Austin Lyons of Chipstrat and Vik Sekar of Semi Exponent break down a huge week for ASML. After TSMC, Samsung, and Intel all publicly committed to a new 6x12-inch photomask standard, they discuss what this means for the adoption of high NA EUV lithography. The two explore the technical tradeoffs of anamorphic optics, the economics of throughput, and why this coordinated move was essential for the entire semiconductor supply chain.
Things we cover:
The rise of personal AI agents like Astra, Instinct, and Muse
ASML’s business model and logic vs. memory sales
The move from 6-inch to 6x12-inch photomasks
Anamorphic optics and the high NA EUV throughput problem
TSMC, Samsung, and Intel’s high NA roadmaps
Why the mask supply chain needed a coordinated commitment
This podcast is lightly edited for clarity.
Personal AI Agents Get Real
Hello everyone, and welcome to another Semi Doped episode. I’m Austin Lyons. I publish Chipstrat, which is strategy analysis across semiconductors, AI, and data center infrastructure. Institutional investors read it, and so do technologists and executives from the biggest companies to startup CEOs. So check it out at chipstrat.com.
Vik: Hey, I’m Vik Sekar. I run Semi Exponent, a research and consulting firm focused on semiconductors and AI infrastructure. I also write Vik’s newsletter on Substack where I cover everything from optical interconnects, networking, memory, packaging, power semiconductors, for a whole range of investors of various kinds and engineers. You can find it at viksnewsletter.com.
All right, Vik. So, hey, what’s going on over there, man? I feel like I haven’t seen you in a while.
Vik: Yeah, it’s been a while actually. I think we needed a little summer break. We had a week off at the end of August and we were like, okay, let’s just recover and regroup in September because stuff’s always going on in this world and if you don’t take a break, it’s not going to work out. So we had a little week off. But in that week, of course, we weren’t quiet or anything. We were pretty much writing our substacks and spending some time with the family.
But what ended up happening was in this week, there have been at least two or three new products that are all personal AI agents that I’ve been messing around with a lot. The first one I messed around with was Grokbot and I was posting on X that I didn’t really like it because it sucked a whole lot of tokens. But then the whole idea was that it has its little CPU, a little VM machine, a computer, right? And then it goes and operates your computer screen for you and handles your logins. We made a whole episode on Grokbot and its implications on CPUs for AI. But then later, several other things have come up. Of course, the most notable one being the Astra 6 from OpenAI, ChatGPT 6 Astra, which is actually kind of amazing.
So I’ve been a cloud user for the longest time. So I was like, okay, here you go, OpenAI, take my 100 bucks and let’s try this thing. And I wanted to try the computer use thing because it’s pretty fascinating. It handles your browser and all that. It’s awesome. So I just put in, “Hey, I need to book travel tickets to OCP Global Summit in October in the Bay Area where I’m going.” And I was like, “Find me tickets. And so on the way back, I want to go to Japan where I’m planning to have my family over and then I want to have a week there and then I want to come back.” So it planned all this stuff for me and it was like, “Oh, here’s your ticket fare.” And then it’s like, “Hey, do you want to fly instead of Tokyo, do you want to try Osaka instead so that your travel itinerary could be from Tokyo to Osaka, you could do your vacation, fly out of Osaka?” I’m like, “Yeah, yeah, yeah, look up that ticket.” So it went and I could see the mouse moving, right? So it’s looking at the thing, clicking buttons and typing in the airport codes and I’m like, this is cool.
That’s amazing.
That’s awesome. And you know, I like the magic of, “Hey, you told me to do this, but have you thought about that?” I think that’s where it’s killer.
Vik: It went one step further too. It tried the Osaka thing and it’s like, “No, it’s too expensive. Your original flight was better in and out of Tokyo. But you’ll have to backtrack your holiday back.” Okay. Then it’s like, “Hey, do you want me to try booking only till Tokyo and I’ll book a separate flight ticket from Osaka to wherever you have to be?” And I was like, “Oh, cool. Yeah, that’s a good idea. Do that.” So it tried out all these scenarios for me. It was awesome. So I’m going to play more with it. What’s 100 bucks if it can help me plan a better ticket? I’ll recover my $100 subscription in no time. OpenAI should be sponsoring us. This is such a great pitch.
