I am moving the Asianometry Newsletter to Passport and adding it to the Stratechery Plus bundle. Released video scripts and the audio feed will be accessible to Stratechery Plus subscribers (and we have a trial offer for you below). YouTube videos will remain free, as they always have been.

As you know, video transcripts were posted somewhat sporadically. Moving forward transcripts and podcasts will be first class citizens: they will be available the same day as my public YouTube videos for Stratechery Plus subscribers.

To manage your account, including adding the podcast feed to your podcast player, and to manage your email settings, go here (link). You can use the same email that you used to subscribe to this Substack.

To be clear, this is in addition to my YouTube channel. That is the core of my work, and if you want early access you can still subscribe to my Patreon. Nothing will change in that regard, other than the fact this is officially now my full-time job.

I have also been working with Ben to make YouTube videos for his public Articles. You can see those videos here.

I am a big fan of Ben Thompson and Stratechery and have long wanted to work with him. We’ve also arranged a one month trial to Stratechery Plus for all of you:

• You can opt in to receive Stratechery emails or add the Stratechery podcast here

• You also have access to Sharp Tech, Sharp China, Dithering, and Greatest of All Talk; click any of those links to add the podcasts to your podcast player

This one month trial is completely free, and you don’t need a credit card, but if you do decide to join the entire bundle is $15/month (including the Asianometry podcast and transcripts). You can of course cancel at any time.

This Substack and its URL will remain open, but it’ll be mostly just for directing you guys to the new Asianometry site. Go there for the latest work.

This is a big transition and there will be some bumps along the way. Hope you guys can bear with us as we make this update, but I think it’s going to be great. And, to be clear, if you just want to watch the videos, those are and will remain free forever.

Starting a YouTube Channel

Jon Yu, welcome to Stratechery.

JY: Hello. Glad to be here.

There's going to be a nice long intro here before this interview, introducing you and your work to the Stratechery readers. Not just your work, but also your work for Stratechery, which we can get to in a moment. But I always like to start these interviews by learning more about the person themself. Where'd you grow up? What inspired you to start a YouTube channel focused on technology in Asia? Was that always a focus? Or did you sort of stumble onto this?

JY: It was always the focus. I grew up in Southern California, and then worked in the Valley for about 10 years after college. I was just like an ordinary guy on the street, and then kind of burnt out and thought to myself, "You know what? I should go to Asia" — I didn't even know, I just called it Asia.

Yup. I wrote an Article a few weeks ago about building things. There's different aspects, some are China-specific, some are in Taiwan or Indonesia and I'm like, "You know what? I'm just going to use Asia, and we're going to roll with it". It's way too broad, but it works sometimes.

JY: I know, right? I would go to these cities, and I had no idea necessarily the difference between these cities. I just picked one and then jumped, ended up — Taiwan gave me the best offer work-wise, so flew over there thinking I would be there for six months, and that was eight years ago. So, I ended up staying for a while.

I think about a few months in, I went on a trip to Japan with my mother, and my mother asked me, "I don't really know what you're doing over there in Asia, in Taiwan, and you should share that with us". I was like, "Okay, mom, I'll open a YouTube channel for you". Asianometry started out as a tourism channel, where basically I would go to some — like the Daxi Statue park, and I would film the video, and I would say, "Mom, this is this statue and that's that statue", and eventually you just run out of things to film, to visit.

If you accrue travel costs while making videos, it gets much more expensive.

JY: Oh, immensely expensive. The return on invested cost is already terrible on a YouTube video, and then suddenly, we're spending hundreds of dollars to travel.

Anyways, you run out of things to talk about, and I still had the channel, so I figured, "Hey, I'll make videos about Asia". I started out only about China and Taiwan, but then the comments, YouTube comments were just so brutal. They're just like, "This channel should be called Chinanometry, not Asianometry." That forced me, I was forced to export my research to other countries.

Well, I mean, you're pretty well-known, I think, amongst — some of my readers know who you are in Silicon Valley, particularly for your work on semiconductors and things along those lines. Was that just a, "This is something that Asia is really good at", you sort of stumbled into it?

JY: Well, half and half, I think. If you're in Asia, in Taiwan in particular, it's really hard to avoid, to not notice TSMC so I covered TSMC pretty early on. But then also secondarily, my father is an analog chip designer, and occasionally, we'll talk about the work that he did and the work that he does for a while. So, it was in the blood but in the environment too.

Yeah. Well, I was going to ask you what's been the most surprising part of your YouTube experience, but I think the fact that you're sitting here and you have hundreds of thousands of views on your videos these days about chips, when you started out doing a sort of tourism channel for your mom, that might be the answer. But beyond that, what has been the most interesting thing about being a YouTuber?

JY: You get really into the dynamics of the algorithm, you get an understanding of the shortfalls of the algorithm, you can grow really fast. But people, it's kind of a dog-eat-dog world out there. I've learned a lot from you, especially in terms of consistency and continuing to bring something to the table every week or every few days. So, that's something that's really surprised me, and I think that's also just surprised that I can keep going at this pace. It's been eight years now, so it's a while.

Yeah. There's a lot I could relate to. I was only going to come to Taiwan for a year, and here we are both sitting here. And yeah, the job I've had longer than anything else, and it's like if you would've told me in 2013, "Are you still coming me up with stuff to write about?", at the beginning, you're like, "Wait, once I cover what I want to write about, then what?", turns out the technology sector is a good foundation for always having new things to explore and explain.

JY: Yeah. But I do feel sometimes I go back to my older videos, I would go to some beautiful topic and I look to myself, and I'd be like, "Man, I wish I could do that again with what I know now", it would probably be three times longer, but it would be beautiful.

Well, that's kind of what I want to do in this interview, because I write a lot about semiconductors, and obviously TSMC. I was going through, number one, everything I want to ask you about, you already have a video about, so we'll have lots of links throughout this interview. But at the same time, I'm like, some of these videos, like your current videos have, like I said, hundreds of thousands of views. Some of these ones that are really pertinent have 50,000 views. I'm like, "Wait, you guys have to go back and go through this library". I don't know, maybe you have an excuse to remake some of them.

JY: Yeah, that would be fun, but there's always new ideas coming up. One thing I'm surprised me, I guess to go back to your question, there's always more ideas. My list is hundreds of topics long, I'm going insane every day I add more.

Has YouTube changed? Is there a bit where chasing the algorithm and that changes, and you mentioned the comments being brutal and bullying you into covering more things. Is that a consistency of the YouTube experience, or has it changed in the eight years you've been doing it?

JY: I think early on, the YouTube channel, it was much easier to just go big with certain topics. I think there's still certain keywords or things or topics, strategies that sort of work. For example, calling out a country still works to some extent. Analyzing deeply something that no one else has looked at, good to some extent. I think that's still tried and true, and I still like to do that. I noticed that I don't do as well if I make an Nvidia video nowadays-

Right. There's a lot of people making Nvidia videos.

JY: Because everyone makes videos of those.

Yup.

JY: But no one else is going to do a video about the [Japanese whisky]. Anyone wants to search that, I'm there. In some ways, it's still the same, but in some ways, it's just more challenging because there are a lot of more people.

Why Chips Are Made of Silicon

Yeah, that makes sense. Like I said, what I think would be interesting is go through — I've noticed as people become more interested in semiconductors, you start out with a very reductive view. It's like, "Well, Nvidia is the most valuable company in the world, oh, TSMC makes video chips, they must be the most valuable. Oh, TSMC depends on the ASML, I've heard that name, that must be the most valuable". I think it'd be interesting to explore and lay out the overall process and market constraints around this that drives differentiation. So I was thinking about, "Okay, where to start?", I want to talk through the semiconductor process. We could start with silicon and the Czochralski process. Let's do this. Sand. How does sand become a wafer? Let's start at the very beginning. And like I said, let's talk about it, and the historical aspect, we can sort of dip in and out on each one.

JY: Yeah. I think if you want to go back to it, a lot of people will start with the transistor, the first transistor, which is a solid state switch essentially. Now, switches existed within history to build electronics, and people have been building electronics since, I would say, the 1800s.

I would say the thing is that the solid state transistor was a interesting thing because it used a semiconductor material, which is at the time I believe it was germanium or something. The germanium one didn't quite work out in the market because it had difficulties at higher temperatures and higher frequencies, and that's why silicon becomes more prominent in that feature.

Now, silicon is not the best semiconductor. It's not perfect for a lot of different things for transporting charge carriers, like electrons or electron holes, but what it is good is that it scales, and it has good — its derivatives are really good for protecting the transistor from outside contamination. I think that helped the industry latch onto it as a process. You would see these introductions of various historical events, for example, like the planar process and Fairchild Semiconductor, that would help develop silicon as this core product within the industry.

So at the very heart of it, at the very beginning, you create the silicon, and you turn it into a melt and you can use as melt the dip using a sea crystal to build what is called a boule and it's this massive thick thing.

The long cylinder of silicon, yup.

JY: Yeah. And that's a special part of silicon that helps make it special in the industry. For example, there are other materials, like silicon carbide. Silicon carbide doesn't necessarily turn into a boule, you can't dip it, you can't use the Czochralski process to create this massive single crystal that basically you can chop up and turn into wafers.

Right. And just for context, silicon carbide, that's the material that Meta is using for the lenses in their AR glasses and actually, what you just said is one of their biggest challenges in bringing these AR glasses to market, which is it's really hard to make silicon carbide. It doesn't have this reproducible way that traditional silicon does where you can make these long crystals that are easily sliced up into wafers, and they're like, "Can other people use silicon carbide, so people can figure out a good process here?", that's one of their big challenges.

JY: I know, right? And I think it reminds you of the fact that silicon is sort of a miracle material and that's why it's been so consistently used throughout history. It really has all these special properties that help make it a unique material for all these scalable processes.

No, it's a good point. This applies to electronics generally. You might have something that is, in a narrow technical sense, not the best, but manufacturability is really important and durability is really important. Those are some of the things that really, really set silicon apart.

JY: Yeah, I agree.

Fabricating a Chip

Well, let's go through this. You have a set of companies, they make the wafers, they deliver them to TSMC. This is just a piece of, it's not glass, but it's a piece of silicon. How is the transistor actually created and put on that? What are the steps? What companies are involved in that? I just want to get deeper into this actual process for people who, particularly people who don't really know how this works.

JY: I think what they'll do is that they'll, at first — silicon manufacturing is basically broken down into a bunch of repeated processes. It's a bunch of recipes put together by TSMC. TSMC is the integrator of all this equipment, and the equipment comes from various different companies like you mentioned. And I found it really helpful to break it down into a framework of what those steps would be. You have deposition.

What is deposition?

JY: I do love this about the semiconductor industry, at least the terms make sense. It's about laying down or depositing a thin layer of material. Now, what that material layer is going to be is then specified in the type of deposition. These tools are provided usually by Applied Materials or Tokyo Electron, and what they do is they have these certain sub-sectors of different deposition techniques. You would have, for example, oxidation, which is where you take advantage of silicon's ability to grow and oxide by reacting it with water to create an oxide layer to protect the transistor or whatever is on the silicon from outside contamination, now that's deposition. And then you would move that to — you would pattern a layer, whatever that layer would be with lithography.

Got it. And this is the one that most people know about, which is where ASML enters the game.

JY: Yeah. But then there's also an underrated product that I should mention is photoresist. Photoresist is provided by company JSR, which is one of the dominant photoresists of our current generation. In fact, Japanese companies almost have, they have a monopoly on photoresist, a complete monopoly I think for 20 years.

And this photoresist, that's another layer that's sort of put on before the lithography?

JY: It's a liquid that you pour onto the wafer and then it absorbs the light in a way to preserve the pattern from the light because you have to preserve it, you can't just dump it on the silicon. If you do it on the photoresist and the photoresist is baked to resist, ergo the name, an acid that comes later, which is the etch layer or step, which we'll talk later.

Got it. This Japanese angle is interesting. And actually the other thing I'm getting from you and I think is underrated is a lot this stuff isn't just equipment, it's a lot of chemicals and it's a lot of material science is probably the more underrated IP aspect, which is very, very well hidden and sort of preserved to your point about these companies retaining very dominant positions.

JY: Yeah. TSMC splits their supply chain into equipment and materials and I've been told the materials is basically almost all Japanese. The Japanese have a very, very tight grip on materials, photoresist, all these different things for reasons that we can go into later, but it's part cultural part economics and stuff like that.

Okay. How does the lithography enter the equation then?

JY: The lithography is, basically the goal is to transfer the pattern of the design or the pattern where the transistors would be, how they'll actually look, and transfer that onto the wafer, onto the deposited layer that you have here and that's one half of another step called etch. They used to call it litho-etch-litho-etch. Litho is you put the pattern down, etch is that you basically solidify it into the layer. Could be silicon, could be another metal layer, could be different things. That's why I think etch is very underrated, but I think litho is probably — it consumes a lot of the value chain, but it probably shouldn't consume so much of the attention.

Etch is like Lam Research. Are they the dominant player there?

JY: Generally it's a bunch of different players because etch is very varied, but yes, Lam is one of the big ones.

I think you actually probably gave the answer. Etch might be really hard, but there's more competitors in the space and especially once we moved to DUV, there was, again, the Japanese dominated that for a long time. ASML was more behind and came up, but then you got to EUV and that's where ASML leaped ahead sort of in conjunction with TSMC, it was really a hand-in-hand thing there.

JY: Well, I think it's important to note that also DUV, one of the breakthrough products of ASML was during the DUV days and their big innovation within the DUV space was the focus on productivity. ASML really focused on not necessarily making the most incredible litho machine, but making a machine that was very productive, which was their TWINSCAN thing where they would be able to process more wafers per hour than even the Japanese could do and that really caught them off guard, especially in the early 2000s.

And this really happened with the shift from 200 millimeter wafers to 300 millimeter wafers?

JY: Yeah. And one of the big failings I think of Canon and Nikon was partially that they didn't really handle that transition well, and also because they were all vertically integrated and ASML is a much more horizontally integrated supplier that was able to take the best stuff and kind of use that where Canon and Nikon kind of fell short there.

When TSMC Passed Intel

Well, I think the other thing too is because Canon and Nikon were big Intel suppliers, and Intel just wasn't that worried about productivity because they made so much margin on their chips, and they were only making it for themselves and so Intel wasn't really pressuring Canon and Nikon to sort of go faster, whereas at TSMC it's like it's a pure direct line from our productivity to our revenue numbers and so there was a real meeting of the minds there where TSMC wanted less vertical integration because they wanted to be able to incorporate best of breed up and down the stack, and they also wanted to go way faster, and so it was a real opening and opportunity. ASML is really important for TSMC with EUV, but TSMC was really important for ASML in terms of ASML getting to its large market share in the first place by capitalizing on that wafer transition.

JY: Yeah, it's a very weird relationship between these two companies. They're technically corporately siblings, and you have the sibling tensions of all that entails. Its fun.

Is that a real thing? Because they're both downstream from Philips, that's sort of the relationship you're talking about. Why isn't Philips a dominant sort of semiconductor name today when you had these two companies come out of them?

JY: Oh, that's a funny story. To hear it from the ASML people, Philips was just a failure of a company. They were too bureaucratic, too solidified in their ways. They're almost like, I guess the Japanese in the 1990s and they really deserve to, Philips kind of lived their life and deserve to wind down and they sold those stakes way too early as it turns out. But they weren't the first, they wouldn't be alone in believing that TSMC and ASML had reached the peak a little bit too early.

Is there an aspect where this 200 to 300 millimeter wafer transition and the increase in efficiency and speed that it allowed for, was that actually, if you were to really zoom out, was that when TSMC started to pass Intel? This was still a good decade before their process surpassed Intel, but is there a bit where just the speed and efficiency, and the learning and the iteration that was downstream from increased velocity was in the long run when this changing of the guard began?

JY: I think it's kind of tough to make a call like that. I think someone at Intel once told me that TSMC wasn't necessarily faster than or ahead of Intel in the sense that their digital logic products, they're making better digital logic products. I think because the market that they were operating in at the time incentivized them to go for lagging-edge and trailing-edge chips, but in that area they were very good at that. They built a flow and a process and nodes that were customized to serving external customers.

This changed when the mobile industry started to become more of a thing and Intel missed that, but once it became more clear that mobile chips needed to get more speed, that's when Intel or TSMC decided that, yes, we need to actually start going up this node process. I think trying to distinguish when TSMC would start to surpass Intel, that's tough to kind of find that because they was always so different from one another in these early years, and you wouldn't really say that they were competing basically until what, 2007, 2008.

Yeah, they're hardly competing today, in some respects. I guess the most direct place of competition until Intel's foundry really gets going is via AMD and Intel to a sense. But here's another question. There's a bit where Intel famously, a lot of these technologies, they brought them to market first, and you think about how do you bring new technologies? Being highly integrated is at least theoretically an advantage, but you have this bit where TSMC is not highly integrated, they're looking for all these different pieces, but they're very process-oriented and they're very efficient and setting it all up, but is there just a bit where the complexity got so high that the integration tax was too large and process actually mattered, iteration speed mattered more than anything else?

JY: Integration tax, I think it's debatable, I think right up until the 10 nanometer or 14 nanometer fiasco. Right before the 14 nanometer fiasco, you could argue that Intel was two generations ahead of TSMC and they were going so fast. I think what happened wasn't less of an organization failure in the sense that they were too arrogant to push ahead of the technology, but I think if there's one aspect where I say the integration of Intel failed them was that when 10 nanometer really started to fall apart, the product side of the company had already built products for a 10 nanometer spec that was not feasible, was not possible. And that basically froze the company as they had to break down the product, redesign it, and that took two years.

You could say that's when the vertical integration of Intel really started biting them in the butt. But prior to that, I would say 2011, as recently as 2011, when they were doing 22 nanometer, the second generation FinFET and all that, they were really killing it.

That's right.

JY: They're going really far ahead.

No, I mean they were so far ahead with FinFET. It's under-appreciated how relatively recent that was. You actually said something really interesting about the sort of arrogance bit, because I think the common perception of Intel is they refused to leap ahead and they didn't adopt EUV. But actually there's an underrated aspect here, which was there were other aspects they were trying to go ahead too far too quickly.

JY: Yeah, well, they really thought themselves as the purveyors of Moore's Law. They were the most advanced manufacturing company in the world, Moore's Law was theirs, and they had to keep pushing the boundaries. And they've already done it many, many times the tick-tock —

The tick-tock strategy, yes. Improve design, improve process, improve design, improve process.

JY: That had really gone well for a couple years and that was really killing it. That arrogance jumped ahead of what the technology was capable of in 2012, 2013, when Intel was working out 14 nanometer, 10 nanometer because they do that in parallel. The EUV was not ready and EUV wasn't even close to ready. EUV because of the power source, the power source wasn't even close to ready and it was really flat, flat, flat for years and then suddenly it was a huge jump. It was a mess-up, but they probably-

What were some of those leaps they were trying to make in a world where EUV was not ready? They were trying to get to EUV eventually, but there was a cobalt issue and some other things along those lines. What were some of the roadblocks that they ran into?

JY: I would generally say it's like we don't really know. Intel has never been really clear. Dylan [Patel]'s pretty clear, pretty insistent that it was cobalt, but I think in generally, you could just say the whole spec was ambitious. There was a slide that I saw from the Intel Technology Investor meeting where they looked at density, and they're saying 14 nanometer is going to be another two-time shrink within two years or something on top of 14, or 10 would be a double shrink on 14 and that simply was too much and it was too far ahead of what was possible. TSMC would never do that sort of thing because they're much more incremental and they would go year after year, very slow and now that sort of taking too much step is kind of the reason why they couldn't do it.

Right. And of course they're trying to recover by taking massive amounts of steps in a very short amount of time. I guess maybe the cure will work out better than the disease.

JY: We shall see about that.

Assembling a Chip

All right, let's go back to the process. You have the silicon wafer, you have this deposition step, you have the mask, which is the pattern for the chip that the light source shines through, you etch it onto the chip. Then you mentioned once you've etched a layer on, it's going through multiple times. How many times is the step generally repeated as you're sort of building up a chip?

JY: Hundreds of times I think, and I've heard N2 is something like 20% more steps than N3 and N3 was something like 700, 800 steps. It's a lot of steps and these include a lot of different things and they also have to have incredibly high yields each individual step or else the end product will be terrible.

Is this where a lot of the yield problems happen, is in this sort of recycling of steps?

JY: I think yield problems — I mean they're never going to tell me, but I generally understand the yield problems come from the new introduced steps or interactions between different steps that don't end up working well or they come out of the end as they have no idea where it is and then they have to go back and they revert. It's a lot of different things that they introduced into — if you think about it, the TSMC node is basically an accumulation of nodes that work basically from the very beginning. They're slowly extending it, testing it, see if it's going to work. If it doesn't work, they'll cut it off and then they're adding these new steps to see if you can achieve a better process.

You still have to connect these transistors together. It used to be different pieces you would actually wire them together. What is that layer made of and how is that actually accomplished where you have all these transistors etched on a chip, but you have to connect them? There's a communication layer, there's a power layer. How does all that work?

JY: That's done using something called metalization and metalization is kind of like, that's a whole journey in of itself. It used to be they used aluminum to deposit, they would pattern where the wires go and then deposit aluminum in between the transistors as if it's just another thing, but then they moved to copper because the resistance of the aluminum wires got too high that you started getting delays in how the signals were sent. So they moved to copper, but copper has its own problems where it leaches into the silicon.

So what they did is that they had IBM develop this wonderful copper lining that protects the copper from leaching to the silicon, and the whole industry followed what IBM did and copper has to be done in a weird inverse way, but that's served us well for 20 more years and now I think the industry is probably looking more at ruthenium, which is going to be the next big jump in interconnects beyond copper.

Is lithography used for these interconnects or is that different? How is the actual pattern established and laid down?

JY: Yeah, it used lithography. EUV is used for what very small percentage of the steps within—

Right. They still use the old DUV for as much as possible because it's much cheaper.

JY: Yeah. Like 80, 90% I would say much higher levels of the metal layer. So it's M1, M2, M3, M4, all that M1, M2 layers might be done with EUV and then the rest would be done with DUV.

So then how does the power layer into this? Right now — we can get to backside power in a moment — but right now you have the transistor layer, you have the communications layer, and then power goes on top of that?

JY: Power is interspersed with the communication wires. They also go through the same wires and they go through the same thing. You have these big interconnects or big wires at the beginning and then they slowly get smaller and smaller until they finally reach the individual transistor, and you mentioned backside power. That is the situation where they got to take that whole network, split it out and move that underneath the wafer and save a lot of space.

But the problem is, now you've done a lot of work. If the most difficult layer is the transistor layer, but you already laid down all the power bit, then you're throwing a lot more away if the chip screws up or if something goes wrong in that middle layer. Is that right?

JY: Yeah, and it is a wafer bonding technology and wafer bonding is relatively young.

What is wafer bonding?

JY: Wafer bonding is where you take two wafers and you smack them together and there's various ways to do it. It really, it blows your mind how much detail they've gone into this, but you can bond them using Van der Waals forces. You can heat them up, you can use glue, you can use a whole bunch of different things.

But at the core of it, the concept is you take two wafers and you smack them together and you would take, in the case of backside power, you're taking this power network and you're just melding it to the bottom of the transistor layer and making it into a thinned transistor layer to connect the transistors to the power. Hope that makes sense.

The Semi Supply Chain

How do you find out about this stuff? And to what extent is the knowledge of how to do this kept internal versus it becomes shared across the industry and people know how to do that? What is the give and take in terms of, "I want to learn from other people, they can learn from me", versus, "This is super important intellectual property that keeps us unique and no one else can know?".

JY: It's a very interesting question. I think it's like, one example that comes to mind is the copper interconnect technique. As you mentioned, the way to keep the copper from leaching into the silicon is very strange, it doesn't really make sense. It's called Damascene and IBM developed it, but all the other American semiconductor companies quickly developed it afterwards.

Partially it's because of diffusion through employees, there's inter-corporate organizations where they all mix and they chat, there's like a feeling within the industry that it's time to adopt this and they all work towards it. But then they all have their own different styles, I guess, to fill their own proprietary specialty in it and you can definitely see this within definitely the packaging industry for sure.

So I would say you have big ideas and concepts that are shared within conferences, within conversations over the years, but then the small details, which is where the juice is squeezed, that is developed within the companies and they'll keep that to themselves.

Actually, I want to tie this into the question about the Japanese companies before. Is there an aspect where almost the more mechanical bits are, "Look, if you know what to do, you can internally then figure out how to actually do that", but is it a bit about chemicals and that nature? You might know what to do, but if you don't actually have the recipe, you're not going to figure it out. Is that tied into why that's been so much more resistant to competition?

JY: Precisely. It's one of the big ones, and the other thing is that no one just wants to spend their life spending 20 years studying a chemical or cycling through all these chemical recipes to find out what might work. A lot of these are like there's one person in America who might know this type of resist, this chemically-activated resist or whatever — maybe okay, maybe 5 or 10. But then in Japan you can take a whole bunch of PhDs and say, "This is your life now, you will study this for 30 years" — and they'll do that. Then that's how these companies like JSR build immense, immense moats that are basically impenetrable because no one else wants to do that work and no one else has an IP.

So is there a case that JSR has a bigger moat than ASML?

JY: Oh, that's a tough one. I would say yes. In some ways Tokyo Ohka and JSR are more moat-y than ASML because there are alternatives.

But as far as the steps go, are the same chemicals used from the chemical perspective? Are those all used in trailing-edge as well as leading-edge, or is there any differentiation in that regard? Whereas obviously for the equipment, EUV is only used on the leading-edge. You're not going to use it for your 28 nanometer chip or whatever it might be. The chemicals is that just — that's a commonality, it's used for everything?

JY: No, the chemicals are tuned for the process. So what will happen with TSMC is TSMC's people will work with the JSR, Tokyo Ohka's people. They'll be like, "Okay, this is what the node's going to do, I have a good idea of what we want with this resist", and the JSR guys will say, "Oh yes sir, yes sir", and they'll come up with this very special formulation that works precisely with the dose that you want it to with all that, and then once TSMC accepts it for the node, then JSR starts raising prices.

TSMC knows what they're getting into. But I think actually I love that example because I think it speaks to what is unique about TSMC. It's this, I've talked about they're a customer service organization for fabless chip companies and they're very collaborative and they're going to work with you to, they have the IP libraries, but they're going to get it working on their process. But what you're speaking to is this deep level of collaboration on the process side and actually you say systems integrator, that sounds low brow, but in fact it's essential to how this all works and comes together.

JY: Yeah, and I think a lot of this stuff is so optimized now. You can't really just pull something off the shelf and I think these chemicals are insane. You can imagine. They look, if you imagine these are insane molecules made of a whole bunch of different things and you have arms of molecules, they're designed to do certain things when the light activates them or something like that.

It's pretty nuts.

JY: It's very moat-y.

The chemical part I think is very underrated, I would say I underrated it, this has been actually pretty illuminating. I do have one question — what is ion doping?

JY: So what you would do is that you would use an ion beam to basically dope the parts of a transistor because a transistor silicon by itself won't send a current from one side to the other. So what you need to do is that you need to embed certain elements within that silicon at different sides to make it electrically active. And this is done with an ion beam.

Who makes the ion beam?

JY: Oh, that's a good question. There's a lot of different companies, I wasn't able really to find a big company that does it, it's not like a new technology.

So you have these three layers. You have the transistor layer, you have the communications layer, you have the power layer. Is this when you can start testing it and see when it actually works? How is the whole testing, dicing, and binning or can you be testing throughout so you know when you can toss it and when you cannot?

JY: My understanding is that they do testing more at end of certain stages, they'll do checkpoints and because they never want to have a situation where you do the whole thing and then you test it at the end and you're like, "Okay, this doesn't work", and we don't know why. But I know that that does happen.

Got it. So with these big five companies, where do they fit in this stack? And I know you mentioned JSR, who I don't have in my big five, but I guess I'm more focused on the manufacturing aspect. In the steps we talked about, so where is, I mentioned it, but where is LAM Research?

JY: These companies, they just all interact with each — they're all merged with each other nowadays, so they all do the same thing. So you could look it up and it'd be like LAM has a deposition product, Applied as a deposition product, Tokyo Electron has a deposition product and that's what I generally have seen. And it seems strange as a deliberate strategy, I would say, on the side of the semiconductor makers, but they all have their own products to step on each other's sides.

So in a typical process, do you go all in with one company or you pick best of breed from all the different ones?

JY: TSMC tends to do, what they will do is that they never want a single supplier. So what they'll do is that if one supplier is working for a deposition, right, they do CVD for a particular deposition product. What they'll do is that they'll go to another company, another one of their, like Tokyo Electron for example. They'll say, "We are using Applied for this" or "We're using LAM for this", "Can you make a product to do this as well and hit this spec at this price, and we'll buy it from you?", so what they do is that they really play each other off in the center of such a way and they do this for almost every supplier with the exception of two.

Got it. So TSMC is really breeding this commonality of competition to keep their costs down. Well, I'm sure that drives them crazy as far as ASML goes. You mentioned two, what's the other one? If ASML is one that's unique, what's the other one?

JY: Within the company they have something called "Shuang A". So "Shuang" means double, so there's ASML and Applied Materials.

What does Applied Materials do that's so special?

JY: I was not told what that was, I'll try to find out.

China, Intel, TSMC, Rapidus

Oh, very interesting. That is interesting. So with this understanding and perspective, what does that mean in the context of China? I will say one thing. I mean there's the chemical aspect, which I think is very interesting and probably under-appreciated. The other thing though is there's a bit where TSMC breeding multiple suppliers in different areas, in some respects it's a diffusion of the technology that maybe makes it harder to control or that also speaks to the possibility of Chinese companies learning how to do these different steps. What's your overall view of China's prospects in coming up in semiconductors broadly?

JY: I think the core concept is that the core concept is not secret. So the product and what the machine and the equipment will do, most people will already know. There's so much documentation on EUV, how it works, how the layers are built on the mirrors, all that is already known. The secret is how to turn that into a commercial venture and how to turn that into a profitable commercial venture. TSMC can guarantee a second source, a potential second source because they say, "We'll buy that from you, if that passes our test, we'll buy that from you and we'll pay money for it".

So I think in the case of China's, they'll probably know how to do this stuff and I've read their textbooks and they seem to know it pretty well. So I think the secret more is, can they move that from an academic sense to an actual doing it sense? That is much more challenging because that requires volume, that requires manufacturing, that requires wafers, and also lots of capital risk on behalf of the buyer.

So I get this question a lot in terms of what China can rise up for. I think nothing's impossible, it's made by man, it can be done. It's required to be done economically, exportably, to be in a way to export, that's much more challenging and I think right now the wafer volumes aren't there yet.

Where can China realistically get to? I guess there's two questions. Number one is if they have access to western equipment, SMIC can make a 7 nanometer chip. They might not be able to make it profitably or with high yields, but they can make it versus the, "We don't have access to equipment anymore, we have to actually make the equipment itself". What are the barriers there? What's the hardest part to catch up with? Or is it actually, we're missing the whole point because it's actually the chemicals and the materials engineering?

JY: The chemicals, I think they're so far behind, I can't even imagine anyone catching up to Japan in terms of chemicals.

It really is interesting to me is the role that the equipment makers play in building the process node. A guy at TSMC's lithography bay, even the manager, even a higher level manager, has no idea how the machine works. Conceptually, they don't know how to use it necessarily, they rely on to be told by other people within the ecosystem, the equipment manufacturer or their R&D guy to say how to use this machine.

So surprisingly little fundamental core knowledge of that is within the industry. So cutting that off to some extent will be immensely damaging because you can't just say, "We're going to just rebuild this node with a different machine", that basically, that introduces so much new uncertainty into the flow that makes even me anxious. So it's pretty crazy.

So do you think that, I don't want to put you on the spot, but broadly from a timeframe perspective, if China did get Western equipment, is it maybe a bit longer to catch up than people might think?

JY: It probably will. I've done videos on how the Soviets and the East Germans managed to procure special banned equipment over the years. They're pretty good at it. So it's like I would not be surprised to see these things starting to pop up and I would not be super surprised to see a gap to be like three years and that three years will be gone before you know it, right? Three to five years gone before you know it. You will never know. You'll say, "Oh my gosh, they caught up so fast". Well, from the beginning you already said it was going to be three to five years and now it's three to five years.

So where is Intel in all this? We touched on that a little bit, where they went wrong. They got stuck at the 14 to 10 nanometer transition, EUV wasn't ready at the beginning. Then had to rework it, the ways they tried to leap ahead didn't work. Can they catch up? Are you feeling good about 18A, can they actually leap ahead as you talked about?

JY: I'm not going to answer that, I feel 18A looks good on a conceptual level, but Intel has always been really good at introducing concepts to the industry that no one will pick up. So it's very, they don't know how to do Foundry, that is for sure.

What's the bigger failing here because when it comes to Foundry? We just talked about two sides, there's the customer service side, which they're not great at, but there's also the integration bit. Is there a bit where they're actually, they're not sufficiently used to leveraging and leaning on and depending on suppliers to the extent that TSMC was because they wanted to do it all themself and know it all themself? Or am I overrating that?

JY: There is an arrogance I think on the case of Intel at least for last couple years, up to a certain couple of years. How is it now? Who knows? I mean, a lot of people say Pat Gelsinger changed the culture, did they change that part of the culture? I don't know, but I think the main issue is that like Morris Chang said it best, "TSMC has learned to dance with 400 different partners, Intel has only danced alone".

So I struggle to really understand how you can really re-engineer an entire company to become so vertically focused and so focused on this one thing to becoming so broad and ecosystem oriented. It's very strange. I am not going to say they can't do it because the Intel bulls will come after me, but I think it's going to be interesting ride.

I really actually think this has been a very illuminating, I think to your point. There is this bit about this integration background, and to your point when you're describing this TSMC process about TSMC, not even people at TSMC knowing how particular steps work, but bringing in and depending on other entities to do that and you can see why they did that before when they were behind because that was the way to catch up and it turned out in the long run that ended up to be the way to get ahead as well. That's a cultural learning for TSMC that's been embedded for 40 years and it's just totally the opposite way that Intel has always sort of operated.

JY: It definitely is, and I think it's like what they'll do is that they're so perceptive on pushing problems up to back to the company. I think I recall that one of the bigger breakthroughs within EUV was discovered, on how to keep the container lenses clean was discovered not within the R&D section, but basically by accident and it's circulated from there. So it's like these bringing in the equipment vendors, bringing in your ecosystem partners to work with them and help develop your process node is such a clear, it's so important and that's very difficult for a lot of different companies and it takes humility with a dose of you have to be humble and confident at the same time and that's very challenging.

Do you think TSMC — or is this going to fall on the equipment vendors and the chemical manufacturers — can push innovation in the long run? Because I think the flip side of the Intel arrogance is a lot of that arrogance was well-earned. Like we mentioned FinFET before, a tremendous breakthrough. You're basically making these transistors 3D and Intel pushed that forward. Intel birthed the idea of EUV way back when and ASML brought it to market. But can TSMC in the long run push us through similar barriers or does this approach only work as fast following to a certain extent?

JY: Yeah, I think that's one of the big questions, even the people within TSMC are aware of this. There's a need for crazy left field thinking that Intel brought to the table, very ambitious, like moonshot thinking that Intel really did bring to the table. Everyone wants to see Intel succeed and become a viable second option. I think the question is can they do it? I don't know. But the ideas I think for that they brought were always illuminating and always pushed the industry forward and I don't think TSMC has the culture for that.

