ASML's EUV machine is the most complex thing

I wrote about TSMC last year and the process of making chips. Since then, I’ve gone deeper. Specifically, I’ve gone deep into the machine that makes those chips possible. ASML’s extreme ultraviolet lithography scanner. The NXE:3400B.

I need to tell you about this machine.

It costs $150 million. One machine. The price of a small fleet of 747s. It weighs about 180 tons. It requires three Boeing 747s to ship (40 freight containers, 20 trucks, 3 cargo planes). It has over 100,000 parts. And ASML, a company in Veldhoven, Netherlands, is the only company on Earth that can build it.

Not one of a few companies. The only one.

I’ve been reading everything I can find about it: ASML’s own technology pages, papers on IEEE Spectrum, Chris Mack’s lithography blog. And the more I read, the more I’m convinced that this machine is the single most impressive physical object humans have ever created. Not the most important (that’s maybe vaccines or the printing press). But the most complex. The most precise. The most “how is this even possible.”

Let me walk through it. I’ll try not to lose you, but I make no promises because this is the kind of thing that makes your brain feel inadequate.

The problem: we need smaller light

I covered this briefly last year, but let me go deeper.

To print circuit patterns on silicon, you need light. The pattern you can print is limited by the wavelength of the light you use. Short wavelength = smaller features = more transistors = faster chips. This has been the driving equation of the semiconductor industry for fifty years.

For decades, the industry used deep ultraviolet (DUV) light at 193 nanometers. Through increasingly clever tricks (immersion in water, printing patterns in multiple passes, mathematically warping the pattern to compensate for diffraction), they pushed 193nm light far beyond its natural limits. Down to 7nm features. Features more than twenty-five times smaller than the light printing them.

But the tricks were getting exponentially harder and more expensive. Each new node required more passes, more masks, more computation, more time. The industry was running on engineering heroics. Everyone knew a better solution was needed.

That solution was EUV. Extreme ultraviolet light at 13.5 nanometers. Fourteen times shorter than DUV. With EUV, you could print tiny features in a single pass instead of four or five. Simpler. Cleaner. Better.

One small problem: EUV light doesn’t really exist in a useful form. You can’t just buy a brighter lamp. And generating it turns out to be one of the hardest engineering problems ever attempted.

Laser-vaporized tin (yes, really)

Here’s how you make EUV light.

Step one: generate tiny droplets of molten tin. Each droplet is about 25 microns across. Smaller than a grain of pollen. These droplets fall through a vacuum chamber at a rate of 50,000 per second.

Step two: hit each droplet with a laser pulse. Not just any laser. A 40-kilowatt CO2 laser. But it’s actually a two-step process. A weaker “pre-pulse” laser hits the droplet first, flattening it into a pancake shape (this increases the surface area for the main pulse). Then the main pulse hits the flattened droplet and vaporizes it completely, turning it into a plasma with a temperature of about 500,000 degrees Celsius.

Five hundred thousand degrees. For reference, the surface of the Sun is about 5,500 degrees.

Step three: this plasma emits EUV light at 13.5 nanometers. The light radiates outward in all directions. A collector mirror, shaped like a satellite dish, captures as much of it as possible and focuses it into a beam.

This happens fifty thousand times per second. Fifty thousand tiny suns, created and destroyed in a vacuum chamber, each one lasting a few nanoseconds.

And that’s just the light source.

The mirror system

EUV light is absorbed by everything. Glass, air, water, basically every material you’d normally use to make a lens. You can’t use traditional optics. The only thing that reflects EUV light (and even then, only about 70% of it) is a multilayer mirror: alternating atomic-scale layers of molybdenum and silicon, deposited with a precision of fractions of a nanometer.

The EUV machine has eleven of these mirrors. Each one polished to a surface roughness of less than 0.05 nanometers. To put that in perspective: if one of these mirrors were scaled to the size of the entire Earth, the tallest bump would be about one millimeter.

Every time the light bounces off a mirror, it loses about 30% of its energy. After eleven reflections, you’ve lost most of the light you started with. That’s why the tin plasma has to be so intense, and why the whole system operates at such extreme power levels. You start with a lot of light because you know you’re going to lose most of it.

The stage

The silicon wafer sits on a stage that moves at incredible speed and precision. The stage accelerates at several G’s, moving the wafer into position for each new exposure. It has to position the wafer with an accuracy of less than one nanometer while moving at high speed.

Think about that. The stage is moving a 300mm silicon wafer (about the size of a dinner plate) with sub-nanometer precision. While accelerating harder than a fighter jet. In a vacuum. Under EUV light generated by laser-vaporized tin.

Every component in this system is operating at or near the physical limits of what’s possible. The lasers are among the most powerful industrial lasers ever built. The mirrors are the smoothest surfaces ever created. The stage is the most precise mechanical positioning system ever engineered. The vacuum system, the vibration isolation, the thermal management, all of it is pushing boundaries.

The contamination problem

Here’s a detail that took me a while to find, buried in a technical paper.

