Photo illustration by Will Joel / The Verge.
A few weeks ago, President Joe Biden was in the Netherlands, where he asked the Dutch government to restrict export from a company called ASML to China. ASML is the only company in the world that makes a specific machine needed to make the most advanced chips. Apple couldn’t make iPhone chips without this one machine from the Netherlands’ biggest company. ASML doesn’t just shape the Dutch economy — it shapes the entire world economy. How did that happen?
Chris Miller, Tufts professor and author of Chip War: The Fight For The World’s Most Critical Technology walked me through a lot of this, along with some deep dives into geopolitics and the absolutely fascinating chip manufacturing process. This one has everything: foreign policy, high-powered lasers, hotshot executives, monopolies, the fundamental limits of physics, and, of course, Texas. Here we go.
Decoder with Nilay Patel
Professor Chris Miller’s new book “Chip War” explains the complicated global politics inside your iPhone
Forward 15 seconds
Back 15 seconds
Share
Subscribe
Description
Chris Miller is a professor at The Fletcher School at Tufts University and the author of Chip War: The Fight for the World’s Most Critical Technology. Welcome to Decoder.
Thanks for having me.
We’re here because of a recent news story that struck me as indicative of all kinds of things happening in the tech world and in the chip industry writ large, which is that President Biden is trying to pressure the Dutch government into not shipping chip-making equipment to China. At first glance, that seems like a very surprising geopolitical situation to be in. Can you explain what’s going on there?
Right now the US is trying to cut off China’s ability to make advanced semiconductors, on the judgment that advanced semiconductors are critical to training AI systems. If you can’t get access to the most advanced chips, then you can’t make meaningful advances in AI.
To make an advanced semiconductor, you need to buy machine tools from just a handful of companies around the world that have the precision capabilities to manufacture these tools. One of the most important of these companies is a firm called ASML, which is based in the Netherlands. It has unique capabilities — which no one else in the world can replicate — to produce a type of machine called an EUV lithography tool, without which making an advanced chip is simply impossible.
Just looking into this, I had no idea that ASML was the biggest company in the Netherlands and in many ways critical to their economy. Most people don’t realize that chip-making is that central to the economy of that country. From reading your book, the process of developing EUV started here in the United States at Intel, then got completely away from Intel. How did it end up that a Dutch company owns this piece of critical chip-making technology?
Listen to Decoder, a show hosted by The Verge’s Nilay Patel about big ideas — and other problems. Subscribe here!
The concept of lithography, which is the process of using light to create patterns on silicon wafers, was invented in the US in the late 1950s and deployed in the chip industry from the earliest days of the first semiconductors. The chip industry was founded in the US, in Texas and in Silicon Valley, so the early users of lithography were largely American firms. In the 1980s and 1990s, the industry was trying to move to a more advanced type of lithography called EUV (extreme ultraviolet) lithography, which is referring to the type of light that’s in these systems.
A lot of the research was funded by Intel and a couple of other US chip firms and was done in US national labs, which had the types of equipment and testing capabilities that were needed to actually make UV light at the requisite wavelength possible. But there was no US firm that could commercialize this equipment. Even though the science and technology was largely done in California. ASML was a company that already made older-generation lithography tools and had the capabilities to turn the science into a mass-manufactured device. That set ASML on the trajectory to where it is today: the only producer in the world of machine tools that can produce EUV light and use it to produce semiconductors.
Can you walk us through the basics of EUV lithography and how that makes chips?
So first off, what is lithography? If you want to make patterns on silicon wafers, you do so by shining light through masks. The masks will block light in certain areas and let it through in others, and that is how you get a pattern in a miniaturized version on a chip. Advanced chips today have millions, or often billions, of tiny circuits carved into them. They are often the size of a virus or even smaller, so you really need ultra-precise carbon capabilities.
EUV lithography uses light at a wavelength of 13.5 nanometers, an ultra-small light far smaller than the wavelength of visible light. You need really small wavelength light because the circuits you’re carving are very, very tiny; they themselves often measure just a couple of nanometers in dimension. Producing this type of light is really hard, because it’s right next to the X-ray spectrum. Production of it is complicated and the development of mirrors to reflect it is also very difficult.
A ball of tin is pulverized by incredibly powerful lasers and explodes into a plasma measuring several times hotter than the surface of the sun.
Here’s how the process works. A ball of tin falls at a rate of several hundred miles an hour through a vacuum and measures around 30 millionths of a meter in diameter. It is pulverized by two shots from one of the most powerful lasers ever deployed in a commercial device and explodes into a plasma measuring several times hotter than the surface of the sun — several hundred thousand degrees Fahrenheit. This plasma emits EUV light at exactly the right wavelength of 13.5 nanometers, which is then collected via a series of about a dozen mirrors, which themselves are the flattest mirrors humans have ever produced. The mirrors reflect the light at just the right angle so that it hits the silicon wafer and carves the circuits on the chips that make your iPhone possible.
