If you’re new to the publication, I recommend checking out the quick start guide. It’ll tell you what to expect from these articles, and get you excited for what’s to come.

👋 Introduction

Today I’ll be starting the deep dive on the third company in the semiconductor series. The previous two, as readers will know, were Nvidia (a chip designer) and TSMC (a chip foundry). The third company to talk about is ASML, a Netherlands-based equipment manufacturer.

Why do I care so much about the semiconductor industry? It’s because of the AI boom. Semiconductors, especially GPU chips, are needed to drive progress forward. Companies that are involved in semiconductor production are the metaphorical “picks and shovels” of the AI movement. If you aren’t familiar with that phrase, it’s an old bit of investing advice for identifying winners in the stock market.

During the US gold rush, some of the folks who became the wealthiest were the ones who sold picks and shovels. They didn’t care whether their customers found gold or not, because their profit was secured regardless. This has become a common investing refrain. Find the company that will make money regardless of which customers become successful.

That’s what I’m doing with AI. It’s likely that Apple, Google, Meta, and OpenAI will make lots of money selling their AI products. It’s likely that some other companies will become very successful selling their AI products too. But I don’t have to figure out which ones will be successful, because I’m looking at the companies that supply all of the above names. Regardless of whoever wins, I win too.

I believe we are still in the nascent stages of the AI revolution, and demand in the semiconductor industry is just heating up. While hype has driven some stock prices to crazy levels, the long term value in companies still exists. In anticipation of correctly-priced investment opportunities, I continue my analysis of the sector as a whole.

Let’s dive in!

Understanding the industry

Where are we in the supply chain?

The last article on TSMC focused on the foundries, who manufacture the actual chips for fabless companies like Nvidia. Today I’m going one link further down the supply chain to the equipment manufacturers.

Consider the metaphor of building a house. The fabless companies are the architects who design the house. The foundries are the general contractors and laborers who build the house. The equipment manufacturers are the ones who supply the saws, hammers, drills, etc.

Of course, the equipment we’re talking about is far more complicated than hammers and drills. A foundry’s “starter kit” of tools costs hundreds of millions of dollars.

There are tools required for etching the chips. Tools required for depositing specific chemicals on the surface of the wafer. Tools required to measure the alignment of transistors on the wafer. Tools required to clean the chemicals off the wafer once production is complete.

But we’re going to talk about one tool in particular today.

The lithography machine.

What does a lithography machine do?

A lithography machine performs the most critical step of chip manufacture - it applies the chip design template to the silicon wafer.

Think back to the last article. Once the silicon has been purified, and the cylinder has been created, and the cylinder has been sliced into wafers, then the wafer goes into the lithography machine. The machine prints the chip design onto the wafer, after which the wafer continues on to the remaining steps of production.

The chip design exists on a reticle. A reticle is basically a stencil. It has the exact drawing of where the transistors need to go. The lithography machine projects this drawing onto the wafer, and then does “addition by subtraction”, essentially removing all the parts of the wafer that aren’t in the design. This is a massive simplification, and we’ll get into some more details soon, but for right now this is good enough.

Lithography machines are needed because of how small the transistors are. The reticle is not a 1:1 scale of the final chip, it’s more like 10,000,000:1. The lithography machine shrinks the design down.

As we covered before, in order to continue increasing chip power, we need smaller transistors. In order to make smaller transistors, we need lithography machines to shrink the chip designs down to the smallest possible size.

To give you an idea of the scale these machines are working at, a 1 cm chip can have around 26 miles of patterns within it. 26 miles in 1 square centimeter! That seems to defy common sense.

So, if we want AI to continue to improve, we need lithography machines that can print smaller and smaller transistors.

How does it actually work?

Lithography literally means “printing with light”. In this case, the light is ultraviolet light.

I’m going to dive into how the machine works, but I promise I won’t get overly technical. Stick with me here - I think it’s really important that you, as a potential investor, understand the complexity of the engineering so that you can appreciate the company’s moat. Being able to understand what’s happening will really make a difference.

Ok, here we go:

1. The silicon wafer is coated with a chemical layer called photoresist.

   

2. The machine generates ultraviolet light and projects it through the reticle (aka stencil).

   

3. The light that gets through the reticle is focused using a projection lens. This is how we get down to the nanometer-sized transistors.

4. The UV light changes the chemical composition of the photoresist that it hits, making it soluble. That soluble area is then washed away. Now you have some silicon that isn’t covered by photoresist.

   

5. The exposed silicon is etched away. Now you only have silicon where you want it to be - these are the transistors that’ll conduct electricity.

   

6. The remaining photoresist is cleaned off and the gaps are filled with silicon dioxide, which protects the transistors.

   

One layer of geometry has been created, and this process needs to be repeated for as many layers are needed in the chip design. And then repeated for the hundreds of chips in a wafer. And then repeated hundreds of thousands of times. Remember, this is all happening at a nanometer-level precision.

It should be clear that these machines are no joke. The level of precision is not something that’s easily digestible to the average person. Sub-nanometer discrepancies will render millions of dollars of product worthless. Every step has to work perfectly.

What are the types of lithography machines?

Lithography machines are classified by the wavelength of the light that they project onto the wafer. Older machines had a longer wavelength, which limited how small the transistors could be. Advancements drove the light wavelength down:

1. 435 nm, called mercury g-line

2. 365 nm, called mercury i-line

3. 248 nm using krypton fluoride or KrF

4. 193 nm using argon fluoride or ArF, also referred to as deep ultraviolet (DUV)

5. 13.5 nm also referred to as extreme ultraviolet (EUV)

From 365 nm and down, the machines are still widely produced and sold. A significant portion of the equipment manufacturers’ revenue comes from the older machines, even today.