Is there a bit, we're a little too hard on Intel where they just kept making these huge leaps and one of the leaps, they didn't cross the chasm, they fell down and maybe we should have a little more grace for the fact that they ran into a wall because at least they were trying to jump over it, as it were?

JY: Yeah, there's one half of that I agree, but then there's other half is that there were also, we should never forget that they were a monopolist, an arrogant monopolist that tried to kill competitors and did anti-competitive pricing things against AMD. We should never forget that part that they were also trying to evade other markets and they victimized Compaq and all these other companies, so let's not forget that part too.

I have a somewhat related question, but what is Japan doing with Rapidus? Do they have any chance to build this other alternative supplier? They're trying to build a two nanometer process in Hokkaido? Any takes on that?

JY: Dylan and I like to say, basically we have a group chat and where all of us together and we like to call Rapidus, the maker of fine artisanal wafers. We're going to make these beautiful artisanal wafers like Samurai swords and each one's going to be absolutely perfect and I have full confidence that Japan can make an absolutely perfect N2 wafer, like 10, but then it's like can they make 10,000? That's a bigger question. I don't know if they're there yet, but they're going to spend a lot of money on it. I think they're going to try, we're going to see. They'll have customers, but I don't think they're not going to be ultra successful. It's not exactly SMIC.

Are these TSMC fabs in America a viable second source or is it one of those things where, "Once TSMC Taiwan goes down, these fabs are not worth nearly as much"?

JY: I'm sure they'll have value. There's so many Taiwanese over there working on it, I'm sure they have some value, I hope so. But you could argue there is a very viable argument to say that without the R&D flowing from Hsinchu, the fab won't be commercially viable within two years. But then, in such a scenario where you're saying what you're saying is happening, then I think there's very different things going on and all the assumptions should be reconsidered.

Then there will be much bigger problems you have to think about. What makes memory different from logic? Why isn't TSMC just manufacturing all the memory? Why is that SK Hynix or Samsung or Micron? What is different about that process versus making logic chips?

JY: Memory is a lot more repetitive and there's a lot more emphasis on materials, there's a lot more emphasis on processes than necessarily innovation. Taiwan tried to build a memory industry, a DRAM industry and ITRI [Industrial Technology Research Institute] couldn't make it happen. The failure of ITRI post-TSMC is actually something to be talked about someday, but it's like they tried to get all these small memory makers to come work together to challenge Samsung and it was a massive failure. Memory is so much more economies of scale oriented and Samsung has Samsung money, so they really carved the market there.

Samsung just famously, every time there was a down market, which happens with memory because it's more of a commodity, they would just invest through it. And so, every time the market came back they'd have a larger and larger share. Is logic at that point? I think one of the most interesting parts of TSMC history was 2008 when the great financial crisis hit, TSMC management wants to pull back and Morris Chang comes back to the company and he's like, "No. Did you see what Apple just introduced? We're actually doubling down", that was a bit where they invested through a downturn and really emerged on top. Was that a transformation where logic was more a specialty thing that's why Intel moved to logic, they abandoned memory, and now it became more of a commodity and that was also it became more memory-like than it used to be or does that not make sense?

JY: It's interesting, this funny thing about that. Morris Chang saw the opportunity from the iPhone, he also didn't think the iPhone itself would be successful. He was one of those guys like Steve Ballmer, this is what I've heard, he was one of those guys that Steve Ballmer were being like, "This is the $600 phone, no one will buy it".

I'm amenable to the belief that logic will always have value at some point, there'll always be differentiation. So you could say maybe you could make an argument that TSMC thought N2 wouldn't be a valuable node, that's why they built a smaller fab for it, but then suddenly AI became a thing and now they're pivoting on that as fast as we can possibly see. So it's really fascinating, I don't buy the fact that logic will become like memory.

But the volumes are becoming almost more memory-like particularly in some regards.

JY: In some regards, yeah, maybe because memory is becoming more like logic in some ways.

That's interesting. How so? Use it like the high bandwidth memory and things like that?

JY: Yeah. They'll have elements of logic where you need to manage the flow of data and stuff like that, and they're very complicated too, they're very complicated stacks and there's a lot of crossover now, tt's very fascinating.

Consumer Electronic Cars

Well, we've dived super deep into the old Jon Y catalog. You've been doing a lot recently about cars and electric cars. Is this just a, "They're computers on wheels", is that the framing or what do you find interesting about this space and that is piquing your interest?

JY: A couple years ago, I think I was, someone who visited from China told me that these cars are coming and these cars are — when Xiaomi started making cars themselves, I was like, they're trying to make cars more like consumer electronics. They're trying to make it like drones or Bluetooth stereos and I think that's something I wanted to call out and I tried to call out as soon as I can, and I've been really interested, I was like I'll pull it on the side because it's a technology that integrates all these different things together and you could argue that turning it into a consumable is kind of anathema to what the United States has really seen what cars would be, and can be, and I think that's something that's very fascinating. That's a very Asia-focused car philosophy in my opinion and these cars are coming. I mean, everyone keeps saying it, everyone keeps talking about it, and I feel like there's not enough movement from domestic old legacy car makers to change and it makes me sad.

And this is thinking about cars as almost more like the OEM model like consumer electronics, "We're going to just build out a standard platform", whatever differentiation there is can come from software and that's how you have a Xiaomi car, right?

JY: Yeah. Have you tried one of the Luxgen n7 cars? I rode in one recently.

I've been in some of the Luxgen vans, but I haven't actually driven any of them.

JY: The new n7 is built on their Foxconn platform or whatever, and it's like an amazing car. When I rode that and it's like $30,000, it's a very good car. It's $30,000 and I rode that car, I'm like, "This is not a Chinese car", Chinese didn't make this — Taiwanese Foxconn made it. They're new to the cars, this is a really good car and I'm just like, I mean you would think that the Chinese know something special, but I would argue that it's the concept of the car as this electronics thing is that's the true differentiator. And if so, it would require so much more work, I would think, on the behalf of the product makers to make a better car here, a competitive EV.

It's almost what they don't know and it's precisely because the traditional automakers, because making an engine was really hard, and so it's very difficult to accept the idea that your highly differentiated expertise and what you're good at is actually now an obstacle to success because of the cost that goes into it and the wear and tear inherent to that. And actually no, the solution is to make it as simple as possible and then consumers just want cool software, that can be enough.

JY: And that makes me sad that in some ways that Apple left the car industry, it would've been really fascinating to see.

I think the Foxconn connection there was clear. It's like, yeah, you build us the standard platform, we will do our software on top. I agree, it's almost like the TV thing, they were rumored to do a TV for years and years, 10, 15 years ago is always the talk and you wonder maybe they should have done that because what platform does that give you to do other stuff in the future? You have to run the same thing about a car. Even if it was started out not being fully self-driving, that gets you down the road to eventually do that.

JY: Yeah, I mean, maybe that's one of the downsides of the Tim Cook era, they did good on extending the life of their single core product, but I don't know, they didn't have the, I felt like maybe they fell short going as crazy as doing some crazy stuff. I think [John] Gruber mentioned that before — one of the losses of Jobs is that they didn't do weird stuff enough.

Yeah, weird stuff is valuable.

All right, well if anyone wanted to jump into the Jon Y library, the semiconductor stuff's amazing. We're going to have a ton of links to it, you're doing more on cars. Are there any other pet topics that just stick out to you? This is some of my favorite stuff to do. You mentioned actually the Soviet economic issues. Those were some really interesting videos, what else is in the must-see list?

JY: I have an extended series on water desalination technologies and that turned out to be looking at the water desalination progression from the economics of desalinating water at a scale to feed a country is fascinating and I think people should learn that.

One of my favorite videos in recent days was looking at the Saudi and Middle Eastern water supply technologies to desalinate water at millions and millions of gallons of scale and looking at a lot of that is basically filled by cheap energy and it's probably not economical, and I think it was fun. I think water is something that a lot of people need to pay attention to and where your water comes from and how it's being made. And I have what, 12 videos on it, it's really fun.

Well, I mean, what I think is clear in this conversation, and I think everyone should pay attention to, is if you want to understand how something works and you want it in a 20, 25-minute or an adjustable video, Jon is your resource. I've learned a ton from you, I started out asking you questions that I already knew the answer to, but I knew the answer because I watched your videos. So everyone should follow, you can now get access. You can listen via podcast, you can actually read the transcripts, but the YouTube videos are the bread and butter, everyone should go and subscribe. Jon, it's great to have you on board. You've done a great job with Stratechery videos and I'm super excited to be working together going forward.

JY: Thank you. I'm really excited to be working with you and it's going to be fun. It's going to be a great 2025 and beyond.

Can you imagine it? Another year is at an end and it is remarkable to think that time has gone by so quickly.

I have done a recap video for the past two years. Let's do it again.

In today's video, I want to reflect on 2024 and the state of the channel. I am still actually working, haha. But a few minutes to reflect.

Year Recap. My favorites.

I have personal favorites this year that I want to share with you. Let us begin with that.

My favorite video to make in 2024 was that of the Nisei interpreters during World War II. Their plight of navigating two different cultures strikes a chord with me. And their stories of heroism - particularly the story of Terry Doi at Iwo Jima - will always stick with me.

Second is the story of Texas Instruments. It has it all. The integrated circuit, silicon transistors, Morris Chang and a disastrous foray into retail. I find myself going back to a few stories and re-reading them for enjoyment. This one I come back to a lot.

Third is the story of Ei-Nis, the semiconductor maker in Serbia. Augmented with thoughts and image records from the son of a former worker there, I found it extraordinary insight into the Yugoslavian economy.

Fourth is my video about how a CVD diamond is made. I came across the company during a visit to a startup fair in Taipei. I asked if they’d be interested to letting me do a video about them and amazingly enough they were. And that is how I found myself in a small village outside Taichung, handling diamonds that might cost tens of thousands of dollars on the market. Lab diamonds are the future. Enough said.

Fifth is my video on RF filters. A miraculous piece of technology that helped create the AI giant now known as Avago/Broadcom. I want to thank the anonymous viewer in Denver who spent the time to walk me through it in a call.

Workflows

In recent videos, I have been augmenting the research with conversations with key people. These might be experts in a field, researchers talking about their life's work, or participants in a historical event.

Perhaps one small advantage of being more well known is that people recognize who you are, and thus are more willing to speak with you. I am appreciative of that. Many of these people have unique life experiences that should be shared. I worry about the knowledge of such experiences vanishing into the ether, and am glad to offer the channel as a way to get it out there.

If you got a fascinating idea, let me know about it. Some of my favorite videos to research and write about have come about at the suggestion of a viewer. For those who have recommended an idea to me but I have not been able to get to it, do not worry, it is on the list.

Someone once challenged me whether I really had a “list”. I sent them a screenshot of it. Yes the list exists. Here is merely a tiny portion of the list.

Milestones and Travel

In 2024, the channel gathered 26 million views, which is a lot.

The channel also passed over 100 million lifetime views. To me, it is stunning to think about how many people the videos must have reached around the world.

There are a lot of milestones around, and most of the time I just ignore them. Work continues on. But for some reason, 100 million sort of resonates with me. I know that is how much your average MrBeast video does, but to me it is worth celebrating.

I did a lot of traveling this year. This year I was in Japan, the United States, Korea, Singapore, Switzerland, Belgium, Netherlands, France, and the UK.

My favorite of all the places we traveled? I really liked my trip to Europe for ITF 2024. It gave me a chance to explore many of the interesting technologies they are working on over there.

And of course, I got to visit the ASML campus, where it seems like I have a few fans.

Since then, more than a few people have asked me how I managed to get a visit. Honest truth? No clue it just happened. The trip also let me bring back like two bags of cheese to Taiwan. I have since eaten all that cheese so I will just have to go back.

Top Videos

The top three performing videos this year were:

"Why is Japan so Weak in Software?" This topic had been suggested to me by a friend in Japan during a visit. I thought it would be fun, but ended up struggling with it a great deal for several months thereafter.

Did I ever really answer the question? I am sure there are plenty of YouTube commenters who claim I did not, but I gave it the best try that I could. The resulting video was a fascinating exercise and I want to do more of these in the future.

"The Rise and the Fall of the Cray Supercomputer". The idea had been suggested to me in a few reader emails, but it was a trip to the Computer History Museum that convinced me that I had to do it.

I travel a lot around Taiwan for family reasons, so I remember many of these videos by where I was when I made them. I distinctly remember working on the Cray video in Kaohsiung, Taipei, and Taichung - a whole variety of places.

I wanted to do a video about the rise of Japanese supercomputers, and how they affected Cray in the marketplace. Still might do it, but after doing that video I was all supercomputer'ed out.

Third, "The Birth, Boom, and Bust of the Hard Disk Drive". I worked on this video throughout my travels in the United States and Japan. When I uploaded it, I clearly remember wondering if I had done the topic justice and whether it might be of interest to viewers. As a guy who bought bunches of these HDDs back in the day, learning about them was fascinating to me.

I did a follow-up video on that video discussing some new technologies in the industry's future. The tech behind what they are doing is amazing. But I remain unsure about the future of the hard disk drive industry in the face of the NAND onslaught.

Videos That I am Sad About

Every year, there are videos that did not perform as well.

First and foremost is the IBM PC video. The story of how Compaq, Intel and Microsoft wrenched the IBM PC away from its originator is absolutely fascinating. Yeah, I know I put an ad on it, but I still think it was a wonderful story.

Some of the profiles of semiconductor companies like LSI Logic and STMicroelectronics did not do as well as I would have liked. LSI Logic had been one of the major American semiconductor companies.

And STMicro in particular involved a retracing of the European semiconductor industry. The story of CEO Pasquale deftly managing two long-standing cultures and actually creating a profitable company out of it is remarkable.

I also was a bit disappointed about the performance about the quantum sensing video. Nitrogen Vacancy diamonds are some of the weirdest things I have thought about in a while, and I think y'all are sleeping on it.

Meetings

Yes, I still meet with people in Taipei! Some of the most enjoyable conversations I have had this year include:

The conversation with a systems engineer at Rocket Lab, who patiently took me through the business of launch systems and space-based services. I don't think much about space, but that was very enjoyable.

I enjoyed a conversation with the Nvidia engineer working in vision AI. A kind soul who was open about what it was like to work there.

And there was the chat with a few students from Yale, who asked a bunch of very detailed questions about Taiwan and the Taiwanese semiconductor ecosystem. I think we talked for four hours!

And the various members of the semiconductor industry who took time out of their busy days to talk to me about what they are doing in their part of the world. Every day, these guys work magic.

For at least 2025, I will be in Taiwan and the US for work - and will try to meet with people as I can. With the way things are now, and recent events, I think it is important that people come here and see what Taiwan is like for themselves.

The Future

So what does 2025 hold in store for the channel?

There remains a whole lot of interesting things to be talking about in chips and semiconductors. I think I will still cover those stories in the future.

But on the other hand, I feel like I am getting a bit too heavy into the semiconductors. And while there are plenty of comments out there saying that when I wander away from semiconductor topics it's not as good, it keeps me fresh and happy.

Personally, running the channel is still exhausting. More than a few times in 2024, I found myself uploading later than midnight, and then immediately prepping the next video. I wish I had more time.

But the YouTube game is such that if you are not consistently posting then people are watching someone else. I get it. That is the life we sign up for. I still love doing this. As always, working on this channel is the privilege of my life and I look forward to having more stuff for you guys in 2025.

Share

I lightly edited some of the conversation to remove filler words. Other than that it’s pretty much as I remember it.

Jon:

Welcome to the show. I actually don't do a lot of these interviews, but I'm excited to make an exception. So welcome to the show Tae Kim. How are you doing?

Tae Kim:

Good, good. Jon, thanks so much for having me. I'm so happy to be here.

Jon:

Yeah. You have a very interesting book. And I just ripped through it. It was very fun. How did it get started and kind of what's the story behind it?

Tae Kim:

I’ve been following NVIDIA for my whole life. From the 1990s, I was a PC gamer, building my own gaming PCs, and I’ve been following it. When I was on Wall Street, I invested in it at a hedge fund, so I’ve known NVIDIA from the early days. For the last 10 years, I’ve been in media—Yahoo Finance, CNBC, Barron’s, and Bloomberg Opinion—so I’ve been following it as a journalist also. I got an email in May of 2023. It was a cold email from an editor at Norton, and he said Matthew Ball, who wrote the famous Metaverse book, said that I could do a book on NVIDIA. Would I be interested in that? The first thing I did was ask, what’s Norton? I didn’t know Norton, and they’re actually a very well-regarded publishing house making books for Nobel Prize-winning academics and Ben Bernanke. The people in the newsroom were like, yeah, Norton’s a really big deal.

The second thing I did was go to Amazon. I said, there’s no way no one has written a book on NVIDIA. It’s the biggest chip designer by market cap in history. Every internet company in the world has multiple books on Amazon—Facebook, Apple, every major tech company. There was no book. I couldn’t believe it. No book on NVIDIA. I’ve been buying their graphics cards for 30 years. I know the business side, the finance side, and the technology side. I’ve been in media for 10 years. I thought, I can do this. I can write this book.

I had a meeting with the editor at Bryant Park, and within weeks, I had a book deal. It went by so fast, and I was off and running.

Jon:

It’s kind of crazy, right? You talked a little about how Jensen reluctantly somewhat agreed to the book. Did he say anything? Has he read it yet?

Tae Kim:

No, he hasn’t read it yet. I’m sure they’ll read it in a few days. I’ve known Jensen from the beginning. I met him 20 or 30 years ago when he was raising capital with Morgan Stanley on the secondary. He probably doesn’t remember me. I talked to him a little bit. He’s a very intense person and takes things personally. I wrote negatively about NVIDIA during the whole crypto boom, so I’m sure that’s in the back of his mind.

For the first few months, I’ve had a good relationship with the company. They let me talk to the CFO every quarter. But when I told them I was writing this book, they were very noncommittal at first. They were like, “Well, we’re not so sure.” So I just went to work and started interviewing all the early NVIDIA employees and senior management. Surprisingly, they were ecstatic to talk to me. They said, “NVIDIA is such a great company. I had such a great experience there. There are all these amazing stories, and no one has told the story. It’s great that someone’s writing a book on it.”

After six or eight calls, word got back to NVIDIA that I was being very serious and substantive about it. They decided to cooperate. They gave me access to their senior management team and helped facilitate interviews with co-founders, which was fantastic. It worked out really well in the end.

Jon:

I did a video about NVIDIA a long time ago during their graphics card rise. This is so much more detailed than what I could have done at that time. There was so little information. It is very strange. A lot of that stuff would almost be lost to the sands of time otherwise.

Tae Kim:

Yeah, it was really difficult the first 10 years because the internet didn’t really take off until the early 2000s. There were websites in the late ’90s, but many have disappeared. There’s some material in internet archives, but it was really hard to even find press releases.

That was difficult, but I got the entire senior management from the first 10 years to talk to me because they were so happy to share their stories. It was an honor to compile all this computer history. I’m a big fan of early computers, business history, and strategy, and being able to talk to the co-founders of 3dfx was incredible. I don’t know if you were into PC gaming back then, but they were the dominant graphics card company for several years. It’s amazing that NVIDIA was able to overtake them. Talking to the 3dfx co-founders about why they faltered was so cool.

Jon:
Yeah, it’s actually one of the things I really appreciate about the book. You get to see the other side—from the competitors competing with NVIDIA. That’s very unique. I don’t see that often, and it’s very thorough here.

Tae Kim:

They’re very honest about how they lost. They got too fixated on features instead of the schedule. They wanted to make the perfect chip. Jensen was the exact opposite—he made the schedule the feature. They aligned with PC makers and how they buy chips, reducing the cycle from 18 months to six months. It was such a genius strategy. They’re doing the same thing now with AI GPUs, going from a two-year to a one-year cycle.

Think about AMD trying to compete with this—accelerating product cycles makes it incredibly difficult to keep up. That’s exactly what happened to 3dfx, Matrox, and dozens of other graphics companies like S3. They just couldn’t compete with NVIDIA’s relentless execution.

Jon:

Did you feel like there were opportunities where other competitors could have done, to sort of beat Nvidia back in video or sort of kind of like, defeat them? Because there were such times that they stumble, right?

Tae Kim:

The two major competitors, I would say, were 3dfx, and 3dfx had really good engineering and made two great graphics cards, but they couldn’t handle the 2D side, which is ironic because that’s the easier part. You need both 3D and 2D, and they couldn’t manage the 2D. Then they made a terrible strategic decision by buying STB Systems, which was 3dfx’s biggest board partner.

When they bought STB, it became hard to manage. Board partners buy graphics chips, put them on cards, and sell them to PC makers and consumers. But when 3dfx bought STB, they wanted to become the brand and couldn’t handle the inventory. It was a complete debacle. Inventory piled up, and they couldn’t manage the supply chain.

And obviously, all the other board makers competing with STB decided to switch to NVIDIA.This actually helped NVIDIA because it gained all the board partners that left 3dfx. That critical mistake directly led to 3dfx’s bankruptcy, as they couldn’t manage supply chain and inventory effectively.

Jon:

You mentioned a second competitor

Tae Kim:

The other big competitor was ATI, and it’s kind of funny how it’s connected to 3dfx. Jensen told me that one reason they messed up after acquiring 3dfx was the NV30, the infamous “leaf blower” chip—bloated, overheating, and requiring super loud fans. Jensen said the NV30 was a disaster because they hired 100 3dfx engineers at the same time, and it was hard to integrate them. The NVIDIA team and the 3dfx engineers didn’t communicate well, and they didn’t work closely with game developers. They skipped features that were really important to game developers, leading to significant communication and cultural issues.

On top of that, NVIDIA had legal disputes with Microsoft, and Microsoft withheld the DirectX 9 tech specs. This forced NVIDIA to design the NV30 essentially blind, which was a complete disaster. Meanwhile, ATI bought ArtX, the team that made the GameCube chips. They were skilled at designing graphics chips and created the R300, which outperformed the NV30 in every way. It delivered better framerates, was faster, quieter, and sold at the same price.

Dierks, who heads software engineering at NVIDIA, told me this was a critical moment. If ATI had been more ruthless on the business side, they could have bankrupted NVIDIA. The R300, particularly the Radeon 9700, had higher margins, and ATI could have cut prices, which would have crushed NVIDIA, as NVIDIA was barely making money selling the NV30 at $399. Dierks said that if Jensen had been running ATI, he would have bankrupted NVIDIA because he would have known exactly what to do.

There were so many near-death experiences like this for NVIDIA, and Jensen’s resourcefulness was key to their survival. In 1998, they were literally weeks away from running out of cash.

They had this manufacturing problem with TSMC. Intel was breathing down their neck, and Jensen was able to convince their three board partners to give them money. He said, “Our technology is good. We’ll give you 10 percent off the IPO when we eventually IPO. Just give us some money now.” He was able to pull that together in weeks. The CFO admitted this was all Jensen, not him.

Another big instance was when they had their first couple of chip disasters. Their first two chips were complete failures, yet Jensen kept raising money. Their third chip, the Riva 128, had a critical fault. Curtis Priem, one of the three co-founders, explained they didn’t have the ability to make a really good VGA core, which was a big problem. Doom was the biggest game of the day, and even though it wasn’t a 3D graphics game, people still played Doom and Doom 2, so you needed a solid 2D VGA core. NVIDIA didn’t have it, nor did they have anyone who could make one.

Jensen somehow licensed a 2D core from one of their main competitors, Weitek. After securing the license, he hired their top VGA expert, Gopal Solanki, who became critical. At a point where the company was about to die, Jensen found a way—convincing someone to give them money, a license, or a solution. This happened repeatedly.

Jensen is relentless, especially in recruiting. People say no to him again and again, but he keeps pursuing them until they agree. He’s constantly learning. One of the three pillars of the NVIDIA way is insane work ethic. People work like crazy. There’s also talent cultivation and speed and velocity. We’ve discussed speed and velocity, but talent cultivation is equally important.

For example, when Jensen was talking to Scott Sellers, co-founder of 3dfx, he was already trying to recruit him before 3dfx even started. He asked, “Who’s really good that you’ve worked with? Who are the rock stars?” Sellers mentioned Dwight Diercks, who Jensen eventually hired. Jensen constantly gathers knowledge and uses it later.

Jensen is the reason NVIDIA has survived in this brutal industry, where nearly every other company either went bankrupt or got acquired. Out of 80 to 100 companies making graphics chips, NVIDIA is the lone survivor.

Jon:

You talked a little bit about Intel. And I actually tweeted about the Intel thing. They had to kill Intel. Kill Intel. What, what was, what was that? What went on with that?

Tae Kim:

Intel, the big Goliath, was 860 times larger than NVIDIA in revenue at the time. As the chip king, they noticed NVIDIA making money with the Riva 128 and gaining traction. Intel decided to act like the dominant player, essentially saying, “We’re going to snuff them out.” They told PC makers, “We’re coming out with the i740. It will have better image quality, it will be faster, and it will be great. Why are you buying chips from NVIDIA? Just wait and buy ours when it comes out in February or March of ’98.”

This completely destroyed NVIDIA’s pipeline. PC makers hesitated, thinking Intel might offer something better. Jensen was alarmed and realized the situation was dire. At a company meeting, he famously said, “We’re always 30 days away from going out of business.” This time it was true. He told employees, “Intel is out to get us. They’re out to kill us, and we have to acknowledge that and fight back. We have to kill Intel.”

That might sound superficial, but the employees really took it to heart. One of the unsung heroes of the book, Caroline Landry, who was almost like a moral compass for NVIDIA in their early years, remembers feeling invigorated and riled up after Jensen’s speech. She started working even longer hours than she already was—past midnight, all weekends, waking up at 4 or 5 in the morning. Even though she wanted to go back to sleep, she’d tell herself, “No, we must kill Intel,” and head back to work to help make the next chip better.

When Intel’s i740 finally came out, it wasn’t that great and faded into oblivion. This was a common tactic in the computer industry, though. Microsoft did something similar in the ’80s—it’s called FUD: Fear, Uncertainty, Doubt, and Vaporware. If a third-party software company launched a great product, Microsoft would announce a similar product and say, “Why are you buying this?” It would kill the sales of that third-party product. Jensen was able to recognize this tactic and figure out how to rally his employees to respond and do better. He’s done this throughout NVIDIA’s history.

Jon:

You also mentioned CUDA. I found that really interesting because when I did my video, I barely mentioned it. You gave so much more depth. What sort of direction did they have when they were making CUDA? What drove its adoption?

Tae Kim:

CUDA is such an amazing story because it took so long to take off. It’s obvious now that it’s a huge part of NVIDIA’s moat and the reason they’re doing so well, but back in 2006, it wasn’t obvious. It was a huge expense, a massive R&D cost. They dedicated a significant part of the chip die to it, which made it expensive and hurt their gross margins.

Early on, it was just a concept—a parallel computing platform and ecosystem. It was Jensen thinking about the future and where the computing industry would go. He believed performance computing would eventually move to parallel computing devices like GPUs that could split workloads across thousands, and later tens of thousands, of cores, compared to CPUs that typically ran serially with only four to eight cores. He knew that as computing became more data-intensive, accelerated computing would win out.

For the first few years, CUDA didn’t generate any significant revenue. Some scientists used it for simulations, weather prediction, or molecular biology analysis, but it wasn’t bringing in big money. Internally, some people thought it was a waste of time, and Wall Street was pressuring him to cut it. Even in 2011, six or seven years in, it wasn’t a major revenue driver. But Jensen kept pushing.

Eventually, it started gaining traction for high-end engineering and scientific simulations. It was a slow grind, but Jensen never gave up. The real breakthrough came when AI started taking off—about three years ago. Even then, it didn’t fully explode until a couple of quarters after ChatGPT came out. That kind of long-term thinking is similar to Reed Hastings at Netflix. Hastings had an intuitive sense that the market would eventually move to video streaming over the internet. He didn’t know exactly when the technology or infrastructure would be ready, but he positioned Netflix to take advantage of the opportunity when it came. Meanwhile, he built the DVD rental business as a bridge.

NVIDIA has done this repeatedly—with 3D graphics, programmable GPUs, CUDA, tensor cores for AI, and full data center solutions with Mellanox. Jensen is always looking five, 10, 15 years ahead. He has an intuitive sense of where the market will go and positions NVIDIA with the best technology, ecosystem, and assets to seize the opportunity when it arises.

Take DLSS, ray tracing—it took 10 years. No other company invests in bleeding-edge R&D like this and successfully commercializes it. There are many examples of companies in history, like Kodak inventing the digital camera but failing to capitalize on it, or Xerox creating the Windows GUI. [NVIDIA, under Jensen, ensures they don’t just innovate but also dominate the markets they create.]

Jon:

You talked in one of the most fascinating parts of the book—the email culture at NVIDIA and the “Top 5 Emails” system. Can you explain what that email culture is like at the company? It seems very distinct

Tae Kim:

It’s called “Top 5 Emails.” This has been around since the beginning. Jensen loves email and is constantly on it. To this day, people who left NVIDIA 20 years ago can email him, and he’ll instantly respond, often within a minute. When Blackberry came out, he was on it constantly.

This system was a way for him to break through the bureaucracy that usually happens in large companies. In most companies, as they grow, managers create status reports, and those reports are cleaned up by layers of management to remove anything negative. By the time the CEO sees it, the report is often useless and doesn’t help steer the company.

At NVIDIA, they started the “Top 5 Emails” system where, every week or two, every employee writes an email to their team, manager, and various distribution lists that Jensen can access. These emails outline the top five things they’re working on or the top five things they’ve observed. It could be a major competitor’s move, a new technology, or a cool AI paper. Everyone knows what their team is working on, and Jensen reads hundreds of these emails every day.

This gives him the pulse of what’s happening across the company. No one can hide whether things are going well or poorly. He knows what people are working on and whether the vision he’s laid out is being executed because he’s constantly reading these emails, often all weekend or on Sunday evenings.

I don’t think other companies do this. In large companies, bureaucracy often filters out bad news because people don’t want it to reach the top and cause trouble. At NVIDIA, there’s a strong emphasis on intellectual honesty. Whether the news is good or bad, you must be honest. If someone hides something and it blows up six months later, that’s considered the worst thing you can do.

There’s a saying at NVIDIA: “No one loses alone.” If you’re falling behind, you need to tell your team, and they’ll help you. It’s one of the Jensenisms at NVIDIA. This “Top 5 Emails” system has been his way of staying informed about what employees are working on for decades.

It’s not one-way communication either. If Jensen reads a Top 5 Email that interests him, he’ll send a quick reply: “What’s this? What are you doing here? Why aren’t you doing it this way?” That’s his management style, and it’s why he has 60 direct reports—something unique in corporate America. Most CEOs have a small group of five or six executives reporting directly to them. Now, for Jensen, because he has this email culture where he can send quick emails very quickly, all day long, and he works more than anyone else, that’s why he can manage 60 people directly.

Jon:

Does he jump, uh, you kind of mentioned this a little bit in passing, but does he jump, um, titles, I guess, does he jump levels, uh, within the company to reach out to people directly?

Tae Kim:

Yes, that’s part of the culture as well. NVIDIA managers tell me that’s just how NVIDIA operates. In most companies, the CEO talks to a manager before engaging with an employee. At NVIDIA, all levels of the company are in the room at the same time, and Jensen loves talking to entry-level and junior employees because they’re often the ones doing the actual work on the ground. He’s constantly peppering them with questions.

NVIDIA managers say, “If you’re an employee under me, it’s fine if you’re talking to Jensen—just let me know what happened after the fact so I know what he’s asking about in my division.” That level of transparency is part of NVIDIA’s culture and avoids siloed mentalities. Jensen is constantly talking to people at all levels of the company.

Jon:

Interesting. What kind of productivity or work practices did you take away after interacting with them?

Tae Kim:

I’ve tried to be more blunt and direct. That’s difficult in a corporate environment, but it really does save time. So much of the workplace involves long meetings, followed by more meetings, and concerns about not upsetting or embarrassing colleagues. Then there are offline meetings and calls, all of which waste time.

If you’re blunt and direct—saying, “I have this issue, and I don’t think we’re going in the right direction”—it saves a lot of effort. I applied this during my book process. It’s better to be straightforward rather than coddling people’s feelings. It works. It gets everyone on the same page, and your colleague will usually align and move in the same direction because people generally want to be effective.

After this process, I feel like I have this “AI Jensen” on my shoulder telling me to be blunt and direct. Two important lessons from the book are: mission is the boss and speed of light.

Mission is the boss means making the right decision for the company and the customer—not what’s good for your boss’s boss. In corporate America, a huge percentage of time—30 to 50 percent—is spent doing things that don’t help the end customer but instead make your boss look good to their boss. Minimizing that and focusing on the mission, not someone’s bonus, is critical.

Speed of light means working with extreme speed and velocity. At most companies, when you do a project, you have KPIs, and you might say, “I did 10 percent better than last time” or “We’re 10 percent better than the competitor.” If you said that at NVIDIA, you’d get dressed down. They don’t care about what you did last time or how you compare to competitors. They care about what’s physically possible. If everything were perfect—no queues, no lag—what’s the absolute limit of physics?

For example, if there’s lag between one factory process and another or inventory issues, NVIDIA asks how to eliminate that lag entirely to the limits of physics. By operating this way, no competitor can match them. If everything is benchmarked to the ultimate possible speed and quality, they’re already operating at the fastest and most productive level.

If you do those two things—mission is the boss and speed of light—you’ll outperform nearly everyone. In large technology companies, there are horror stories. At Google, you need approval from five different stakeholders. At Intel, you might do the same PowerPoint presentation to half a dozen people, and nothing moves forward. At NVIDIA, you do one presentation to Jensen, with everyone relevant in the room, they make a decision, and you go execute. There’s no delay, no indecision paralysis.

Jon:

Is it a presentation or a whiteboard?

Tae Kim:

It’s primarily a whiteboard, right? Jensen loves the whiteboard and hates PowerPoint. At most companies, you go through slides and talk through them. At NVIDIA, they want you to take a marker, go up to the whiteboard, explain everything, and defend every single thing you say. If you propose a product or strategy, you must defend it with data and facts. It’s a very active process, and it’s completely different. I’ve been involved in companies where, as a management consultant, the end product was just PowerPoints. That’s not NVIDIA.

Jon:

You mentioned he likes a specific brand of marker sold in Taiwan. Do you know what it is?

Tae Kim:

I do have a photo of it. I can send it to you afterward. It doesn’t have English on it, so I can’t read the name, but I’ll share the photo.

Jon:

I asked all my friends in Taiwan if they knew the brand or recognized it. They were so confused. Do you have a favorite Jensenism other than the two you mentioned? He’s so quotable.

Tae Kim:

He’s very quotable. One Jensenism I like is: “Second place is the first loser.” I think he took it from the founder of Ferrari. He does not want to be in second place.

This story comes from a marketing manager who came from S3, another graphics chip company. In PC gaming magazines, they rank graphics cards based on benchmarks—naming the best, second-best, and so on. One magazine ranked NVIDIA number two. At his prior company, being number two was fine. The manager showed it to Jensen, and Jensen didn’t like it. He said, “No, we have to win. We have to be number one. Second place is the first loser.” That shows his extreme competitiveness and why he works so hard—he always wants to win.