That’s how you get to an A13 chip, right?
That’s right.
TSMC has to buy this machine from ASML, which has to assemble all these components from the flattest mirrors ever produced to the most powerful lasers ever deployed in a commercial setting to balls of tin. I imagine the balls of tin are somewhat easy to acquire. It has to make that machine, then it sells it to TSMC, which then uses it to make iPhone chips or whatever else. Does ASML just wash its hands of this machine when it sells it to TSMC? This sounds like a very complicated thing to operate.
It’s extraordinarily complicated. Just shipping the machine alone takes multiple 747s to move and they cost $150 million a piece. There are ASML staff on site next to the machine for the entire lifespan of these tools. ASML is the only company that knows how to service them when something goes wrong, and they are the only company with the spare parts in case something breaks. You just can’t operate them without ASML staff.
They’re so sophisticated and so precise that learning how to operate them in a mass production facility requires not only the semiconductor companies like TSMC to have done a lot of research into using them, but also a deep partnership with ASML, because they have really unique knowledge about how the optics work and how the light reflects and refracts in different contexts. You need to partner very, very deeply with ASML to understand how to actually use these machines in mass manufacturing.
It sounds like ASML has a monopoly on this fab equipment. Do they sell to other vendors? Can Intel buy these machines? Can other foundries? Can Samsung buy these machines?
Yes. ASML sells to customers all over the world — except in China, which we can discuss — but there are only a couple of companies that can really plausibly use an EUV machine. It’s TSMC, Samsung, Intel, and a couple of memory chip makers as well, like SK Hynix and Micron. There are very few other potential customers out there, because the price tag is so high and the level of precision manufacturing skill needed to actually make use of them is really so niche and unique that ASML knows it will only ever have a customer base measuring half a dozen or maybe at most a dozen firms.
Why doesn’t ASML just make the chips itself?
Well, ASML has no idea how to make chips. They’re an extraordinary firm, but one company can only do so much. This machine is just one of multiple ultra-complex machines needed to make chips. In addition to shining light at exactly the right wavelength through this really complicated optics, you also need different machines that can lay down thin films of material just a couple of atoms thick or etch canyons in the silicon just a couple of atoms wide. These machines are produced by different companies that have their own unique capabilities, about which ASML knows nothing. And these companies know little about lithography.
We really need a partnership of the toolmakers like ASML and the chip makers like TSMC to actually produce effective semiconductors.
The chip makers themselves also have unique capabilities. TSMC is better than anyone, including its suppliers, at using the machines to actually effectively make chips. We really need a partnership of all of these different firms, the toolmakers like ASML and the chip makers like TSMC, to actually produce effective semiconductors.
Yeah. People can argue about whether Milwaukee or DeWalt makes the best power tools, but they don’t make you a carpenter. Is that the vibe here? That you can buy the tool, but you have to actually know how to use it?
That’s absolutely right. Knowing how you use it is a process that not only requires starting with a PhD in electrical engineering or material science, but really requires years of working with the tools. The process of developing an EUV tool took 30 years. That just gives you a sense of the scale of precision that was needed to actually harness it.
I want to get to that, because that history with Intel is really interesting. We actually had Pat Gelsinger, the CEO of Intel, on the show recently and I asked him about EUV. We can get to his answer and why he thinks it was dumb that Intel didn’t win this race, but I want to come back to where we started, which is that President Biden is in the Netherlands and he is pressuring the prime minister to restrict exports of ASML devices to China. You said there are only a handful of companies in the world that can use these machines and none of the companies you named are Chinese companies. Why is it a concern that the Dutch government will allow exports to China?
If you have access to advanced chip-making tools, like those produced by ASML and the small number of other advanced tool makers in the world, you have a reasonably good shot at making advanced chips. Now, it’s still not guaranteed you can do it, but the tools are one of the key choke points that only a couple of companies can produce, and none of them are produced in China. The US is thinking about next-generation military and intelligence systems that will increasingly rely on artificial intelligence. AI systems are trained in vast data centers that are full of sophisticated chips like GPUs, which are the type of chips that are used to train AI systems.
If you can’t mass manufacture cutting-edge chips, then you can’t get the data center capacity that you need to train AI systems. The US is ultimately trying to accomplish stopping China from developing advanced data centers. It’s using the machine tools as the choke point, preventing US firms, and also Japanese and Dutch firms, from transferring this equipment to China.
