not the only lithography maker — Nikon and Canon make them too. ASML is unique only for the cutting-edge EUV machine capable of single-digit-nanometre resolution.
EUV mechanism: a CO2 laser hits a falling tin droplet 50,000 times a second, vaporising it to 220,000°C plasma; the EUV flash bounces off Zeiss mirrors smooth to within a sub-millimetre bump if scaled to the size of Germany.
Most chips (toasters, routers, washing machines, vaccines) don’t need bleeding-edge nodes and run fine on older Canon/Nikon machines.
The “big five” equipment makers: ASML (lithography), Applied Materials (deposition/implant/polish), Lam Research (etching), KLA (inspection/metrology), Tokyo Electron (track/clean/thermal). Three are American and stay low-profile.
Four foundries run the factories — Samsung, GlobalFoundries, Intel, TSMC — and (except Intel) they manufacture other companies’ designs, not their own.
TSMC makes 90%+ of leading-edge chips, almost all in Taiwan; its Arizona/Japan fabs lag a generation or two because the “soft skills” of running a fab don’t transfer easily.
Bare silicon wafers: ~half the world’s supply is two Japanese firms, Shin-Etsu and Sumco. A 2021 fire at a Shin-Etsu supplier helped cause a chip shortage that stranded half-finished cars for ~2 years.
Memory matters as much as logic: SK Hynix is the sole qualified supplier of Nvidia’s top HBM3e stacks, using bonding-wire processes almost nobody outside South Korea can do at scale.
CoWoS packaging (attaching memory to GPU) was, in 2025, almost entirely inside one TSMC facility — rate-limiting the entire global supply of frontier AI chips for ~18 months.
Chip design software is a duopoly: Cadence and Synopsys, at >$500k/seat/year, both American, both on the US export-control list.
The system’s stability comes from monopsony offsetting monopoly: every monopoly supplier sells to only a few buyers who also can’t go elsewhere, so everyone is forced to cooperate.

Claude’s Take

This is a genuinely good explainer that earns its central paradox. The “robust and fragile” framing could easily have been a cheap hook, but the video pays it off with a real mechanism — monopoly suppliers held in check by monopsony buyers — which is a more sophisticated point than most chip-supply-chain content bothers to make. The toilet-and-MSG hook is a gimmick, but it’s a true one, and it’s used to illustrate a real economic logic rather than just to be quirky.

The accuracy is solid where I can check it. The big-five equipment makers, the TSMC concentration, the SK Hynix HBM3e position, the CoWoS bottleneck, the Cadence/Synopsys duopoly, the 2021 Shin-Etsu fire — these all match what’s broadly reported. The video is also honest about its own scope, repeatedly flagging that it’s a “very narrow look” and that the specifics are classified or beyond the host’s expertise. That restraint is a good sign; channels that overclaim usually don’t pause to say “there are more qualified people to explain this.”

Two small caveats. First, “220,000°C” and similar figures are the kind of round numbers that get repeated across the chip-explainer ecosystem; treat them as order-of-magnitude, not gospel. Second, the GenSpark mid-roll is a full minute of unrelated AI-tool advertising that has nothing to do with the topic — skip it. Docking nothing for it since it’s clearly marked, but it’s filler.

Score: 8/10. Clear, well-structured, technically careful, and it leaves you with a mental model (the monopoly-monopsony lock) rather than just a list of trivia. It loses points only for being a survey — it admits it’s skating over photoresist, gases, mask blanks, and bonding wire — so it’s a map, not a deep dive into any one territory.