There’s also a funny story about one of NVIDIA’s CFOs in the early years, Jeffrey Barr, who was ranked in the top 50 chess players in the U.S. earlier in his life. Jensen knew this and wanted to beat him at chess. He studied like crazy, memorizing openings and strategies, and played against him. But Barr could tell what Jensen was doing, so he’d play unpredictably, doing things out of the ordinary. Jensen couldn’t react and got beaten every time.

Jensen couldn’t handle losing. Every time he lost, he’d flip the chessboard over and say, “Jeff, we’re playing ping pong now.” Jensen had been a ping pong champion growing up and was very good at it. He’d play Barr in ping pong and beat him badly, just to feel better. Jensen needs to win—he’s extremely competitive. It shows his character. Even when playing chess with a co-worker, something relatively unimportant, he needed to win.

Jon:

When you talk to other people at NVIDIA, do you feel like you’re speaking to mini Jensens? Do they repeat what he says?

Tae Kim:

There are a couple of people who adore Jensen and want to be like him. For them, it’s a life goal. But beyond that, there’s a wide variety of personality types at NVIDIA. The one consistent thing is the work ethic. People who stay at NVIDIA have an unbelievable work ethic. Those who don’t usually burn out within a few years.

Personality-wise, there’s diversity in how people relate and work, so the company isn’t full of “mini Jensens.” But everyone appreciates his work ethic and the culture at NVIDIA. Many people who leave NVIDIA struggle to adjust at other organizations. It’s hard to be blunt and direct at other companies because people get easily offended. It’s also hard to move fast with decisions because you have to get buy-in from multiple people.

That’s the big takeaway—people often have a tough time working elsewhere after working at NVIDIA.

Jon:

Do you think NVIDIA survives without Jensen?

Tae Kim:

I’ve been asked this a few times. I think it really depends on the CEO who takes over. I don’t think that’s happening anytime soon because Jensen keeps saying he loves NVIDIA, it’s his life, and he loves working. I don’t see him leaving NVIDIA anytime soon. But eventually, someone will take over, and it really depends on who that person is.

Are we going to get a Satya Nadella? Or are we going to get a Steve Ballmer or John Sculley? Every major tech giant’s future depends on the quality of its CEO. It’s hard to speculate on who could fill Jensen’s shoes at this point.

Jon:

What do you think motivates him to keep going? He’s one of the longest-tenured CEOs, and now NVIDIA is one of the biggest companies in the world. What drives him?

Tae Kim:

I think he’s just very competitive, and he loves what he does. He feels he’s having a huge positive impact on the world. He’s super excited about how NVIDIA is enabling digital biology, potentially curing diseases and cancers. I think that drives him.

He’s also extremely proud of what he’s built with his corporate family. He truly sees NVIDIA employees as his family. He’s very hard on himself about the few layoffs NVIDIA has had in its history—he still feels bad about them. But he’s proud of how he’s helped his employees and their families succeed. He knows his work enables them to pay off college educations and support their families financially.I think that pride in helping his employees and their families keeps him motivated.

Jon:

When speaking with him, when you're interviewing, what sort of questions you feel got him most worked up?

Tae Kim:

He didn’t like how the early employees kind of look back at NVIDIA's early years nostalgically. He did not like that. He even joked that he was kind of embarrassed by all the mistakes and mismanagement.

He doesn’t view NVIDIA’s first 10–15 years as a great origin story, even though I think it is. It was strange having to defend those early years to him. I told him, “You grew faster than any other chip company in history, and you’re calling it a mismanaged company.” But that’s part of why he does well—he’s so afraid of becoming overconfident and complacent that he tricks himself into believing, “You suck. You’re terrible. You’re not good enough.”

There’s a story where they had a blowout quarter, and Jensen walked into the room and said, “In the mirror today, I told myself, ‘You suck, Jensen.’” On the flip side, when things are really hard and he’s worried about how to tackle something, he tells himself, “How hard can it be? We’ll figure it out.”

He uses these mental tricks not only on himself but also with others, using catchphrases like “How hard can it be?” It often turns out to be pretty hard, but by convincing himself mentally to avoid both overconfidence and despair, he finds balance. It’s almost like he’s his own personal psychologist.

Jon:

One of my favorite conversations you had was the one with Rick Tsai from TSMC. You mentioned the concept of “rough justice.” Can you explain more about that?

Tae Kim:

That conversation with Rick was one of the coolest things, and we had a nice, long conversation about other stuff too. That’s one of the coolest things about this book—you get access to these major historical figures, and you can ask them questions about everything, and they’re very open.

Rick said that when he was the main contact at TSMC for NVIDIA, he came to meet Jensen pretty often. They went to a mid-level restaurant—not quite Denny’s, but above that. They were having some issues with manufacturing. This was one of the early meetings between Jensen and Rick, and Jensen said, “Let me tell you this, I have this business philosophy of rough justice.”

What that means is sometimes, in any particular deal or interaction, one side might get 60 percent of the benefit, and the other side gets 40 percent. But next time, it might flip, where the other side gets 60 percent, and you get 40 percent. It’s never going to be perfectly 50/50. But if you work together and have a strong, trusted relationship, where over time it balances out to about 50/50, that’s a great partnership.

Rick thought that was really brilliant. He said sometimes TSMC might need to get the better end of an interaction, or NVIDIA might need extra help, like during low yields or other problems. The key was being there for the other partner at different times, and over a few years, it balances out to even. Rick said that’s exactly what happened.

Rick also said this is one of the strongest partnerships in technology history, lasting nearly 25 years. He compared it to a friendship—when your friend is having issues, you’re there for them, and they’ll be there for you when you’re going through problems. That’s how they did business.

Jon:

What things do you think NVIDIA didn’t do that were significant to their success? What missteps did they avoid?

Tae Kim:

They did try to get into mobile chips. There are stories about them trying to get Samsung to use their chips. I don’t know if they ever got the Samsung Galaxy contract, but I don’t think being a supplier, like Qualcomm in smartphones, would have helped them. That’s not really what they do. Their products tend to be high-performance and high-power, not low-power.

It’s kind of funny because NVIDIA is likely to do some AI PC stuff within a year, but mobile wasn’t their strength. It would have diverted focus and management if they had succeeded with Tegra. They didn’t, aside from the Nintendo Switch, which is a very small part of their business.

Jon:

Do you find it’s a challenge now that the company is so different? How does Jensen instill fear and drive when the employees are so successful, with many likely being multimillionaires from NVIDIA stock?

I think he just retains the right people. Many senior employees I’ve talked to have stayed for 15, 20, 25 years. They’ve been wealthy for a long time, yet they still work like crazy. These are people for whom this is their life. They love designing and engineering great products that are changing the world in so many ways, especially now with AI.

We’re not talking about video games anymore. We’re talking about robotics, digital biology, and other technologies that could change the world. Jensen brings in people passionate about the mission. From what I’ve seen, the people I’ve talked to are not slacking off at all. And if you do slack off, Jensen will be all over you. People who do want to slack off retire—many have retired.

Jon:

What do you want readers to take away from the book? What do you want them to think about and reflect on after reading it?

Like I said, the ideas of “mission is the boss” and “speed of light” are more effective ways to work and be productive. I believe that. I also want people to reflect on corporate culture. In many large corporations, the people who become CEOs are nothing like Jensen. They’re schmoozers, MBAs, and finance types. I wish every company would put technically competent people in leadership roles to make decisions, not just marketing or finance professionals.

Finally, NVIDIA’s story shows there are no shortcuts. Jensen talks about how success requires pain and suffering. It’s funny but true. It’s through trials that you learn, pocket those lessons, and improve. I want people to understand that. The internet is full of ads and influencers promoting “get rich quick” schemes. The Jensen story and NVIDIA’s history show you have to work hard, endure pain, and put in effort. Only then do you get the reward.

That’s the main takeaway: no shortcuts—just hard work and perseverance.

Jon:

Thanks. That really stuck with me as well. Tae, thank you so much for taking the time. I really appreciate it.

Tae Kim:

Jon, thanks so much for having me. It was a blast. Thanks again.

A datacenter with 15 megawatts of IT capacity is estimated to use about 80-130 million gallons of water each year.

That is as much water as three hospitals, or two 18-hole golf courses.

The current AI boom has the world's tech giants building big data centers across the world.

We all know they need a lot of energy. But what about water?The two are very closely tied together.

This channel has a not-so-secret obsession with water. In this video, we look at the data center's massive water footprint.

Types of Data Centers

Data centers come in all sizes and functions.

Some small enough to fit into a closet. Or massive custom-built facilities that span entire football fields. Generally we classify them by floor space, the number of servers inside them, or how much power they consume during operations.

The biggest facilities are referred to as "hyper scale" - having about 5,000 servers and being 10,000 square meters large.

Facilities built by global-scale operators like Google or Meta are often used for their global-scale applications - Gmail or Facebook or something like that.

But data centers can also be leased out to customers. For security, latency, and reliability reasons, cloud providers need to build these cloud data centers in a range of zones. These hyperscale facilities are so large because of energy efficiency scaling and economics.

We measure a data center's energy efficiency through power usage effectiveness or PUE. PUE is calculated by dividing the total energy delivered by the total energy going to the ICT equipment.

So the most efficient possible efficiency rating would be 1.0 - meaning that 100% of the energy going in is used on ICT.

The larger a data center is, the lower their PUE tends to be. Google and Microsoft claim that their hyperscale data centers have a PUE of 1.2 or 1.1. A closet data center on the other hand might have 2.5 or so.

This gap is almost entirely attributable to cooling. Hyperscale data centers can afford more efficient cooling systems, modeling airflows through the aisles or even employ liquid cooling systems. So let us talk cooling.

Cooling

Almost all of a data center's consumed electricity is converted to heat.

Even if a data center isn't working at its full capacity - and they rarely are - it is still withdrawing 60-100% of its maximum power. That is a whole lot of heat.

To prevent long term damage, electronic equipment have to be kept cool. Hard disk drives need to kept at temperatures of about 45 degrees Celsius or 113 degrees Fahrenheit. Solid-state compute and storage chips have higher limits - 85 degrees Celsius.

Maintaining that stable temperature - and humidity - is the job of the cooling system. There are two types of cooling systems - air and liquid cooling systems. Most data centers use the former as part of a raised-floor system.

The racks are elevated about 2-4 feet above the ground. The underfloor area might have some cables running through it, but it is mostly cold air.

Cold air comes out of the data center's Computer Room Air Conditioner or CRAC. The air is shepherded into cold aisles, sucking heat from the servers.

This hot air rises and is then re-collected to send to fluid heat exchangers for transfer out of the room. There are a few systems but the one most commonly used has two fluid loops - a process and condenser loop.

In the process loop, heat is taken off the data room floor by a fluid refrigerant - usually a mix of water and glycol. The heat is then transferred to the water-based condenser loop for final transfer to the outside.

This two-loop system is reminiscent of those for nuclear power plants and prevents contamination between the inside and outside. It also grants greater operating flexibility and efficiency at the cost of being more expensive.

At the end, we have a set of cooling towers to cool the water in the final loop. Inside the tower, the hot water flows down and some of it evaporates - releasing energy. The cooled water returns into the system for reuse.

The evaporated water leaves the tower as steam. About 1% of the water evaporates for every 10 degrees Fahrenheit of cooling (~5.6 degrees Celsius) - though this depends on ambient temperature and humidity conditions.

How much water is used each day depends on the size of the CRAC it must support. Regardless, evaporated water must be replaced with new, make-up water.

Energy Usage

The second major way that data centers use water is an indirect one but also far larger: Energy.

In 2022, data centers made up about 1 to 1.3% of the world's electricity consumption.

Water is withdrawn to generate certain types of power. Anything thermoelectric like coal, natural gas or nuclear, where we are boiling water to spin a turbine needs water.

In 2021, 73% of the US's power came via thermoelectric means - per the US Energy Information Administration. That has been trending down recently as the mix has moved away from coal and towards natural gas.

Per a study by the Berkeley National Laboratory, water usage via energy generation can be as much as 2-3 times larger than what is directly consumed by the cooling systems.

Evaporative cooling towers might consume thousands of gallons of water. But without them, data centers must resort to air conditioning systems which use even more power and thus water, indirectly.

Companies like Apple are working to have their data centers use more renewable energy, or run programs to offset their use. But for the most part, a data center pulls power from its local grid.

The Big Companies

We can take a look at a few companies and their water usage.

In 2022, Google said that they used about 4.3 billion gallons of water for cooling. They say that 25% of the water they used is reclaimed wastewater, or seawater.

Digital Realty is a large real estate investment trust that specializes in data centers, with over 300 facilities around the world.

They report their water sources with a great deal of helpful breakdowns. About 36% of their water comes from municipal non-drinkable sources.

They also report that roughly half of their 2022 water consumption were in areas experiencing some form of water strain. I laud them for reporting that stat.

AWS does not mention how much of the water they use comes from potable sources, but their 2022 Sustainability Report does say that 20 of their global data centers use recycled water for cooling.

16 of those data centers are in Virginia, 2 in California and 2 in Singapore.

Microsoft's 2022 sustainability report discusses their commitments to getting to water positive, which includes sponsoring projects that replenish more water than they consume - reaching a million people with water access. Which is always good.

Water Sources

Of course, this does means that for Google, Digital Realty, and probably Microsoft and AWS, the majority of the water directly withdrawn and used for cooling is drinkable water from the local water supply.

Some 40-50% of the global population lives in areas suffering water scarcity, and many data centers need to be located near large population centers. Which means adding a new and competing demand for water.

SemiAnalysis reports that one of the leading states in the US for data center buildouts is Arizona. Builders include Microsoft, PayPal, Meta, and more.

This is probably because it is so sunny - the state grew its solar capacity by 20% in 2023. Land is also cheap, the government is rather business-friendly, and there is relatively low risk of natural disasters like earthquakes.

But the area also experiences drought-like conditions from time to time and is highly dependent on the Colorado River for its water.

The concern seems to be top of mind for companies like Meta. For their part, Meta announced that they will fund projects to restore water in the Colorado River and Salt River basins and source its water using long-term storage credits so no water is taken from the municipal area.

And it will also leverage direct free cooling to cut water usage by 60%. What's that?

Free Cooling

Okay.

So is there any other way to cool down a data center without running a huge HVAC system or evaporating thousands of gallons of water each day?

Probably the most widely adopted water-free cooling system is "free cooling", or utilizing nature to cool things down.

The simplest implementation of free cooling is to leave the windows open: Direct free cooling.

But outside air is rarely of high enough quality to be indoors - it has smoke, dust, gases, what have you. It will ruin the electronics. So we need dehumidification, air filtration, and cleaners.

So certain areas that would require the use of dehumidifiers or filters will partially defray the cost benefits of direct free cooling. So total energy savings are dependent on location.

But generally, free cooling works. Studies done of data centers in various locations around Europe found that direct free-cooling cuts some energy consumption no matter where the center is. Average savings of about 5.4-7.9%.

Studies done in Australia found that direct free cooling can save up to 60% in the southern capital cities. A good thing since 60% of the country's energy comes from burning fossil fuels.

The potentially significant advantages of a cold climate are why several Nordic governments are chasing data center foreign investment ... and in recent years, getting it too.

Waterside Free Cooling

I suppose I should mention here Microsoft's Project Natick.

This when Microsoft sealed a data center into some metal Tupperware and dumped it underwater for a while. It works, but I doubt many people will be comfortable with putting $30,000 H100s under the water.

But there is something similar that might make a lot of sense: Waterside free cooling. Data centers located near cold water seas can use that water to cool their systems.

Google's Hamina data center in Finland is an example. The data center is a converted paper mill factory that takes in fresh seawater from the ocean using the paper mill's existing pipes.

Thanks to the two-loop cooling system, the seawater does not mix with the fluid used to directly cool the datacenter. After cooling the process loop fluid, it is then mixed with seawater again for cooling down before being sent back out into the sea.

Heat Recapture

Another idea would be to recover the heat and reuse it for something else: Heat recapture.

Recaptured heat can be used many ways: Desalinating water, pre-heating water in thermoelectric plants, direct power generation. Or, we can simply pipe it to people's houses for general heating or to produce their hot water.

This is certainly possible. Captured heat from air cooled data centers comes out hot enough - about 35 to 45 degrees Celsius.

And space/water heating is the single largest end use of energy in a home, accounting for about 6% of America's total energy consumption.

The issue is that we cannot efficiently move heat as far as we can move electricity. The demand sources - i.e. the households - need to be relatively close to the data center.

A second issue concerns infrastructure. For instance, district heating infrastructure. Not every city will have the network of insulated pipes needed to bring heat to homes, and few are willing to pay to have it built despite the financial benefits of doing so.

Score another point for the Nordics, which has done a lot of work here.

Anyway, if data centers do do heat recapture - and they should, despite the payback issues - then they will likely also adopt liquid cooling solutions too.

Liquid is far better at capturing and transferring heat than air is. That cuts down on overall energy consumption and also improves processor performance to boot.

There are a number of ways to do it, but it could involve directly bringing cold liquid to the chips. I am sure there is some Linus Tech Tips video about such a setup.

Temperature Ranges

Finally, there is a point to be made that we do not have to cool the data center down all that much.

The American Society of Heating, Refrigerating and Air-Conditioning Engineers or ASHRAE maintains a set of data center cooling standards.

They recommend a temperature range between 15 to 32 degrees Celsius. Some operators run even cooler, assuming that electronics perform better in colder conditions.

Raising a data center's temperature by even 1-2 degrees has significant financial benefits. So it is worth testing the boundaries.

Yet there is not a lot of information on this and what has been observed seems to be rather contradictory. Google has run a few tests showing that their data centers are operable at even higher temperatures.

Yet at the same time, there are some Google tests which seem to show that hard drives perform worse at cooler temperatures. So who knows?

AI Boom

One thing is for sure: The ongoing AI boom will have these massive data centers ramping up faster and using more energy.

Microsoft's 2022 sustainability report mentions that their 2022 water consumption leapt 34% from 2021. Considering there was only a 13% increase from 2020 to 2021, it appears that this jump is due to ChatGPT and other generative AI products.

Bob Blue - love this name - CEO of the utility Dominion Energy, serves the Northern Virginia data center cluster and said in an earnings call:

> For some context, historically, a single data center typically had a demand of 30 megawatts or greater. However, we're now receiving individual requests for demand of 60 megawatts to 90 megawatts or greater, and it hasn't stopped there.

> We get regular requests to support larger data center campuses that include multiple buildings and require total capacity ranging from 300 megawatts to as many as several gigawatts.

SemiAnalysis did an excellent job analyzing this massive coming demand. They believe that AI will propel data center share to 4.5% of global energy generation by 2030. Water and power consumption are correlated. Just imagine water consumption growing that much too.

Nvidia's coming GPU products like the B100 and beyond are getting even more power hungry. Training a leading edge model will run these GPUs to the max, which means more power and water consumption.

At the same time, deploying them will require immense scale which ... again means more power and water consumption.

Conclusion

So no matter how you slice it, it seems like the future of compute and AI will require far more electricity - and water.

First at the semiconductor fabrication stage. Here in Taiwan, TSMC alone is responsible for 6% of all the energy that Taiwan will consume. By 2025, that is estimated to grow to 12.5% - thanks in part to new EUV lithography machines.

Then after the chips are fabbed, we put them into these data centers and run them to full tilt. They then generate heat that we need even more energy to move out of the room.

If things develop as they seem to be developing, future data centers will need to rapidly adopt a combination of free-cooling, waste heat recovery, and renewable energy like solar. The faster this transition happens, the better. It is both a question of sustainability and money.

Sign up for AiSalon Taiwan @ Appier

At the end of World War II, Japan was at zero.

Then Godzilla came and made it Minus One.

In the wake of this ruin, an extraordinary company was founded. And through a series of fortuitous events, they came across an extraordinary technology.

The discovery of the first transistor set off a race around the world to produce and use it. What can it be used for? How do we manufacture it?

Amidst this race, this scrappy company - led by an extraordinary pair of founders - defied the odds and used the transistor to make a breakthrough radio.

In this video, we look at how Sony mastered the transistor.

Beginnings

I know the Acquired podcast did an episode on Sony's beginnings already. They did a great job, but stick around for this one.

Akio Morita and Masaru Ibuka first met in 1944. They had been assigned to a team to produce a heat-seeking missile - the Ke-Go - and save the country from a losing war.

Ibuka was then 37 years old. Taller and heavier, he wore thick glasses and spoke with a working class accent. He was the chief engineer of a company supplying vacuum tube-based voltmeters and radar frequency control devices for the military.

Morita was then 24 years old. Shorter, but thin and with an aristocratic side profile. He was then serving as a lieutenant in the Imperial Navy, a naval liaison officer.

Working on this missile project, the two men spent a great deal of time together. The project failed, but the two became friends.

When the Allies started bombing Tokyo, Ibuka relocated the company to an apple orchard in Nagano. After the war ended, Ibuka brought his seven engineers back to Tokyo and set up a little lab in a room on the third floor of the Shirokiya Department Store.

Return to Tokyo

Times were hard back then. The first autumn, they took rucksacks to the countryside to get rice and potatoes.

Ibuka - who long considered his greatest asset to be his team and worked hard to keep them engaged - paid his staffers out of his own pocket.

They thought about making anything from golf equipment to sweet bean soup. But they soon found work repairing old radios during the military era.

It got them featured in a column in the Asahi Shimbun newspaper, which caught the eye of Ibuka's former colleague, Akio Morita.

Morita reached out and the two rekindled their bond. Ibuka wanted Morita to come to Tokyo and join him, but since that was risky, Morita thought to split that with a part-time teaching job.

But then the American occupation General Headquarters or GHQ banned former military officers from teaching. That fateful decision let Morita go all in on working with his friend.

Tokyo Telecommunications

There was one more obstacle, Morita's family.

Morita's family ran one of Japan's biggest and oldest sake distilleries - been around for over 300 years. Akio's father took over the business while it was in a sorry state and built it up again.

Thusly, he groomed Akio - his first-born son and 15th generation heir - to eventually take over and run the business. But Akio wanted to be an engineer, building vacuum tube radios and dissembling appliances in his youth.

After a momentous meeting, the elder Morita agreed to - for now - release his son from the obligation to run the family business on a day-to-day basis. Though Akio did become its chairman and dutifully attended every board meeting.

The Morita family also agreed to invest what is said to be about $60,000 - which even for them was a lot of money - into the newly founded Tokyo Tsushin Kogyo Kabushiki Kaisha.

The name literally means "Tokyo Telecommunications Engineering Company".

Later, whenever the company sought more capital for growth, Akio did like as any good millennial does and went back to his parents for more money in exchange for stock. At one point, the family owned 17% of Tokyo Telecommunications.

The private funds of the Morita family plus the connections of the company's first president Tamon Maeda - who also happened to be Ibuka's first father-in-law - kept the little company alive where many other similar businesses died.

Morita and Ibuka

Morita and Ibuka worked well together. Correction. They were absolutely perfect for each other.

Both were extremely technical. I already established Ibuka to be a talented and creative engineer, but Morita was no chopped liver. He was a trained physicist who ghost-wrote a science column for the Asahi Shimbun in his early days.

Ibuka was 13 years the senior. Quiet but impulsive. A dreamer type with an idealistic vision. Maybe a bit naive. His goal in life was to make electronics for everyday people, but honestly he lacked a good sense for what those people wanted.

In the beginning before Morita arrived, Ibuka invented and tried to sell an electric rice cooker. It looked like a wooden tub and didn't cook rice well, which is a fairly important prerequisite. The company made a hundred and sold none of them.

Morita on the other hand is animated and energetic. He largely ran the business, and liked to move fast to make things happen. Morita is strict, but towards Ibuka he was astoundingly protective, saying:

> Ibuka is a such a warm and honest person, that I saw at once that I had to become tough and shrewd to protect him. My mission is to realize Ibuka's dream.

Without Morita, Masaru Ibuka would have never sold any of his inventions. And without Ibuka, Akio Morita would not have anything to sell. They really were perfect for each other.

And unlike another famous engineer-sales co-founder pair, people at Sony never recalled Ibuka and Morita fighting. It is a bond described as something like love.

Tape Recorder

The company started off producing whatever it could to survive.

One of their first successes was a vacuum tube-based voltmeter first invented by an engineer on their team. They sold maybe about 30 to 40 of these each month, which kept the doors open and the workshops busy.

The company also grew close ties with the broadcaster NHK, providing high quality broadcasting equipment and other professional equipment to replace that which had been lost in the war.

These projects kept food on the table, but Morita and Ibuka wanted more - a machine for the consumer, the ordinary man. During a visit to the GHQ, they came across a paper tape recorder used for radio transmission.

The team knew audio, and they felt that a tape recorder might be well received by the Japanese. They acquired a few patents, hired some engineers, and with the help of some nifty reverse-engineering, produced the G-type tape recorder.

Tape recorder performance heavily depended on the quality of the tapes. But plastic was then still hard to come by. So the team made magnetic tape from paper covered in iron oxide powders and lacquer.

In the beginning, they had people run back and forth across the paper with an hairbrush. The first tapes were so bad that you could not even hear someone say "Moshi Moshi" or "Hello" in Japanese. Not to mention the chonk baby weighed 35 kilograms and cost 170,000 yen or about $14,400 in today's dollars.

With the help of a trading company owned by the descendants of the Tokugawa Shogunate - yes, really - they sell their first tape recorder to an Oden shop in Tokyo station where it played a jingle to attract customers.

The Ministry of Justice bought the G-type in higher volumes, leading to new versions like the H-type and the more portable P-type. Morita, a talented salesperson, helped sell plenty of tape recorders to companies and government. But yet again, consumers spurned the big bulky machine.

Ibuka remained vexed. Whilst in an antiquities shop, he saw someone buy a carved ivory thingy that he felt was worthless, paying a crazy high price. Frustrated, he thought "Why would someone buy this when they can buy Japan's first tape recorder?"

The answer, of course, is obvious. The thing was too chonk and too expensive. It needed to get smaller and cheaper. But what technology can possibly help make that happen? Yeah, I wonder ...

Licensing the Transistor

Japanese companies in the early 1950s looking to get into transistors had two options.

The technology had to come from America, obviously. But from which American company? You can sign a patent licensing agreement with either RCA or Western Electric, Bell Labs' commercial telephone device subsidiary.

The first transistor that Bell Labs publicly announced in 1948 was the "point contact transistor". It can amplify signals like a vacuum tube can, but it turned out to be too unstable for commercial use.

Soon thereafter, Bell Labs' genius William Shockley announces the junction transistor, a more resilient and commercially viable transistor. RCA's physicists immediately recognize it as the one.

So RCA did their own crash program on transistor development and made rapid progress. One of their big successes involved a variant of the junction transistor first invented by GE, the alloy junction transistor.

You made it by heat-fusing or alloying two very small pellets of indium about a millimeter large on opposite sides of a thin slice of germanium. RCA improved GE's processes and made it more commercially viable.

By 1952, they started licensing their patents and holding symposiums to teach people how to make their transistors. RCA did this because it was their business model. Until 1957, they licensed their patents as part of a big expensive bundle, "package licensing".

Patent licensing as a business model incentivized RCA to also sell agreements to teach people how to use their technology. This made them the preferred licensor of transistor technology - some 80 companies attended their first symposium.

By contrast, Western Electric's main business was making telephone product. They were only licensing out the transistor technology because of their anti-trust agreement with the US Government. Such an agreement did not mean they had to teach people how to adopt said technology.

It reminds me of how Bell Labs treated Unix's early licensees. They just sorta threw it at people and said “Here, go nuts”. And kind of like the Unix situation, this turned out to be a blessing in disguise.

Western Electric

In 1952, Ibuka takes his first trip to the United States to check out the American tape recorder industry for learnings.

But while in New York City, he learns through a stockbroker friend Shido Yamada about the Western Electric transistor licensing opportunity. $25,000 US dollars paid upfront against future royalties for some transistor knowledge.

Why Western Electric and not RCA? Most likely, Ibuka just couldn't get in touch with them. Moreover, Tokyo Telecommunications probably would not have been able to afford the RCA package license anyway.

At the time, Ibuka knew as much about transistors as I know about European soccer, but he is immediately convinced that it is the next big thing. Ibuka reaches out to Western Electric. Suspicious, Western Electric asks for a lot of detailed information like company history and recent financial reports. But eventually he convinces them the sincerity of his outreach.

$25,000 happens to be a lot of money - about $300,000 today. Worse yet, you needed it in dollars at a time when US dollars were hard to come by in Japan. The government tightly controlled its disbursement.

Upon returning to Tokyo, Ibuka goes to MITI, Japan's top industrial policy body, for the money and was ridiculed. A small company with no experience even in making vacuum tubes. Why do they deserve the foreign currency for this?

A Risky Move

But Ibuka is undeterred. He has now come up with something to use the transistor technology for: Radios.

Radios small enough for a single person to even put in his pocket - "pocketable". Ibuka gets fired up over the idea and when he is like this no one can stop him.

In August 1953, Ibuka takes a risk and sends the charismatic Morita over to New York City to secure the license. Morita barely knows English.

But he is so persuasive and anxious to get going that Western Electric grants him a license, provisional to the Japanese government releasing the funds. Western Electric's VP of licensing Frank Mascarich recalled later:

> I wasn’t terribly pleased with the arrangement ... but he was so persuasive, and so anxious to proceed with his plans, and after all, it took considerable time and expense for him to travel to the United States from Japan, that I decided to give him some of the technical information so that he could take it back with him and immediately embark on his project to manufacture transistors.

Thusly, Morita came back with a few sample transistors, Germanium crystal, and a copy of the book "Transistor Technology", edited by members of Bell Labs.

In July 1953, even before Morita left for New York, Ibuka formed an elite five-person transistor task force led by Morita's brother-in-law, Kazuo Iwama.

Iwama grew up with Akio, their two families having known each other for a long time. Despite originally trained as a geophysicist, Morita had him join Tokyo Telecommunications soon after its founding.

35 years young, rational and very level-headed, he ran the tape recorder division before being switched into transistor production. He too knew nothing about transistors, but had the tremendous energy to go learn.

But first, MITI. MITI is obviously annoyed as heck that Tokyo Telecommunications signed a deal without first running it by them. However, Ibuka cannot be denied. He goes to MITI and lobbies the bureaucrats, concluding:

> "We intend to move forward with or without you ... but if you approve our deal with Western Electric and give us some development money, you will look smart!"

A change in leadership shifted MITI's opinion on the discretion, and they finally approved the funds in January 1954. Iwama and his team spent the time reading Shockley's 1950 book "Electrons and Holes in Semiconductors", and then the Bell Labs "Transistor Technology" book after that. Neither book was very useful.

As soon as MITI approved the funds, Ibuka and Iwama boarded a plane to New York to take the next step in transistor technology. Meanwhile, the rest of the team back in Japan starts cobbling together a primitive semiconductor manufacturing line - furnaces, crystal pullers for making crystal, assembly tools, and so on.

Learning It

Unfortunately, the transistor as transferred over from Western Electric was not good enough to be used in a transistor radio.

Western Electric originally thought Morita and Tokyo Telecommunications were going to use their transistors for hearing aids - a reasonable thought. Some of the first transistorized products were hearing aids.

Upon learning that Morita wanted to use it for a radio, they told him that twelve other licensees were then trying to make transistors capable of the higher frequencies for radios. None had yet succeeded.

Nevertheless, Ibuka and Iwama got on a plane and flew to Western Electric's transistor factory in Allentown, Pennsylvania. But there, faced with what it means to run a transistor business for the first time, Ibuka wondered if he might have gotten himself into a pickle.

But Iwama did his homework. He walked across the factory floor, taking in everything. Whenever he saw something of interest, he stopped and tried to ask about it in the best English he possibly could.

At night, he writes and sketches everything he possibly could remember. Pages of documents - referred to within the company as the "Iwama Report" - start flowing back to the team in Japan in February.

Until April 1954, Iwama visits several Western Electric, Bell Labs, and even Westinghouse transistor factories - eventually filling four folders with his letters. He and his team compared what they saw in the factory with what was laid out in the Transistor Technology book for clues.

Which Transistor?

What Tokyo Telecommunications was trying to build is a superheterodyne receiver. It works using the principle of the same name and is what all the modern FM radio receivers do.

A radio signal is put out by the radio station.

The superheterodyne receiver receives that signal, amplifies it, and then mixes it with another high frequency signal created by an oscillator.

The resulting mixed signal, called the Intermediate Frequency or IF, is more convenient to process and amplify before turned into an audio signal.

High frequency transistors are needed for the oscillator and mixer to create the Intermediate Frequency as well as to amplify the IF itself. The core issue in transistorizing the radio is creating those high frequency transistors with good yield.

Iwama had to decide which of the two existing variants of junction transistors to pursue: Alloy, which we discussed earlier, or grown junction transistors?

The junction transistor is a sandwich of three components - the emitter, base, and collector with the base in the middle. The PN junctions or barriers in between these areas give the junction transistor its name.

Pioneered by Bell Labs, the grown junction transistor produces the sandwich using the Czochralski process - where you pull out or "grow" a whole crystal out of its melt from a small seed.

By selectively adding dopants as we pull out the crystal, we can produce the transistor sandwich at the same time as the crystal.

During his visits, Iwama drew these crystal-pulling tools and how they carefully pulled the crystals out of the melt. He asked Western Electric for detailed schematics of the machines but was rebuffed. Sad.

The difficulty and cost in recreating the crystal-pulling machine forced Iwama to decide for the team to start making PNP Germanium alloy junction transistors.

Some of Tokyo's Telecommunications' first alloy junction transistors were completed in June 1954. A few were inserted into a prototype portable radio, but failed to make the cut. High frequency was not achieved.

Grown Junction

So what now? The Tokyo Telecommunications team went back to their notebooks.

If you recall, you make an alloy junction transistor by soldering two beads of Indium onto the sides of a crystal slab of Germanium. This particular arrangement got you a positive-negative-positive or PNP transistor.

It was theorized that it might be possible to achieve faster frequencies by reversing the polarities, so negative-positive-negative or NPN rather than PNP. This was based on the notion that negative charge carriers or electrons travel faster than positive ones, or electron holes.

At the same time, Iwama came to believe that the alloying process cannot be scaled to high volume production. His conceptual reasoning was that the grown junction process was simpler.

With that, you have one step, the crystal pull. By contrast, the alloy junction process had two: First the crystal pull and then the alloying of the beads onto the sides.

So the team shifted to making NPN Grown Junction transistors, which meant finding the right dopants to add to a growing Germanium crystal so to produce this unique NPN arrangement. The team tried adding many different dopant impurities to produce the right effect.

During this process, a scientist named Tetsuo Tsukamoto - a Japanese born in Taiwan - tried phosphorous and antimony, which failed. He then tried phosphorous and indium, a weird combination.

But amazingly, the resulting transistor had a higher frequency when tested. It was incredibly encouraging ... except Tsukamoto could not replicate the result despite how many more efforts.

Another team went to the library and found an early issue of the Bell System Technical Journal, the Bible of semiconductors, with an article saying that phosphorus doping did not work. It was incredibly discouraging for the team members, who saw Bell as the "voice of the gods".

Right then, Iwama dropped by to see on things. He listened to their efforts and the discovery of the Bell journal article ruling out Phosphorous and said:

> But I remember you said you succeeded once. I still hope you will succeed again. I will take all the responsibility, so don't worry if you use a little more time. Why don't you try again?

So Tsukamoto kept at it. Three weeks later in defiance of Bell's expectations, he discovered a phosphorous doping recipe that produced a suitably high frequency. Initial yields were at 5%. And to Western Electric's horror, Ibuka skipped a pilot plant and went straight to mass production.

The Transistor Radio

The 1954 news of Texas Instruments producing the first miniaturized transistor radio, the Regency TR-1, spurred the team at Tokyo Telecommunications to work even harder. Now, they knew that a radio was possible.

In early 1955, they completed their radio, the TR-55. It used five transistors, three were grown junction and two were alloy junction types. The grown transistors for the oscillators were still so unstable that factory workers had to pick the right one for each individual radio set.

In August that year, the TR-55 went on market in Japan only, selling for the equivalent of about $980. It did not sell that well at first. People docked the relatively weak sound compared to vacuum tube based radio receivers.

But in the summer of 1956, young people started warming up to the little device when they realized that the cost of its tiny battery was only a twentieth that of tube-based receivers. Improving transistorization resulted in smaller and more successful radios.

Later, they realized that Tokyo Tsushin or Tokyo Telecommunications did not make much sense as a name, especially in the United States. So they coined a new name for the product - and later the company - Sony.

Conclusion

In producing a proper transistor for a commercial transistor radio, Sony had to master both the theory and engineering.

Even the Bell Labs' famous "cookbook" was not detailed enough for Sony to reproduce a semiconductor line. They had to send one of their own to the United States to review and document everything. Yes, in some ways you can say they copied the Americans.

But mere copying wasn’t enough. They still had to dissemble what they thought they knew and go back to basic first principles. Iwama's decision to scale grown junction rather than alloy junction. The realization to switch to NPN rather than trying to make PNP work.

These are not possible without thoroughly understanding semiconductor physics. When you are breaking new ground in electronics at commercial scale, this mix of theory and practicality is necessary to succeed. And a lot of sweat, blood and tears of course.

Like as I wrote before, I want to thank Professor Shintake as well as Dr. Patrick Naulleau of EUV Tech for their help and consultation. They are the true experts of this domain.

I do want to note that ASML chose the number of mirrors they did for a very real reason. It gives them full ability to print specific features on the mask. Simplifying the number of mirrors that away so we should keep this in mind. Perhaps fabs in the future adopt a mix of lithography machines - doing some very complicated layers with those machines and others with simpler mirrors.

A mirror inside ASML's EUV lithography machine reflects just 70% of the EUV light it receives.

With 10-12 reflections in the machine, this can get inefficient. Just 1% of the photons hit the wafer. Electrical power efficiency is said to be less than 0.2%.

It also contributes to troublesome stochastic defects, since not enough EUV photons hit the resist to overcome quantum effects.

So a recent paper from Professor Tsumoru Shintake at the Okinawa Institute of Science & Technology caught my eye.

It proposes a simplified setup with radically fewer mirrors. But Shintake makes it clear to me that his system no way challenges ASML's. In fact, it should complement it.

I think this thing can work. In today’s video, I want to walk you through this interesting new thing cooking up in beautiful Okinawa.

Beginnings

We should begin with a brief overview of a commercial EUV lithography system. I am not going to cover everything, just enough to get you through this video.

First, we need EUV light, 13.5 nanometer wavelength light. The light source creates it in a number of ways - lasers hitting tin droplets, particle accelerators, whatever you want.

A mirror then collects the light and sends it through the Illumination module, which spreads out the light and makes it as uniform as possible for the mask.

That light then bounces off a photomask, a special mirror with the chip design printed onto it.

The reflected light then goes through an Optical Projection module that reduces the size of the pattern on the mask's field by some ratio and focuses it.

Finally, the light hits the resist-coated wafer. Ideally at the exact same angle across the machine's whole wafer exposure field so to avoid distortions.

I must pause briefly to explain Wafer Exposure Field again. It is a fancy word meaning the size area of the wafer exposed at once by the lithography machine.

We want the field to be as big as possible. A smaller field means we need more exposures to cover the whole wafer. That both hurts productivity and offers more chances for things to go wrong like alignment errors.

The industry standard has been 26 millimeters by 33 millimeters. It is important to be backwards compatible with that, or else your solution is not economically viable. The bare acceptable minimum would be a width of half that - 13 millimeters.

That can barely print an Apple SOC chip. The A18 Pro's longest side is just under 13 millimeters. Thanks to Max of the YouTube channel High Yield for the help on that.

Anyway. ASML offers two variants of the EUV machine - a Low and High-NA. The Low-NA machine's NA is 0.33. The High-NA machine, 0.55.

NA stands for Numerical Aperture. Roughly speaking, it describes the range of angles at which a lens can accept light. So a higher NA lets you pattern at higher resolution. Though as we will discuss, this benefit does not come free.

Let us count the mirrors for the Low-NA machine. One near the light source to collect the light, four in the illuminator, the mask-mirror itself, and then six mirrors in the projection module. That makes 12 total reflections - each of them absorbing 30% of the photons.

Schwarzschild Optics

Why so many? Over the 20+ years of EUV development, teams have experimented with the number of mirrors.

One of the first EUV systems was designed by EUV pioneer Professor Hiroo Kinoshita.

His system had a flat reflective photomask. And for the projection module, there was a two-mirror Schwarzschild optics system.

Schwarzschild optics have been used before in microscopes. The classic system has two sphere-shaped mirrors - a primary and secondary - aligned on a single axis. Thusly, the secondary mirror is in the light's path.

Professor Kinoshita's main goal at the time was to prove that you really can use mirrors to reflect EUV light onto a chip design pattern and then onto the wafer. This two-mirror arrangement succeeded in doing that, but also presented serious limitations for high volume production.

The mirrors’ spherical, curved shapes mean that they reflect the chip design image in a spherical, curved manner. This curved reflection gets more pronounced the further away you get from the center, or off-axis.

But since both the wafer and the mask are flat, the curved image is not uniformly in focus. This "field curvature" distortion as it is called limits how large we can make our wafer exposure field.

It gets worse when we start scaling the NA - which per the famous Rayleigh Criterion is one way to improve patterning resolution. Because a higher NA means capturing light at more angles, it also worsens the field curvature effect and amplifies the optical distortions.

NTT's Aspherical Systems

As a result, people moved away from Schwarzschild spherical mirrors.

In 1992, the Japanese telecom company NTT presented an early EUV system they had been working on for the past three years. Its projection module used two aspherical, equal-radii mirrors.

Aspherical, meaning not spherical. Note that these are sometimes also referred to as Schwarzschild as well.

And Equal radii, meaning that they share the same radius of curvature, but in different directions.

One primary convex mirror curving outwards and a secondary concave mirror curving inwards. This lets them cancel out each other’s field curvature distortions.

Both mirrors had holes in them to allow light to enter and exit the larger projection system.

This setup was called the "equal radii two-mirror" projection module.

In addition to the two mirror projection module, the NTT system had a reflecting photomask, two-mirror illuminator, and a synchrotron - a type of particle accelerator - for its light source.

The MET Tools

If you recall your EUV history, and it should be part of the high school curriculum if you ask me.

Then you know that in the late 1990s a consortium of American IC makers led by Intel greatly helped refine the EUV technology.

The EUV LLC consortium's product was the Engineering Test Stand. Its 4-mirror projection system had an NA of about 0.1, which was low compared to a production machine. But its exposure field of 24 by 32.5 millimeters was large enough to meet high volume industrial requirements.

But that 0.1 NA. Afterwards, the American national labs and the famous technology consortium SEMATECH decided that if we wanted to test the other aspects of the EUV ecosystem, we should do it on a small scale exposure tool with an NA more comparable to that of a production machine. Like 0.3

So in 2004, they joined together to build the Berkeley Micro or Microfield Exposure Tool or MET, which had an 0.3 NA. It adopted the two-mirror equal radii projection system and had an exposure field of 0.6 by 0.2 millimeters.

The Berkeley lab has continued working on the MET for many years since, upgrading the optics and mechanics to produce MET5 in 2020. MET5 has an NA of 0.5, a field size of 0.2 millimeters by 0.03 millimeters, and a magnification factor of 5.

By then, it was decided that if we wanted a bigger field size and better resolution, then we had to have more than just two mirrors for projection. ASML's first EUV Alpha Demo Tool - shipped in 2006 after six long years of development - had six.

More mirrors seemed the only way to achieve a suitably large field. However, the additional mirrors necessitated a more powerful light source to deal with the substantially greater power loss - from something like 10-20 watts to 200+. Doing that took roughly another ten years and it got dodgy at times.

The Shintake System

I asked Professor Shintake about how he came up with his system.

Professor Shintake began studying the nuclear sciences before switching to electron accelerator science.

He has spent thirty years researching such systems and their related technologies at places like Stanford.

In 2011, he helped build the world's second X-ray free electron laser.

The Spring-8 Angstrom Compact Free-Electron Laser. It is Japan's first such device.

He has also designed an underwater propeller for generating electricity from tidal currents. That is very cool. Anyway what I am trying to say is that he - unlike me - is not some random dude on the Internet.

In 2022, he came across a diagram of the ASML EUV system in the newspaper and wondered why the EUV light takes such a strange path. Such an off-axis path would ordinarily degrade optical performance.

Over the next two years, he studied over a hundred papers and books, and spoke with various people in the industry. During his research, he came across the Petzval field curvature theory.

Some background. The Petzval field curvature theory was made by the Slovak mathematician Joseph Petzval. Joseph is perhaps most well known for inventing the first photographic portrait lens.

Petzval's formulas suggested a way to use the two-mirror equal-radii projection system to increase the field size. To confirm his hunch, the Professor purchased some optical simulation software and ran it for three months using a number of parameter combinations.

How It Works

Shintake humbly says that his system evolves from the existing ASML 6-mirror projection system. Let me give a brief overview.

So the stages are largely the same - light source to illumination to mask to projection to wafer. So let us trace the path the light takes through the system.

First, we collect light from the light source with the goal of sufficiently illuminating the flat EUV photomask with EUV photons.

After reflecting off the photomask, the light enters the projection module chamber with the M1 and M2 mirrors inside through a hole in the M1 mirror.

Once inside the chamber, the light goes down to the M2 mirror.

Then it reflects back up to the M1 mirror, and reflects off that.

Finally after that, it goes down to the resist-coated wafer, exiting the chamber through a hole in the M2 mirror.

In total, we have five reflections just like the old NTT EUV prototype system back in the early 1990s.

Projection

Did that make sense? I hope it did. Because now we are gonna do a deep dive into the projection system.

A presentation from Carl Zeiss and optics legend Dave Shafer first inspired Shintake to make this projection module. One particular slide showed how a two-mirror, equal-radii setup can project a perfectly flat field - as in no field curvature - while avoiding other issues.

But his presentation also noted that the mask and wafer would need to be positioned inside the center of the two mirrors. So most everyone presumed that this arrangement would be impossible or impractical to achieve in practice.

Shintake played around with this and found that we can get similar-ish results by making the M2 mirror as thin as possible and as close as possible to the wafer. It still leaves us with some field curvature distortion but within parameters.

The MET also had to put the mirror as close as possible to the wafer. It can be bit tricky to execute, since it can be hard to make that mirror as thin as possible while keeping it rigid. But this can be done.

To enlarge the small field size, we make the tool much larger. In the case of the first MET, the distance between the photomask - the object producing the image - and the wafer - where the image is being formed - was about 276 millimeters or about 10.8 inches. Which as you undoubtedly remember, gave us a field size of 0.6 by 0.2 millimeters. Too small.

Shintake's proposal increases the distance between the object and the image - and thusly also the tool's height - to the very limit allowed by modern semiconductor fabs' ceilings:

About 2 meters or 6 foot, 6 inches.

Or just a bit taller than a Michael Jordan, the semiconductor industry's widely-accepted standard for height increment.

The semiconductor fab people do not seem particularly phased by this, at least nowadays. I feel like ASML's High-NA EUV tool has sort of reset people's expectations for tool sizes. But I do wonder about the second-order effects of such a big device. Shintake will likely rejigger this down the line.

Illumination

So this two-mirror projection system has been known. Based on the history we just reviewed, it is not particularly novel.

So I think what is actually most special about the Shintake system is the module located above the projection, the illuminator.

The EUV lithography tool is a modern scanner device. So the mask is exposed through a fixed "exposure slit".

You "scan" or, like, drag the mask through the light beam while moving the wafer at the same time in the opposite direction. Once the scan is done, you step on to the next part of the wafer.

It makes more sense to be honest when you look at a transparent mask rather than a reflective mirror one.

The challenge with the illumination system is that the reflection geometries are complicated. The EUV light must be positioned in such a way to bounce off the mirror and into the projection chamber.

This was a great challenge. Professor Shintake told me that he designed three previous illumination systems that did not work. Then, while cleaning his room, he came up with a clever something he calls "dual line field".

With dual line field, four smaller light sources bundled together into two larger light sources are positioned to illuminate the mask. But how to keep the sources from reflecting onto each other? As Shintake says in an interview:

> If you hold two flashlights, one in each hand, and aim them diagonally at a mirror in front of you at the same angle, then the light from one flashlight will always hit the opposite flashlight, which is unacceptable in lithography.

> But if you move your hands outward without changing the angle of the flashlights until the middle is perfectly lit up from both sides, the light can be reflected without colliding with the light from the opposite flashlights

What makes this really clever is that unlike the ASML machine, there are two fixed exposure slits.

So as the lithography machine does its scan, a spot on the mask gets exposed to the light cone from the first slit, projecting a spot onto the wafer.

But the mask is still scanning. That spot on the mask moves out of the first light cone and then gets exposed to that second light cone, projecting to the wafer. Such an arrangement can only work with a scan-and-step machine, with a moving mask.

This is very clever. This two-slit arrangement lets the two illumination cones to be offset, thus avoiding a reflection back at the sources. I am guessing this is why the Okinawa Institute has a patent application out on the dual line field. It’s key to making this all work.

The Possibilities

So what is possible with this simplified design? Let us first look at the critical dimension.

Run the variables through the famous Rayleigh Criterion equation and you find that a 0.2 NA and a 13.5 nanometer wavelength gets you a 24 nanometer half-pitch - assuming that the K(1) process factor is the same 0.35 like as with traditional EUV machines.

To compare, a leading edge 193-nanometer immersion lithography machine has an NA of 1.35 and a K(1) of 0.27. This works out to about a 40 nanometer half-pitch.

Now, there is more to consider than just resolution - throughput, power cost, overlay, maintenance, and so on. But just based on critical dimension, this looks good.

How about ASML's Low-NA EUV machines? They have about a 13-nanometer half pitch. Their High-NA EUV machines' theoretical half pitch is around 8 nanometers. So the ASML EUV machines can still do much better here.

But if we raise the NA of the Shintake system to 0.3, then the critical dimension half pitch works out to 16 nanometers. That is interesting.

0.3 is probably where this will go. Shintake told me that the 0.2 NA machine proposed in his first paper did not match up with industry standards. Namely, its field scan size of 20 millimeters and its reduction ratio of 5.

So he is rewriting the paper with more industry-standard parameters - a 13-millimeter field scan size and reduction ratio of 4. The device might also be smaller somewhat.

This would require the adoption of field-stitching techniques like those the industry is making in preparation for High-NA EUV. But we should note that most smartphone SOC chips might already be printable.

The Bent Mask

So what are the catches? There are a few complications and the paper lays them out.

But one that caught my eye is that with two mirrors we cannot avoid the projected photomask image having some bit of field curvature - which can lead to print errors.

So for additional help, he suggests that we slightly curve the mirrored photomask to correct the remaining curvature.

Yes, as in taking a regular flat mask and mechanically bending it (Bend It Like Beckham!) by anywhere from 30-120 microns - depending on NA and tool size - when mounting the thing onto the chuck.

I asked someone in the mask industry about trying to bend these quarter million dollar EUV photomasks, and he sounded concerned. The mask blanks are made from a very stiff substrate of ultra low expansion glass and they might crack.

My friend suggested making curved blanks and then printing the chip design onto that. This will also be challenging since the multi-electron-beam mask writers aren't equipped for that right now. The depth at which they can write features - depth of focus - is limited, and will need to be made variable.

More consultation is needed, but the Professor assured me that the curved mask is not absolutely necessary for the system to work.

An Imminent Breakthrough?

This paper was somewhat covered in the overseas media, but I think in an unfair way.

Overseas commentators exclaimed in typical bombastic, uneducated format: "Japan on edge of EUV lithography chip-making revolution", with the subtitle "Okinawa Institute of Science and Technology claims breakthroughs that could break ASML’s monopoly on advanced chip-making equipment" - based on what I presume to be a five minute reading of the first paragraph.

I abhor such statements because if you read the whole paper, you can see what it actually is. It is a creative re-imagining that builds upon the decades of work already done in the EUV ecosystem. It complements, not displaces.

Right now, the machine only exists in a computer simulation. I think it can work, theoretically, but we are far away from a real device. There remain many more technical hurdles to overcome, including those not yet foreseen.

If someone wanted to found a startup around this from scratch, there is so much more they would need to figure out. Like how to make the EUV light source, the mirrors, the photomask plus their blanks, the precision mechatronics to manipulate the wafer stage with nanometer-accuracy, the special new photoresists, and the software to control all that.

So Shintake intends for his system to help, not challenge the ASML systems. He re-iterated this several times during our interview.

Conclusion

Before we end, I want to thank Professor Shintake and Dr. Patrick Naulleau of EUV Tech for taking the time to speak with me, and answering my uninformed questions.

Let me wrap this up. I admire Shintake's system because it represents a rethinking of EUV as we know it today. He went back to the drawing board and came up with something wonderful. I don't see any reason why this can't work, though a commercial machine needs a lot more than just that.

So what do we do with this? The paper proposes that next we should do a proof-of-principle experiment. A real-ish lithography machine with requirements more in line with the semiconductor industry, like a 13 millimeter-wide field for instance. I think someone should try to help him make it.

Two semiconductor companies. Both alike in dignity.

In fair Europe, where we set our scene.

From ancient roots break to new silicon.

In 1987, Italy's SGS Microelettronica and France's Thomson Semiconducteurs decided to join forces. It was a historic alliance that created a European semiconductor giant - a globally competitive one at that.

The stories of these two star-cross'd silicon lovers are fascinating. In today's video, we talk about the merger that created SGS-Thomson, now STMicroelectronics.

Oh, and I do want to sincerely apologize for totally mangling these European pronunciations. Feel free to make fun of me in the comments.

An Urgent Need

Let us start in Italy. Because the history of SGS is far simpler. And also because I like pasta.

SGS begins with the iconic Italian industrialist Adriano Olivetti. Adriano was the son of the founder of an iconic Italian typewriter company. A visionary leader with immense drive and energy, he pushed his company into the computer space in the mid-1950s.

Their first prototype - the Elea 9001 - had been built with vacuum tubes. It worked fine, but Adriano and the brilliant Chinese-Italian leading his computer team Dr. Mario Tchou decided to redesign and rebuild the computer using solid-state transistors.

The resulting success convinced Olivetti and Tchou that transistors were the future. But Adriano did not want to import these critical items from the United States - where they were first invented - or West Germany or the Netherlands. They wanted to make them here in Italy.

However, the Olivetti Company had its hands full with the work of producing computers. Fundamentally, they saw themselves as a mechanical company ill-equipped to handle the challenges of solid-state device mass-manufacture.

While debating the merits of starting a full research laboratory themselves, the two hear about another Italian company working on transistors.

Starting SGS

Born in 1906, Virgilio Floriani grew up believing in the value of technology and innovation.

During the war, he worked as a designer and engineer for a military radio factory. After the war ended, he founded the Italian telecommunications company Telettra in 1946. The name is an amalgamation of "Telephony, Electronics, Radio".

Telettra introduced a series of new technologies to post-war Italy. So when Floriani realized that nobody else in the European telephone industry thought these transistor things were important, he saw a chance to seize the high ground.

So Telettra bought a license from Bell, and then in 1956 traveled to the Murray Hill laboratories to join a hundred-person symposium to learn transistors.

He did not learn much, since his English and theoretical solid-state physics knowledge were limited. But upon his return, he set up a very small laboratory in Milan to build solid-state diodes and transistors.

More than a year later, Telettra publicly present their work in early April 1957. Olivetti and Mario Tchou hear about this and paid him a visit. After some discussion, Olivetti and Floriani decided to join forces - combining research from Tchou's lab in Barbaricina and Telettra's processes.

Thus in October 1957 we have the "Società Generale Semiconduttori", or SGS.

The name literally means the “General Semiconductor Society" and they presented themselves as the first Italian company expressly founded for the "research, study, and manufacture of diodes and transistors". Their products would be available to any company.

SGS would set up a research lab and factory in the small town of Agrate Brianza along the Milan-Bergamo motorway to take advantage of Milan's big industrial base.

Starting with a license from General Electric, they start producing their first germanium diodes for their parent companies in 1959. Their transistors accelerated Olivetti's work, helping them ship the Elea 9003 computer.

SGS-Fairchild

SGS was founded the same time as Fairchild Semiconductor.

Fairchild quickly distinguished itself with several innovative moves. For instance, making silicon transistors when everyone else - except Texas Instruments - was doing Germanium.

General Electric was a famous name, but their transistor technology quickly grew outdated in the fast-moving industry.

Between 1957 and 1961, the company lost nearly 700 million lire. Which I think is over a million USD in 1960 dollars, assuming an exchange rate of 600 lire to 1 USD.

SGS soon realized they needed a new technological partner. So Olivetti and Floriani reached out to Bob Noyce of Fairchild through Adriano's brother Dino and asked about a partnership.

The resulting deal made Fairchild an equal partner to Olivetti and Telettra - a third each. Fairchild transferred their silicon technology and sent over their staff to start converting those Germanium lines over to silicon. SGS began building factories across Europe - England in 1963, West Germany in 1964, France and Switzerland in 1966, and so on.

SGS-Fairchild was a fairly successful endeavor. They turned a profit in 1963. And in 1967, they generated about 17.5 billion lire - or about $30 million USD. 75% of that revenue came from outside Italy. They had about 3,200 workers and owned 30% market share of the Italian market.

However, the arrangement did not last. Fairchild wanted SGS to just stick to sales and manufacturing. But SGS and the Italians wanted to do R&D right in Agrate, Italy where they were founded. They argued that the European semiconductor market was different from America's.

For instance, the American semiconductor market in the 1960s leaned heavily towards the military - space and missiles. Europeans on the other hand were focused on telecommunications and consumer items.

This disagreement led to a split in 1968. Fairchild allowed SGS to keep the licensed technology, but sold their shares. A year after that, Telettra was sold to the massive Fiat Group, dumping their SGS shares as well. Thus by 1970, Olivetti was left the sole shareholder.

Despite these troubles, SGS emerged from the 1960s and the transition to integrated circuits as one of the few European semiconductor successes. Their partnership with Fairchild gave them a valuable planar IC technology targeting consumer and specialty devices.

Going to France

Let us step away from the guys in Italy and pick up a new - very messy - thread over in France.

The company that we will eventually call Thomson is the result of a series of mergers that occurred throughout the 1960s and 1980s.

It gets messier than the Habsburg family tree so we will do our best.

Keep in mind this is not a general history of French semiconductors. I highly recommend these two resources on whom I leaned greatly for this section - Andrew Wylie's French Vintage Semiconductors Page, and Mark Burgess's French transistor history page.

There are two dominant semiconductor groups that will eventually merge to create Thomson Semiconductors - COSEM of CSF and SESCO of Thomson-Houston.

We shall start with COSEM. Yes, I know we are going two levels deep here. Bear with me.

COSEM & CSF

COSEM was the semiconductor manufacturing arm of "Compagnie Générale de Télégraphie Sans Fil" or CSF.

CSF was a radio and telephone company with roots dating back to the 1880s. By the 1950s, they produced military and telecommunications electronics hardware. Being vertically integrated, they also produced their own vacuum tubes, ferrite cores for memory, and so on.

CSF was one of the very few transistor companies that did not start their journey with a Bell Labs license. Rather, they relied upon technology transferred over from the state's telecommunications laboratory - "Centre national d'études des télécommunications" or CNET.

In addition, they leaned on the connections of critical members in the French semiconductor community to the US. Notably, Claude Dugas, a French-born physicist who studied at Carnegie Mellon and had been good friends with scientists in Bell Labs and others.

In the mid-1950s, CSF founded "Compagnie Générale de Semi-conducteurs" or COSEM in Saint-Egrève, a western suburb of the town of Grenoble. The area was chosen for its nearby university as well as its dextrous female workforce - deemed suitable for delicate transistor manufacture work.

In 1961-1962, COSEM had become France's largest semiconductor-maker with 17 to 45% share of the French market. While they largely serviced the local market - selling to French computer makers like IBM France and Bull - up to 30% of its output was sold abroad.

SESCO & Thomson

SESCO hails from France's other major semiconductor maker - Compagnie Française Thomson-Houston or just Thomson-Houston.

Thomson-Houston was founded all the way back in 1893 as the French subsidiary of the American Thomson-Houston company - one of the predecessors of General Electric. They sold electric trains, railway signaling, telephones and the such.

A wild series of restructurings, mergers, transfers, and one 1936 nationalization later, we have the French Thomson-Houston company. French is in the name because of the nationalization.

During this period, they spin off their heavy engineering divisions. That eventually becomes Alstom, the French Rail company. What a coincidence, right?

In the 1940s, Thomson-Houston is making radars and radio transceivers for the professional market, as well as consumer radio sets and telephones. Transistors make a lot of sense for radios. However resistance from the vacuum tube people prevented them from going as hard into transistors as they should.

They fell behind. So in 1956, Thomson-Houston teams up with its former American owner General Electric to work on semiconductors. Later in 1961, the two found a semiconductor subsidiary, "Société Européenne des Semi-Conducteurs" or SESCO.

In 1966, Thomson-Houston merged with the electronics arm of the French arms company Hotchkiss-Brandt. Hotchkiss is somewhat famous for their Hotchkiss Revolving Cannon, capable of firing 43 shells per minute. The new company is called Thomson-Brandt.

Okay, just to recap. We have CSF, the telecommunications company, and Thomson-Brandt ... which itself was a merger of the electronics divisions of a French-nationalized American company and French cannon company.

Merger to SESCOSEM

French theoretical solid state physics knowledge was very strong.

However, French companies struggled to translate that knowledge to the market. Notably, their telecommunications and electronics companies stuck with vacuum tubes and then Germanium for longer than they should have.

France is known for having numerous foreign companies operating there, which offered substantial competition to domestic firms. So with Thomson-Brandt struggling to compete and CSF over-leveraged thanks to its expansions throughout the 1960s, the two decided to merge in 1968 to create Thomson-CSF.

Soon afterwards, their semiconductor divisions - COSEM and SESCO - joined together to create SESCOSEM, "Société européenne de semi-conducteurs", or literally "European Semiconductor Company". Whew. SESCOSEM began as a second-source producer for companies like Motorola, helping the latter more easily enter French markets.

The Pull of Computers

The end of the 1960s saw a spate of European countries trying to domestically produce their own computers.

The most famous of these efforts was France's Plan Calcul, which Charles de Gaulle initiated in 1966 after General Electric acquired the last French computer-maker, Bull.

One of the companies involved was Compagnie Internationale d'Informatique, or CII. This was a non-vertically integrated company partly owned by Thomson-CSF and subsidized by the French government.

That same year, the United Kingdom government also midwifed their own domestic integrated computer giant - ICL. I covered it in a previous video.

Calcul should have presented a great opportunity for SESCOSEM. After all, they were CII's sister company. And Plan Calcul did allocate 20 million francs for SESCOSEM to build semiconductor TTL plants for CII.

But CII - not to mention ICL and the other European computer companies - opted to build around Texas Instruments. SESCOSEM did manage to receive the second source contract, but nevertheless suffered damage from CII arguing that if they were forced to buy local then their products would be inferior to IBM's.

Throughout the 1970s, SESCOSEM struggled to turn a profit. Their volumes were too small, which management naturally blamed the domestic market for. But one could also point at SESCOSEM's own business practices, which forced it to offer too large a range of niche products.

For instance, its "controlled delay" ("retard contrôlé") strategy, which targeted a 2-year lag behind American leaders. Ostensibly to avoid taking excessive risks on potential flops. In 1972, Thomson-CSF absorbed SESCOSEM into its own corporate structure for tax reasons. Meaning, SESCOSEM were losing a lot of money.

Italy Intervenes

Let us step away from France and go back to SGS.

It is 1970, and after Telettra and Fairchild left the business, SGS suddenly found itself in a rough spot. Fairchild provided a good portion of its technology and there were too many unprofitable factories across Europe that needed to be closed.

Its sole controlling shareholder was Olivetti, which was suffering its own turmoil. Adriano Olivetti and Mario Tchou had died in the early 1960s, leaving them in a bit of a lurch, leadership-wise.

They were also in financial straits due to intense Japanese competition and an ill-advised acquisition. So Olivetti could not cover for a struggling semiconductor company, and they started looking for a financial partner to take the business off their hands.

Motorola sniffed at a possible acquisition, but the Italian government - recognizing SGS's strategic importance - stepped in. In 1972, they bought out Olivetti via a state-owned telecom/electronics holding company called STET.

Then STET merged SGS with a small Italian semiconductor company ATES to make SGS-ATES.

ATES ("Aquila Tubi e Semiconduttori") was founded back in 1959 as a vacuum tube manufacturer. Two years later, they got a license from RCA and Marconi and built a factory in the city of Catania in Sicily. Their management team took over running the combined organization.

SGS-ATES

The Italian government saved the company from falling into American hands, but did not improve their loss-making financial status.

The company did have some interesting products. They held a license from Zilog to produce the Z-80 and Z-8000 microprocessors. Zilog cofounder Federico Faggin once worked at SGS-Fairchild before he joined Intel.

That was basically the only digital integrated circuit they sold. But they also had some good RF germanium, analog integrated circuits, and power stuff.

Yet despite these good products, the company lost money throughout the entirety of the 1970s - $20 million in a good year, $50 million in a bad one. Each year, the government compensated them for exactly the amount of that loss.

The government did this because they - like many other people in Europe - thought that semiconductors would always lose money. You must do them because they support critical things like nuclear weapons and other strategic industries. But don't expect to make money from it.

In an atomic bomb, who cares if a transistor costs one dollar or a thousand dollars? So just make the best darn transistor you can make, and forget about the cost. So volumes cratered, product lines expanded, and costs rose.

This arrangement made sure that everyone stayed employed - not unimportant in Italy those days with its leftist bent - and some high quality products got made.

But it also left the company in limbo. They were perfectly satisfied to sell small numbers of chips solely into Europe - itself a declining semiconductor market. So the 1970s saw SGS-ATES gradually lose share and money in the overall market.

Turnaround

In 1979, the company hired a new CEO - Pasquale Pistorio.

Pistorio was born in 1936 in one of Sicily’s poorest villages. He studied electrical engineering at Turin and then became a salesman for Motorola - turning down a job offer from Olivetti.

Over many years, he worked his way up to be vice president of Motorola's international semiconductor division and the most senior non-American in the world's second largest semiconductor company.

The SGS job offered a 40% pay cut and required him to take an armored car thanks to terrorists. But Pistorio embraced the professional challenge of turning around what he called a "basically broken" company that still held national importance to his home country. It would also allow him to move home to be with his aging parents.

Upon arriving, Pistorio set out three goals for SGS-ATES: To make a profit, to enter the US market, and to become a billion dollar company. And to show that things really had changed, he started firing people. He recalls in an oral history later:

> A guy ... says, "Look, yeah, our company's a disaster. The quality's terrible. The cost is high. The service is a disaster. It's still a miracle that we get some orders."

> I said, "You are the marketing manager?" He says, "Yeah." What can you do? So you simply ask him to leave, because it cannot be.

Pistorio started from the top and worked his way down. His first month, he fired 20 out of the 80 top managers. Once that was done ...

> The second month, I attacked the lower level. Absenteeism was 22 percent. Can you imagine 22 percent absenteeism? And then I fired some 17 ... absenteeists, and their track record was more than 50 percent of the time for three years in a row

Over a thousand employees were fired and money-losing factories like the one in Scotland were closed. Naturally this led to a huge huff. SGS-ATES was a government-owned company, and firing people seemed unimaginable.

Many of the lost jobs were in Catania, and involved the poorest, least educated workers. But the company had lost money for a decade. Italy in 1980 was itself going through political upheaval over these questions. Pistorio stood his ground, and the government backed his plans. They knew that things had to change.

Return to a Profit

Their new culture would be centered around hard work and globally-minded expansion.

Pistorio is known for having what Businessweek calls a "passion for his work". He gets to the office at 630 AM, when everyone previously came in at 930 AM. And then he would work until 10PM.

SGS quickly adopted American-style management techniques, and dramatically upped their work hours. They were only working 1,500 hours a year rather than 2,000 in the US and 2,200 in Japan or Singapore.

To convince the unions to accept this, he assigned all top management including himself to work a few graveyard shifts. The tactic succeeded and in doing so, SGS became the first company in Italy to employ women at night.

Another thing crucial to the SGS turnaround were its strategic alliances with big companies like IBM or Hewlett-Packard for computers. These close relationships allowed them to co-develop special custom chips that go into products. These were high-margin chips in popular, high-volume electronics.

Going into the 1980s, all the big four European semiconductor companies - Siemens, Philips, SGS, and Thomson - were losing money. They had lost money all through the prior decade.

In 1983 SGS became the first of the four to turn a small profit. Pistorio had promised to do it in four years, but managed it in three. The other three European companies now saw that it can be done, and pushed for the same thing as well.

All this time, Pistorio wanted SGS to become big. He knew that volume and scale were key in the global semiconductor game. The higher the volume, the better the yields, the better the profitability.

They were one of the first European semiconductor companies to build a wafer fab in Asia - in Singapore in 1984. Its Southeast Asian facilities would eventually contribute 30% of total volume in less than half a decade.

SGS revenues had grown over three times what they were when he took over - hitting $375 million in 1986. But looking ahead to the 1990s, Pistorio became convinced that the future belonged only to the biggest companies making at least a billion dollars in revenue. The fastest way to achieve that was a merger. But with whom?

Thomson and SGS

Thomson was looking to scale up throughout the 1980s as well.

In 1982, the Mitterrand Socialist government nationalized five of France's biggest companies, including Thomson.

Alain Gomez was appointed by the government to be chairman of the Thomson parent company. Harvard-educated, Gomez brought sweeping change and focus to the electronics company just like what Pistorio brought to SGS.

Pistorio knew the management of Thomson Semiconductors - including its CEO Jacques Noels - quite well. The two companies seemed like a natural fit. Both were largely government-owned, were similarly large at about $400 million in revenue, and were following the same strategy.

Over time, the idea of a combination began to percolate. They started to cooperate in a formal manner in 1986, when they jointly applied for funds for a big memory chip project via EUREKA - Europe’s big R&D initiative.

Gomez approved, though they did keep some bits. Thomson Semiconductors was one of the world's largest military chip suppliers. But for national security, this stayed with the larger Thomson company. Anyway. So in 1987, it was announced that SGS would merge with Thomson Semiconductors.

Digestion

The new company - named SGS-Thomson - made about $800 million in revenue.

This made them the second largest European semiconductor company after Philips, and the 12th largest in the world. It tripled the product offering and added many customers in segments like industrial and telecommunications.

But the new company needed extensive reform. In 1987, SGS-Thomson turned about $200 million in losses - all from Thomson - and had $350 million in debt - all from SGS. Critics in the media called it a case of the "blind leading the blind".