Yes, yes, yes. OpenAI, if you’re listening, I use both Claude and Codex. My son has been working on his own indie game all summer and he’s working on getting it in the App Store. And it’s really great. He has Codex do all of it for him, which is awesome because then he can focus on the storytelling and the story itself and it does all the programming. And then of course, I use Claude and do a lot of Claude coding stuff. But I love your use case of, “Try all these permutations of travel to save money,” because I tried to do that. I know there might be better things out there, but it takes so freaking long that I’m just like, screw it, whatever. I’ll just book whatever. I just get fed up.
Vik: And one more thing I haven’t even gotten down to is telling it, “Hey, I have points on this airline. I want you to select this or that, prioritize this or that.” So it can be like, “Oh, wait, if you select this, you will get points on this airline.” So much is possible and it’s all AI based. It’s amazing actually. So I really find it useful stuff.
And I did another crazy thing which I shouldn’t admit on camera, but here goes, right? Because YOLO, you YOLO AI all the time. So there is this other tool called Instinct, which I sent to you. It is this little tool that has no information. So if you go to instinct.co, you see a web page that’s like a letter written to somebody, “Hey, we’re building this.” No information about the company or what it is or what, literally no pricing even and it’s only an invite-based thing. And so I happened to chance upon an invite which I sent to you. So thank you to whoever is listening to this who gave me that invite. And so what I did was—it’s so dangerous. Do not try this at home. Okay, do not try this at home. So what I did was I gave it access to my stuff and email and all that because I’m curious. People are saying this is amazing. This tool is amazing. So if it’s not really useful to me, I want to know. But for it to be useful to me, I need to give it some access. So I looked up beforehand, can I revoke access if I give it? And I looked, yes, you can revoke access. I’m like, okay, fine, I’ll revoke it if it’s bad.
So what I did is I gave this tool all these details to my email and all that and it actually keeps a track. And the way you use it is you use it with your phone. So you can text it on iMessage or WhatsApp. And so it sent me a message saying, “Hey, you just got an email, by the way, you’ve got your press pass for the OCP Global Summit has been approved. Here’s the code you can use to get access.” And I’m like, “Oh, cool, this is cool. I should do it sometime.” But then I was working on something else. And then I was like, “Hey, you’re supposed to be a personal assistant, right?” I’m like, “Okay, go ahead and sign me up with the code. Do it for me. You take care of it.” And then it went off and I went back to doing my work. After some time it sent me another message and it’s like, “Here you go, I registered you. I put your name down. This is your affiliation I put down. I put your meal preference as no meal preferences. And by the way, I ordered you a medium size t-shirt.” I’m like, how did you get my size right? Or did you just guess it? Or did you know I’m an average Indian size? I’m like, how did you guess a medium shirt? By the way, I ordered you a shirt. I’m like, what the hell?
That’s amazing. That’s pretty good. That’s pretty good. So did you revoke access?
Vik: Yes, I did. I saw the email come through with the QR code too. I’m like, “Oh, here’s the QR code.” I just click star on my Google Drive. I’m good to go now. But then I’m like, holy shit, this is amazing. But it’s also too amazing for me right now. So I’m like, okay, that’s it. I revoked all access to it. I deleted it. They allow you to delete your particulars. There’s a way you can delete on both sides. You can revoke access on the Google side and you can delete any hashed information that is stored on Instinct. They allow you to do that. So I cleaned everything up. But it’s pretty good. And then I find out that Muse, Meta’s Muse Park is another personal AI agent that’s come up. So I installed that, but I haven’t tried it out yet. But I’m trying all these personal agents and I think it can be really useful to be honest.
Yeah, totally. Okay, lots of thoughts. So when you sent me Instinct, you sent me just a very short text. It was late at night here and it was just, “Hey, check this out.” And it’s a link to a website. I click on it and it’s instinct.co. It comes up, it’s white background with black text, just raw HTML. It looks like the very first time I made a website. It looked just like that. No CSS or anything. I scroll around. All it says at the bottom is “San Francisco, California.” It doesn’t say anything about who it is, whatever. And I’m like, “Oh, Vik got hacked. Vik totally got hacked.” This is a .co website. He said something very generic, just “Check this out.” It’s got nothing here other than “sign up.” And I’m like, this could be someone in North Korea. This could be someone random, who knows?
Vik: It’s one of us you load here. Yeah, exactly.