There are a number of Chinese firms that have tried to become advanced enough to buy ASML’s EUV tools. SMIC, China’s leading foundry, is the best example of this. But the Dutch have been imposing controls on EUV exports for a couple of years, not letting ASML ship these tools to China. Now the US wants the Dutch to impose controls on a broader set of lithography tools — not only the most advanced, but also the second most advanced set of tools. That is something that is a new ask of the US government and it is requiring pretty extensive negotiations and discussions between the US and the Netherlands about whether this will be allowed or not.
What is the argument for and what is the argument against?
The argument for is that even the second-generation most advanced tools can be used to produce some pretty sophisticated chips, which is certainly true. The argument against is that it will be expensive for ASML and other companies to lose market share, because Chinese customers have been investing very heavily in chip-making capacity, subsidized very heavily by the Chinese government over the past decade.
For many chip toolmakers, China has become a really important market for non-cutting-edge tools. This will be a costly stop if the Netherlands implements the types of export controls that restrict not only the most cutting-edge, but also the second generation most cutting-edge tools. The cost will be in the billions of dollars in euros for leading toolmakers.
You’re describing an escalating regime of sanctions against China, in terms of making chips, in terms of technology transfer from United States companies to Chinese companies, and in terms of other international companies to Chinese companies. Is that all of a piece? Is that a strategy? Or is that, the Trump administration was mad at China, so they imposed sanctions against Huawei, and now the Biden administration has woken up to the chip shortage because of the pandemic? “We need a national chip foundry. Okay, it’s Intel. Do something in Ohio. Here’s some other stuff.” Is there coherence to all of these moves, or is it just reactive?
No. I think there really is a coherent strategy, and I would differentiate what’s happening now from Trump’s trade war on the tariffs. It’s really a separate track of discussions. I would also differentiate it from the semiconductor shortages. The shortages were not about which country is capable of producing the most advanced chips. What you find inside the national security bureaucracy, the NSC, and the intelligence agencies over the past seven or so years, is that there has been an increasing concern that China is making real advances in chip-making capabilities, just as it’s becoming clearer and clearer the ways in which advanced chip capabilities, and especially the types of chips that go into data centers, will be critical in training next-generation AI systems.
From the late Obama administration all the way up to the present, including the Trump administration, there has been a fair amount of coherence in terms of the policy with regard to restricting China’s access to advanced chip technologies. It’s something that has been done not only by the US, but also Japan, South Korea, and Taiwan. A number of different countries have taken steps to impose new investment screening mechanisms or to restrict technology or knowledge transfer to China when it comes to advanced semiconductors.
Does China have the ability to catch up on its own or does it actually need the technology transfer, the equipment that would otherwise be sold to them?
Well, this is the big question. The US strategy will succeed if China can’t catch up on its own. The US is betting that the answer is that China can’t catch up, or at least can’t catch up anytime soon. But there is some uncertainty around this. It’s hard to predict whether China will find ways to produce some of the necessary technology domestically, or if they will find ways to split apart the Western coalition and acquire some pieces of tech from countries that are unwilling to follow the US lead on export controls.
My best guess would be that the controls that the US and Japan are pretty clearly going to impose will be really problematic for China over the next couple of years, and potentially over the next 10 years or so in terms of making advanced semiconductors. The more countries that are on board with these controls — and that’s why the Dutch are so important — the more likely these controls are to work.
How did we even end up in a situation where there’s one Dutch company that we have to get a hold of in order to make sure China doesn’t gain these capabilities? Again, in a functional market, especially for something like chips, which are so important to everything, there would be multiple companies with multiple different approaches to making chips at the scale that modern chips are required to be made at the process nodes that we operate at now. Instead, there’s just one and it’s in the Netherlands. How did that happen?
Well, across the chip industry, over the last couple of years you’ll find that there has been a real trend to concentration, with in many cases just a handful and in some cases just one company capable of producing the types of software and machinery involved. There’s two reasons for that. One is that many parts of the chip-making process are just brutally capital-intensive. It’s extraordinarily expensive to make this machinery. That really disincentivizes competition, because a new entrant has to spend billions of dollars before they can see if their product even works.
The second reason is that the types of knowledge and expertise you need to produce these types of tools are something you can’t study in the abstract. You have to hone it over the course of your manufacturing. There is no amount of training or a PhD program that is going to let you understand how these systems work when they’re actually manufactured. You have to have your hands in the machine tweaking it over time.
That means that people who are working on these tools in companies have really unique knowledge that is hard for anyone else to acquire. That provides a really strong moat around those companies, because there is no straightforward way for anyone who is not working at these companies to develop the requisite knowledge. So the combination of capital intensivity plus really unique knowledge makes it very difficult to set up any sort of competing firms.
