Turtles All the Way Down

In Part I, I traced the physical supply chain of the AI buildout — the machined cold plates, the copper busbars, the shrinking workforce that builds all of it — and ended with a claim: the AI supply chain is hitting a physical wall.

You thought power transformers were bad? Let me show you what's underneath the chip supply chain.

Micron-level tolerances have a texture that's hard to convey. The thermal drift that ruins a morning's work. The way a cutting tool deflects differently on the fifteenth hour than on the first. The sinking feeling when a coordinate measuring machine tells you a casting warped during heat treatment. EDM wire snapping at 2am on a part with three weeks of work in it. This is the world of precision manufacturing. And when you dig into the supply chain of EUV lithography — the machine that makes every advanced chip on Earth — what you find is that world, taken to its absolute extreme.

I want to take you on an archaeological dig. We start at a machine that costs more than a Boeing 787, and we keep digging — past the obvious bottleneck, past the bottleneck's bottleneck, past the bottleneck's bottleneck's bottleneck — until we hit bedrock. It's turtles all the way down. And what sits at the bottom changes how you should think about every "scaling" narrative in AI.

The $350 Million Machine

Every advanced semiconductor chip made on Earth in 2026 passes through a machine made by a single company in Veldhoven, the Netherlands.

The company is ASML. The machine is an EUV lithography system. It is the most complex device humanity has ever manufactured.

An ASML EUV scanner contains roughly 100,000 individual parts, sourced from approximately 5,000 suppliers, of which 800 are critical tier-1. The next-generation High-NA system — the EXE:5000, which started shipping in 2024 — costs upwards of $350 million per unit. It arrives disassembled in 40 shipping containers and takes months to reassemble on-site.

There is no alternative. No other company makes EUV lithography systems. No other technology can print the features required for leading-edge chips. This is not a market monopoly. It is a physics monopoly. Nobody else has figured out how to build the thing.

ASML plans to ship roughly 90 EUV systems in 2026, including its first volume run of High-NA tools. Every conversation about chip supply chains, semiconductor sovereignty, the geopolitics of AI — all of it converges on this machine. The question everyone should be asking: what constrains how many ASML can build?

This is where the turtles start.

First Turtle: The Frames

If you had to guess the bottleneck for building a $350 million machine with 100,000 parts, you'd say chips. Semiconductors. Rare earths. Something that makes headlines.

Wrong.

Start with the frame sets that hold the optical components in the High-NA system. These are manufactured by VDL ETG, a Dutch precision engineering firm. Each frame set begins as two blocks of aluminum, each weighing 20 tonnes. Over 1,300 hours of CNC machining — more than 54 continuous days per block — 85% of the material is carved away, leaving a 3-tonne precision structure. VDL had to build entirely new production halls and raise their ceilings to accommodate 9-meter-tall measuring machines just to verify the finished parts.

But machining is only the start of it. Those 3-tonne frames go through stress-relief heat treatment cycles so they don't warp under their own internal stresses once the metal is removed. Critical interfaces are ground and lapped flat — not CNC-finished, but hand-finished on surface plates by people who can feel microns with their fingertips. Mounting surfaces for optics are wire-EDM'd to geometries no end mill can reach. The whole assembly is then measured on coordinate measuring machines the size of rooms, in temperature-controlled environments held to fractions of a degree.

I know machining. 1,300 hours is extraordinary, but the tolerances for these structural components — micron-range — are achievable. Any good five-axis shop can do this work. You can parallelize it: more machines, more shifts, more suppliers. ASML pre-orders aluminum months in advance precisely so machining doesn't become the constraint.

The frames are a turtle. But there's a turtle underneath them.

Second Turtle: The Mirrors

EUV light has a wavelength of 13.5 nanometers. At that wavelength, light is absorbed by everything — including glass. You cannot use lenses. There is no transparent material. The entire optical path uses mirrors instead.

These are not ordinary mirrors.

The projection optics in a standard EUV system contain six mirrors, each polished to a surface figure error of 50 picometers RMS. Fifty trillionths of a meter. A silicon atom is about 200 picometers across. These mirrors are smooth to a quarter of an atom.

Zeiss's own analogy: if one of these mirrors were scaled to the size of Germany, the largest bump on its surface would be 0.1 millimeters.

These optics are made by exactly one company on Earth: Carl Zeiss SMT, in Oberkochen, Germany. Zeiss SMT employs approximately 8,500 people and generates €4.1 billion in annual revenue — essentially all of it serving ASML. ASML owns 25% of Zeiss SMT, an equity stake taken specifically to ensure Zeiss had the capital to expand. The math is simple: ASML ships exactly as many EUV systems as Zeiss can produce optics for. No more.