Pistorio was appointed the new CEO. He held a meeting with people from both companies. There, he told them that they were one company now and needed to integrate as fast as possible. The longer it took, the worse it would get.

Within a year of the merger, Pistorio closed five plants - three in France and two in Asia - and started moving more production over to the remaining Asian facilities. All the sales offices were merged together and top management was cut by a third - with an eye to balance French and Italian.

These changes were tricky with SGS-Thomson now having two national governments as its shareholders. It caused issues with things like choosing a headquarters. In exchange for accepting an Italian CEO, the French government mandated that the headquarters be in Paris for ten years.

From 1987 to 1991, SGS Thomson lost a cumulative half billion dollars. They would have gone under had it not been for a strategic $800 million loan from their governments. That money was eventually paid back with interest after their later IPO.

But in 1992, the company finally turned around, making a tiny profit of about $4 million. The year after that, 1993, the company made $100 million in profit. The worst was over. Pistorio was quoted in a news article at the time: "I am glad to say that we made it, that this has been a successful merger."

Conclusion

A year later, SGS-Thomson went public on the Paris and New York Stock Exchanges.

In 1998, they changed their name to STMicroelectronics - an abbreviation of their respective halves.

The company's corporate alliances with Nokia in wireless, Western Digital and Seagate in disk drives, and Bosch in automotive had given it a highly diversified portfolio of products powering MP3 players, smart cards, and phones.

They hit a particularly big home run with their bet on analog and mixed-signal chips. Such chips have important roles in cars, industrial machinery, and power applications.

In 2005, Pistorio retired from STMicroelectronics and became its honorary chairman. He has since spent his life focusing on environmental and social issues like water conservation and inequality.

As of this writing, he is 88 years old and still kicking. His semiconductor legacy is secured. The company he once led back from the brink is now one of the largest chip companies in Europe. What a life!

Subscribe now

Broadcom is the second largest AI chip company in the world.

Thanks to that, the company is the 11th largest company in the world. Over $600 billion as of this writing, bigger than Visa and just behind TSMC.

It is a bit crazy considering that in 2009 the whole company was worth $4 billion. 150 times growth in 15 years is kind of wild.

But what is Broadcom? How did they get this big? What do they do? In this video, how a little chip division grew to be a $600 billion AI juggernaut.

Beginnings

The company now known as Broadcom started as a spinoff of a spinoff.

In March 1999, the iconic California-based computer-maker Hewlett-Packard decided to split into two.

Everything unrelated to computers, IT or printers would be put into the new company - Agilent Technologies.

The new publicly-traded company covered HP's former test and measurement, medical products, chemical analysis, and semiconductor businesses. Accounting for about $8 billion of HP's $47 billion total revenues.

Analysts saw this as a necessary re-focusing of a business that had grown too large. In an interview at the time, Agilent's new CEO Ned Barnholt positioned it as a coming-out of sorts from the shadows.

Agilent's debut on the markets was one of the biggest IPOs in history up until then. The stock popped nearly 70%.

But the new company struggled to grow in the rough years after the bursting of the Dotcom and fiber optic bubbles. The company's revenues shrank nearly 50% from 2000 to 2001.

And soon they started laying people off and selling entire divisions to raise money and simplify the business. So in June 2005, they put up their chip division - called the Semiconductor Products Group - up for sale.

Spinning Off the Chips

There followed a brief auction... Which PE firms KKR and Silver Lake Partners won for $2.65 billion in August 2005.

The government of Singapore also co-invested in the deal through their Temasek and GIC sovereign wealth funds. Singapore had been a partner with Hewlett-Packard since 1970, when HP first chose to set up a factory there.

It was a glorious time to be in private equity. Some of the biggest private equity deals in history were closed during this 2006-2007 period before the Global Financial Crisis.

Big deals included the utility company TXU (bad deal), the financial services company First Data, and the hospital company HCA. These regularly topped $20 billion in total value.

This Agilent carve-out deal was nowhere as large as those monsters. That same month, KKR closed a $10 billion deal for a majority stake in Philips Semiconductor.

That was later merged with Motorola's internal semiconductor division Freescale to become what is now NXP.

Nevertheless, the Agilent deal seemed like a good arrangement. KKR and Silver Lake believed that the division had the potential to thrive.

Hewlett-Packard's semiconductor division dates back to 1961 when the electronics giant believed that it had to bring component production in-house.

Over the years, that group expanded to a range of industries. Some of their more prominent products were FBAR RF filters used for mobile phones’ radios (more on that later);

Optical transceivers;

And custom chips for HP's lucrative printers to lock you into their ink cartridges;

But the semiconductor business changed a great deal in the 1990s. The rise of foundries turned the tables on the economics of owning your own internal semiconductor division.

And despite an admirable attempt to diversify, the division’s cyclical nature and chronically low margins exposed Agilent to losses when times were bad. 17% of the semiconductor division's 2004 revenues still depended on HP itself.

In the end, Agilent's management realized that this business was not "core". Thus in December 2005, Avago Technologies emerged into the world. It was the largest privately-held fabless semiconductor company in the world.

Hock Tan

Avago's first CEO was the division's original general manager Dick Chang.

Chang originally joined Hewlett-Packard Labs and became Vice President and General Manager of the Semiconductor Products Division.

But a few months after its founding, Avago announced that it had a new president and chief executive - Hock Eng Tan (陳福陽).

Hock Tan is of Chinese descent, born in Malaysia. At the age of 18, he won a scholarship to MIT, where he received a bachelor's degree. Then he went to Harvard Business School to do an MBA. After that, he returned to Malaysia.

There, he worked at a number of companies. First, Hume Industries, a cement maker affiliated with a large Malaysian conglomerate. Then a venture capital firm called Pacven Investment.

After that, he returned to the United States. There, he took finance roles at Pepsi and General Motors before landing at the personal computer company Commodore International in 1992.

At Commodore, Tan served as vice president of finance and then CFO during the iconic computer company's final days as it struggled with huge losses and low sales. That must have been brutal.

After Commodore announced its bankruptcy in 1994, Hock Tan left to become SVP of Finance at a chip company called Integrated Circuit Systems or ICS.

ICS

ICS was founded back in the late 1970s in Valley Forge, Pennsylvania.

The company was a design house, meaning that they designed semiconductor products but did not sell them. They did contract design work for larger companies like United Technologies or GE.

In the late 1980s, they developed a hit new product - a silicon timing device or frequency timing generator. Every electronic system needs a timing element, and many older ICs used to use crystals to do that.

The silicon timer replaced those crystals with an all-silicon solution, and that was a big deal. It was a huge success and in 1991 they went public.

Hock Tan joined ICS in 1994. Then in 1995, he became CFO. A year after that, he became COO too.

Then in 1999, ICS announced a leveraged buyout by its own management team for $257 million. The PE firm Bain Capital and the bank Bear Stearns helped provide funding. Hock Tan became the CEO.

In 2005, ICS merged with another company producing mixed signal chips - Integrated Device Technology in a $1.7 billion deal. Tan became the combined company's chairman. It was then that he was recruited by Silver Lake to be Avago's CEO.

The Franchise

Bloomberg quoted Kenneth Hao;

A partner at Silver Lake and now chairman and managing partner - saying about Tan:

> He starts with a point of view that the semiconductor industry has matured ... The businesses must be run differently than when they were growing up

Tan really likes the idea of focusing on a company's core "franchises" - a word he has used multiple times. In a 1999 interview while he was still CEO of ICS, the silicon timer company, he likened ICS to a franchise:

> We have essentially in our business almost a franchise ... in the sense that anybody who wants timing solutions ... would immediately think not just of crystals but of silicon, and not just silicon but of Integrated Circuit Systems

> Our market share in the PC space is extremely high, and that effectively gives us almost a franchise ... We get to be selected in most situations as the first source

Years later, he brings up the phrase again in a 2018 interview with the Wall Street Journal:

> It's about putting together a very good portfolio of product franchises to create a lot of value ... There’s no long-term vision or ambition other than that ... Frankly, we overinvest to ensure we are way ahead of No. 2 or No. 3

The franchise metaphor is illustrative. The best competition is no competition. Fast food and car dealership franchises maintain small localized monopolies over their areas.

In the old days, the industry's cyclical nature and early growth motivated the old guard semiconductor companies like Motorola and National Semiconductor to launch as many products in as many different fields as possible.

But things have since gotten more complicated. Designers are now often more interested in solutions for their systems rather than bespoke components. And foreign competition has flooded the market with options.

So Tan's thinking is that today's leading semiconductor businesses need to build and maintain "franchises". To be the instinctive number one choice whenever a designer needs a thing for their system. A small localized monopoly.

Strengthening the Franchise

When Tan takes on a new business, he wants to strengthen that "franchise" as much as possible.

That means a focus on cutting-edge products within the franchise as well as new verticals and applications to expand into. When asked about ICS's growth strategy he said:

> We’ll stick to our knitting, essentially, which is silicon timing solutions. Our strategy is to expand applications, and increase market share in digital consumer and communications

This means cutting speculative projects like the ones the old guard semiconductor companies used to do. If it does not help the core franchise, chuck it.

This approach has more than a few critics. Cuts to R&D expenditure or sales of non-leading divisions have long term implications, since these things take time to bear fruit. The franchises have to come from somewhere, right?

Tan correctly points out that he invests considerable resources into existing franchises - milking the cow so to speak. Nevertheless, a company of Avago's style must - like a hummingbird - be eating constantly in order to appreciably grow.

Road to IPO

Avago began life as a very diverse business.

At the end of 2005, the company had five different end markets generating over 10% of sales. This was the convoluted legacy of forty years of diversification at HP.

So Tan and his team started organizing this mishmash to create a coherent company focused on its analog, mixed signal, and optoelectronics franchises.

In March 2006, they sold their storage business to PMC-Sierra for about $420 million.

In May 2006, they sold off their printer ASIC business to Marvell for about $250 million.

Then in December that same year they sold their CMOS image sensor business to Micron Technology for $53 million.

They did one more sale a year later - selling their Infrared operations to the Taiwanese electronics company Lite-On for $20 million.

These sales not only raised money to pay down debt, but also slimmed the company's headcount from 6,500 in 2005 to around 3,600 in 2008.

By 2008, though Avago's revenue and earnings had not changed all that much since the acquisition, the company had paid off a billion dollars of net debt.

The improvement was enough for KKR and Silver Lake to take the company public in August 2008, two years after it first went private.

The two PE groups would later sell down their shares in the years after the IPO, eventually realizing a 5x return on their shares. Not bad for a couple years' work.

The Mobile Revolution

The IPO came about at a fortunate time.

2007 and 2008 heralded the introduction of the Apple iPhone and the subsequent smartphone mega-boom. And Avago parlayed their early work in FBAR RF filters to make the most out of the surge.

As I detailed in a prior video, RF filters are critical but unheralded parts of the mobile phone. They help the modem separate the data signal from noise - saving on power and improving the user experience.

These filters are actually more like MEMS - tiny tuning forks - rather than Nvidia GPUs. But they are highly specialized - requiring special thin film techniques for certain unconventional materials like gallium arsenide.

As I mentioned earlier, HP had been working on FBARs since the 1990s. In 1999, they released their first commercial FBAR filter.

Two years later in 2001, they released the first FBAR filter for the 1900 megahertz band, a GSM band.

In 2008, Avago paid $30 million USD to acquire Infineon's BAW RF acoustic filter business and their patents. FBARs are one major subtype of BAWs filter. Infineon's filters were SMRs, another sub-type.

In 2010, Avago was first to market with an RF filter for the 4G-LTE bands. LTE sparked a mobile data revolution across the world, forcing RF filters to have to accommodate increasingly more bands PLUS Wi-Fi and Bluetooth.

This greatly increased their complexity, and thus their prices. The RF filter industry grew from about $100 million in 2004 to over a billion in total revenues a decade later. Avago earned a good chunk of that money.

At the same time, smartphones got skinnier and more power efficient. This necessitated the consolidation of the many various discrete components of what we call the RF front end - power amplifier, filter, switches, and antennas. Everything short of the transceiver and modem, which is dominated by Qualcomm.

So the mobile smartphone and LTE booms were good for Avago throughout the late 2000s and early 2010s. But by 2013 mobile was some 50% of their revenue, raising concerns of over-concentration. So Hock Tan and Avago went out to buy a new franchise.

LSI & The Data Center

In December 2013, Avago announced that it would acquire LSI Logic for $6.6 billion.

I covered LSI Logic and its journey in a prior video. What follows is a brief summary.

LSI had been founded by the former CEO of Fairchild Semiconductor - a no-nonsense Brit named Wilf Corrigan.

They raised a little money, engaged a Japanese fab to make wafers for them, and came out with an innovative “master slice" system that helped customers replace entire circuit boards with custom chips.

Revenues exploded to nearly $2 billion. However this space had many competitors and times got tough after the Dotcom bubble burst.

After Corrigan retired, LSI's new CEO shifted to a fully fabless business model and started developing custom silicon for running data centers and helping to make them more efficient.

LSI Logic already had a foothold in this business through their work on hard drive controller chips. They then bulked it up in 2007 with a $4 billion merger with Agere Systems.

Agere was a spin-off company from Bell Labs/Lucent that made semiconductors and components for wired and wireless communications networks.

Together with Agere's business, LSI Logic built up a proficiency in enterprise storage systems - networking and storage hardware connected to SSDs or HDDs.

In a time when cloud storage startups like Box and DropBox were raising at billion dollar valuations, this franchise attracted Hock Tan's team.

And for LSI, they had been in the desert for a few years anyway, a payout would be a good end for them.

Avago's acquisition of LSI Logic was like one of those deep sea gulper eels that try to swallow fish way bigger than them.

In 2013, Avago produced $2.5 billion in revenue. Their market cap was just about $12 billion.

Only $1 billion of the $6.6 billion purchase price was their own cash. Their PE partner Silver Lake invested $1 billion and the rest, $4.6 billion, was borrowed. So again, Hock Tan and his team started cutting expenses and paying off debt.

In May 2014, they sold LSI's flash memory solutions - which includes their PCI Express flash solutions as well as controller chips for solid state drives - to Seagate for $450 million.

Then in November 2014, they sold LSI's network chip unit, called Axxia, to Intel for $650 million. The unit's chips help internet service providers and data centers monitor and manage traffic over networks.

Broadcom & Brocade

A year after that, Avago struck another, far larger deal - the venerable analog chip company Broadcom.

Avago had tried once in late 2014, but couldn't agree on price.

But the LSI acquisition drove up Avago's stock and gave it a richer currency to do a better deal.

So in April 2016, Hock Tan called up Broadcom's management again with an offer, then tried to hash out a deal right then and there. From start to end, it took a month and a half.

The $37 billion acquisition was the biggest technology deal up until then. Avago paid a nice 28% premium on Broadcom's share price, which was then already at a 9-year high.

Broadcom is a storied name in semiconductor history. Founded in 1991, they produced analog system-on-chips for high speed communications equipment like cable modems and TV set-top boxes.

So throughout the 1990s, they grew with the rise of cable operators like Comcast and broadband internet. In the 2000s, they entered a series of other markets like enterprise switches and networking.

Broadcom was another gulper eel type deal. Its revenues in the prior year were nearly twice the size that of Avago's.

As part of the deal, Avago agreed to take Broadcom's name and add Broadcom co-founder Henry Samueli to its board. It reminds me of those type marriages between a new-money family and a family that is poor but with a long and prestigious name plus a castle.

Little over a year later, Broadcom acquired another storage networking company Brocade for about $6 billion. Again the cycle restarted, cutting expenses and paying off debt. We don't need to rehash this.

Qualcomm

So in just a few years, Avago/Broadcom gathered a great bunch of assets.

It is a big company version of buying properties on mortgage, fixing them up, renting them out, and then paying down the debt aggressively. Rinse and repeat for a few times over.

The issue though was that after Broadcom and Brocade, there were not many targets left big enough to move the needle. But you miss 100% of the shots you don't take, so Tan went right for the prom queen - Qualcomm.

I also covered this saga in a prior video, but this is the summary.

On November 6th 2017, Broadcom offered $103 billion - an offer immediately rejected by the Qualcomm board as too low. They tried to engage but was rebuffed - so they went direct to Qualcomm shareholders in an attempt to swap out the board.

In January 2018 Qualcomm filed a notice with CFIUS to block the acquisition. CFIUS is a committee overseeing national security issues on mergers.

CFIUS highlighted concerns about Broadcom's reputation of cutting R&D, feeling that it would extinguish Qualcomm's chances of competing in 5G technologies against Huawei and other Chinese companies.

In response, Broadcom re-domiciled itself out of Singapore to the United States. Then-President Trump held a public conference with Hock Tan on this.

The re-domiciling would have presumably invalidated CFIUS's jurisdiction. But the Trump Administration ended the whole kerfuffle when it issued an executive order blocking the deal.

That happened six years ago - time really flies. Qualcomm's market cap has since risen to a number above Broadcom's offer price due to surging revenues from the COVID technology boom.

But the stock price hasn't moved all that much since in part due to the post-COVID hangover, Huawei competition, and what not. As of this writing, Qualcomm's market cap is $190 billion.

Looking back, I feel like the Qualcomm shareholders would have preferred the growth from having Broadcom shares.

But on the other hand, CFIUS's point was that they wanted to retain Qualcomm's ability to compete in certain wireless markets like 5G. Broadcom would have supposedly snuffed that out. It wasn’t a money thing.

Yet Huawei is still dominant in 5G and Qualcomm's modem work seems to be more in the news for fighting Apple on royalty payments.

But I certainly feel that the Broadcom acquisition would have snuffed out Qualcomm's attempts to compete in Arm-flavored laptop chips like Snapdragon X. On the whole, I still think it's a good thing that Qualcomm stayed independent.

The AI Boom

Blocked from further growth in silicon, Hock Tan and Broadcom have since turned to acquiring software companies.

This includes CA Technologies, Symantec, and VMWare. Basically running the same processes, pulling out the franchises, cutting expenses, and paying off whatever debt incurred for the buy.

But until ChatGPT came out, there were always questions of what was next. What thing can they get next to take them to the Holy Land? The answer had been inside them all this time.

When Avago/Broadcom acquired LSI Logic in 2013, they also acquired a small custom silicon design division that helped external customers produce their own chips for the data center.

When Google started on the first iteration of their own chips - the Tensor Processing Unit or TPU - they did not need that many of them. So in 2016 they engaged Avago/Broadcom to help design and produce the chip for them.

Why did Google do this? Because custom silicon design teams are expensive. You not only pay a lot for talents, but also EDA tools and IP for them as well. Unless your volumes are massive, it is better to engage LSI or Avago to do the design based on your spec plus other value-adds like testing and packaging.

Back in 2016 when the TPU relationship presumably began, custom silicon was quite small for both sides. Broadcom's revenues from that year was estimated at about $50 million or so.

But it grew over time and by 2020, Broadcom’s TPU revenues were estimated at $750 million. Together, the two companies have made several new TPU iterations on faster nodes.

When ChatGPT blew up in November 2022, it started the current AI boom, with everyone from startups to tech giants getting into generative AI. Many of these players hit compute bottlenecks, because nobody not even OpenAI expected ChatGPT to be such a hit.

Google realized that their TPU ASICs gave them a huge competitive advantage in compute over Microsoft, Oracle and others. So they started buying more TPUs, which benefits Broadcom.

SemiAnalysis projects that Google will pay Broadcom far more in 2024 for TPUs - something like $8.5 billion. Only some lame computational lithography company called Nvidia makes more from AI chips.

That is not counting whatever they make from their relationship with Meta or other giants like Microsoft.

But Nvidia isn't going to let Broadcom have that all to themselves. In February 2024, Reuters reported that Nvidia is setting up their own AI chip design unit. The idea being to leverage their IP and expertise to keep customers from designing alternatives to their profitable H200 and B100 accelerators.

Good. Let them fight.

Conclusion

I want to thank Dylan Patel of the aforementioned SemiAnalysis for some clarifying comments on this video.

Dylan is great. He did a piece prior on Broadcom's history that I think would be good companion reading to this one.

So what is Broadcom? Friend of the show Digits to Dollars lays it out pretty clear: Broadcom is a publicly traded private equity fund masquerading as a semiconductor company.

They have a portfolio of strong technology franchises. They acquire new ones, slim them down, focus them, and then leverage them to buy new franchises. They grow whenever they buy something new.

This time, one of their franchises has stumbled into the heart of the AI boom. That does not mean they won't fight like heck to stay there like with the mobile and cloud booms prior. I reckon they will invest whatever necessary to keep and extend their position in this gold rush.

Subscribe now

I recently returned from a two week trip to Europe for IMEC’s ITF World 2024. It was a wonderful whirlwind of a journey and had a great time. I received a great deal of ideas for videos and will be working on them over the next few weeks. The trip had many upsides - sorry I didn’t organize any meetups as I was too tired - but the highlight was undoubtedly this trip to ASML.

I recently had the opportunity to visit ASML's headquarters in the small village of Veldhoven in the Netherlands.

It was a wonderful experience. I learned a lot about the company's history, culture, and more.

And I was honored to have the rare opportunity to see the $300 million or whatever High-NA EUV tool. I call it the Beast.

So in this short video, just a few random thoughts on seeing ASML in the flesh.

Eindhoven

My hotel was in Eindhoven, which is short ways near Veldhoven.

The town is quite nice, albeit rather slow-paced. By the way. For what it is worth, I paid my own way including transportation and hotel stay. And ASML did not sponsor this video.

Eindhoven will forever be known as the home of Philips, the once-massive consumer electronics tech giant.

They first built the lithography machine technology and later spun off ASML as a joint venture with ASM International.

They were also TSMC's cornerstone foreign investor, owning 28% of the company at the start and 36% later on.

Philips later sold off all the shares they got from both those efforts. As of this writing, the two companies have a combined market capitalization of $1 trillion. Would have been nice to have those stakes now.

But such is life. Investment managers in the mid-1990s were pressuring Philips to sell those stakes to "unlock" shareholder value.

ASML in Veldhoven

Over the past five years, ASML has grown incredibly fast in almost every possible way.

For instance, the stock has gone nuts - to the utter surprise of virtually everyone who has ever worked at the company.

The company has gained a huge and unexpected new following in the popular consciousness. That includes a great deal of people who believe that the company's products make chips right out of the box like a pasta machine. Not the case, by the way.

And then headcount. Today the company has over 42,000 employees across Europe, Asia, and the United States. A bit more than half work at the Dutch headquarters.

And some 20,000 of those employees joined the company in just the past five years. That is kind of staggering.

All this past growth - plus anticipated future growth - has caused some issues. For instance, where to put them. The company's current campus sits between two highways, which constrains its growth.

On one side you have natural protected areas that cannot be developed. Not to mention that there is a very large highway in the way.

On the other side you have residential development and homes, which I assume belong to the village of Veldhoven. I don't think it is in the company's best interest to irritate those people by surrounding them with big industrial buildings. So what can you do?

The company wants to expand their factories to add more capacity for EUV and DUV machines, particularly the former. There are also concerns about having enough water and power.

Plus transportation. Most people have to drive in because Veldhoven is a literal village and the roads are not adequate to handle all that. So the company is negotiating with the local governments about that.

Debate

Those negotiations have caused a bit of a ruckus recently.

A few stories made it into the news about the company potentially "leaving" the Netherlands. You have stories of something called "Project Beethoven", a $2.5 billion incentives plan by the government to "keep" the company in the Netherlands.

The name is rather dramatic - fit for a movie. But it does no favors to the core issue. To me it seems fair that one of the things that governments should do for their residents, businesses included, is to provide infrastructure.

But the negotiations have stirred up a debate in the government. A group of Dutch people have the sentiment of "Just go then!"

And indeed, several previously dual-headquartered companies like Shell and Unilever have left the Netherlands for the United Kingdom.

But based on what I have seen and heard, I don't think leaving the Netherlands is an option. Veldhoven has been ASML's home since they were unceremoniously dumped there after the Philips spinoff.

So much of the company's supplier base and talents live and work here. ASML is not going to pull an Oracle and force them to relocate. They want to stay. But if there is nowhere to expand, then they seem to be open to doing that expansion elsewhere.

Fortunately, that doesn't seem necessary. In April, the company signed a letter for land for a second campus somewhere north of the Eindhoven airport - close to the current campus.

Culture

Let us talk about company culture.

TSMC and ASML are like siblings. They share a corporate parent. TSMC turned to ASML and their lithography machines to defeat the then-dominant Japanese semiconductor makers. And over the next thirty years, they grew up together.

I can feel how their cultures align. Both companies have a strain of “get it done” and "no excuses" culture. They are both very technical. When discussing issues, their focus is almost always on the science and technology. They like geeks and builders.

At ASML, people have to be ready to argue and back their points because their peers will challenge them on it. "Challenge" is one of their core values.

Perhaps it has a bit to do with the Dutch attitude, which is famously blunt.

So the company's recent massive hiring wave presents a big challenge:

How to integrate all that new talent and bring them into ASML's unique culture?

I am reminded of the big hiring sprees that the American tech giants like Google and Meta experienced during the pandemic. With all that new hiring, there inevitably has been some loosening of standards.

Right now, times are good. But as they say in a famous show, "winter is coming".

Early on, ASML lost millions struggling to compete in the stepper market, surviving only thanks to government subsidies. And even later on, the industry's frequent cyclical downturns have meant layoffs.

The company's struggles are engrained in its literal architecture. Their office buildings were designed such that they can be hived off and rented out if necessary.

As the Old Guard luminaries start to retire - like CTO Martin van den Brink, who has been around since the very beginning - I can’t imagine that people aren’t wondering about how to retain the things that made the company successful.

The fate of ASML's old parent company Philips serves as a stark reminder of what can happen if people get complacent and the old culture is lost.

Philips is still around in Eindhoven. The name is everywhere. ASML’s parts labeling number system comes from Philips. When I visited ASML, Philips was then doing a popup shop, selling lights and coffeemakers.

But today Philips is a mere husk of what they used to be - taken down by suffocating bureaucracy and excessive invention.

If you want to learn more about ASML's history and rise, I do again recommend the book "Focus" by Marc Hijink. It is very thorough and on the point.

Competitors

ASML has a monopoly on EUV. I don't think that is changing.

The technology cost billions of dollars to produce. A competitor would start 20 years behind. And with the number of buyers so limited, they would never make their R&D investment back.

But most of what ASML ships in a given year are standard DUV TWINSCAN machines. These DUV machines are the industry's workhorse and pattern the vast majority of an integrated circuit - everything but its deepest, densest parts.

The DUV machines do indeed have a lot of competition from the Japanese - Nikon and Canon. Historically, they were the market leaders, and the companies are still around today. They are a formidable crew.

And of course, the Chinese lithography-makers like SMEE are also trying to get there too. They are far behind, but highly motivated.

For this reason, ASML focuses a great deal on productivity and improving cost of ownership. It is the idea behind the TWINSCAN’s twin scan. ASML's goal is not to make the prettiest or even cheapest machine.

Or the machine that most satisfies the armchair semiconductor engineers on Twitter or HackerNews.

ASML's goal is to make a machine that offers the best value for its customers - the most accurate and productive machine possible. Even a billion dollar machine can make sense for the fab if it can accurately do enough wafers per hour.

Building Machines

I was curious to see how they assemble the lithography machines in the cleanrooms. But I probably should not have been so surprised.

The cleanroom factory reminds me for some reason of a very bright, very clean Home Depot. It is a bustle of activity with workers in blue or white bunny suits walking about - sometimes with equipment.

By the way. Despite the fact that ASML’s product is a physical thing, it seems like only a minority of its employees work the factory floors. Most work in offices - sales, software, and the such.

Anyway. Down a few hallways, we have these medium sized, well-lit rooms. Inside each room, there is an EUV or DUV system being assembled, like a patient in a hospital room. The machine gets put together using the customer's requested modules, and then tested.

The customer reviews the data and the simulation tests and signs off.

After that, the machine gets pulled apart into modules again over the span of eight to ten days. ASML then carefully ships those parts over to the customer's fab. After the customer gets the machine pieces, another team carefully puts it together over a much longer period of time. They run a final spec check and if that is all good then it's the customer's.

During my visit, ASML was in the midst of assembling some High-NA EUV machines. So I got to see some of those modules first. The easily identifiable modules are the laser, light source, illumination optics, reticle stage, projection optics, and wafer stage.

Very quickly, here is how it works. The drive laser generates EUV light in the light source. The light gets evenly spread out in Illumination. It then bounces off the reticle/mask and gains the chip design information. That light then gets shrunk down in the Projection module before finally hitting the wafer.

The Beast

I was honored to have a chance to see the Beast - a fully-assembled High-NA EUV machine, the EXE:5000.

The Beast sits in its own room somewhere within the ASML factory. It is about two stories tall - a massive conglomerate of wires, pipes, tubes, flashing lights, and metal frame. The first thing I said when I saw it was: “this thing is insane!”.

It wasn’t even everything. The drive laser, a massive CO2 laser from the German company Trumpf, sits in a nearby room. Those and other sub-fab components take up two whole floors.

The metal is particularly striking - shiny stainless steel that looks fresh out of the CNC milling machine. Everywhere you look on this thing, you see luminous steel and rough, sharp edges.

When you are in the room with the Beast, you hear pumps constantly whirring - like vacuum cleaners. I am told that you get used to the sound.

The Beast is so large that it has its own ladder and platform attached to it. You can climb up on top of it and look down on the various modules like the light source.

The old Low-NA EUV machine had its light source near the machine's bottom - I guess where you'd call its belly.

The Beast raises its light source up to the end of its spine. This lets them take out one mirror, a big boon since each mirror reduces the EUV light power by 30%. I expect that innovation to eventually filter into the Low-NA EUV machines too.

The machine was extensively re-engineered from Low-NA. Now knowing that the EUV light collector mirrors need to be replaced every so often, the designers rejiggered things to make that mirror easier to access.

The wafer stage is the same as the older machines. But because the mask field size has been cut in half, the reticle stage has been engineered to move twice as fast to overcome that productivity hit.

It moves so insanely fast you cannot believe it is also making movements that are precise to the nanometer.

And the machine's whole top can be lifted up and set down on itself. I like that little design touch. That is done using a custom-built, automatically-guided crane inside the cleanroom. Yes, it needs a crane.

I am not sure if the Beast was running at the time I visited it. But the Belgian research institute IMEC and ASML are working together on a High-NA EUV ecosystem and that includes offering a pilot line with the Beast for potential customers to make test wafers.

And to add, I do think TSMC will eventually buy these machines. Most analysts agree with me here - even the initially skeptical ones. TSMC likes the technology, just not the price.

If that is all there is, then I am sure they will close the gap and bring the Beast to TSMC fab floors.

Hyper

Where are we going next?

There is no light wavelength after EUV. You are basically at X-rays now and that stuff is hard to project and manipulate using lens. Optical is everything at ASML. They don’t seem to believe in e-beam direct.

So it is Hyper-NA EUV, then. But does that mean the machine is about to get even larger? In a few years, am I going to be standing in front of an even bigger Beast?

ASML people are aware that that’s not feasible and they talk of ways to keep the size from blowing up again one more time. We shall see what brilliant tricks they can come up with.

It took ASML 10 years to do High-NA EUV. If ASML decides to do Hyper-NA EUV right now and it takes a similar development cycle, then the company - and the industry as a whole - has some form of vague clarity of its lithographic future into at least 2040.

Conclusion

I want to sincerely thank the people at ASML for their kind hosting and thoughtful replies to my idiotic questions.

We focus a lot on ASML's machines, but I think there should be more attention on ASML itself - the machine that builds those machines.

The people at ASML are aware that High-NA EUV does not work, right now. There are many problems. There is more than a little skepticism floating around, particularly on the Internet.

But the company has faced similar doubts before. TWINSCAN, 193 nanometer immersion, and EUV all did not work at the beginning too. The first EUV machine took 23 hours to pattern a single wafer. Today’s best EUV machines can do 180 wafers per hour.

The key thing was to get going on it, and to work closely with the fabs and the rest of the semiconductor ecosystem to eventually get there. If there is a pathway to making it work, then that is enough.

To me, that is the sort of technological optimism that we all need to have more of in our lives.

Subscribe now

Computers that run the Von Neumann architecture store their programs and data in the same memory bank.

Since both have to travel the same road to get and from the CPU, we find that the system is ultimately limited not by the CPU or GPU's computational limits but by said road.

This is the famous Von Neumann bottleneck.

In a previous video I talked about in-memory computing as a way to bring a computer’s memory closer to the compute.

But making a computer that thinks like the brain - a neuromorphic system as it is called - entails far more than just memory.

For this video, a look at these brain-inspired systems and the fundamental differences between computer and brain.

Why the Brain

First things, first. The brain.

Computer scientists have long had the desire to replicate the brain. But why? What is so special about the brain? Aren't computers just better?

Computer scientists have long admired the brain's ability to function at very low energy. The brain operates at about 12 to 20 watts of power - 20% of the body's metabolic rate.

The desktop computer on the other hand does about 175 watts. Leading edge AI accelerators like the Nvidia H100 use anything from 300 to 700 watts.

The brain also does not operate at a very fast pace - with a clock frequency of about 10 hertz. Though this varies depending on what the person is doing at the time and their mental state.

So the brain does not use a lot of power and doesn't operate very quickly. And yet it is capable of so much.

Imagine a bee. A bee's brain has less than a million neurons and runs on less than a watt of power. And yet it can fly. It can navigate to flowers and back home. It can socialize and maybe even calculate things.

A bee's brain is just as capable as a 18 billion transistor system on chip, and it can do all that with just a million neurons and virtually no power.

Perhaps we should start using brains and not silicon chips for learning? Oh wait ...

The biological brain's powerful capabilities - achieved at low power - is the single most significant motivation for building neuromorphic hardware today.

Insane Parallelism

A brain accomplishes these with parallelism.

Your brain's 86 billion neurons operate without a central clock. By which I mean that they fire their signals - referred to as "spikes" - on their own time, based only on the spikes they receive from their neighbor neurons.

Measured in floating-point operations per second, the brain is an exaflop-level compute device on par with the most advanced supercomputers.

An Nvidia H100 by comparison can only do something like 60 teraflops - depending on the variant.

The brain’s lack of synchronization is in contrast to digital circuits like a CPU. A CPU relies on signals from a central clock as a reference by which to coordinate their calculations. It lets them crunch certain tasks very quickly, but synchronicity has costs.

A central clock system costs time and energy to distribute the clock signals across the system. And there is waste as each system component does not execute its task until it is told to do so as per the central clock signal.

Chaos as a Feature

No coordination implies that neural activity is chaotic, to say the least.

No doubt about that. But brains make that chaos a feature not a bug. The neural environment is full of noise - spikes firing from neuron to neuron but going nowhere.

When a neuron receives a spike from a synapse, the majority of the time it does nothing - noise. But there are so many extensive connections - with so many synapses - that enough spikes get through to the right neurons to carry on.

Again, in huge contrast to the digital computer, which works hard to make sure that every signal matters. A modern 3.2 gigahertz Intel CPU sends a "noise" signal once every 24 hours.

But as always, there are tradeoffs. You use a lot more power to achieve this very low signal-to-noise ratio. Imagine the work you need to put in to synchronize billions of transistors.

That is how the brain achieves its efficiency. Neurons literally throw things at the wall and are not afraid to make mistakes. In doing so, they find something that works. I hope that made sense.