So I’m going to quote Pat Gelsinger from when he was on Decoder a few months ago. I asked him about EUV, because Intel famously bet against EUV, and now they’re going to buy machines from ASML to put in Ohio to try to build next-generation chips. I asked, “What happened with EUV?” He said, “We were betting against it. We took a lot of risk in Intel when we were like, ‘Hey, we don’t need EUV. We’ll go to advanced quad patterning of the lithography.’ We were doing other things to avoid needing EUV, and those things just weren’t panning out. At a minimum, we should have had a parallel program on EUV that said, ‘If we get this wrong, if we get quad patterning or other techniques that we’re doing in self-aligning wrong, we should have had a program for that.’ We didn’t. We were betting against it. How stupid could we be?”
Now in hindsight, the answer is that they were extremely stupid. He’s the new guy and he is fixing it, so I think that’s why he’s allowed to say that they were being stupid. Were they actually stupid in that moment? Was it correct to say, “Oh, you can bet against EUV and maybe something else will pan out?” Or is that just Intel being run by a series of accountants instead of engineers, and now they have an engineer?
Well, in defense of the accountants, I think you could say the following…
I think that’s the first time anyone has ever said “in defense of accountants,” but by all means.
So EUV was a technology that was supposed to be production-ready a decade before it actually emerged. The development process had repeated delays, billion-dollar cost overruns, and for a long time in the late 2000s and early 2010s, it seemed like something that might completely fail. Up until 2015, it was really very uncertain whether or not this would ever work, and if it did work, whether it would be remotely cost-competitive.
In that context of uncertainty, you can understand why there are people at Intel who wanted to bet against the EUV and instead bet on what they called quad patterning, which means using existing lithography machines that everyone knows works and doing more lithography runs to carve ever more precise circuits. That was obviously going to be more expensive than doing fewer runs of lithography, because you have more steps, but everyone knew the machines worked.
Trying that was the low-risk option in some ways. In hindsight though, it didn’t work. It was a terrible bet, but you can understand why they took that bet. To point some blame back at the accountants, I think there was a bit too much risk aversion and an unwillingness to, as Gelsinger said, prepare multiple pathways for the R&D to see which one worked. There was probably a bit of cost-cutting that seemed wise at the time in terms of spending resources efficiently but in hindsight had a tremendous cost for Intel, because it left them unprepared when quad patterning turned out to be inefficient and in some cases completely incapable of producing the precision that Intel needed.
Is that why Intel has been delayed behind TSMC at every successive process node? Because their techniques just weren’t working out?
That’s part of the reason. I think it’s a complicated answer, but certainly the delay in EUV is an important part.
Just walk me through that process. They have a big competitor at TSMC that is obviously embedded with ASML. That’s very good, because its business model is just manufacturing chips; all of its energy is focused on manufacturing chips. Intel is at a place now that I asked Pat about. I said, “Do you think of yourself as a national champion in the United States? You’re it. You’re what the Biden administration has at this moment in terms of a large chip-making concern that can insulate the United States from the global supply chain.”
He was like, “Yeah, I don’t know about that,” but he has to step it up and become a foundry. That’s what he wants to do. He said he would put an AMD logo on the side of an Intel fab if AMD wants to manufacture chips there, but he has to go and buy an EUV machine from ASML. He has to learn how to make ARM chips, all while developing the next generation of Intel’s own x86 chips. Is that possible? Does that seem like we’re putting too much pressure on this one company?
Well, I think there’s no doubt he has a challenging job. Now, he’s obviously an impressive guy. If anyone can do it, he can do it.
No lack of confidence from Pat.
Well, I think there’s no doubt that he has turned the culture around at Intel, but I think you’re right to outline the challenges that Intel faces on the process technology manufacturing side, on the design side, and on the business model side with creating this new foundry business. It’s going to be hard.
“I think a lot for the United States depends on whether or not Intel succeeds.”
I think tackling each of these three challenges simultaneously is the only choice he has, because Intel has to deal with all of them. But you’re right to say that it’s a tall order ahead of him. I think a lot for the United States depends on whether or not Intel succeeds.
The reason I ask about Intel specifically is because it is the only choice. There’s not another scaled American producer of chips. There is a little bit of TSMC activity here and a little bit of Samsung activity here, but it’s not their leading-edge process nodes. TSMC hasn’t even really started at scale yet. Intel is what we have. It feels like if they were much more successful, then the national security conversation, the supply chain conversation, and the export control conversation might be very different. But because Intel is where it’s at in this moment of pretty dramatic transformation, it has downstream impacts on how we’re dealing with China.
I think that’s right. I think of the three leading firms in producing processor chips, Intel is the one that would naturally invest in the US, because it is home-based in the US. I think it’s also worth noting, though, that Intel doesn’t have an existing foundry business in the US. In terms of building up foundry capacity in the US, we’re starting from a pretty low base across the board. Samsung has a facility and GlobalFoundries has facilities, which are not the most leading-edge, but have some impressive capabilities and meaningful scale as well.