Tin. The same tin that generates the EUV light. It goes everywhere. When you vaporize fifty thousand tin droplets per second, some of that tin ends up on the collector mirror. Even a thin film of tin on the mirror surface degrades its reflectivity. So ASML engineered a system where hydrogen gas flows over the mirror surface. The hydrogen reacts with the deposited tin to form stannane (SnH4), a gas, which gets pumped away.

They’re cleaning the mirror while the machine is running. While fifty thousand micro-suns are being created and destroyed per second, mere centimeters away.

And the masks (the patterns that the EUV light shines through) have the same problem. Any contamination on the mask gets projected onto the wafer, magnified by the optics. A particle on the mask surface, invisible to the naked eye, becomes a fatal defect on the chip. So the masks are stored in special containers with pellicles (thin protective films) that are themselves technological marvels, thin enough to transmit EUV light, strong enough to protect the mask, resistant to the intense radiation.

Every layer of this system is solving a problem created by another layer. It’s problems all the way down.

The history

Something that struck me while reading about ASML: they’ve been working on EUV since the late 1990s. The first research prototypes were built around 2006. The first machines shipped to fabs for testing around 2010. They weren’t reliably producing chips in high volume until just recently, 2018 or so.

That’s roughly twenty years from concept to production. Twenty years of engineering. Twenty years of “this might not work.” Twenty years of solving one impossible problem only to discover the next one.

There were points when the industry nearly gave up on EUV. The light source was too weak. The mirrors degraded too fast. The masks were too dirty. The throughput was too slow. At various points, the consensus was that EUV would never work in production. ASML, TSMC, Intel, and their partners kept pushing.

I think that’s worth noting. The most important technology in the semiconductor industry, the one that enables every chip being designed today, nearly didn’t happen. Not because of physics. Because of patience. It works now, but only because enough people decided to keep going when the rational calculation said stop.

The supply chain

ASML doesn’t build this alone. The machine is an integration of components from about 5,000 suppliers across the world. The lasers come from TRUMPF in Germany. The optics come from Carl Zeiss in Germany. The wafer stages come from various precision engineering firms. ASML is the integrator, the company that takes all these impossible components and makes them work together.

TSMC in Taiwan, Samsung in South Korea, and Intel in the United States are ASML’s primary customers. They buy these machines (with years-long lead times) and integrate them into their fabs. Each fab might have dozens of EUV machines, running around the clock, producing the chips that end up in your phone, your laptop, your car.

The geopolitical implications of this supply chain are staggering. A single company in the Netherlands, dependent on suppliers in Germany and Japan, selling to fabs in Taiwan and Korea and the US, produces the machines that make the chips that make modern civilization function. If any link in that chain breaks, the whole thing stops.

I don’t want to be alarmist about it. Supply chains are resilient, and there are redundancies. But the concentration of capability is remarkable. There is no backup plan for ASML. There is no alternative supplier. If EUV lithography is the printing press of the digital age, then Veldhoven, Netherlands is Gutenberg’s workshop.

The numbers that make me dizzy

Let me collect some figures that I keep coming back to, because I think the sheer accumulation of them tells a story.

The EUV source creates plasma at 500,000 degrees Celsius. The mirrors are smooth to 0.05 nanometers. The stage positions wafers to sub-nanometer accuracy at multi-G acceleration. The vacuum system maintains pressures comparable to outer space. The entire machine consumes about 1 megawatt of power (enough to power about 750 homes). The light source produces only about 250 watts of useful EUV power at the wafer, which means the machine is less than 0.03% efficient at converting wall power to useful light. And this is considered a triumph. Because even 250 watts of EUV is enough to expose 170+ wafers per hour, and each wafer becomes hundreds of chips.

ASML ships about 50 EUV machines per year. At $150 million each, that’s $7.5 billion in revenue just from EUV. The company’s total revenue is higher because they also sell DUV machines and provide maintenance and upgrades. But EUV is the product. The reason ASML is worth more than most banks and airlines combined.

Why I’m obsessed with this

I’ll be honest: I don’t fully understand all the physics. I’ve read explanations of multilayer mirror design and plasma dynamics and I can follow along, but I couldn’t derive any of it from first principles. This is well beyond my technical depth.

But I think that’s part of why it fascinates me.

This machine represents the absolute peak of human coordination. Tens of thousands of people, across hundreds of companies, in dozens of countries, all contributing to an object so complex that no single person understands all of it. The laser physicist doesn’t fully understand the mirror polishing. The mirror polisher doesn’t fully understand the stage dynamics. The stage engineer doesn’t fully understand the plasma physics.

And yet it works. It produces chips with features smaller than a virus, at a rate of more than 100 wafers per hour, day after day, in fabs around the world.

I think there’s something in there about what humans are capable of when we coordinate properly. Not through force or hierarchy, but through shared standards and complementary expertise and a collective will to push the boundaries of what’s physically possible.

Or maybe I’m reading too much into it. Maybe it’s just a really impressive machine. But at 1am, reading about tin droplets and plasma temperatures, it feels like more than that.

It feels like a cathedral. Not one made of stone. One made of light.

Related thinking: