Lithography: the machine that prints the modern world

Every layer of a chip begins as a pattern that must be transferred onto the wafer. Photolithography does it like a photographic enlarger running in reverse:

1. Coat the wafer with photoresist, a light-sensitive chemical.
2. Shine light through (or off) a mask, a stencil carrying one layer of the circuit design.
3. A lens system shrinks the pattern ~4x and projects it onto the wafer, exposing the resist.
4. Develop, then etch or deposit material where the resist allows; strip; repeat for the next layer, 60-100+ times per chip.

The physics constraint that rules the whole industry: the smallest feature you can print scales with the wavelength of the light. Sharper chips need shorter wavelengths, or ever more elaborate optical tricks. That single relationship, features scale with wavelength, explains four decades of lithography history: visible light, then ultraviolet at 365 nm, then deep ultraviolet (DUV) at 248 and 193 nm, each transition forcing new light sources, new lenses, new chemistry.

And it explains why the industry eventually faced a wall: below roughly 20-nanometer features, 193 nm light is like painting miniatures with a broom. Something drastic was required.

Before the drastic solution arrived, engineers spent fifteen years extracting impossible resolution from 193 nm light. Two tricks carried the industry from the mid-2000s into the 2020s, and both still matter:

• Immersion lithography: put ultra-pure water between the lens and the wafer. Water bends light more than air (higher refractive index), effectively sharpening the optics. This gave roughly one generation of improvement.
• Multi-patterning: if one exposure cannot print features close enough together, print half the pattern, shift, and print the other half (double patterning), or quadruple it. Resolution improves without new physics, but costs explode: each extra patterning pass means more masks, more process steps, more alignment risk, more yield loss, and more time per wafer.

Multi-patterning is the reason "can you technically make it" and "can you make it economically" are different questions. A chip layer that needs one EUV exposure might need four to six DUV exposures to replicate, viable in principle, ruinous at scale, and this asymmetry is precisely what makes advanced lithography access such an effective policy lever (lesson 3 returns to this).

By the mid-2010s the arithmetic was closing in: every further node on DUV alone multiplied cost and complexity. The industry needed the wavelength jump it had been postponing for twenty years.

Extreme ultraviolet (EUV) lithography drops the wavelength from 193 nm to 13.5 nm, a 14x jump that restores single-exposure printing at the leading edge. The number sounds like an incremental spec change; the engineering it demands borders on the absurd:

• The light source: nothing lases at 13.5 nm conveniently. The production solution fires a droplet of molten tin 50,000 times per second, hits each droplet with a high-power CO2 laser pulse (twice, a pre-pulse to flatten it, then the main blast), creating a plasma hotter than the sun's surface that radiates EUV light.
• The optics: 13.5 nm light is absorbed by everything, including air and glass. The entire beam path lives in vacuum, and lenses are replaced by mirrors coated with ~100 alternating atomic layers of molybdenum and silicon, polished so precisely that, scaled to the size of a country, their largest imperfection would be millimeters tall. Even these mirrors absorb ~30% per bounce, and the light bounces about a dozen times.
• The machine: ~100,000 parts, several shipping containers to transport, roughly $150-200 million per unit (the newer high-NA generation, with larger optics for the next nodes, runs around$350-400 million).

Each machine prints wafers for exactly the handful of fabs pursuing the leading edge, and the world's entire supply comes from one company: ASML, in the Netherlands.

"One company makes the machine" invites conspiracy-shaped explanations. The mechanical ones are better:

• A thirty-year bet. EUV research began in the late 1980s in national labs; ASML committed in the late 1990s and shipped economically viable production tools around 2018-2019. Two decades of negative return on a technology that repeatedly looked like it might never work: competitors (in Japan and the US) evaluated the same bet and folded. Monopoly here is what surviving looks like.
• Customers funded the finish. In 2012, the three leading chipmakers took equity stakes in ASML and co-funded EUV development, the industry collectively decided it needed exactly one winner to exist at all.
• The moat is the supplier tree, not the blueprint. The machine assembles irreplaceable sub-monopolies: mirrors from one German optics house (Zeiss SMT), the laser from one German firm (TRUMPF), mask blanks and photoresists dominated by a few Japanese suppliers. Replicating ASML means replicating that entire tree, plus the thousand-engineer install-and-service operation each fab depends on. (High machine uptime is a service achievement; the machines are never really "sold and forgotten.")
• Tacit knowledge. Decades of accumulated adjustments live in people and processes, not documents. This is yield learning (lesson 1) applied to the equipment layer.

The generalizable rule: monopolies of this type are made of time. Money compresses many things; thirty years of co-evolved supplier ecosystems is not one of them.

The two tool families define the industry's most important capability boundary:

Three readings of the table matter later in this path. First, DUV is not obsolete: even the most advanced chip uses EUV for only its finest layers; the bulk are printed with DUV. A fab is a fleet, and chokepoints apply fleet-wide. Second, the capability boundary is quantifiable: with EUV, a given node is a cost problem; without it, the same node is a multi-patterning yield problem that compounds per layer. Third, the supplier column is the entire export-control story in miniature: two governments' jurisdictions (Dutch, Japanese) cover every machine on the list, which is what makes the controls of the next lesson administratively possible at all.

Because every advanced fab depends on one vendor's machines, several unusual dynamics follow, each a preview of the policy story:

• The order book is a map of the future. Fabs order tools two to three years ahead; a record backlog (dozens of EUV systems queued by the major foundries) is the most reliable public signal of where leading-edge capacity will exist in three years. Analysts read ASML's backlog the way meteorologists read pressure systems.
• Allocation is power. When tool supply is the binding constraint, who gets the next machine determines which fab, and which country, expands first. This allocation is decided by delivery queues, long-term agreements, and, increasingly, governments.
• Service is a standing dependency. An EUV tool needs continuous vendor support: spare parts, software updates, resident engineers. A machine already installed can degrade into a very expensive sculpture if service stops, which means export restrictions can reach backwards into machines sold years earlier.
• Single-source risk cuts both ways. The same concentration that makes the supply chain controllable makes it fragile: one factory fire, one supplier failure, one logistics break at the wrong node of the tree, and the entire industry's expansion slips.

None of this required anyone to seek geopolitical leverage; the leverage emerged from industrial structure. The next lesson is about the legal machinery built to use it.

Lithography condenses this path's core ideas into one object:

1. Physics sets the terms (features scale with wavelength), but economics picks the winner (whoever survives the decades-long R&D valley owns the layer).
2. Capability boundaries are sharp. The DUV/EUV line separates "expensive but routine" from "heroically difficult": a discrete threshold, not a gradient. Sharp thresholds are exactly what regulations can be written around.
3. Deep moats are ecosystems. The monopoly is not a machine but a web: national-lab research, one optics house, one laser firm, materials suppliers, service engineers, and thirty years of tacit knowledge. This is why "just build a domestic equivalent" programs measure progress in decades.
4. Chokepoints are dual-use. The same single point that gives policymakers a lever gives the industry a fragility, and gives the firm holding it a strategic importance out of proportion to its size.

Keep the sharpness of point 2 in mind: when governments sat down to write rules restricting "advanced" chipmaking, they needed legally definable lines, and lithography's discrete capability tiers offered the cleanest ones available. How those rules actually work, entity lists, direct-product rules, node thresholds, and the coordination problem among three governments, is next.