But in terms of building up scale and foundry, everyone is starting from a pretty basic starting point. In some ways, that is why the US is probably not going to end up betting solely on one firm, but on all three — Intel, TSMC, and Samsung — to try and get all of them to invest more in the US, and see which one is able to develop the biggest facilities, the most functional business model in the US, and who wins the race.
That’s a very American way of doing it, right? It’s like, “We’re going to subsidize the creation of a market, and then whoever wins, wins.” Don’t you kind of just end up with someone who is going to win the race, and then we have another weird monopoly in the United States? Is there any thought to, “actually, what you need is diversification at every level of the supply chain”?
Well, the challenge is that diversification is very expensive. If you want to pay for additional capacity that you’re not going to use in the chip industry, you’re going to spend a ton of money. Just a single new cutting-edge chip-making facility costs $20 to $25 billion, and it’s cutting-edge for just a couple of years. I don’t think there’s really a lot of appetite in the US for undertaking the tremendous capital expenditure that you need to build surplus capacity. At the margin, we’re going to get some through the CHIPS Act, but we’re not going to get a ton of surplus capacity.
We need companies that are going to have functional business models after we help them get off the ground with their foundry businesses in the US. That’s why I think it does make sense to bet on multiple companies and see which ones are able to produce that.
I don’t think it’s necessarily the case that we’re going to end up with one company winning and the others losing. It could well be that we end up with multiple commercially viable foundries with capacity in the US, and that would be a great outcome. There’s no necessary reason why this particular market has to end up with one dominant firm and others far behind.
Is there any thought to investment beyond the foundry players? “Hey, we should fund a competitor to ASML,” or, “Hey, we should look for the next technology beyond EUV and have the government subsidize it so that we can diversify that layer”? Again, as we are having this conversation, President Biden is walking around in clogs being like, “Please don’t sell this machine to China.”
Well, when it comes to, “Should we have a competitor to ASML and Advanced Lithography?” I think the answer is that the cost-benefit there just doesn’t play out. We’re likely to get the Dutch implementing controls that are pretty similar to US controls. Over the next couple of months, I think we’ll see that be the result of these conversations.
There’s not much supply risk with the Netherlands; nobody is worried about the Netherlands not shipping to US firms. The cost of setting up an alternative lithography firm would be very expensive, because, again, ASML has unique capabilities that they have built up over 30 years. It would be very hard to replicate that or to build a competitor. So I think our production and our R&D dollars are much better spent elsewhere.
But when you talk about next-generation lithography and next-generation tools, that’s a great place to invest. If you look at the way the Commerce Department is planning to spend the CHIPS Act money, what you’ll find is that three-quarters of the funds are going to go to incentivizing more manufacturing, and an additional 25 percent is going to go to funding R&D. Part of that will go to the next-generation tools, including potentially next-generation lithography systems, which will be needed in five or 10 years.
What are those next-generation lithography systems?
Well, ASML itself is planning two further generations of EUV tools. Right now they have basic EUV.
Wait, the thing you described with the ball of tin falling down, being hit with lasers, and producing plasma hotter than the sun, that’s basic EUV?
That’s basic, yes. More is to come. The next generation is going to be called high numerical aperture EUV, which is going to have more sophisticated optics that will let you carve more precise chips. Those machines are supposed to be available in three or so years. They’re going to cost twice as much as the basic EUV tools.
Then beyond that, there’s R&D underway for what ASML calls hyper numerical aperture, so even more specific optics, which is unclear if it’ll work. It’s a decade away from production, but that is where R&D is already happening, and that is what we need if we’re going to make smaller and smaller transistors on more and more sophisticated chips.
You have a chapter about smaller and smaller transistors in the book. I’m sure Decoder listeners are very familiar with the concept of Moore’s Law, which is just a prediction that the chip industry will double the density of transistors on a chip every year. We are already talking about having to fire lasers at a ball of tin through the world’s flattest mirrors and then building hyper-specific optics to make them even smaller. Is there a limit? I feel dumb being like, “Is there a limit to Moore’s Law?” Anyone could have said that at any time in the past 40 years and been proven a fool, but we are now talking on the level of atoms. Is there a limit where the chip industry is like, “Okay, we’re not going to get past the level of individual atoms”?
At some point, the answer is yes, but we’re not talking about individual atoms yet. We’re talking about layers of materials measured in individual atoms, but transistors themselves have lots of atoms in them, even at their current microscopic scale. I think we have a pretty clear line of sight — at least through 2030 or so — in the existing plans of firms like TSMC and Intel as to how they’re going to keep shrinking transistors, stacking them on top of each other, and using more tricks to get more of them on chips. It’s harder to say beyond 2030 or so. It has always been hard to look too far into the future, so I don’t know how meaningful that is.