The High-NA mirrors are worse. Several hundred kilograms each — twice as large and ten times as heavy as standard EUV mirrors. Zeiss built 150-tonne measuring machines housed in 5-meter vacuum chambers just to verify them. Those VDL frames — 2,600 hours of machining, new production halls, 9-meter measuring equipment — exist solely to hold these mirrors still.

Making a mirror smooth to 50 picometers is not a single process. It is a months-long iterative siege against imperfection. CNC grinding gets the mirror roughly shaped — that's the easy part. Then ion beam figuring takes over, removing material atom by atom. The mirror is measured with custom interferometers. Errors are calculated. The ion beam corrects. Measure again. Correct again. Ten to fifty iterations, each taking days. Each correction fixes one set of errors while introducing new ones. The process converges like Zeno's paradox made physical: halving the remaining error each time, asymptotically approaching perfection.

I've watched optical engineers describe this process. There is a particular look they get — a mixture of pride and exhaustion — when they talk about a mirror that took six months and still needed two more correction cycles. One Zeiss engineer I spoke with compared it to "sculpting fog." You remove atoms you can't see, to fix errors you can barely measure, on a surface you must never touch.

You cannot parallelize this. You cannot skip steps. You cannot rush the thermal settling between measurements. One mirror. One months-long serial process. One company.

Zeiss is the bottleneck. It is always the optics. But that pushes the question one level deeper — and this is where it gets truly unsettling.

Third Turtle: The Suppliers' Suppliers

Behind Zeiss sits a layer of suppliers so specialized, so small, and so sole-source that when I mapped them, I felt actual vertigo.

The mirror substrates. They must be made from a material with near-zero thermal expansion — because nanometers of thermal distortion would ruin a 50-picometer surface. Two materials exist on Earth that qualify. The first is Zerodur, made exclusively by Schott AG in Mainz, Germany. Founded in 1884. Sole-source for 140 years. Making Zerodur requires a months-long ceramization process governed by thermodynamics — crystal nucleation proceeds at a rate set by physical constants, not manufacturing parameters. One furnace, one blank, months per cycle. The second is ULE glass, made exclusively by Corning in Canton, New York, through a flame hydrolysis process that is equally slow for equally non-negotiable thermodynamic reasons. Corning received $32 million under the CHIPS Act to expand production. That number tells you everything about the scale: $32 million to expand the material foundation of a trillion-dollar industry.

Two materials. Two companies. Two campuses. Both constrained by thermodynamics that does not care about your capital expenditure plan.

The polishing machines. The ion beam figuring systems that polish mirrors atom by atom are made by two small German companies: NTG in Gelnhausen (roughly 200 employees) and scia Systems in Chemnitz (50–100 employees). These are the kind of companies where the managing director knows every engineer by name. Production volumes are measured in single-digit units per year. Their order books in 2025 are full. These tiny firms — invisible to the financial press, unknown to every VC in San Francisco — make the machines that make the mirrors that make the chips that train the AI models that underpin a $4 trillion market cap shift.

The measurement instruments. The interferometers that measure mirror surfaces to sub-nanometer accuracy are made by Zygo Corporation in Middlefield, Connecticut, a town of 4,200 people. Roughly 500 employees. Here is the recursive trap that stopped me cold: a Fizeau interferometer works by comparing the test surface against a reference flat that must itself be polished to at least the accuracy you're trying to measure. The reference flat is verified against another reference. You need optics to measure optics to make optics. It's turtles inside the turtle.

The measurement uncertainty directly limits how far you can correct. If you can't measure below 30 picometers, you can't figure below ~50 picometers. The metrology doesn't support the manufacturing — it defines its ceiling.

And making those reference flats? They are ground, lapped, and polished in a process that itself relies on precision-cast housings, heat-treated structural components, and thermally stable mounting systems machined to micron tolerances. Every layer of this supply chain, when you crack it open, reveals the same substrate: advanced manufacturing processes — machining, casting, grinding, lapping, heat treatment, EDM — stacked on top of each other in recursive dependency.

This is the third turtle. A handful of companies, each with fewer than a thousand employees, each sole-source, each producing things measured in single digits per year. The entire AI scaling narrative — the $100 billion data centers, the geopolitical jockeying, the breathless earnings calls — rests on the output of workshops where individual craftspeople know individual machines by their quirks.