And by living amidst chaos, brains also become shockingly resilient and flexible - even in situations of massive damage. It is not hard to see the value of this for computer hardware systems.

Neuron

People who want to copy the brain often start with its fundamental unit - the neuron.

Neurons are cells in the brain and the rest of the nervous system. When they receive spikes from their neighbors, they can - if they so choose - send their own spikes on to their neighbors.

All neurons share several common features - dendrites, soma, and the axon. Spikes enter the neuron's cell body - formally called the soma - through dendrites.

Dendrites are the neuron's input pathways. A typical neuron in the outer layer of the cerebrum - the largest part of the brain - has 10,000 inputs.

A typical neuron in the cerebellum - the second largest part of the brain, but trying hard to get to number one - has up to a quarter of a million inputs.

If sufficiently stimulated, the neuron sends a spike to its neighbors. Or a series of spikes - a neural code, though how the code works remains somewhat of a mystery.

This spike is sent through an out-path called an axon. The axon's primary goal is to ensure the signal is forwarded faithfully - though it does not always do this.

And then from there, it enters the neighbor neurons' dendrites through what we call "synapses" - an electrochemical structure for connecting two neurons.

Such a description understates their role in computation. Variations in synaptic structure over long time periods - a thing referred to as "synaptic plasticity" can subtly change the neuron spikes and how they are received. We should not ignore them.

Diversity

So those are the common structures, but the brain's 86 billion neurons exhibit incredible diversity. As you might expect with 86 billion of anything.

Some neurons - due to the complicated chemistries of their axons - can send signals faster than the Taiwan High Speed Rail. Other neurons, slower than a Taiwan sea turtle.

Some motor neurons can stretch up to a meter long. The ones in a giraffe can be multiple meters long! Other neurons on the other hand might just be 0.1 millimeters long.

And neurons - just like us - have their own preferred stimuli. In 1981, David Huber and Torsten Wiesel won half of the Nobel Prize for Medicine for showing that some neurons fired most rapidly when shown lines going in one direction over others.

As I mentioned above, the majority of the time a neuron receives a spike from a neighbor neuron, things get left on read.

So if we are to sum it up enough to get us to the next section of the video, the brain works by propagating huge, random waves of spikes throughout its billions of diverse neurons. Many of these spikes end up as wasted noise, but many are relevant too.

Neurons and Synapses merge together the work of computation and memory - doing a bit of both. Memories are stored in the relative strength of the synapses between neurons. But those synapses can do calculations as well.

That is why the brain does not suffer the Von Neumann bottleneck. That's their secret, Cap. They aren't separate.

Going Neuromorphic

So in order to create a proper neuromorphic computer,

we not only have to implement artificial versions of neurons and synapses - But also the way they very tightly bind together the memory, compute, and communications between the two.

Now, you might be thinking, "What about artificial neural networks like those running in ChatGPT? Can we call those neuromorphic?"

It is true that these neural networks started from our understanding of how the human brain works. So many of the concepts overlap.

Perceptrons for instance are a simplified mathematical model of the meat neuron. It approximates the neuron's behavior by taking in a weighted sum of inputs, applying an "activation function" to mimic the neuron's stimulation, and fires off an output to its neighbors.

But virtually all of these artificial neural networks - especially the ones running out there in the real world - run on Von Neumann hardware - which means dealing with the bottleneck.

Changes can be made to the hardware in order to improve performance and power consumption, of course. That is why we GPU and TPU. But the tantalizing possibility remains of getting game-changing benefits by running this neural "software" on actual neural "hardware".

The neural software - often referred to as "spiking neural networks" to differentiate from Von Neumann-style ANNs and Deep Neural Networks - shall be discussed some other day. Let's talk hardware.

Silicon Neurons

Many industrial and academic players have demoed neuromorphic hardware created with traditional CMOS transistors.

So normal semiconductor manufacturing processes like that for Von Neumann computers. Let me note a few of these "silicon neurons". Many are - like today's artificial neural networks - programmatic approximations of the neuron's behavior.

We have IBM's TrueNorth project from 2014 - the first widely distributed neuromorphic chip. It is capable of running inference - recognizing that a person is doing something in a video or controlling a robot.

TrueNorth is a special CMOS integrated circuit with 4,096 cores - each with 256 programmable neurons. The whole chip has 256 million programmable synapses.

A big point that IBM made was how it uses far less power than most computer systems - the chip's 5.4 billion transistors consume about 70 milliwatts.

Other semiconductor companies have shipped silicon neurons of their own too. Intel Labs has their Loihi and Loihi 2 neuromorphic AI hardware.

The European Human Brain Project was a massive ten year science project initiated in 2013, looking to do groundbreaking work in the neural sciences.

One of their projects was the BrainScaleS project, which was first released in 2011. A second version was released in 2020 with improved local learning capabilities.

It is a mixed-signal ASIC chip that uses analog electronic circuits to mimic the spiking neurons of the brain. Very interesting though its analog neurons are not as flexible as its digital counterparts.

And then in 2019 we have the Tianjic project, a hybrid chip created by scientists at China's elite Tsinghua University. It is a hybrid platform that attempts to run various types of neural networks - including those designed for neuromorphic hardware.

Crossbar

How do we implement a neuron network in hardware? Let us briefly look at how IBM's TrueNorth does it.

TrueNorth is fabbed on a 28 nanometer CMOS process. As I mentioned, it has 4,096 of what they call neurosynaptic cores. It is capable of running widely adopted convolutional neural networks.

Each basic TrueNorth core has 256 axons, 256 neurons, and synapses fully connecting them in a 256x256 crossbar structure, implemented using SRAM.

Each of the axons are given a synaptic weight depending on the neuron that they are connected to. And each neuron has a state or "membrane potential" as well as a particular threshold for sending on a spike.

The system works on cycles. During the cycle, spikes travel to the neurons through axons, which affects the spike's value. The neuron collects the incoming spikes into buffers and then evaluates them.

They then update their own membrane potentials accordingly. Once having done that, they compare it against their threshold. If the membrane potential meets or beats the threshold, the neuron sends a spike of its own. Spikes can be sent to local neighbors or outside the core itself.

Just like the real brain, TrueNorth does not have a global clock. The elements in each of the cores work asynchronously in cycles, doing things only in response to events. They also operate at low frequency - 10 hertz.

And to implement the neuron's inherent randomness and noise, each core has a random number generator that can raise the thresholds for creating a spike. Or randomly delay or even halt a spike's transmission through a synapse.

It's like that episode of It's Always Sunny in Philadelphia. Wildcard!

CMOS Shortcomings

A significant benefit of building these "silicon neuron" chips using CMOS processes is that we get to draft in the wake of the chip giants.

They share some of the benefits of the brain. TrueNorth for instance looks to be very scalable, sips relatively lower power compared to GPUs, and can run widely used neural network software to get accurate results.

But there are drawbacks. A test of TrueNorth's performance found tradeoffs between energy efficiency and accuracy. If we want the model to give us competitively accurate results - which can make a big difference in a commercial context - we need to use more power.

These circuits are also quite large, since each of the cores have to have their own memories. Von Neumann machines benefit from having a single, very dense central memory bank. Having denser, higher capacity allows us to accommodate the larger models that are so in vogue.

And lastly, the brain is an analog device - digital devices are an ill fit for replicating their behavior. So we need to incorporate the analog element, which limits the system's flexibility.

These disadvantages have pushed scientists to look beyond CMOS for new ways to implement neuromorphic devices. The most popular such approach is the memristor.

Memristors

In 1971, Leon Chua - then a professor of Electrical Engineering and Computer Sciences at UC Berkeley - published an article proposing a new type of circuit.

By looking at the relations between the three major circuit elements, he proposed the existence of a fourth - the "memory resistor" or memristor.

This paper was very difficult to read and I will freely admit that I did not get it. So I am not going to explain any more than that. Anyway, The idea fell on the wayside until 2008, when scientists at Hewlett-Packard announced a physical implementation of the memristor.

The original 2008 memristor was a simple metal-dielectric-metal sandwich with two terminals, or points of connection. In this case, the metal electrodes were made from platinum and the dielectric was titanium oxide.

If we are to apply a voltage pulse to the memristor, the film can switch its electrical resistance - flipping between an insulating and conducting state in a non-linear fashion.

What is it about the memristor that makes it so suitable for neuromorphic computing? Their key characteristic is that the value of that electrical resistance is dependent upon the history of the voltage passing through it - ergo the name "memory".

Even better, it can remember that history even when the power goes off. This makes the memristor a form of "non-volatile memory" like flash memory or a hard drive.

People quickly drew connections between memristor behavior and that of biological synapses. Scientists envisioned them as a non-volatile memory for in-memory computing. For the past 15 years, such devices have been at the center of neuromorphic study.

But there are a few challenges. The first has to do with manufacturability. It can be very difficult to produce enough of these memristors uniformly and at scale. Can they handle many cycles of resistance switches? And how long can they store their data? Engineers are still working through these questions.

Also, the changes in the memristor's resistance are non-linear, which makes it somewhat challenging to program for. New software paradigms in neural network programming may be needed for memristor-based neuromorphic computing to work. The work goes on.

Conclusion

I think the biggest challenge with these neuromorphic systems though is the competition.

Nvidia and other adherents of "Huang's Law" are leading the way. According to Huang's Law, silicon chips powering AI more than double in power every two years.

In the past 10 years, AI inference performance in the GPU has improved 1,000 times. It might be hard to compete against this. But maybe we don’t have to. Considering each system's advantages, we might start seeing more hybrid systems.

Perhaps we can put neuromorphic and Von Neumann chiplets together so to give us the best of both worlds. Like a mullet. Business in the front. Party in the back.

As for TSMC Arizona, that fab will eventually make chips. Local news just reported that they added their last steel beam in February 2024. But as Mark Liu said, the future depends on the size of the subsidy.

If TSMC wants 40% support like with Japan on a $40 billion project, that's $16 billion in subsidies. Will the administration want the PR optics of giving $16 billion of taxpayer money to a foreign company in an election year?

Maybe. Maybe not. I should add that that $16 billion doesn't have to come in the form of cash subsidy. The investment tax credits will do some heavy lifting too.

I should not have been so skeptical. Uncle Sam came through with the chips and salsa, baby 🎉! From Bloomberg:

The US plans to award Taiwan Semiconductor Manufacturing Co. $6.6 billion in grants and as much as $5 billion in loans to help the world’s top chipmaker build factories in Arizona, expanding President Joe Biden’s effort to boost domestic production of critical technology.

Under the preliminary agreement announced by the US on Monday, TSMC will construct a third factory in Phoenix, adding to two facilities in the state that are expected to begin production in 2025 and 2028. In total, the package will support more than $65 billion in investments at the three plants by TSMC, the go-to chipmaker for companies such as Apple Inc. and Nvidia Corp.

TSMC’s third fabrication site, or fab, will rely on next-generation 2-nanometer process technology, and is slated to be operational before the end of the decade.

I am frankly tired of doing videos about TSMC for now. So this newsletter is all that we are going to get. Here are some raw thoughts about the TSMC CHIPS deal, based on what I have seen. Add disclaimers - things are subject to change. Memorandum of Intent etc.

$11.5 billion is a lot of money. And then if we add the investment tax credit, which I think is underrated. It credits you 25% of capital investments in semiconductor equipment and facilities. So TSMC upped their total investment to $65 billion. Assuming that is all capital, that can be up to $16 billion in tax credits - which are deducted directly from tax liabilities. TSMC did about $30 billion USD of pre-tax profit in 2023. Tax credits will be nice.

So all in all, my crappy napkin estimate is that TSMC gets something like $28 billion in US government support. That well clears the 40% requirement that I had mentioned in the Japan video. I cannot imagine another non-American domiciled company ever getting this much government support before. Can you?

(Though Samsung is probably going to get something comparably monstrous. Samsung is upping their Texas investment to $44 billion. So my guess they eventually get a total package worth 40% of that as well all-in with credits, loans, and taxes.)

Nikkei Asia also posed this issue to a US government official and they said:

TSMC is the most advanced leading semiconductor company in the entire world. What taxpayers' dollars are achieving here is of the utmost importance for our country and for our economic and national security

Arizona Ahead of Schedule?

TSMC is taking the Arizona plant very seriously. As soon as I heard that they sent their VP of Operations Wang Ying-lang (王英郎) over to Arizona like the freakin’ Wolf, I knew that things were going to get DONE.

I briefly mentioned him at the end of another video.

Wang is TSMC’s youngest ever deputy general manager, and a crazy guy who got his PhD during the day while also working the TSMC night shift. Nobody at TSMC knew.

Dude and his team has been chugging. UDN, who I generally find pretty reliable, published a report saying that TSMC is pushing as hard as it can to get some chips out by end of 2024.

Now, you might want to say - “Oh TSMC was sandbagging until the CHIPS money comes out”. To that, I would say: Eh, progress has been going all the while. They just raised the last steel beam in Feb 2024.

Considering 6.5 months for fabrication and 1 month for testing, the wafers have to go into the oven like right now if they want to be out by end of 2024. Right now they are rushing to connect the power and get that first production line up and running by that deadline. Two-ish months from last steel beam to a N4 wafer entering the production line does not sound like sandbagging to me.

(Reader Paul pointed out that the LinkedIn post points out that this beam raising is for an auxiliary building for the second fab. I stand corrected. The first Fab is already done so it is not really a mad dash anymore.)

If Wang pulls this off, the man is a legend and deserves to be a future CEO.

N2 Fab

The Third phase of the Arizona will be an N2 fab, and TSMC says that this will be operational before the end of the decade - 2030.

(I will be in my forties by then. NO!!!!! REVERT)

The N2 part gets the headline. But I should again point out that TSMC will be on to their A14 (and I guess A10?) nodes by then. Guess that one comes out in Taiwan some time in 2028?

All in all, this is a great thing. We get more silicon to build towards AGI - especially since indications seem to point to this N2 node being particularly well suited for AI calculations. What with the GAAFET and all. And TSMC seems to be building it literally everywhere they can - Hsinchu, Kaohsiung, Taichung. Maybe even Japan.

CHIPS Act

I asked before, What was the CHIPS Act Office doing? What was taking so long? Someone finally got back to me and the answer to me seems pretty reasonable.

As it turns out, they were working hard to disburse the limited amount of money and it takes a long time to do that. They were modeling out the cost structures of the fabs, basically to sanity-check whether the proposals coming to them were being reasonable or not. And they were working hard from a point of zero, setting up a new office and staffing it. That takes time, unavoidably. That’s fair.

I wonder if these announcements - and their sheer size - put the ball back in Japan’s court. They are undoubtedly negotiating for a Phase 3 Fab of their own over there, and I wonder if the CHIPS Act affects whatever aid package TSMC will get.

Taiwan

I think there might be some consternation back home in Taiwan. Oh, is the US stealing TSMC again? They all need to get over this. TSMC is now building fabs in Japan, the US, and Germany - and the governments are writing massive checks to make it happen for them. It is an amazing opportunity for Taiwan!

I hope it allows those countries to see the best side of Taiwan. I hope it lets them appreciate what Taiwan can do. And at the same time, I hope this lets TSMC and their Taiwanese suppliers see some of the best of what those other countries can offer - with open pathways for both sides to learn.

Let’s make some chips!!

Share The Asianometry Newsletter

For the past three weeks I have been traveling through Japan and the United States.

This trip has been part sight-seeing and part learning tour with the goal of understanding more about the AI and AI chip boom landscape outside of Taiwan.

Now that I am finished with the big conversations, I wanted to sit down and share a few thoughts.

The Trillion

There’s been a lot of news about a chip venture thing that Sam Altman of OpenAI is working on for the AI industry.

The news came slowly. First there was the report - released at around the time he was temporarily ousted from OpenAI - that he was raising money from Middle East investors for some chip venture. Back then it was just "billions".

Since then, you have also gotten news that Sam Altman and his team have talked to TSMC, people at Abu Dhabi, and the US Government.

Just as I was leaving Tokyo for San Francisco, the WSJ broke the report with that eye catching $1-7 trillion figure that has got everyone buzzing. It was one of the first things that I started asking people about when I arrived in Silicon Valley.

Many people agreed with me that the number is a bit too much. Perhaps it can be a negotiating tactic to set expectations for future talks. Once you start talking trillions then hundreds of billions no longer feels as substantial.

Then you have the Information adding a bit of context to Altman's remarks, saying:

> But in reality, Altman privately has told people that figure represents the sum total of investments that participants ... would need to make, in everything from real estate and power for data centers to the manufacturing of the chips, over some period of years.

The ecosystem concept makes a lot more sense. Total semiconductor sales in 2023 were about $520 billion. Total capital expenditures - according to the trade group SEMI - were about $140 billion.

But even if we were to assume that this trillion is spread out over five years and diluted with real estate expenses and power, that is still another COVID-like step function upwards in capital expenditure.

The semiconductor industry is a conservative one. Many of the people there are old-heads who have seen many a cycle of booms and busts downstream. Sam isn’t really the first guy to knock on a foundry’s door saying that his use case is going to change the world.

Scaling Laws

The concept driving this investment is something called the "scaling laws".

The name dates to a 2020 paper posted by OpenAI titled "Scaling laws for neural language models". I am not going to go over the fine details, but the gist is that if we combine more data and compute, we get better results (i.e. less loss).

This was one of the core ideas that helped OpenAI make the GPT-series a reality. Ilya Sutskever, one of the company's cofounders, mentioned something like this in an appearance on the No Priors podcast back in November:

> I was very fortunate in that I was able to realize that the reason neural networks of the time weren't good is because they are too small. So like if you tried to solve a vision task with a neural network with a thousand neurons, what can it do? It can't do anything.

> It doesn't matter how good your learning is and anything else. But if you have a much larger network then it can do something unprecedented.

GPT-3 is big. GPT-4 was even bigger, performing far better. And GPT-5, whenever it comes out, will be even bigger. So far there are no indications that the scaling laws have broken apart.

I find the parallels with the semiconductor industry's Moore's Law very interesting. The two "laws" do not say the same thing, but they might have similar effects on their respective industries.

In the 1980s and 1990s, Moore's Law became the rallying cry of the entire semiconductor industry - the metronome keeping time for every company from Santa Clara to Tokyo to Hsinchu.

There is a chance that the Scaling Laws can make a similar impact on the AI industry. A simple set of easily understandable rallying cries that drive R&D roadmaps for whatever years to come.

There are a variety of arguments against the Scaling Laws. For instance, people have commented that we have basically pulled all the existing data across the entirety of the Internet.

But there are ways around such things, if the money is there to make it happen. Physics problems have cropped up ahead of Moore's Law, and the industry derived new engineering solutions: the High-K Metal Gate, FinFET, DUV Lithography, Dry Etch, Ion Implantation, SEM, and so on.

The bigger question is whether the money is there to continue driving this investment. More on that later.

For a way deeper review of the technical issues behind scaling, I recommend Dwarkesh Patel's post - "Will scaling work?" His podcast is quite excellent too.

AI Chips & Nvidia

Where there are profits, there are competitors coming out with ideas to take them.

There has been a lot of ink spilled on Nvidia and their competition - particularly the multi-pronged attack on Nvidia's AI accelerator profits. I don't want to add too much to this.

I do question whether the Nvidia fortress is going to be as assailable as it might at first seem. Jensen is responding to the threat by aggressively ramping up the annual updates to their accelerator lineup.

For more information on this, I refer you to the report from SemiAnalysis. It goes into much more detail. Dylan's great.

But this relentless pace reminds me of the days of old Nvidia during the Graphics Cards Wars, when Nvidia released a new product every six months to the market. It worked for them then. Why not try it again now for the AI Chip Wars?

Nvidia can maintain this rollout speed because - as Jensen implied during his appearance at the Acquired podcast - they ship before they test.

They use the latest computer software design and emulation tools to model and ship new GPU designs to the market without first fabbing a physical chip:

> The reason why we needed that emulator is because if you figure out how much money that we have, if we taped out a chip and we got it back from the fab and we started working on our software, by the time that we found all the bugs because we did the software, then we taped out the chip again. We would’ve been out of business already.

The Nvidia teams will try to tape out "perfect chips" as Huang said, but this inevitably will cause problems for customers. For instance, deployments with drivers and other affiliated software that won't work all that well at first.

But the Nvidia brand can take a hit like that, whereas it is unlikely that the AI chip startups will be able to. It takes several years to build a team and then get that team good enough to ship a working product.

This high speed of iteration will be rough on customers. Many have no choice but to buy what is available, but it stings to spend tens of thousands of dollars on a machine, only to have a vastly better one coming out just a short time later.

Giants and Verticalization

Nvidia is less concerned about the startups and even the established silicon players than they are about the tech giants.

The tech giants - Microsoft, Google, and the like - are the ones driving the current investment spend in AI today. They are also the ones with the most incentive to cut Nvidia out of their margin.

They would do this using custom-designed chips or ASICs. Large companies have sought these since the good old days of the computer - commissioning an ASIC to replace an entire board of discrete electronic components to save money.

An example of this would be the Apple IIe, a third-generation version of the Apple II PC powered by a custom ASIC.

A more modern example would be the Google TPU, which right now is in its fifth iteration.

It makes up a big part of Google's compute advantage over OpenAI and the other AI labs.

Microsoft seems to be the giant pushing hardest on vertical integration. CEO Satya Nadella said in the Q2 2024 earnings call that most of the usage in Azure AI services is coming from inference rather than training. Those activities are easier to push to a custom-designed ASIC.

In addition, Microsoft has been working on vertically integrating other parts of the database stack. For instance, the Information's recent report that they are working on a network card to shuttle data between servers.

This type of vertical integration implies either that the scale of the AI database stack is so large and costly that every penny counts/will count. Or that the growth in the industry has petered out, leaving players to compete on price. These scenarios both feel weird to consider.

What does this vertical integration trend mean for Nvidia? It will take some time for the other tech giants to ramp up their AI chip designs. Apparently the first Google TPU was a very bare-bones product. So in the short term, things will be as they are.

But as those chips get better in the medium to long term, it makes sense that Nvidia push as hard as they can to always win the performance-cost crown. Who knows.

Financially Sustainable?

I want to move away from cost. I think the cost benefits of vertical integration are significant enough to matter. Now, I want to move to the other side of the financial equation.

One of the really big questions that I wanted to answer coming to the United States was to learn more about whether or not this OpenAI boom is Real. Real, with a capital R.

There has been so much progress in the AI industry in the past year. For instance, the OpenAI text to video Sora model. The improvement in the models' output quality has been impressive.

But even if the technology is amazing, it does not seem like a hit product by itself. We need to embed them into compelling products. Are there indications of such products breaking through into the public? Who is making money from actual consumers with them?

The massive investments required by Moore's Law and the semiconductor industry were driven by real demand from various downstream industries. First the military, then consumer electronics like radios and calculators, then the PC, smartphones, and then cloud computing.

These were all things that ordinary people wanted. To me, it feels unlikely that the truly large investments in AI will happen unless those ordinary consumers start buying these services in a major way.

The Financial Times did report that ChatGPT hit a $2 billion run-rate revenue in December 2023. That is basically the only solid piece of news indicating that people are paying real money for an LLM service.

Now, we should note that it is still early in the generative AI boom. ChatGPT is a little over a year old.

But I also want to note that the iPhone was first released in June 2007.

About two years later in 2009, Apple sold 20 million iPhones for like $13 billion in revenue. Growth continued, with 40 million iPhones sold the year after that, and then 72 million and 125 million after that.

Okay fine, the iPhone is the greatest consumer technology product that history has ever seen. It is a difficult example to live up to.

ChatGPT is still one of the fastest growing consumer tech products in history, depending on what you think about Threads. And people are still using it a lot - Altman recently said that ChatGPT generates 100 billion words a day.

And Facebook took a while to produce revenue - at first investing to build up the audience. Perhaps ChatGPT is doing the same Aggregator approach, gathering users to monetize later. But few products cost as much to use. AI probably has shorter a leash than we think.

Other than ChatGPT, the one product that everyone in the industry seems to have their eye on is the Microsoft Copilot subscription service. We want to see if this $20 a month offering catches on with people and enterprises.

If it does, then we are really off to the races ... the sky really is the limit ... we are so back ... and so on, insert bombastic metaphor here. If not, then we have to adjust some mental models.

Conclusion

There is one last thing that I should mention. Maybe where the LLM revolution will be most Real is when it is embedded into the products we know today.

For instance, a few people I spoke to insisted that I am overlooking the impact that AI automation will have in supercharging advertising sales.

Like Google's Performance Max ads, which use AI to automate the creation, deployment, and targeting of new ads. Ads like Performance Max apparently generate tens of billions of dollars.

Or how when Apple disabled cookie tracking in 2021 with the App Tracking Transparency thing. Meta/Facebook at first said that this would cost the company over $10 billion in sales.

But over time, Meta somehow managed to claw back its targeting accuracies using AI, collecting data from its Conversions API.

So maybe the killer app for AI is just more ads.

But Ben Thompson and others have been saying for a while that AI is an enabling technology - a technology for making the rich richer, rather than making a whole new class of rich people like the PC or electrification was.

So for me, if this is all that it is, then it is a bit disappointing. But it makes the AI boom nevertheless real.

Anyway, these are my reflections from my trip to the United States. I hope to make more frequent trips in the future and have more interesting conversations with people in the space.

If you are interested in having a chat, shoot me an email. Maybe I will be coming to town again and can have that chat in person.

I want to call out a Substack written by a friend. Tanj boasts a monstrous boatload of knowledge about semiconductor manufacturing - particularly in memory. He recently started his new Substack and I want you all to go subscribe to it. Like now.

Poratbo

Exploration of new ideas in technology and science of the small. "Plenty of room at the bottom" - Feynmann, Caltech, 1959

By Tanj

This below post on FinFETs is extremely informative and I loved it.

Poratbo

Pipe.3: The Chip Base and the Basics

In the last post the scope of the cell library was described. It will be for static logic design, with around 30 of the most useful and optimizable logic building blocks, including some of the circuits for clocking and latches, and the largets cells will be full adders. Now we get started…

Read more

2 years ago · 2 likes · Tanj

Almost every integrated circuit will have an analog component. That component is often pretty small. Sometimes as little as 3% of the total thing.

Yet this tiny little piece of analog circuitry regularly takes up over half of the integrated circuit's total design cost.

That is because analog chip design is a goshdarn art. Hand-drawn by a chip-Michelangelo staring up at the ceiling night after night.

And if there is one thing we all know, it is that computers can't do art like artists can ... right? In this video, we are going to talk about the art of analog circuit design and AI’s potential to help.

Beginnings

The digital world is made up of 1s and 0s.

But the real world is analog.

Transistors are well suited for digital signals. They either open the floodgates and let the current saturate: 1. Or they shut the gate and let nothing through - 0.

Analog signals aren't like that. They are continuous - a waveform where each value in the signal is not abruptly different from the previous or the next value.

Common analog signals include those for sound, temperature, pressure, and electrical voltage. These things come from the real world.

Digital on the other hand is all about numbers - a concept that humans made up.

Analog Systems

Analog electronics take in these analog signals and do things to them in order to create a desired output.

Every integrated circuit or system-on-chip has an analog component. Information from a sensor or a transducer might first enter the circuit as an analog signal.

Transducers are devices that convert one form of energy into another - things like microphones. Other analog inputs include button press-down actions or clock timing signals - think a vibrating quartz crystal.

The analog circuit then accepts the input signal and manipulates it. That can mean filtering, counting, adding, or subtracting. If necessary, we might convert the signal from analog to digital or vice versa.

Once the work is done, the system's output signal is sent to where it can interact with the real world again using another device like a display or another type transducer.

The recent trend has been to digitize all the things and that has yielded real benefits. For instance, digital audio signal processing for noise cancelling features in hearing aids.

But some functions will always be analog - if only at the borders between digital and the real world. You hear me, you digital dweebs? Y'all ain't ever killing off analog!!

Designing

Before we talk analog chip design, I want to briefly go over how we design semiconductors in both the digital and analog worlds.

I recommend watching some of the other videos I did on semiconductor design. But I am going to hit the main beats just so that we are all on an even footing.

First, you meet with the customer and decide their specifications. You translate these specifications into sub-systems and components.

Next up, we have a circuit designer choose the right circuits, their sizes, and their interconnections to fulfill the specifications.

Once that is done, we then do the physical design layouts. We need to figure out where to put these circuit devices and lay out the interconnections between them.

After that, we verify the whole design to make sure it meets all the foundry's manufacturing specifications.

The foundry provides design rules to you that you cannot violate.

Then we ship. A circuit design - just like a video for this channel - is never really done. We just tear ourselves away from it and ship it so that we can move on to the next one.

Digital Versus Analog Design

So why is this design thing so hard for analog? Digital and analog chip designers both complain that their jobs are hard.

They are both right, of course. The jobs are hard. digital designers have to deal with the interactions of billions of transistors - and all their tradeoffs.

But analog chip design is harder in a different way. In digital designs, every transistor is roughly the same size. The challenge is dealing with how they interact with one another.

Analog designs have far fewer devices in them - tens and hundreds rather than billions - but each of those devices can have different physical sizes and electrical properties. And those circuits are highly sensitive to their surroundings - neighboring circuits, the environment, and so on.

Parasitic Extraction

Doing a physical design layout is particularly nasty. You need to run the layout through a simulator to perform a process known as "parasitic extraction".

Which sounds like something you do to a tapeworm, I know, but hear me out.

Every analog design’s components and connections bring some amount of parasitic elements. These elements can take various forms but the two major ones are parasitic resistance and capacitance.

Parasitic resistance refers to reductions in the electric current as it flows through a device or an interconnect. Both those things are made from real world materials, and all real world materials have some amount of resistance.

Parasitic capacitance on the other hand refers to unwanted storage of an electrical charge that exists simply because two components or interconnects in the circuit are close together.

This exists because an integrated circuit is made up of conducting layers - the metal interconnects or the metal silicon substrate - and insulating layers - silicon dioxide.

When you put an insulating layer in between two conducting layers ... well, gee. That kind of looks like a capacitor - an unwanted capacitor.

Parasitic resistance and capacitance are like taxes - unavoidable but very much unwanted. They delay the speed of the signals, consume more power than you wish, and degrade the circuit's overall performance.

When we do parasitic extraction, we are trying to calculate how much parasitic resistances and capacitances the whole circuit has. This means modeling the electromagnetic effects of all the various devices and wires in said circuit.

This is complicated and computationally intensive. Capacitance is particularly nasty. Imagine all the many devices and the multitude more interconnects on the chip.

Knowledge Intensive

This all makes analog design an evolutionary, effort-intensive, knowledge-intensive process that starts from the top down.

When preparing his netlist, the designer manually calculates the variables for things like the size of the devices.

Then he runs it through a basic analog circuit simulator - the 50-year old SPICE program is still widely used - to predict how the circuit performs in silicon. Not good enough? Back to manual calculations.

When done, he passes it to a physical design engineer who does the layout, calculates the parasitic effects, and checks the layout against the foundry's design check rules.

Once he is done with the layout, he needs to pass it back to the circuit designer so he can recheck whether the design still achieves its stated goals.

This happens over multiple cycles and for each of the analog blocks and sub-blocks in the whole circuit. Even the smallest specification change in either the layout or the circuit netlist requires a full review.

This process is long and arduous with as much backtracking as the first Halo game. Compare that to digital chip design, which can be highly regimented with multiple teams sequentially handling the design process through each stage.

The wide variety of schematics, the number of conflicting requirements, the many parasitic elements and the higher order effects between all these devices/interconnects. That all matters. Design is all about optimizing for certain variables within a space.

And in analog design, that space is galaxy-sized. Without straightforward, reliable heuristics to fall back on, analog and mixed signal designers rely on experience and knowledge built up over time.

Famed designer Bob Dobkin likens it to learning a language. You start by learning the grammar rules and the dictionary.

And after learning for many years, you finally can read or write documents without having to consult a dictionary or grammar checker all the time. You simply intuit what it means.

It is also why these designers care so much about a design's "aesthetics". Lacking solid benchmarks, the aesthetics of a particular design is an indicator of its correctness.

Leading Edge

When we think about today's super-sexy 10, 7, and 3 nanometer process nodes, analog does not often come to mind.

Yet the analog problem is often in the minds of today's leading edge system-on-chip designers. Particularly those designing chips for mobiles like iPhones.

Many analog designs struggle to work well with leading edge processes due to these aforementioned parasitic effects. Particularly, parasitic resistance in both the transistors and the interconnects between them.

Parasitic resistance occurs in leading-edge FinFET transistors at two places. First where the current travels up to the 3D fins from the rest of the circuit and vice versa.

And second, the current hits resistance as it passes through the FinFET gate itself. This is in part due to the gate's very thinness. There is just not enough conducting metal for the current to travel through.

A similar effect is behind the parasitic resistance in the metal interconnects. Leading edge interconnects are made from copper surrounded by a barrier layer of tantalum nitride.

As the interconnect shrinks, the tantalum nitride layer stays the same size. So the proportion of low-resistance copper to high-resistance tantalum nitride decreases - increasing the interconnect's total resistance.

For example, cutting the metal pitch from 80 nanometers to 48 nanometers - a 40% reduction - raises the line resistance by 6 times.

Anyway, this long-winded digression is basically to illustrate the importance of analog design at the 7 nanometer level and below. Parasitic resistance levels worsen with every new node and dealing with it goes on top of all the other challenges of analog design just like a cherry on a delicious Black Forest Cake

which I haven’t eaten in such a long time because I am on a diet right now but I am just really badly craving it what with the chocolate shavings and the frosting ... oh wait where was I?

Uh ... so like as I said before, every SOC has to have an analog component and it is becoming a big bottleneck in the design. Automating analog design would be sweet.

Automate This!

A great deal of digital EDA tools blew up in the 1990s.

And now, digital hardware designers can write high-level specifications in a language like Verilog. The EDA takes that and automatically translates it into circuits.

EDA's success in digital design prompted people to do the same for analog. And we have been trying ever since. These historically targeted individual parts of the design process:

First, circuit sizing or design parameter optimization - the work of picking the value and sizes of the circuits.

The general approach has been to create an overall cost function and then try to solve for it. Usually using something like simulated annealing or genetic algorithms.

Second, laying out the circuits and their interconnects. Procedural layout software does exist - SLAM and ILAC are prominent examples - but they heavily rely on either templates or pre-defined rules. And their outputs still fall far short of what an experienced human can do.

There are a few attempts to create an end-to-end analog design product - like the Berkeley Analog Generator. Designers don't manually draw anything but instead lay out several high-level principles.

So all in all, software packages either need the human designer still in the loop. Or, they impose a great deal of constraints solely to cut down on the number of possibilities and make the computational task easier.

People in the industry haven't warmed up to these tools yet, preferring to do it largely by hand to get maximum control. This and the overall smaller size of the analog semiconductor industry has meant that analog EDA tools remain niche.

Machine Learning

Go used to be a game said to have too many possibilities for traditional computers to work out.

Powerful machine learning models trained on huge datasets helped computers gain the intuition to conquer Go. Now what about analog chip design?

In recent years, academics have explored the use of neural networks to help place the devices in an analog circuit. You show it a whole lot of human-done layouts and train the model on them.

There are a number of AI startups working on this problem. I recently had the chance to speak to one such team out of Toronto - Astrus.ai. They are trying to use a game-playing AI kind of like those behind AlphaGo to play the "game" of analog layout.

There are a few other open source solutions. Recently, there is ALIGN - which stands Analog Layout, Intelligently Generated from Netlists. It uses a combination of imposed constraints and machine learning models to route and place devices in 24 hours without a human in the loop.