But we’re going to focus our attention even further - from all the tools and all the lithography machines - down to just DUV and EUV. These are the technologies that are responsible for the advanced chips (<10 nm) of today, and will likely define the advanced chips of the future.

What is DUV and who are the key players in the market?

DUV was first commercially viable around the year 2000. It was a natural evolution of the previous models of lithography machines. Its development was driven mainly by trial and error - it took many iterations to get a working machine that could be deployed at scale.

A quick history on DUV machines: the first DUV machines produced chips with transistors in the 60 nm range. Incremental improvements to DUV allowed the machines to produce smaller and smaller chips. Immersion, which involves shooting the UV light through water, was a big breakthrough that allowed for production of chips in the 20 nm range.

The next big breakthrough involved patterning. The challenge was that transistors were too close to each other and the DUV laser wasn’t precise enough to maintain integrity between adjacent cuts. So the solution was to do the cuts in two separate passes. This came to be known as double patterning. Let me explain it with a visual:

Basically, because the lasers couldn’t get too close to previous cuts, the reticle was separated into two layers that each required less precision. The laser could then cut out each part of the reticle, and the two parts could be stitched together. This introduced its own complexities - namely, chip designs needed to be optimally and automatically split into multiple layers, and the two produced layers needed to be aligned perfectly. Entire companies sprung up with the sole purpose of addressing these challenges.

Double patterning worked for 20, 16 and 14 nm nodes. But DUV plus double patterning was insufficient for 10 nm nodes.

So what did they do next? Triple patterning of course. It worked for 10 nm nodes, but at great cost. See, while double patterning was of a reasonable computational complexity, triple patterning was exponentially more difficult. Let’s consider an example.

I have a square and I have to decide which of the four points will go on each layer in a double pattern. There are only a few possibilities to consider:

But when I add a third layer, suddenly the possibilities increase drastically:

This is a very simple example, but it highlights how quickly the number of options explodes. Imagine instead of a simple square, we had thousands, millions, billions of points.

The production process using DUV machines became much slower and more costly for foundries across the globe. It was clear that we had reached the absolute physical limitations of what DUV could produce.

Nonetheless, there were still chips to be produced and money to be made. Three companies came to dominate the DUV market (and still do): ASML, Canon and Nikon.

Looks like a pretty good market share for ASML, right? Sure is, but more on that in a bit. First let’s regroup on the current state of DUVs.

We know that we’re dealing with a mind-bogglingly complex engineering problem, and we’re reaching the limits of what physics can offer us. Around 2016, a lot of people started proclaiming Moore’s Law to be dead. How could we keep doubling the number of transistors on a chip if we couldn’t produce a smaller transistor?

There’s one other takeaway I want you to get from this section. There was a reason I went into the details of immersion, double patterning, triple patterning, and so forth. It may be harsh, but I view these innovations with a degree of hesitation. It feels like a “band-aid” solution; in other words, the DUV lithography machine has certain capabilities, but those capabilities are being extended beyond their intended use, via clever engineering. There’s nothing wrong with clever engineering, but I’d be more confident if the lithography industry wasn’t relying on stitching multiple solutions together.

Partly this was driven out of necessity, because the next-gen tech (EUV) kept getting delayed for years and years. These incremental solutions to DUV were the only commercially viable option. But it’s worth remembering how DUV was extended, because it becomes a recurring theme.

What is EUV and who are the key players in the market?

EUV was long thought to be impossible. Not impossible in a “that seems difficult” sort of way, but impossible in a “the laws of physics don’t allow for that” sort of way.

The idea was initially proposed in the 90s by a team of researchers funded by the US government. It was far from anything usable, more academic than practical. In a stark departure from DUV and previous technologies, innovation couldn’t be driven by trial and error. The scale of innovation was such that a deep understanding of physics was required.

ASML purchased the rights to the results of the study, and began to pour billions of dollars of R&D into EUV. Other competitors moved on after not being able to get access to the study. Why waste time and money chasing something that seemed impossible?

But then ASML caught what they were chasing.

In 2016, right as the doomsday predictions for Moore’s Law were reaching a peak, ASML proved the commercial viability of EUV. TSMC bought a machine and the first 7 nm chips were produced.

A breakthrough in chip production. Moore’s Law was saved, thanks to ASML’s EUV lithography machine.

Today EUV machines are used to produce 7 nm, 5 nm, and 3 nm chips. Remember the market share of DUV machines from earlier? Let me one-up that with a market share chart of EUV machines, globally, for the entirety of history:

That’s right. No one else has figured out how to make EUVs. ASML is the only company in the entire world that knows how to make these machines. That’s been true for the last 8 years, and it looks like it’ll hold for the foreseeable future.

How did ASML pull this off? How did they do what scientists and engineers all over the world thought was impossible?

That’s where we’ll pick up next week. And a little preview - all of the engineering that I talked about today is just the setup. What goes into the development of EUV machines is far more complex (if that’s even possible) than anything DUV-related.

Takeaways

Takeaways? Uh, engineering is insane.

The story so far is that lithography machines are responsible for printing the ever-smaller transistors onto chips. Advancements in these machines are what keep the innovation cycle going. When it seemed like Moore’s Law was dead, the EUV lithography machine saved the day and unlocked new futures of computing.

We know ASML did the impossible. But how did they pull it off?

Let’s summarize what we’ve learned today:

This chart visualizes the things we want to check before making an investment decision. Ideally, we’d get 8 green lights and could have confidence buying into the company. If we get 8 red lights, then it’s clear to stay away. If it’s somewhere in between, then the decision gets tougher.

Next week, we’ll dive in to the Product section of the article and get answers on the second row of this table.

Thanks for reading, have a great week, and I’ll see you soon!

Agree? Disagree? Would love to hear your thoughts - leave a comment.