History says the appropriate prediction is that they’ll figure it out. I mean, that is why it’s called Moore’s Law.
That’s right.
At some point you’re running out of things to even measure; you’re running out of units. Like, Intel had to move to angstroms instead of nanometers. Maybe that was some branding, but they definitely did it.
The reason I ask in this context is because we’re talking about limiting advanced chip-making equipment to China and we’re talking about the next generation of GPUs or other AI acceleration chips. 2030 is tomorrow on the scale of industrial policy and foreign policy. Even in 2040, that is still kind of tomorrow on the scale of industrial policy. Isn’t it inevitable that China will catch up, even if they’re restricted in all these ways? Are we just buying time, or are we actually creating a durable sustaining advantage?
Well, I think it’s inevitable that China catches up to the current status quo at some point. Whether that’s in 2027 or 2035, I don’t know. It’s not going to be in 2024, it’s multiple years out. Will China ever catch up to the cutting edge? I’m not sure that the answer to that question is yes, even if Moore’s Law…
Even if Moore’s Law expires, right? If the cutting edge becomes fixed.
Well, it depends on what we mean when we say, “Moore’s Law expires.” At some point, it will be impossible to shrink transistors further, but that doesn’t necessarily mean that the computing power you can get out of an individual chip will necessarily come to a halt. You can package them in different ways, you can put memory closer to the processing power, you can improve your interconnects, you can put photonics on a chip. There are a lot of different techniques that in many cases are just in their infancy, that are creating new ways to get more computing power out of chips, and all those will require both creative design and really precise machine tools to make.
Even if you were to tell me transistors won’t shrink by a single nanometer after 2030, I would still say that we’re going to get more computing power out of a square inch of silicon well throughout the 2030s and beyond using all of these other techniques. So if you take this most expansive view of Moore’s Law and say all the different things you can do to tweak a chip to get more out of it, I think there’s a really long runway that goes well beyond 2040 in terms of the things we can do to produce more computing power. For that reason, I would be pretty skeptical of the thesis that we’re ever going to hit a brick wall.
When I think about the three companies that are doing the best at pushing computing power forward for a given chip, it’s Apple, Nvidia, to some extent AMD — so maybe it’s three-and-a-half companies — and it’s TSMC, which is the manufacturer for those two-and-a-half companies there. Apple is really good at packaging, really good at optimizing their software for their own hardware, and really good at pushing the limits of ARM. Nvidia is obviously the leader in GPUs. AMD is less of a jump over the average Intel chip, but is doing better just because they’re using TSMC’s manufacturing capabilities to improve their battery-life-to-performance ratio.
I look at those three-and-a-half companies, and I think, “Okay, what they are dependent on is TSMC.” If TSMC was not able to push forward manufacturing, their techniques for building and designing better chips would actually be for naught, right? They are wholly dependent on TSMC, which is in turn wholly dependent on ASML. What is that relationship like? Does Tim Cook wake up in the morning worried about Dutch restrictions on ASML exports? Or is he a TSMC customer? Or is that just an API where he places an order and the chips come out?
I think most of TSMC’s customers have gotten used to the fact that TSMC has an extraordinary track record in managing their own supply chain and making sure that problems are solved before they actually happen. One of the reasons customers love working with TSMC is that they don’t have to wake up in the morning worried about what’s going to happen upstream of TSMC’s production.
Are people thinking more about the upstream supply chain than they ever have before because of these restrictions? Absolutely. But if you’re Apple or Nvidia though, they don’t really apply to you, because all of the inputs that TSMC relies on for its production are produced in the US, Europe, or Japan. There’s no chance that those countries are going to control their transfer to Taiwan anytime soon, so you’re actually pretty secure in terms of your upstream. It’s your downstream — the assembly of chips that is often taking place in China and the assembling of the final goods — where you have the most political risk, both in terms of what Beijing does and also in terms of what Washington does.
Well, there’s real political risk because the T in TSMC is Taiwan. The fabs are in Taiwan, especially the leading-edge fabs. If there is strangeness between the United States, China, and Taiwan, the iPhone economy grinds to a halt, right? The chips just go away, the Nvidia GPU economy grinds to halt. Is that something that we should be more worried about?
If there were a war or a blockade between China and Taiwan, the impact on not just the tech sector but all of manufacturing would be close to catastrophic.