Another notable one is MAGICAL, an open source, fully automated analog and mixed signal layout system. It uses gradient descent - an optimization algorithm for training machine learning models - to iteratively place the devices and wires.

There are two big issues with these open source solutions, however. First, the design needs to go to the foundry. The foundry has a set of design rules as part of a Process Design Kit, and that PDK tends to be proprietary. That doesn’t jive with open source.

Second, the models needed to do industrial-sized analog designs have to be quite large. This means having to spend a fair sum of money on GPU training time and good data sets. Open source might have troubles getting these.

Conclusion

The complexity of analog chip design helped put food on my family's table. My father was a chip designer. For whatever reason, he fell into designing analog chips.

A long time ago, he told me - without explanation - that analog chip design never got automated. And because of that he was lucky, because a lot of designers in other parts of the industry lost their jobs.

It also explains why he so rarely used a computer while he worked - sitting there in his home office with some paper, a ruler, and a pencil.

I am glad to have been able to do this video and more deeply learn the nuances of what he meant back then.

I wonder if these new machine learning techniques can ever conquer the analog layout problem. Good luck to everyone trying it. But in some little way, I hope it won't happen. Because analog is the real world. And its design is an art.

Share The Asianometry Newsletter

I recently read an interview with ASML's (now retiring) CEO - Peter Wennink - about the issue of banning Dutch exports of lithography machines to China.

In it, he talked about the notion about ASML restricting exports of older, DUV lithography machines.

> Maybe they think we should come over the bridge, but ASML has already surrendered. We are no longer allowed to supply EUV to China. And EUV is half of our turnover ...

He also pointed out that the lithography ban directly benefitted American semiconductor equipment makers of items not covered by the bans:

> If China still wants to do something with advanced chips, they have to buy very advanced deposition and etching machines to 'shrink' even more. And they mainly come from the US.

His feelings illustrate the delicate, difficult challenge of keeping valuable, militarily powerful dual-use technologies out of the hands of the enemy.

The China semiconductor issue is not new. The United States and the West have long encountered these same fractures and sentiments during its long engagement with the Soviet Union.

In 1987, a massive scandal broke in Japan that illustrated how weaknesses in the system caused a major lapse.

In this video, we are going to look at one of the most serious geopolitical technology scandals of the 1980s: the Toshiba-Kongsberg Incident.

Beginnings

World War II greatly stimulated Soviet industrial capacity. One can argue that - even despite the titanic war damage - the Soviet economy was stronger after the war than before it.

The reason for this expansion in industrial and economic capacity was technology acquisition. Before the War, trade with Europe was tiny - just 1% of GNP.

The years during and immediately after the war, however, the Soviet Union acquired and received a wave of Western technologies via the American Lend-Lease Act, reparations and the wholesale confiscation of German assets.

Lend-lease provisions alone infused the Soviet Union with over a billion dollars worth of modern American equipment. This equipment played a significant role in building up the Soviets' indigenous technological base and military after the war.

Post-War

Export controls in the United States date back to 1917 during World War I, but it was during the Second World War that they became what they are today.

During the 1930s, they were used to enforce American neutrality. Moral outrage over Japanese bombing of civilians in China disallowed the exports of military equipment or related industrial supplies to warring nations.

After World War II ended, the United States rapidly began normalizing its wartime restrictions, moving back towards free trade.

In 1944, there were 3,000 restricted items in the United States. Three years later in 1947, just 352 restricted items.

But various parties including the American military grew concerned that unrestricted free trade would enhance a growing competitor.

One particularly nasty issue regarded the export of American petroleum. In February 1947, those became unrestricted and Russian tankers showed up to California, loading up over 600,000 barrels of oil.

This evoked memories of pre-war oil shipments to Imperial Japan, a tender subject. Furthermore, it fanned inflation concerns due to the potential of domestic shortages.

Anti-communist feelings were building in congress. Various Congressmen wanted to end what they called the "economic appeasement" of the Soviet Union, and repeatedly attempted to pass legislation on their own.

The 1948 Prague Coup - in which a Communist Party captured absolute power in Czechoslovakia - and the Berlin Blockade only intensified these feelings.

This eventually led to the Export Control Act of 1949. It carried over many wartime powers into peacetime - creating a licensing system to protect national security and advance foreign policy objectives.

Marshall Plan

Imposing export controls solely on American nationals would weaken their effectiveness and disadvantage American businesses. So Europe had to get on board. Luckily, there was a mechanism.

In 1947, the Marshall Plan was about to head into action, bringing over $12 billion of reconstruction aid to sixteen European countries.

The USSR rejected this aid - which was obviously not being done out of the Americans' good hearts - and proposed their own trade bloc. This would eventually become the CMEA and include Bulgaria, Czechoslovakia, Hungary, Poland and Romania.

American policymakers saw this rejection as a reason for wide ranging export controls. Why give aid to any country that traded with the Soviets?

The Americans drew up several lists for controlled items and sought voluntary export restriction agreements from the countries receiving Marshall Plan aid.

Rather than negotiating with each of those countries in a multilateral forum, they decided to first pass it to the British and ask them to build up support amongst their European allies.

COCOM

The United Kingdom circulated the list amongst representatives of the Organisation for European Economic Co-operation or OEEC.

The OEEC was formed for administering Marshall aid and is the forerunner of today’s OECD.

The biggest concern amidst the representatives was the American request that Marshall Aid recipients impose export controls against non-Marshall Aid recipients - essentially the CMEA.

For what it is worth, the actual items on the lists back then - things like bismuth metal and broaching machines, which can be used to create patterns in metal - weren’t a huge concern.

The biggest holdouts were Sweden and Switzerland, who valued their neutrality. France, the Netherlands, and Belgium were also concerned.

But developments in Communism caused new alarm. The explosion of the first Soviet nuclear bomb. Shortly after that, the end of the Chinese Civil War and the establishment of the People's Republic of China. All of this reinforced the notion that an economic blockade was necessary for the containment of Communism.

Thusly, representatives of seven countries came together in November 1949 for a meeting - Belgium, Canada, France, Italy, Norway, the United States, and the United Kingdom.

Here begins the informal international export control group called the Coordinating Committee for Multilateral Export Controls, or COCOM.

A titanic victory for the Americans and one that could not have happened without Marshall Aid. As one official grumbled, "Western Europe sold out their trade principles for good American cash".

Effectiveness

Unlike NATO, there was no formal treaty for COCOM and it was all voluntary. As in there was no enforcement mechanism. So as you might expect, the whole thing was like herding cats.

Even after COCOM formed, it took the outbreak of the Korean War in 1950 for the United States to get consensus on a second, less controversial list of some 200 items.

Over the next few years, COCOM expanded to include Denmark, Portugal, West Germany, and more. After Japan regained its sovereignty in 1952, they requested to join and were accepted.

Despite the addition of these new partners, the United States was always at COCOM's heart - the driving force of its effectiveness.

Managing export controls is a balancing act. People love exports, and serious differences can develop between allies over it. Excessive restrictions can lead to cynicism and laxity in enforcement.

On the other hand, technology transfer strengthens strategic enemies. Lenin once said, "The imperialists are so hungry for profits that they will sell us the rope with which to hang themselves." He wasn't wrong.

Evading COCOM

Unbeknownst to the West, the Soviet Union over time developed countermeasures to COCOM. They have an incentive to. Acquiring advanced western technologies saves them millions of dollars of R&D and years of research.

Better yet from the Soviet perspective, integrating stolen Western technologies forces the Americans to spend even more resources to figure out ways to counteract their own handiwork.

The Soviets have legal and illegal methods to acquire this stuff. Legally, this can be through scientific or technical exchanges and conferences as well as open literature and trade.

Illegally, they can divert legitimate trade to new locations or use espionage. Or they can just purchase technology through dummy corporations, middlemen or foreign actors.

This must have worked well as the Soviets employed nearly 2,000 such agents, smugglers and middlemen for this task.

A Propeller Milling Machine

In the 1970s, the Soviet Union received information through a network cultivated by the convicted spy John Walker Jr.

Walker Jr. was an US Navy chief warrant officer and communications specialist who started spying for the Soviets in 1967. Mostly for financial reasons, it seems.

Walker told his Soviet handlers that the United States was capable of accurately tracking Soviet submarines through the excessive noise from their propellers.

The Soviets now knew that they needed quieter propeller blades. Propellers are cast out of bronze or steel. But to reach their final finished state, they need to be milled - meaning ground down with a power tool.

Usually, such a thing would be done by hand. But a precision milling tool would save a lot more time and offer a great deal more precision. This was exceptionally difficult - the Americans did not expect the Soviets to do this for several more years.

As was their preference, the Soviets searched the West for an equipment provider to help them along with it. They found a great one: Toshiba.

Toshiba

The Toshiba Machine company is a subsidiary of the Toshiba Corporation.

They are one of the world's biggest providers of manufacturing equipment - generating some $700 million in sales in 1986. Communist countries made up some 12% of their total export revenue.

Part of their portfolio were machines critical to producing very quiet submarines: Propeller milling machines.

In late 1979 or early 1980, the Soviet foreign trade organization Tekmashimport contacted a small Japanese trading firm with a presence in Moscow.

They told this firm - Wako Koeki - that they wanted to acquire Toshiba's MBP 110 nine-axis propeller milling machines.

Let me briefly explain this "nine-axis" concept. In these machines, a numerical controller sends commands to an automated milling machine, telling it where to move its milling head along its path.

The more axes the machine is capable of, the faster and more precise the milling can be done.

Toshiba's initial response was that these milling machines cannot be exported to the Soviet Union. But soon after they began looking into whether or not it was possible to temporarily modify the milling machines to evade the export restrictions.

Toshiba and the Soviets also reached out to a second trading company in Norway - Kongsberg Vapenfabrikk - so to purchase computer technology. They wanted to combine the Norwegian computers with the Japanese milling machines. This was all banned for export by COCOM.

A few months later in March 1980, Toshiba's president met with his export sales manager. The sales manager told him that in order to close this sale, Toshiba might have to violate COCOM regulations in Japan.

The president told him to "do what had to be done to get the business". Everything they did thereafter was done knowing that it was illegal.

Resentment

You might be wondering why Toshiba - a large, storied company - was so willing to do this.

It goes back to resentment over lost sales in the past. Back in 1974, the Soviets were looking to buy from Toshiba older propeller milling machines that were less sophisticated - capable only of two-axis simultaneous control.

The Soviets asked for something more sophisticated - more axes - which Toshiba declined to fulfill due to COCOM export trade restrictions.

Toshiba then later learned that the Soviets managed to acquire ten such multi-axis milling machines anyway from a French company called Forest Line. France, if you recall is also subject to CoCom restrictions.

Toshiba eventually determined that these export controls were being more rigorously enforced in Japan than they were in France. This put Japanese companies at a disadvantage. So in order to properly compete for Soviet orders, they needed to violate COCOM.

Japanese businessmen saw COCOM as a joke - a restriction unilaterally imposed onto Japan by the Americans. The prevailing feeling was that you couldn’t help violate COCOM if you wanted to do business in the Soviet Union.

Those feelings had some merit. Other countries did violate COCOM - even the British. For instance, the UK's Rolls Royce in 1975 sold its Spey engines to the People's Republic of China without considering COCOM.

The Sale

In order to get this illegal sale done, Toshiba built up an elaborate scheme.

Toshiba and Kongsberg needed an export license from the Japanese government - MITI. Their license application contained several false statements - including that the exports will not be sold to Communist bloc countries and that the machines will not be used for manufacturing military technologies.

Toshiba presented a document - stamped with the President's own seal - misrepresenting the machines' capabilities and certifying that the exports were appropriate.

MITI employs about 30 export control inspectors and they have to review some 200,000 applications a year. The MITI inspector approved the Toshiba application - Japanese newspapers said that the inspectors themselves recommended ways to evade COCOM - and by December 1983 the two MBP-110 machines were installed in Leningrad.

Toshiba earned $17.4 million for this first sale. The Soviets purchased a few more machines after that in 1984 for another $10.7 million.

Discovery

Three years later near the end of 1986, Hitori Kumagai (熊谷独), former chief Moscow representative at the Japanese trading company Wako Koeki resigned and informed the United States government of the transactions.

The Americans had recently noticed abrupt and unexpected improvements in the quietness of the Soviet Union's nuclear submarines and determined that the sale of these propeller milling machines was the cause.

They reached out to their counterparts in the Japanese government, who did an initial investigation. Toshiba stuck to their story and the investigators accepted it - determining that no export violation occurred and everything was fine.

Then in March 1987, things changed. The story broke in the American and Japanese national media, bringing a great deal of heat onto Toshiba.

Amazingly, Toshiba's president was at first defiant, saying:

> "When we received these charges before, it was found that we were innocent. All exports receive the approval of MITI. There was no violation of COCOM. Why are they bringing this up now? MITI also took the position that "There was no violation. The case is closed."

The chairman of the Japanese Machine Tool Industry Association echoed this sentiment:

> What has brought this problem back is the deterioration of relations between the US and Japan ... [This investigation is] a human sacrifice due to American pressure. It is a false charge.

Consequence

The news was a big loss of face for the Japanese government - which launched a more thorough investigation. Various senior executives were arrested. Toshiba's chairman and several presidents under him resigned their posts.

Prime Minister Yasuhiro Nakasone had to formally apologize, saying that the government would pass new legislation to make sure that it never happens again. But such laws lacked teeth. Toshiba's fine for its $17 million sale was a tiny $14,000.

They did however revamp the export licensing process - increasing staff to 100 inspectors, raising the budget five times, and computerizing the licensing process.

Furious, several American congressmen proposed aggressive measures against the Japanese, arguing that these quieter propellers put American seamen at risk and that $30 billion would be needed for equipment and new subs to detect these Soviet submarines.

They proposed heavy legislative measures, including import sanctions against Toshiba - including barring the entirety of the Toshiba group from bidding on American contracts.

It was delicate times for US-Japan relations and the Reagan Administration intervened. They instead got a lighter sentence including a 1-year ban on Toshiba sales to the Soviet Union. The US Army also canceled the purchase of a Toshiba missile system.

Toshiba Corporation ran a large public relations program apologizing to the American people for their error. It worked.

Norway did an arguably better job than the Japanese. Perhaps because Kongsberg was state-owned. They banned Kongsberg from any future sales to the Soviet Union and did a deep dive Coffeezilla style investigation.

In it, they discovered that French, Italian, West German and British companies also sold sophisticated milling machines to the Soviets and China in the 1980s. This in turn triggered a chain reaction of investigations by the French and so on.

Aftermath

Toshiba-Kongsberg not only demonstrated the weaknesses in COCOM's systems - which were loosened after the detentes of the 1960s and 1970s - but also the breadth of the Soviet technology acquisition machine.

The end of the Soviet Union eliminated much of COCOM's reason for being and it was eventually dissolved in 1994. It was replaced by the much looser Wassenaar Arrangement, where members of one country cannot influence exports of another.

One last thing. In light of recent events, it is interesting to see how the United States and its western allies loosened COCOM controls on the People's Republic of China in the 1970s and 80s.

This not only included dual use items but also direct military equipment like jet engines, ground attack aircraft, rapid gun systems, and computer systems.

This was because of the Sino-Soviet Split. The United States sought to strengthen China and bring it closer to the West as an ally against the Soviet Union. Just found that interesting.

Conclusion

The debate over COCOM's effectiveness was and remains voluminous. There have been hundreds of takes of all kinds and from all angles. Many of these debates are re-emerging in the wake of the recent chip technology bans.

The actions behind Toshiba-Kongsberg show us why international trade restrictions are so hard. Players feel unfair treatment when they are subject to such restrictions, often bringing up questions of sovereignty.

It is why the United States went ahead with the chip sanctions first without any agreements from its Western allies - a demonstration that this was no ploy to win sales for American businesses.

But more work needs to be done. Toshiba-Kongsberg showed that it is not enough to place sanctions and licensing. In order to be effective they need to be fair, backed with deep investigation and resources, and seriously enforced.

Share The Asianometry Newsletter

It is December again. The year has gone by so fast. Where did all the time go?

So I think we should spend some time and look back at the Asianometry channel in 2023.

Popular videos. Videos I thought would be more popular. And big themes I noticed throughout the year.

The Year's Videos

I went into 2023 with a few things in mind, I wanted to maintain a good proportion of semiconductor videos but also explore other technology topics like computers.

The year's top videos were "How China Got the Bomb", "How Semiconductors Ruined East Germany", the Crony Capitalism in Indonesia video, and "That Time the Soviets Abolished Money".

Bonus points to the YKK video, which got a shoutout by Lupe Fiasco, who made one of my favorite songs.

I notice that there was a strong focus on the Soviet Union for last year's videos. I don't know if it is because of current events, or just because the Soviet topics I chose were more "intriguing" than some of the others.

2023 also marks the first video I did about South America, namely the Brazilian computer. I was not sure how it was going to do, but am glad to see that some people liked it. You might see a few more South American videos coming out in the future.

Two of my favorite videos came out this year. One was the copper interconnect video, "TSMC's First Breakthrough", which covered the momentous move to copper interconnects, as well as the 130 nanometer node which TSMC first brought to market.

Another was the "Rise and Sad Fall of Wang Labs". I can remember myself working on this video, sitting in a cafe in Taichung, when it occurred to me that this one was going to be special. The story of An Wang, who is basically forgotten today, is inspiring and should be told more widely.

To be honest, I am still not sure how a video does when I release it. In the end, I am just still doing the same thing. Focusing on what things happen to interest me. I did every video because I enjoyed the topic, so in reality all of these are my favorites.

These Should be More Popular

I did a few videos in 2023 that I thought should have done better. This video won't do as well as any of the others, but I hope at least I can highlight a few of them.

"Silicon Shrank the Hearing Aid", I felt was one of my personal favorites. Many of the people I know wear hearing aids, and I think they are a fascinating technology. The first transistorized electronic product was a hearing aid, and they have always been at the leading edge of miniaturization tech.

"Shaw Brothers: King of Hong Kong Cinema", was a rare story about Chinese entertainment - starting from the mainland to Hong Kong to Singapore. And though this one did not do very well on the view-count basis, it got a lot of responses from Chinese-Americans youths and elders who watched their movies growing up.

And the story of National Semiconductor, where my father used to work. That one did quite poorly and I do wish more people knew about it. National Semiconductor used to be a giant in the industry, and analog remains a critical part of semis.

And of course, the Etch video. A lot of people care about lithography because it's seemingly the sexiest part of the semiconductor manufacturing process. But they call it litho-etch for a reason. You are missing 50% of the duo, if you just care about litho!

People I Talked To

Again this year as with the last two years, I tried my best to meet as many people one on one.

It is a lonely thing, YouTube, and meetings help me learn about who you guys (and 95% of you are indeed guys) are. And of course, I want to learn from your experiences as well.

I want to call out just a few of my notable meetings and sharings. (If I met you and didn’t mention it, then sorry about that, I still remember it!) Like the Microsoft AI researcher who marveled about GPT-4 with me;

Or the analog chip startup guys from Silicon Valley who patiently explained analog design to me;

Or the Los Alamos researcher who helped me understand just how popular Prime Minister Modi is within India. Thank you all for taking the time.

I still encourage you to reach out to me if you find yourself in Taipei. But do try to give me a longer heads up and also introduce yourself so I can get a sense of who you are.

Events

Asianometry did some events this year for the first time.

The first one was in Tokyo, during my trip there in February. It was held in a small conference room offered by a kind viewer.

After that, I did another event in May in Taipei when Dylan Patel of SemiAnalysis visited. It was one of the first times that I set something up within Taiwan itself, and I had a blast.

And of course, the last big one of the year was the September 2023 AI and Semiconductor symposium, which we held before the big SEMICON East 2023 show. That was huge and I did a recap video on it.

I had a great time with all of these events. I was able to meet some great people. And though I was on edge the whole time, it ended up okay. I want to thank everyone for taking the time to come and encourage you to come to the next one - whenever it should happen.

Travels

In 2021 and 2022, I largely stayed at home due to COVID restrictions and what not.

But in 2023 things opened up, and so I took the opportunity to go traveling around a bit. I got to visit Korea, Japan, Thailand, and the United States (for a bit).

In Korea, I had the chance to take a "train to Busan" as they say. Stayed in a hotel near the ocean and got to visit their local aquarium. Also ate some of that famous Korean barbecue.

When people in Korea found out that I was the "Taiwan semiconductor" guy, I was bemused by their consistent asking - "How can Samsung catch up?" The answer of course, is that Samsung is diversified, and few companies are as intensely focused as TSMC is.

In Thailand, I ate mangoes and did massages. I am baffled by Bangkok's insane street layout, but also impressed by their incredible malls and the wonderfully kind people. I have not done a Thailand video in a while, but I hope to soon.

I try not to do that much traveling, since I have work to do and videos to make. But I do enjoy it. And I hope in 2024, I can make time to perhaps come to your country as well.

Conclusion

As with before, I want to thank you all for taking the time to watch the videos.

Looking back on 2023, I marvel at how fast time passed. I look at each video and can remember where I was and what I was doing when I was making them.

Personally, I am feeling better. I am still working and stressed and doing all I can to keep up. But so far I am handling it okay. I hope I can continue this as we go into 2024.

I have some great ideas that I want to work on for the coming year. I hope you will enjoy them along with me. Every day, I am still amazed that I get to do this and that people care for it. Take care and I will see you next time.

Share The Asianometry Newsletter

For the past thirty years, contributing my efforts at TSMC has been an extraordinary journey for me. I am especially grateful to the company's extremely talented team. Together, we have shaped the company into today's industry benchmark. It is a great honor to take over as Chairman of TSMC after the retirement of the company's legendary founder. I now hope to retire but not to rest, to give back in different ways based on my decades of semiconductor experience, to spend more time with my family, and to start a new chapter in my life. I will continue to work with the Board of Directors on corporate governance until the last day of my term, and I am fully confident that TSMC will continue to achieve excellent performance in the foreseeable future

Right now I am swamped by work so I am not able to do a whole video on the topic, but I thought it would be interesting to at least talk a bit about TSMC’s succession issues and what might be next.

Founder Morris Chang has long struggled with the issue of succession. He was already in his late 50s when he first founded the company. So famously in 2005, he appointed Rick Tsai as his successor. Tsai had served in fab operations and then as the COO for a while.

Despite this, TSMC's 40 nanometer node had troubles ramping up its yields. The Global Financial Crisis struck, hurting the industry. Worse yet, TSMC fumbled a series of layoffs. The layoffs were deemed necessary considering the economic situation. But Taiwanese business media reported that those layoffs were implemented with insufficiently rigorous criteria. Workers who supposedly left voluntarily were protesting outside of Morris Chang's house.

TSMC met the expectations of the financial community. Nevertheless, on June 2009, just four years after the CEO handover, Morris Chang removed Tsai, reassigning him to a smaller division. Tsai is now the CEO of MediaTek. He restarted the succession process all over again, and then plunged the company into a massive capital expenditure that paved the way for Apple.

Chang did not completely retire from the company again until 2017 he was 86 years, when he brought Mark Liu and CC Wei as co-CEOs. According to Nikkei, Chang delayed his retirement since he wanted the two to think more like businessmen rather than engineers.

When he promoted Liu and Wei to be co-CEOs in late 2013, Chang said he expected his successor to be more of a "good businessman," not merely a capable engineer. Back then, he said both executives were still more like engineers and that their business acuity needed time to be cultivated. He added that a CEO must be open-minded, worldly and wise enough to develop strategies for the company.

So when Chang retired as Chairman, Liu became Chairman, and Wei remained as CEO. Technically, the Chairman in an Asian company is a bit more outwards facing. He would represent the company in ceremonies and such. During that TSMC Arizona ceremony, Liu is the one popping the champagne, not Wei.

Liu was ideal for the position since he seems to be the more philosophical and thoughtful of the two. Wei is the more practical and hard-charging of the two - a bit more of a go-getter it seems.

Reading this news, I harken back to something that I read a while back. I was doing research for the Nvidia CuLitho video, which included quotes from ASML, Synopsys, and TSMC. The thing that caught my eye was that the quote supporting CuLitho was given by CC Wei, CEO, rather than Mark Liu, the chairman.

It was probably just nothing. But I remember thinking at the time that a quote like this is the such type of thing that the Chairman would be giving, rather than the CEO - who is more busy running the whole company. But it does seem like Liu has wanted to retire for a long time. Wei taking more of a step into the public spotlight - looking back at it - seems like part of that. Liu will probably return to his home in the US.

Wei is the likely successor for the Chairman, unifying the CEO and Chairman positions. It is going to be a big challenge. In the past, the TSMC chairman was about handling business and relationships, and not so much geopolitics. Wei is not known to have done this government and geopolitical stuff before. I wonder how hard it might be to run and keep a massive engineering company at the leading edge while also doing it.

TSMC has gotten to where it is on the basis of a few big bets - some technical, others business-related. The technical ones, I think can be handled by the engineering bench, which is so deep James Cameron wants to send a submarine down there. But I do wonder if there is someone who can unite everyone at the company and make those hair-raising decisions that put the whole company on the line. Like when Morris raised capex in the wake of Global Financial Crisis and recession, or when he orchestrated the cunning acquisitions of TASMC and WSMC in 2000 to boost its capacity far past its rival UMC.

Wei also needs a successor of his own. Based on what was reported back in 2017, Wei is probably around 70-71 now, actually two years older than Liu. I feel like he needs to start grooming a successor. There is a generation who deserves to rise up, but in doing of course there will always be concern about who is taking the controls in the future. TSMC is still a relatively young company with an extremely unique culture which has given it success. It will need to work hard to preserve that while also adapting to new realities.

Post note: I have been working hard for you guys. Sorry I have not been putting out so many newsletters, I suffered a bout of migraines and was stuck in bed. Next week I will have a “Year end review” for 2023.

Share The Asianometry Newsletter

Long time no see.

I have been thinking about this recently. The idea of what goes next after EUV. The current presumption in the wider media is that EUV works and is worth the investment. The former is true in a literal sense. I am not so sure about the latter.

I want to do another video about this down the line. But I am feeling increasingly uneasy about the state of EUV lithography right now. And I am feeling especially … especially nervous about High-NA EUV. The new senior leadership of ASML have a rough task ahead of them.

Also, one of the feedbacks that I got about this video from an industry professional was that “this was trash”. So take that as you may.

Recently, I have been thinking about TSMC's N3 process. The one popularly called "3-nanometers", whatever that means.

There is no doubt about it. TSMC's N3 process family is going through some issues. Currently there are two versions of N3.

The first one is N3B. The B stands for "based". Which according to TikTok means it has a lot of swagger or something.

This is the one that is in high volume production right now. My guess is that it is producing an Apple chip though as of this writing, nothing has yet come out using it since the commencement ceremony at the end of 2022.

Then we have this second process called N3E. It is entirely different and is known as the "real" N3. This one is heading into high volume production later in 2023.

One of the significant differences between N3B and N3E is the number of layers done by EUV.

David Schor of WikiChip Fuse estimated that N3B has 80% more EUV layers than the N5/N4 process nodes. N3E apparently scaled that back, going from 25 EUV exposures to 19.

It is a bit strange, right? TSMC has more EUV machines than any other fab in the world, but their newest and biggest N3 process is nevertheless pulling back on EUV. Why?

Nobody knows but TSMC and ASML. But I have a theory. EUV was supposed to take us to the Promised Land. It hasn't yet because the amazing, double-tin-shot-with-a-laser EUV light source everyone loves to talk about is not powerful enough.

And that is why scientists are tinkering with something else. And that something might cost half a billion dollars. In this video, we are going to look at the experimental idea of Free Electron Lasers for EUV lithography.

Stochastics

The current method of generating EUV light for EUV lithography is laser-produced plasma, or LPP.

We fire a high-intensity carbon dioxide laser at many tin droplets. The droplets turn into plasma which emit the necessary 13.5 nanometer EUV light for lithography.

Imagine EUV photons spraying out towards the wafer like shot pellets fired from a shotgun. The photons bounce off various mirrors, and then the photomask, before getting to the wafer.

Upon hitting the wafer, these photons react with the molecules of the photoresist coated on that wafer. Now the wafer has the chip design printed onto it and we can move on with the rest of the fabrication process.

Well, that is when things go right. The issue is that both these photonic movements and photon-resist molecular reactions are random processes that follow quantum mechanical principles.

These photons will not travel exactly where we intend for them. There is random variation in their distribution as they go. And the molecules on the resist layer are also distributed randomly and will interact randomly upon being hit by the EUV photon.

What all this means is that stochastic effects in EUV lithography leave various print failures on the wafer.

Broken lines, microbridges, and rough edges on designs. We call them "failures" rather than "defects".

A defect is something that happens due to particle contamination, pattern collapse, or otherwise.

This is different from that. Stochastic print failures are far smaller, completely random and thus non-repeating, and come via a law of nature. The randomness is unavoidable.

Dealing with Randomness

So folks, how do we usually deal with randomness caused by the law of small numbers?

We increase the numbers, right? Just send more photons! There are two issues with this.

First is that it takes 14 times more energy to output an EUV photon than a 193-nanometer photon. So for each exposure "dose", a 250 watt EUV light source will create 14 times fewer photons than a 250 watt 193-nanometer light source.

Not to mention that only a small percentage of those photons ever make it to the wafer. Much is lost as it bounces from mirror to mirror or through the pellicle protecting the photomask, or whatever have you.

So we either need to give the wafer more EUV doses, or increase the light source's EUV power to make the dose bigger.

The first option is not a long-term solution. It slows the machine's throughput, negatively affecting its productivity. The fabs won't pay $150 million for a slow machine - not even the dumb fabs like REDACTED.

So we need to increase the power. This is not easy. In a recent presentation in 2023, ASML San Diego showed a potential future pathway towards 600 watts or even 800 watts.

I talked about how they might achieve this in my update video on High-NA EUV. It involves lasers and tin.

800 watts sound like a whole lot. Except that is still far short of what is ideal. A 2019 estimate by KIOXIA estimates a range of 1.5 to 2.8 kiloWatts for the 3 and 2 nanometer nodes.

That is disruptively far more than what existing LPP approaches can do. At the end of the day, there are fundamental issues with the LPP approach - particularly when it comes to tin contamination on the mirrors.

So maybe we can try a Free Electron Laser.

Lasers

Maybe the best way to start is to explain lasers.

The word "Laser" stands for "Light Amplification by Stimulated Emission".

First, we have a gas or element used to create the laser light. This is the "laser medium" or "gain medium". Perhaps the most widespread such medium is Neodymium-doped yttrium aluminum garnet.

We stimulate that medium's atoms using energy from some source like a krypton lamp or LED.

The atoms' electrons absorb the energy and ascend like the Buddha to a higher state of being.

After which, the electrons descend back down to a lower energy level. In doing so, they release energy in the form of light, photons.

That is where the "stimulated emission" part of the LASER acronym comes into play. We stimulated the medium into emitting light.

By doing this to a whole bunch of laser medium atoms at the same time, we can amplify a bunch of photons. That is the "light amplification" part. The light comes in all directions so we use mirrors and an optical cavity to focus that light into a directional beam.

Because the electrons ascending up and down the various energy levels are still bound to the laser medium's atoms, we call this type of laser a "Bound Electron Laser".

So if that is "bound", what do you call a laser where the electrons are unbound from their atoms? Where they are ... free? A free electron laser.

Free Electron Laser

Okay let's do this.

A Free Electron Laser generates light by oscillating charged particles - electrons, mostly - traveling at near the speed of light.

We start with a powerful electron gun to create a beam of electrons.

Then we use a linear accelerator to accelerate these electrons to near the speed of light.

After that, we send the accelerated electrons into an alternating magnetic field - like north, south, north, south, so on.

Each magnet inside the field attracts the electrons, causing them to swerve towards it. This movement is an acceleration - because acceleration can not only be a change in speed, but also a change in direction.

An accelerating electron gains energy. Later on, the electron has to release that energy, and does so in the form of light.

If we are to accelerate the electrons back and forth - essentially "wiggling" them - in the right way, then we can force it to generate huge bunches of photons. This wiggling is done by a long series of alternating magnets called an undulator.

By tweaking the electron beam, its speed as it enters the undulator, as well as the strength of the undulator's magnetic field we can produce 13.5 nanometer light.

SASE

Since 13.5 nanometer light is very easily absorbed by almost every substance in the world, we cannot use an optics system like we do with traditional lasers to collect and intensify the emitted light.

So instead, we intensify the light inside the undulator itself based on a theory called Self-Amplified Spontaneous Emission or SASE.

Every bunch of electrons traveling through the undulator has electrons traveling through at different velocities and states. The distributions of these are random. We don't want this randomness because it causes the electrons to emit light incoherently.

We don't have the order and alignment we need for laser-like behavior. It is like a bar full of people randomly talking - just noise.

But if the electrons are similar enough to each other, and fired in the right way ... then something remarkable happens.

The electrons will interact with each other's own emitted radiation. And that will cause them to self-classify into buckets based on their relative positions.

Now everyone is singing in harmony. The light emitted from the bunches of electrons starts to amplify itself, creating the self-amplifying effect referred to in the name.

There, I think I got that right. If I am wrong, I am sure I'll get some comments.

Synchrotron

Those who have seen the previous EUV light source video might recall that early experiments worked with synchrotron radiation.

A synchrotron is a class of donut-shaped particle accelerator descended from the cyclotron. It keeps a particle beam continuously circulating, emitting radiation. The industry ultimately didn’t use a synchrotron to generate EUV light for photolithography because it cannot generate enough power.

Though a synchrotron is used to generate light for the actinic blank inspection tools. These tools test the EUV mask blanks for defects using actual 13.5 nanometer light - the "actinic" in the name.

While the synchrotron might seem kind of similar to the Free Electron Laser - and they are relatives - they are not the same.

A synchrotron light source is essentially a storage vessel, particles churning around in circles endlessly like as in a 7-11 slushee machine.

On the other hand, the electron beam's requirements for an EUV-class Free Electron Laser are so demanding that only a linear particle accelerator can do.

KEK

A team at Japan's High Energy Accelerator Research Organization

or KEK recently presented an interesting proof of concept.

This Free Electron Laser POC kind of looks like a race track. The electron gun first creates the electron beam. The beam then passes through two linear accelerators.

First, an injector accelerator that adds 10 mega-electron volts of kinetic energy to the electrons.

This is kind of like how a race car speeds up in the pit lane to get to a decent speed before entering the main track.

After that, the electrons go through a main accelerator that adds another 800 mega-electron volts of kinetic energy. This brings them to near the speed of light.

After that, they go through the first turn.

And then after that, into a 200 meter long chain of undulators.

At the end of the undulator chain, out comes the EUV light we so desire and the electron beam.

Normally, the electron beam is dumped right after exiting the chain. The KEK team does not do that. Instead, they send the beam through a second turn, returning them to the main linear accelerator.

There, the beam is decelerated before finally hitting the beam dump which safely absorbs the electron beam and its energy. These dumps are usually made from a non-reflective substance like graphite or concrete.

A beamline will be necessary to bring the EUV light to the lithography machines. This is likely to be a set of mirrors in a vacuum chamber to expand the beam cone and split it for entry into the steppers.

## Photolithography Advantages

A Free Electron Laser has a number of advantages as a photolithography light source.

First, because there are no tin droplets, we do not have issues with tin contaminating the collector mirrors like with the LPP approach.