Bedrock

Beneath the third turtle, there are no more turtles. Just three irreducible constraints:

Physics. Thermodynamics does not negotiate. Crystal growth in Zerodur proceeds at a rate set by nucleation kinetics — physical constants, not engineering parameters. Ion beam figuring removes atoms one at a time. Thermal settling between measurements is governed by heat diffusion. You cannot throw money at heat diffusion. You cannot disrupt nucleation kinetics. These are the speed limits of reality, and they apply equally to a startup and a nation-state.

People. The global pool of humans who can work at this precision is in the low thousands. I have met some of them. They are quiet, meticulous people who speak about measurement uncertainty the way sommeliers speak about terroir — with a specificity that sounds obsessive until you realize it's just competence. The training pipeline: a physics or engineering degree, often a PhD, then years of apprenticeship under someone who already has the craft. That is 10–14 years to produce one fully qualified person. There is no bootcamp for sub-atomic polishing. The critical knowledge is not in textbooks — it lives in people's hands and judgment, built over thousands of hours of practice. A Schott glass technician who has spent twenty years managing ceramization furnaces carries irreplaceable knowledge about how Zerodur behaves under conditions that have never been written down. When that person retires, that knowledge walks out the door.

Sole-source facilities. Schott in Mainz. Corning in Canton. Zeiss in Oberkochen. NTG in Gelnhausen. Zygo in Middlefield. If Schott's Mainz facility had a fire tomorrow, there would be no way to source EUV mirror substrates from anywhere else on Earth. The raw minerals are common — silicon dioxide is sand. The decades of process knowledge to turn sand into the most thermally stable solid ever made? That is locked in a few buildings, in a few heads, in a few small German and American towns that most people have never heard of.

This is bedrock. Physics, people, and a handful of irreplaceable facilities. No amount of capital changes any of it on a timeline that matters for the current AI scaling race.

The Punchline

Here is what I did not expect to find.

When I started this dig, I thought the bottleneck would be something exotic. Some deep physics thing completely outside my experience. A material I'd never heard of. A process from another world.

Instead, every single layer of this supply chain runs on the same foundation: precision manufacturing. Not just CNC machining — the full stack.

The VDL frames: 1,300 hours of CNC machining, then heat treatment and grinding and lapping. Schott's Zerodur blanks: ceramized in custom furnaces, then processed in a €30 million CNC Competence Center. The vacuum chambers for ion beam machines: machined, welded, stress-relieved. Zygo's interferometer housings: precision-cast, machined, thermally stabilized. The initial shaping of Zeiss mirrors: CNC ground. The structural members: forged. The complex internal geometries: wire-EDM'd. The mating surfaces: lapped to optical flatness by hand. The thermal management systems: brazed, machined, tested.

Machined. Cast. Forged. Ground. Wire-EDM'd. Heat-treated. Lapped. Brazed. Coated. Measured.

This is not one bottleneck. It is an entire ecosystem of advanced manufacturing disciplines, layered on top of each other in recursive dependency. Remove any one — take away precision casting, or grinding, or EDM, or metrology — and the chain breaks. Not just for EUV. For everything.

Go dig into the supply chain for gas turbines: blades investment-cast in single-crystal nickel superalloys, coated with thermal barrier ceramics, machined to airfoil tolerances. Fusion reactors: vacuum vessels electron-beam welded from forged segments, then machined to tolerances that make EUV frames look easy. Satellite optics, surgical robots, quantum computing cryostats: machined, ground, EDM'd, brazed, lapped. Every advanced manufactured thing on Earth, if you dig deep enough, rests on this same stack of precision manufacturing disciplines — CNC machining at the center, but forging, casting, grinding, EDM, additive, heat treatment, and metrology holding everything together.

The chip supply chain has a bottleneck, and that bottleneck has a bottleneck, and that bottleneck has a bottleneck, and at the very bottom of the most advanced manufacturing process humanity has ever devised — you find a factory floor. Not a single machine. An entire civilization of manufacturing knowledge, distributed across thousands of small companies, tens of thousands of skilled workers, and a web of processes that took a century to develop and cannot be replicated by writing a check.

The trillion-dollar question — the one that Leopold Aschenbrenner's Situational Awareness never asked — is how many other supply chains look exactly like this.

The answer: all of them. EUV is just the one where the precision is so extreme that the structure becomes visible.

The turtles go all the way down. The question is whether we keep building on top of them without looking — or whether we finally count the atoms.

Part III: Zero Sum →