Yes. In terms of a war or a blockade between China and Taiwan, the impact on not just the tech sector but all of manufacturing would be close to catastrophic. TSMC produces 90 percent of the world’s most advanced processors, but more than that, it produces over one-third of the new computing power the world adds each year. If you add up all the transistors produced on processor chips, over one-third of them are produced in Taiwan. It certainly would be catastrophic to Apple or to AMD if we were to lose access to TSMC’s facilities, but it’s also dishwashers, microwaves, and autos that would face tremendous disruptions.
I mean, we would be back to a manufacturing crisis that would feel equivalent to 1929 in terms of its shock. You wouldn’t be able to buy a car for a year or two, and the same level of disruption for all sorts of manufactured goods. It’s a huge problem. It’s such a big problem that companies really struggle to get their heads around how to insure against it, because the costs of finding alternative solutions are tremendous. There is no other alternative to TSMC in many cases, but obviously the downside risk is substantial and arguably growing every day.
How did we end up in a position where the world’s most important chip fabs are in Taiwan? How do we end up with TSMC?
TSMC emerged in 1987 thanks to Morris Chang, who is the founder of the company. He had actually spent his career at Texas Instruments and lived in Texas for most of his life before that point. He was the person who had the visionary idea to create a foundry business that didn’t design any chips, only manufactured them, which at the time seemed like a crazy concept since there were no fabless chip design firms. He had no customers when he started, but he began convincing companies that he would do all the manufacturing for them and take all the production risk; all they had to do was give him chip designs and he would return functional chips.
That model proved extraordinarily successful because it let TSMC scale by serving lots of different customers. That scale in turn let TSMC hone its production processes, because the more chips they produce, the more they learn from the process of actually manufacturing each chip. There’s a direct relationship between the fact that TSMC is both the world’s largest chip maker and the world’s most advanced, and both of those stemmed from the foundry model that Morris Chang invented.
You said to us before we started recording that Morris Chang is your favorite character in this entire book. There’s a line in the book where you say he is arguably more Texan than Taiwanese. Why is he your favorite character?
Well, I think he’s the most underrated business person in the last 100 years. Most people have never heard of him, even though we all rely on products that his company produces every single day. I think his life is a fascinating microcosm of the chip industry as a whole. He was born in mainland China, moved to the US right after the revolution, enrolled in Harvard and was the only Chinese-American student in his class, and then personally built the chip industry working on production lines at Texas Instruments before founding TSMC.
All of the big shifts in the chip industry and computing technology over the last 75 years are shifts that not only did he illustrate, but he actually made them happen. We all owe a lot to Morris Chang, and I wish more people had heard of him, because I think his importance is really profoundly underrated.
I think TSMC is profoundly underrated. It is a very opaque company. They’re very proud of themselves and they’re very opaque. It’s hard to know how they work. What is your sense of TSMC? I mean, Morris Chang isn’t there anymore. How does the culture persist? What are its new leaders like?
Well, Morris Chang is officially retired, but he regularly shows up in TSMC’s offices and at TSMC’s events, so I’m not sure if we should really say that he’s no longer there. I think the culture he put in place endures in terms of the willingness to take big bets in terms of R&D decisions, in terms of capital expenditure decisions, and in terms of the relentlessness with which TSMC hones its manufacturing processes.
Morris Chang’s coworkers from the 1950s would talk about the ferocity with which he would find inefficiencies in the manufacturing process and then push them out of the assembly lines as fast as he could. I think that commitment to manufacturing excellence is what made TSMC what it is today, and that stems in no small part from Morris Chang and the culture that he instilled.
Why put it in such a perilous geopolitical region? When I think about TSMC now, they have to employ as many foreign policy experts and as many lobbyists as they do people who are relentlessly focused on manufacturing inefficiency, just because of their location more than anything else. Why choose Taiwan?
Well, in hindsight, it would have been great if they had established it in New Zealand or Switzerland, but Morris Chang had spent some time in Taiwan as a Texas Instruments executive. He had helped TI set up a facility there in the late 1960s, so he had gotten to know some of the Taiwanese government officials.
They wanted more US investment and to be more plugged into US supply chains as a way of guaranteeing their security. They bet that integration was the best way to ensure that the US would help defend Taiwan. Today, we’re seeing the fruits of that strategy born out. Taiwan’s important not only because of its geopolitical significance, but also because if there were a war, it would be catastrophic for the world’s tech sector and manufacturing.
Just to state that clearly, you’re saying the decision to put it in Taiwan was incentivized by the Taiwanese government in order to guarantee the United States’ defense support?
TSMC was a direct project of the Taiwanese government to make Taiwan more indispensable in electronic supply chains. And it has worked.
That’s right. There was a direct linkage in the Taiwanese government’s mind between more US investment, more criticality in US supply chains, and more credible US security guarantees. That’s why the Taiwanese government put up over half the capital in TSMC when it was founded. It was a direct project of the Taiwanese government to make Taiwan more indispensable in electronic supply chains. And it has worked.