Second, a single laser can provide light for multiple lithography machines. A laser can generate 10 kiloWatts of EUV power for one machine, or 1 kilowatts for ten machines. This is the same advantage the synchrotron approach offered.

Third, the free electron laser is adjustable. Tweaking the electron gun and the undulator's specs can upgrade the 13.5 nanometer light to an even shorter wavelength like 6 nanometers.

Fourth, there is a possible cost advantage. The laser itself will cost a lot - roughly $400 million along with $40 million of annual maintenance costs. But again you can distribute its fixed costs across multiple lithography bays.

The Free Electron Laser is also likely to use less electricity. The KEK machine uses about 7 megawatts of electricity to generate 10 kilowatts of EUV power, so about 0.7 megawatts per 1 kilowatt of EUV power.

By comparison, the current LPP approach uses about 1.1 megawatts of electricity for 250 watts of EUV power or 4.4 megawatts for 1 kilowatt. ASML is working hard to improve that, of course, but this is a sizable gap. And electricity is one of a fab’s biggest variable costs.

Conclusion

Free Electron Lasers were first invented in 1971 by Stanford's John Madey.

But early machines produced only infrared or microwaves. It was not until the early 2010s that they started being seriously put forth as potential EUV light source - too late to be involved in early EUV development.

But then GlobalFoundries put out the first serious proposal in 2015, and it has been kicking around ever since.

Including the Japan project, Free Electron Laser technologies are also being researched in Europe - Euro-XFEL - and the United States - SLAC.

I reckon ASML has probably looked at it. I wonder what they thought. There remain considerable issues to be overcome, but the concept is tempting. Someone should get the semiconductor Avengers together to talk about it.

Dumping the tin laser approach might cause some disappointment to tech enthusiasts. But the thought of TSMC putting a $400 million, 200-meter long linear accelerator underneath their next fab is kind of awesome too.

Share The Asianometry Newsletter

Also, there is a new video coming today. I think you’ll find it fun at an atomic level.

The big question that I left viewers with at the end of the video was: “How can TSMC be so ‘productive’ with their R&D when their spend trails that of other companies?”

The comments on the video point out a few things:

• First, the lower cost of living and real estate in Taiwan. I don’t buy this as much as the other reasons. Some of these houses in Hsinchu cost as much as a condo in wider LA County - $1+ million USD.

• Other companies like Intel cover more of the spectrum than TSMC. Intel has to do design and other stuff in addition to the manufacturing. They are also making GPUs, TPUs, etc. TSMC “just” does lithography, packaging, and PDK development.

  • And to add about Google, Amazon, etc. Those FANG bros are using R&D spend to build software, which lack the patentability of “hard” processes. So patents - already a flawed metric - are ever more so.

• TSMC benefits from R&D spending across its ecosystem. I like that idea, but is there that much crossover between their fabless partners who are staying up in the “design” area?

On July 28th 2023, TSMC held a ceremony to inaugurate its newest R&D center out in Hsinchu, Taiwan.

The whole thing made the news here in Taiwan so I thought that I should do a short video about it, TSMC's newest "fab".

Bell Labs

TSMC calls the new building the Global R&D center and it is their version of Bell Labs.

It will be where the company experiments with the various processes and operations for its coming leading edge nodes - N2, N1.4 and so on.

Once the R&D fab masters the recipe, they roll it out to one of the big wafer fab centers in Taichung or Tainan. N2's big rollout will be in Hsinchu and Taichung. I think N1.4 is planned for Longtan, Taoyuan.

They are calling this building the Global R&D Center. I mean, that is what is on the sign. But semiconductor factories are built over phases. So I have seen this building also be referred to as Fab 12, Phase 8. Fab 12 being the previous R&D fab.

TSMC's current plan is to add seven more factories around the R&D center - all of them leading edge. The first will be the upcoming N2 fab - Fab 20.

I actually visited the construction site in July 2023. Right now it is just a steel shell. It is scheduled to go high volume some time in 2025.

TSMC will continue building out the R&D Fab. It doesn't look like everyone has moved in yet and crews have not yet finished the work of moving in all the equipment.

Baoshan

The new manufacturing zone will cover the area of Baoshan.

Baoshan is a hilly suburb adjacent to the main Hsinchu area. It is a township covering 10 villages. The land used to host a century-old Hakka village: Daqi Village, formerly home to about 2,000 residents and a temple.

TSMC started acquiring the land in 2019. This acquisition required the company to relocate all of the villagers, the temple, and a local cemetery. Many of the villagers protested, demanding a higher price for their land and it took some time to get that all sorted out.

They did not break ground on the fab until 2020, but it would not open until three years later due to pandemic delays and the difficulty of the construction.

As you can see, the design is very striking. It is reminiscent of the design of their big Nanjing Giga-fab. And also kind of like a spaceship.

The Baoshan land is very hilly. In order to make the land suitable for a fab, TSMC basically leveled the mountains and hills. Near the construction sites you can see huge piles of dirt from that work.

Right now the infrastructure and the roads are still being built out. They will need to expand the roads to prevent big traffic issues, but people are already starting to make the area their home. I can see expensive-looking houses on the distant hills.

Chang's Remarks

Morris Chang, TSMC's legendary 92-year old founder and long-time Chairman, showed up to the commencement event.

According to Taiwanese media, it was the first company-affiliated event he has attended since his retirement in 2018. Though I have heard that he still sometimes goes to the company offices.

Side note. I heard that Morris Chang recently announced at his birthday party that he had completed the second volume of his autobiography. Looking forward to reading it.

Anyway, as he started his speech, some women standing on the balcony cried out "Grandpa I love you!" I thought that was cute. Morris is really loved here in Taiwan.

Chang's speech hit on a few key points. First, he thanked the TSMC R&D team for their contributions to the semiconductor industry. He also noted that their contributions have put TSMC into the thoughts of many military planners. He then drily mused aloud whether that part actually deserved a thanks.

Second, he emphasized that TSMC continue to internally and independently produce its own techniques and technologies.

Chang recalled that back in the early days, Philips - TSMC's big co-investor - asked TSMC to adopt their CMOS technologies.

TSMC licensed those for 5-10 years as a patent umbrella, but insisted on developing their own. It took 30 years for TSMC to turn that technological independence into technological leadership, marked by their 7 nanometer node.

His remarks on this also reminded me of the early 2000s, when a legendary six-man R&D team at TSMC worked to produce the company’s own variant of copper interconnect technology. This eventually allowed TSMC to be the first in the industry to ship a working 130 nanometer process node.

Third, Chang emphasized that R&D and manufacturing needed to work closely together. It is not enough that a team produces innovations. Scale is what matters. Which company will be the first to bring their innovations to high volume production?

And finally, he cautioned the company to never slacken in its efforts. He told a story from his Texas Instruments days about how the British Navy was once seen throughout the 19th century as the world's most powerful.

Then in the 20th century, the navy inaugurated a massive new building - the Admiralty extension. Little did they know it, but the building marked the Navy's historical high point. So with this in mind, Chang said:

> 別像英國海軍，有了大樓就開始沒落

> Don't be like the British Navy, starting to decline once they got a big building

Conclusion

In 2022, TSMC spent about $5.5 billion in R&D investment.

This represents a little over 7% of the company's 2022 revenues.

It sounds like a lot. But it greatly lags the R&D spend of the big tech companies in America like Alphabet - which spent $34 billion - or Apple - $23.5 billion - or Amazon - a staggering $46.5 billion.

And just in case you are curious. In 2022, Intel spent about $17.5 billion for R&D.

Compared to those guys, TSMC doesn't spend all that much. But they do seem to be quite productive with that smaller spend. Statista finds that in 2022, TSMC received 3,024 US patents - third most after Samsung and IBM. Intel had about 2,418 patents.

It makes me wonder a few things. Like whether R&D spend as reported by these big American tech companies is comparable to the R&D spend reported by TSMC. It's probably not. I wonder maybe because of tax incentives?

I wonder how much of TSMC's technological advantages are rooted in Taiwan’s lower costs of living and just paying those R&D workers less. It's what everyone in the West (and YouTube) thinks about first.

But personally, I think that though that factor is significant it is probably not overwhelming. Instead, there seems to be other things at work here like automation, scale, fast cycle time, or accumulated unspoken knowledge.

And finally, I wonder just how resilient TSMC’s independent approach can remain in the face of the intense economic, technological, and geopolitical competitions that lie ahead in the semiconductor industry's future. That question remains open.

Share The Asianometry Newsletter

I was asked by a fan in Germany to give a few thoughts on the news. I love my German viewers so let's do it!

But first …

The AI and Semiconductor Symposium in Taipei

Are you going to SEMICON 2023 in September? If you are coming to Taipei, then you should come to the AI and Semiconductor Symposium that me, Dylan Patel of SemiAnalysis and Doug O’Laughlin of Fabricated Knowledge are hosting.

The three of us will give talks about AI, the semiconductor industry, and semiconductor history. And Dylan tells me that there’s going to be a secret special guest. I have no idea who it is, right now. He is keeping it secret.

Get Tickets

Deal

There are a few things about the structure of TSMC Dresden - ESMC as they will call it - which are interesting to me.

ESMC is a joint venture between TSMC and three European electronics companies - Robert Bosch, Infineon, and NXP. TSMC will own 70% and the three European companies will each own a 10% slice.

TSMC will operate the fab, which will produce two nodes - a 28/22 nanometer planar node and a 16/12 nanometer FinFET node.

The "planar" and "FinFET" parts of the node's name is a reference to the type of transistor gates that the nodes will produce. FinFET nodes stick up over the silicon landscape, which is why they are referred to as being 3-D.

Intel first introduced FinFETs in 2011. TSMC and Samsung first started using FinFETs. So this fab trails the leading edge by about 10 years or so. By the time construction is estimated to be complete in 2027, it will be even older than that - 14 years old.

The German government will provide about $5 billion of the estimated $11 billion needed to build the plant. TSMC's press release did not mention how much that they are putting in, but Reuters said it was about $3.8 billion USD.

So all in all, TSMC only needs to spend $3.8 billion to obtain majority control of a $11 billion fab. Of course, I say "only" as in $3.8 billion is nothing, which it isn't. And they still have to run it.

The European Commission needs to approve the deal. There are a lot of subsidies in it, which makes me wonder if it will be a problem. But considering how big of a deal the EU thinks chips are, it should go through fine.

Deal Structure

The first thought that came to mind was that this looks more like the TSMC fab out in Japan's Kumamoto Prefecture than TSMC Arizona.

TSMC Arizona is fully owned and run by TSMC. Meanwhile, that Kumamoto fab - Japan Advanced Semiconductor Manufacture or JASM - is a joint venture with a collective of Japanese companies like Sony and DENSO.

Sony makes image sensors and DENSO makes automotive chips. Both of these companies already own semiconductor fabs in the country. Sony's big image sensor plant is right next door to JASM.

The construction of the Kumamoto Fab looks like it has been going about pretty smoothly. There hasn't been much reporting on it. Some domestic companies in the area have been complaining about the effect of TSMC's hiring push. But overall, no news. And no news is good news.

TSMC has done a few equity partnerships prior to the Kumamoto and Dresden deals. Most notably is their joint venture in Singapore, SSMC.

In general, they prefer having full equity control of their big factories. So I wonder if TSMC sees having these foreign partners as being beneficial in getting the lay of its new land. Having a local partner to smooth out disturbances would have been super helpful.

STMicroelectronics

I had thought that the effort would involve more of Europe's semiconductor companies.

Europe's largest semiconductor makers are STMicroelectronics, NXP, and Infineon. NXP is a Dutch company - descended from Philips' semiconductor division. And Infineon is German - descended from Siemens semiconductor division.

I had expected all three of these big European companies to be involved in the deal. TSMC might present a threat to them, after all. But currently, TSMC's partners are Bosch, NXP, and Infineon. Where is STMicroelectronics?

My thinking is that it has to do with subsidies. Bosch and Infineon are German companies. STMicroelectronics is a French-Italian fusion. France and Italy did not contribute subsidies to the deal. That's my best guess for why STMicroelectronics is not involved.

Post-send note: A helpful reader pointed me to a recent announcement by STMicroelectronics that they are partnering with GlobalFoundries for a new 300-mm fab in France. So another reason why they aren’t joining is that they already have their partnership in hand.

Dresden

Same as with Japan and Arizona, TSMC chose to site their German fab in a rather crowded spot - Dresden.

Dresden already hosts fabs from GlobalFoundries, Wolfspeed, and Infineon so they are entering a bit of a crowded space here. The competition for talents and resources like power will be pretty fierce.

TSMC probably evaluated whether or not they should go their own path and build further away from Dresden. Intel is taking this approach with their German fabs in Magdeburg.

They would have to do more work especially with the infrastructure, but they would be the only game in town.

Demand

One of the big questions that I had when it came to building new fabs in Europe was demand drivers. Who will be buying these chips?

ESMC will produce about 40,000 300-millimeter wafers each month. This makes the fab larger than a megafab - 25,000 wafers - and a gigafab - 100,000 wafers.

That is a lot of wafers. Infineon and Bosch already have factories in Dresden. Infineon says that their fab there is their largest front-end site. So what products are these wafers going to go into?

I was hoping for some sort of surprise application but it seems like everyone is sticking to the obvious - automotive. TSMC's automotive revenues have been on the rise after the pandemic - presumably driven by an electrification trend.

I did find it interesting that NXP also announced that they are partnering with TSMC to develop a new type of FinFET based Magnetic RAM. NXP said that this RAM is more reliable and can update far faster than flash memory.

So I reckon that there is going to be a lot more of this R&D going on between the customers and TSMC.

Conclusion

For the most part, I think this Germany fab is going to go relatively smoothly. The project is less high-profile than Arizona, with various partners involved. That $30 billion Intel project is probably under more scrutiny here.

Furthermore, it is under less time and economic pressure than Arizona. It is a lagging edge node, so there isn't so much time pressure to get the thing out.

I'm looking forward to seeing the fabs! Maybe make a trip out there once it is done. I am also a big fan of Sauerkraut.

Share The Asianometry Newsletter

The mechanical pencil. The calculators. The TVs.

Over the years, Japan's Sharp Corporation has delivered a bevy of iconic products.

But starting at the turn of the century, the company found itself increasingly trapped in a single, incredibly competitive business.

Attacked on all sides by foreign competitors and running out of money, the company sold itself to a Taiwanese electronics giant. The first such foreign takeover of a major Japanese consumer electronics company.

For this video, let us talk a bit about the rise and decline of the 100-year old Sharp Corporation.

Beginnings

Founder Tokuji Hayakawa (早川 徳次) was born in 1893 in Tokyo, the third son of a furniture-maker. He lived a harsh childhood. His mother suffered health problems and could not take care of him so he was put up for adoption.

Treated harshly by his new step-mother, Tokuji was forced to drop out of the second grade for work. But at the age of eight, an elderly lady got him a position as an apprentice for a strict but compassionate metalworker.

After training in metalworking for seven years, Hayakawa graduated and joined the workforce. Soon after came his first invention, a belt buckle dubbed the Tokubijo.

What was special about this buckle was that it can be fastened without needing to punch holes into the belt. It was a hit, selling 4,752 pieces and necessitating a serious expansion in capacity.

Thus at the age of 19, Hayakawa opened a metal workshop of his own in Tokyo with just 50 yen of invested capital. 40 of which was borrowed. The Workshop

A few years later in 1915, a manufacturer placed an order at Hayakawa's workshop for the metal innards for a mechanical pencil.

Struck by that pencil's poor quality, Hayakawa worked day and night to produce a superior version: the Hayakawa Mechanical Pencil.

This pencil was a hit. Most pencils at the time were imported from Germany. With Germany fighting World War I at the time, the Hayakawa Mechanical became the next best alternative.

The business rapidly expanded. Hayakawa's long-lost older brother Masaharu joined the business. In 1923, the two brothers built a new, 990 square-meter factory staffed with 200 workers to handle all the orders.

Earthquake

Then on September 1st 1923, 11:58 AM, the Great Kanto Earthquake struck the country.

Hayakawa, his family, and his factory survived the initial earthquake, which is said to have lasted over 4 minutes. However, Hayakawa's wife and two young children perished in the massive fires that followed the initial quake.

Over 100,000 people died in this national disaster. That does not include the various mob-driven massacres of ethnic Koreans that occurred thereafter. A horrific and under-covered secondary tragedy.

Amidst this devastating societal and personal tragedy, the Hayakawa Brothers pencil factory burned down. Hayakawa still owed 20,000 yen worth of pencils to his pencil sales distributor which he could no longer produce.

To compensate for this, Tokuji transferred over all his pencil-making equipment and granted them free use of his patents. He then moved to Osaka for six months to help them produce these pencils themselves.

Rejuvenation and Radio

Tokuji enjoyed his time in Osaka, and resolved to rebuild his business there.

In 1924, after fulfilling his six month contract, he established Hayakawa Metal Laboratories and started tinkering with a new product - the crystal radio.

Radio was only recently introduced into Japan, and most radio receivers were basically regarded as mere toys. The dominant telephone and telegraph makers did not deal with them, leaving the market to tinkerers and hobbyists.

Hayakawa bought an imported radio for 7.50 yen and tore it apart. Despite knowing nothing about electricity or radio principles, he and his team reverse-engineered the device.

In 1925, Hayakawa released their own domestically-assembled crystal radio - called the "Sharp". At 3.50 yen, it was drastically cheaper than its foreign competitors.

The Sharp radio came at the perfect time. That same year, the Tokyo Broadcasting Station - an ancestor of NHK - started Japan's first radio program. Radio channel subscribers grew steadily throughout the 1920s and 1930s.

Radio Dominance

Despite the Sharp radio's popularity, Hayakawa knew that crystal radios were an inferior product.

A crystal radio relies on the radio signal's power to make sound, limiting its reception and volume levels.

Vacuum tube radios on the other hand were capable of boosting the signal's power and volume.

After reverse-engineering an imported Neutrodyne, Hayakawa eventually produced the Sharp Dyne series of radios.

Depending on the model, the Sharp Dyne had 3-8 vacuum tubes, ran on AC power, and cost a fraction of the foreign imported models. It sold like hot cakes - sales went from 58,000 in 1936 to 130,000 in 1939.

To make enough radios to fulfill demand Hayakawa built a new factory and set up a production line with 23 minute-long steps on a conveyer belt.

In 1942, the company changed its name to the Hayakawa Electric Company, producing all sorts of radios from simple receivers to advanced two-way radios for airplanes in the military effort.

This business supported the company through the difficult war and post-War reconstruction periods. In one case, the company only survived thanks to the Korean War spurring massive radio purchases by the US Army.

Television

In 1951, NHK and other big Japanese companies started experimental television broadcasts. To demonstrate the technology's potential, they set up receivers in public areas.

These large screen projections were well-received so the NHK moved forward with public television broadcasts. Same as with radio, Hayakawa Electric wanted a television receiver ready for the big event.

With technical assistance from RCA, the company eventually produced a device in time for NHK's first television broadcasts in February 1953 - the TV3-14T, Japan's first commercially available television.

The TV3-14T still cost a mega-ton. It cost 175,000 yen at a time where the average civil servant made 8,700 yen a month.

Nevertheless, the product’s early and timely entrance secured Hayakawa's dominance in the domestic television market.

Manufacturing costs fell and monthly salaries grew.

And before long, a household television set was no longer so expensive and sales skyrocketed.

Hayakawa Electric quickly started moving on to other categories.

In 1962, they released Japan's first microwave, the R-10.

As well as Japan's first solar-powered consumer electronics product, a transistor radio. Hayakawa remained a dominant solar cell producer into the 1990s.

But it would be another product that put the company onto the global electronics map.

The Calculator

In 1960, excited about the future possibilities of the growing semiconductor industry, Hayakawa Electric started a team to design logic circuits and eventually produce a computer.

However, Japan's Ministry of International Trade and Industry decided not to enroll the company into its mainframe computer development program.

Without government support, Hayakawa decided to focus instead on sophisticated but non-programmable consumer electronics like voucher printers, cash registers, and calculators.

Calculators, especially. In 1962, the British company Bell Punch had started importing and selling its vacuum tube-powered desktop calculator - ANITA. So Hayakawa set out to produce their own 20-key desktop calculator, powered by germanium transistors.

Thus came the CS 10-A Compet, one of the first all-transistor calculators. It used 530 transistors, weighed 25 kilograms and cost about 535,000 yen or $15,000 in today's USD. Despite the price, it sold out its entire 300-unit supply in just a few months.

The next year, silicon transistor prices rapidly started to drop, allowing Hayakawa to produce the CS 20-A calculator, first released in September 1965. It was more power efficient, reliable, and best of all cheaper.

Newer models quickly followed. 1966 saw the CS-31A, the first desktop calculator made using integrated circuits. The calculator had half the parts, half the weight, and sold for half the price.

That year, Hayakawa led the Japanese calculator industry with 44% market share, selling 24,000 units.

About half of that was exported to the United States through its Sharp Electronics subsidiary as well as global distributors like Sweden's Facit.

I did a video about Facit earlier in this channel. Worth watching.

Sharp

In 1970, Hayakawa Electric changed its name to Sharp Corporation. The brand's name derived from a pencil, but soon came to encompass everything the company did.

That year, Founder Tokuji Hayakawa also stepped down as the company's president. Tokuji passed away ten years later in 1980 at the ripe old age of 86.

New President Akira Saeki took on the challenge of guiding Sharp through new, uncertain economic conditions.

Sharp's calculator's success triggered over 50 Japanese manufacturers - not to mention various American ones - to enter the market too, setting off what we now call the "Calculator Wars".

In about a decade, the calculator went from this desktop-bound monstrosity with hundreds of vacuum tubes to a portable, battery-powered gadget doing all of its calculations on a single, integrated chip.

The calculator was the mobile phone of its time. And the calculator wars pushed Sharp and its competitors to pull every possible trick to make their calculators smaller, cheaper, faster, and better.

A Better Display

Up until now, such calculators used fluorescent display tubes or Nixie cold cathode tubes to display their numbers. However, these were expensive, large, and power inefficient.

In January 1969, a Sharp researcher watched a NHK documentary about RCA. In that documentary, RCA demonstrated a technology they announced back in 1968 - Dynamic Scattering Mode Liquid Crystal Displays.

Scientists first observed liquid crystals in 1880. They were a strange material. They sat in a phase between liquid and solid, and refracted light like a crystal. Thus, the name.

In the early 1960s, RCA researchers noticed that liquid crystals had interesting electro-optic characteristics, and eventually invented a way to electrically control the way they reflect light as a technology for creating a display.

But in order to be suitable for a calculator, the crystal display needed a better contrast ratio and had to work at room temperature. More work had to be done.

From RCA to Sharp

Sharp asked RCA's General Manager Bernie Vonderschmitt to develop and produce DSM LCD screens for their calculators, and even offered to pay for development. However, RCA and Vonderschmitt declined.

The reasons for this are unclear. RCA wasn’t too interested in LCD, seeing it as a distraction from their core Cathode Ray Tube or CRT display business. And Vonderschmitt personally had concerns about LCD's immaturity, cost, and yield.

By the way, Vonderschmitt was a Silicon Valley giant. He helped teach RCA's semiconductor manufacturing methods to Taiwan's first teams and later co-founded Xilinx with Ross Freeman in 1984. His Japanese contacts were essential in helping to secure manufacturing capacity for the first FPGA.

Anyway in 1971, Sharp licensed some of RCA's patents and set up a team called S734 to put a DSM LCD display into a calculator in 12 months.

After tediously mixing over 3,000 kinds of liquid crystals and synthesizing 500+ mixtures, the team eventually produced the right mix for a suitable screen.

Thusly in June 1973, Sharp brought the Elsi Mate EL-805 to the market. Made up of just an LCD and five ICs on a single glass substrate, it is hard to convey just how revolutionary this product was.

The calculator was 2.1 centimeters thick, 12 times thinner than its closest competitor. It used 1/9000th the power of existing calculators and can go for 100 hours on a single charge. And weighed just 200 grams, a 125 times reduction.

More calculators followed, including a solar powered one in 1975. By 1985, the company had cumulatively sold over 200 million calculator units.

An LCD Giant

Sharp continued its studies on LCD technology. In the mid-1980s, they found out that IBM had identified a new full color silicon-based display technology known as TFT-LCD as a potential replacement for CRTs.

They got to work producing their own version of it. And in 1987, the company leapfrogged its LCD competitors in successfully producing a 14 inch full color, full motion LCD display.

Then in 1991, they debuted the first wall-mounted television: the 9E-HC1. People recognized this electrifying technology as the future and it started replacing CRT in televisions and other displays.

Unfortunately, this strong position was in marked contrast with the rest of the company.

Struggling into the 1990s

Following the unraveling of its real estate bubble, Japan's domestic economy weakened from the influence of a strong yen, increasing wages, and the effects of manufacturing offshoring.

This recession fundamentally shifted the direction of Japan's big electronics firms. Cheap foreign imports challenged Sharp's older products like fax machines, calculators, and microwave ovens.

To replace them, Sharp launched a bevy of new, more advanced products. Some like the camcorder were relative hits. The Viewcam had 20% of the Japanese market after its release in 1992.

But other categories like the Wizard electronic organizer never took off, and saturated categories like notebook PCs failed to make enough money.

The Crystal Clear Company

Profits declined throughout the 1990s.

In 1998, net income fell to 24.8 billion Yen, representing just a 1.4% profit margin and a near 50% decline year over year.

This bad performance resulted in Sharp’s president resigning.

The company's one bright spot had been LCD Panels and LCD-related products, responsible for 30% of company revenues.

Thusly in 1998, new president Katsuhiko Machida sought to streamline Sharp and reposition its LCD business at its very center.

The "Crystal-Clear Company" as he called it. In his first press conference, he said:

> Sharp’s problem is low brand perception. For branding, we have to make a "clear face" for customers. To this end, we will set LCD technology as our face. From now, every Sharp product will be related to LCD technology.

Televisions

Sharp's pivot towards displays was risky considering the emergence of intense competition from the rest of Asia, particularly Taiwan and South Korea.

Regardless, Sharp had high confidence in their LCD technology.

They ditched their venerable CRT television business - which dated all the way back to 1953 - and went all in on LCD televisions, launching its AQUOS TV brand.

This bold move seemed to have paid off. In 2001, AQUOS quickly captured 80% global market share in the very young LCD TV market, generating nearly $400 million USD in sales worldwide.

Sharp seemed to retain its dominance even after 2001, when a bevy of Japanese and Korean firms like Sony, Samsung, and LG followed Sharp into the television market.

By mid-2003, AQUOS dominated the local Japanese market with 60% share.

The next year in 2004, Sharp still held a leading 33% market share in the overall LCD industry. But Samsung and the competition were catching up faster than anyone at the company anticipated.

Notably in early 2002, Samsung caught Sharp and the rest of Japan's display industry by surprise by being the first to produce leading edge fifth-generation LCDs.

Kameyama

Looking to cement their lead, Sharp began investing billions of Yen into new TFT-LCD factories across Japan. The largest of which was the Kameyama Plant in Kameyama City in the Mie Prefecture.

This massive factory had two phases - completed in 2004 and 2006 - and cost a cumulative $4 billion USD. It produced over a million panels a year, the world's biggest at the time.

Kameyama represented a big bet on Japan-based production. By now, many other Japanese companies had offshored their LCD production to cheaper countries.

President Machida and the management however had high confidence in the superiority of Sharp's technology, and believed local production was necessary for the company to retain its secrets.

Local media praised the factory as a model of Japanese manufacturing. Sharp positioned its panels as a premium product. The "Kameyama Model" brand was to be like the Kobe beef of TFT-LCD televisions.

Financial Strain

The billions of yen spent on the Kameyama plant however strained company finances.

Like many other conservative Japanese companies, Sharp had a long standing financial rule restricting any single division's investment to a certain percentage of its sales, operating income and cash flow. But because of LCD's importance to the company, these rules were waived.

Unfortunately, the plant came online the same time as other massive facilities built by Taiwan's AU Optronics, Sony, Samsung, and LG.

The LCD panel industry fell into a situation of over-supply. In 2004, a 32-inch TV panel cost about $865. By 2011, that same sized panel would cost about $149.

Throughout the 2000s, Sharp's TV market share would fall from as high as 80% to under 10% while its rivals caught up to the company's first-mover advantage.

The LCD television was starting to turn into a cheap commodity.

Small Success

Looking to retain leadership, Sharp announced that they focus on selling more LCD panels to other businesses. As President Machida said back in 2004:

> We will try to keep greater than 50 percent share in world LCD panel production. Half of our products will be sold to other LCD TV manufacturers, and the rest will be used for our own TVs. Such a plan can be achieved if we continue to invest in production capacity.

This strategy meant even more financial investment - leading to Phase 2 of the Kameyama Plant.

In order to pay for it, Sharp cut its investment in home appliances. By 2006, LCDs made up half of the entire Sharp Corporation's revenues.

You can say that the LCD push had some positives. LCD-related division sales steadily grew from 2002 to 2006.

And in the fiscal year ending March 2008, Sharp as a whole showed record high sales of $34.5 billion and a billion dollars USD in profits.

But was this healthy revenue? Machida mentioned that he wanted 50% of production sold externally, but that ratio never cleared 20%.

And with competition from Taiwan, South Korea, and now China catching up to and surpassing Sharp, the massive investments they made into LCD production were not paying off as originally expected.

Sakai

In 2007, president Machida stepped down and a new President took office - Mikio Katayama. However, Machida retained his influence as Chairman. Together, the two oversaw the financial blow that crippled the company.

In 2007, Sharp announced that it would build a new set of factories in the city of Sakai, Osaka Prefecture.

The Sakai factories would cost $3.4 billion USD and were projected to be almost as large than the latest Kameyama Phase 2 location - 13 million units of 32-inch equivalent panels compared to Kameyama's 20 million units.

Who was going to buy all of these new panels? Sharp reached out to fellow Japanese TV makers like Pioneer, Toshiba and most importantly, Sony.

In February 2008, Sony, then the second-largest LCD TV manufacturer, announced an agreement with Sharp in which Sony would take a 33% ownership share in the Sakai factories - then named Sharp Display Products.

However, Sharp failed to properly manage relations between its TV and Display panel divisions. It is the same issue with Samsung and Apple, where one division of the company competed with the customers of another division.

Sony found itself relegated to second place behind Sharp's own AQUOS TV sales, leading to shortages and delays that pissed them off.

In the end, Sony did not go through on its 33% purchase - it never owned more than 7% - and eventually sold its entire stake in the joint venture back to Sharp.

Sharp tried to sell Sakai's larger panels to the United States, but these 60 inch TVs never really caught on there.

Crash

Then came the Global Financial Crisis. Sales collapsed and losses increased. In March 2009, Sharp announced that it had turned a net loss for the first time since the stock went public in 1956.

Management admitted that they made a mistake in investing in local production within Japan. Yet they still continued investing in Sharp Display Products - which changed its name to Sakai Display - and kept LCD at the center of their strategy.

This was despite the TV business division continuing to deteriorate. By 2009, Sharp's TVs held a market share of just 6-7%, good for fifth place and a third of Samsung's market leading 18.8%.

Operating utilization for the new $3 billion dollar Sakai plant fell to less than 50%. Not being able to decide between the LCD panel supply and TV businesses led to Sharp losing in both of them.

Sharp's sales in the fiscal year ending in March 2012 declined by $10 billion from 2008. The company turned a net loss of over $5 billion, and the company nearly fell into bankruptcy.

In March 2012, Katayama stepped down from being Sharp's president for this terrible performance, making way for Takashi Okuda. However Katayama and Machida remained influential through positions as members on the board.

Foxconn

At this point, Foxconn enters the story. In 2012, the $150 billion Taiwanese electronics giant bought a 9.9% stake in Sharp for about $854 million. They also bought a larger share in Sakai Display.

Why would Foxconn be interested? I mentioned in a previous video that Foxconn has a land-and-expand business model that seeks to take over a client's whole manufacturing supply chain.

Electronics assembly is an extremely low margin business. Foxconn squeezes profits through economies of scale and vertical integration up and down the supply network.

The more components that Foxconn's clients source from a Foxconn affiliate, the more profits Foxconn gets to keep. And for electronics gadgets like your mobile phone, the most expensive component tends to be the display.

Three years earlier in 2009, Foxconn helped Taiwan's Innolux acquire two other Taiwanese display companies to create one of the world's biggest TFT-LCD panel makers. Adding Sharp's capacity and technologies enhances this rollup strategy.

Foxconn did request a more substantial partnership - perhaps even an acquisition. But Sharp declined such a tie-up and would do so for several more loss-making years.

Tailspin

Chairman Machida led the negotiations with Foxconn for the investment. And with Katayama and Okuda still influential in the company, rumors of power struggles wafted through the Japanese business press.

Okuda eliminated 11,000 jobs as new smartphone makers like Xiaomi became big buyers of Sharp's small and medium-sized panels. But the company remained deep in the red.

There was still too many players in the LCD market. In 2012, Sony, Toshiba, and Hitachi merged together their LCD divisions to create Japan Display Inc or JDI. JDI and Sharp competed for several years, causing pain to both.

And Sharp for all of its supposed technological prowess missed several critical industry transitions including that towards new display technologies like LED-backlit LCD screens and OLED.

In March 2013, the company announced a staggering $5 billion loss, leading to all three leaders and former leaders - Katayama, Machida, and Okuda - exiting their positions.

The company never regained its footing thereafter. In 2013, Sharp and Fujitsu began losing their once-solid grip on the Japanese domestic smartphone space - perhaps one of their last hopes in the consumer electronics space. Apple surged ahead and now holds dominant market share.

Takeover

As early as the first investment in 2013, Foxconn had an offer out to Sharp for a more substantial investment. At the time, Sharp declined.

But in 2016, after three more difficult years, Sharp finally accepted Foxconn's offer to purchase a controlling 66% stake for $6.2 billion. However Foxconn then revised that offer down to $3.8 billion upon learning about a bunch of new financial liabilities.

Several other suitors came up, including those organized by the Japanese government. However, Sharp's bank creditors eventually chose to side with Foxconn.

Sharp's acquisition represents the first foreign takeover of a leading Japanese electronics firm. A sad end for a venerable 100 year company.

Post-Takeover

Foxconn appointed a Taiwanese - Tai Jeng-wu - as its new CEO. He has sought to cut costs, closing various unprofitable overseas ventures, cutting even more employees and moving production out of Japan.

In its first full year under Taiwanese management, Sharp turned a profit - though small - for the first time in four years.

Tai re-oriented the company back to producing the branded electronics products like it once had many years ago.

Foxconn is also rebuilding Sharp's withered semiconductor manufacturing business, which include a very nice 130 nanometer fab.

I mentioned this briefly in my video about Foxconn's semiconductor ventures in India.

In February 2022, Tai retired and a new CEO from Foxconn took the job - Wu Po-Hsuan. He previously led the company's brand products division, another indication of its future direction.

Conclusion

Sharp Corporation became a consumer electronics giant not necessarily because it had the best technologies, but because it used technology to produce the right products at the right time.

As the company fell deeper and deeper into LCDs, the company clung to a confidence in the superiority of its technology. But as it turns out, display panels are a brutal, grinding business with low margins and high insecurity. And it ground up Sharp too.

We shall see where Foxconn takes Sharp in the future. The brand still has a 100 years of value behind it. I hope I can see it rise to its former glory, once more.

Share The Asianometry Newsletter