Clearly, it has worked. I mean, we should talk about Russia and Ukraine. I think there is a parallel there to potential conflicts with China and Taiwan. But I just want to finish the thought here. You say the Taiwanese government was like, “We need this. We’re going to spend the money on having a chip industry.” Earlier, you said that to build the next generation of chip-manufacturing companies, you need an enormous amount of capital, long-term vision, and to subsidize a bunch of stuff. That’s what the Chinese government does. They will just happily build oversupply. “Here are 95 bridges, an economy will arrive here one day.”
The United States does not do that. We’re horrible at that, at almost every level. We’re successful maybe despite it — I think there are some people who will tell you that we’re successful because of it — but we don’t do that. Is that what is absolutely necessary, for the United States to say, “We are going to build the chip-making industry. It has strategic purposes that we’ll realize decades from now, the way the Taiwanese government did, and we’re picking a handful of firms to become national extensions of an industry that we think is strategically important for years to come.”
If you want to attract chip-making firms to your country, you have to make it cost-competitive. For the US, cost has been higher for a variety of reasons. Land is more expensive, environmental regulations are more strict, and the tax regime is less generous. If the US wants more chip-making, it has to spend the money to make it more attractive for chip firms to invest.
I think capital expenditure is necessary, but it’s not sufficient, because in addition to the CapEx, you need the expertise. That is something that Taiwan realized very early. It wasn’t just that they put forth a lot of money, they also made sure that hundreds of Taiwanese engineers were doing PhDs in electrical engineering at Berkeley and Stanford from the 1950s.
Although Taiwan seems a long way from Silicon Valley, there were in fact very few places in the world that had so many deep personal connections to Silicon Valley as Taiwan. When Morris Chang moved to Taiwan in the 1980s, he found former classmates he had studied with at Stanford and former colleagues he had worked with in the US who were there in Taiwan working in the chip industry. That deep interconnection was absolutely critical to Taiwan’s success.
If you look at other companies that have caught up in chip-making, Samsung for example, it’s a similar story of a lot of CapEx, but also a lot of integration into supply chains. The challenge that China faces today is that there is no doubt that it has the willingness to spend money, but it’s being cut out of the exchanges of information, components, and expertise that have made catch-up possible in Taiwan and South Korea. That’s the big risk that China faces, but that’s also the explicit US strategy, to cut them out of those relationships and therefore make it harder to catch up.
Just to wrap this up, make the comparison to Russia and Ukraine for me. I know this is also an area of your expertise. We have seen this now. “Even though it’s always on the precipice, no one would ever actually invade Ukraine,” but then they just did it, and now it’s a disaster for them on many levels. We kind of feel the same way about Taiwan. This would be a disaster on many levels. Is there a chance that China is looking at Russia and Ukraine and saying, “We could probably do that too”?
Well, I think on the one hand you would say the Russians overestimated their military capabilities, so surely the Chinese are now wondering if they’re overestimating their capabilities too. That’s one lesson you could draw, but there are other lessons that China might be drawing that are less reassuring. One example is that nuclear weapons work, you can threaten nuclear use and keep outside powers from intervening. The Russians showed that very clearly, because Biden has made it very explicit that he’s not going to do anything close to what would trigger Russia to escalate the war. So nuclear threats work. That’s a lesson that is directly relevant to a China-Taiwan scenario.
A second lesson is that in Ukraine, it has been absolutely critical that Ukraine borders Poland, because you can very straightforwardly ship arms, equipment, and supplies over a land border to a neutral country. Taiwan has no such land border, so you need to find ships that would go into Taiwan to resupply them. That’s a very, very different proposition — not impossible, but very difficult.
If you’re China looking at Ukraine, on the one hand, you think the Russians screwed things up militarily. On the other hand, it’s not obvious that a lot of the other lessons you are drawing might not make you a bit more optimistic that you could keep the US out in a meaningful way. I’m not confident that China is looking at Russia and Ukraine and thinking, “This makes us less confident about Taiwan.” I worry that some of the key lessons that China is drawing are making it more capable of structuring an intervention in Taiwan that would keep the US out.
If everyone else knew that the US wasn’t going to intervene on Taiwan’s behalf, it would be very hard for Taiwan to put up a serious defense. That’s why I think you see China rapidly expanding its nuclear forces, to make those nuclear threats more credible and to try to keep the US out if there were a crisis in the Taiwan straits. I remain quite worried, despite the fact that Russia’s military has profoundly underperformed in Ukraine.
I feel like we could do an entire other episode on Russia alone. We’ll have to have you back. Chris, you have been great. The book is Chip War: The Fight for the World’s Most Critical Technology. Thank you so much for being on the show.
Thanks for having me.