So thank you for YOLOing it and maybe now I’ll try. And I think, zooming out, I’m very excited about Muse. I’ve heard very good things. My son this morning, we saw in the Wall Street Journal and he was like, “Dad, you got to try it.” And I was like, “I know, I know.” I just can’t keep up, dude. This stuff’s happening every day. But I want to try it. And I think what this is is this is the evolution of OpenCog, which was the very early adopter tinkerer hacker, “I want agents. I want them to do useful things. I want them to have memory,” but you have to do a lot of work. And now this is the abstraction to more of a product experience where more people can just go in and they can just use it. They don’t have to think about all that stuff under the hood. So obviously, growing up from home-brewed computers to, this is a Dell computer that you order and it just shows up for you. So I think obviously, as we’ve said before, this does show even if it’s not Grokbot, even if it’s not Instinct, Muse—I mean, think about the reach of Muse. Meta basically has Instagram, they’ve got Meta, they’ve got the whole family of apps.
WhatsApp is there too, by the way. So literally, we could talk about hundreds of millions of people trying this, using hundreds of millions of cores on server CPUs that those people weren’t using yesterday and now they’ll be using in the next handful of months. So obviously, there’s huge implications for server CPUs, for memory and so on. I think just this product experience of agents is worth tracking.
Vik: It’s genuinely useful. If you tell my mom, “Hey, just text it stuff, it’ll do its job,” that is the lowest barrier to entry for agentic AI. Literally everybody can text. My mom texts me. If she can do that, everybody can use AI if it’s through a text message. Everybody’s on WhatsApp too. It’s amazing. I think it’s come to a level of accessibility and usefulness. Actually, Grokbot was not in my opinion as amazing as what I saw ChatGPT 6 do with its computer control. It was actually pretty amazing. I think we are only starting a new era of really accessible AI. So far, people didn’t really know what to ask in a text box. I tell my family, “Why don’t you use Claude to do this?” They’re like, “What do I do?” “Oh, go to chat.openai.com or whatever.” “And then what?” “Oh, there’s a text box.” “Okay, now what do I ask it?” So many problems. Now, it’s getting to a point of usability that is very much for the common person, not the geek.
Totally. I just thought of a use case for my wife to try it, which is we’re trying to plan a family vacation with my in-laws and with siblings. And so there’s going to be several families. And she raised her hand and she’s like, “Oh, this is going to be for my parents’ anniversary. So let me schedule it. This will be fun.” And then they won’t have to worry about it. And once you start getting into that, you realize everyone has calendars. Everyone has preferences. Every time you throw out, “What about this location or this location?” someone’s like, “Oh, how about that location?” And it’s something different. And it’s a never-ending thing and you’re never going to make anyone happy. What if you could just have Muse or Grokbot or Instinct or someone do it? They can manage the calendars. They can throw out the ideas. And plus, obviously, it’s nothing personal. I think it’s super annoying when someone comes back and is like, “Okay, I see that you’ve put time in here, but here’s 10 other ideas.” But the infinitely patient AI can do it. And these are real problems that I think real people would latch onto and be like, “Yes, please help me here. I will entrust some of my information to you and pay for it so that I don’t have to deal with it.”
ASML’s Business and the Move to High NA
Vik: Yeah. And you know what we need to make all of this happen? Chips. And you know who makes them? TSMC. So that’s the segue to the actual part of our episode today. I love it. And who does TSMC rely on? ASML. So let’s talk ASML. They had news this week. News with TSMC, news with Samsung and news with Intel. So we’re going to talk all about it in the next 30 minutes or so.
So first, I’m going to zoom back out. And I think everyone knows ASML by now, but for anyone who’s brand new to the space, obviously, ASML is the lithography company. And lithography is that step where you project the chip pattern onto the wafer, which allows you to etch or deposit or ultimately create the physical implementation of these transistors. And why ASML matters is they’re the only company on earth that sells EUV lithography machines, extreme ultraviolet. We have a whole podcast that we’ve talked about before. We’ll link it in the show notes about lithography. You should check it out. There are some competitors in the space, Nikon and Canon still sell older DUV tools, but at the leading edge, it’s ASML or nothing.
And so what does that business look like right now for ASML? Just in Q2 2026, they reported in July, ASML’s total net sales were 9 billion euros, roughly 10 billion or so dollars. Gross margin of 54%, which is pretty good when at the end of the day, they’re making a manufactured heavy equipment machine, right? So to have gross margins of 54%, net income of around 3 billion euros, so 31% net margin on the quarter. This is pretty crazy. Again, I think in my head I go, “Oh, well, Nvidia makes 70% margins and these memory makers have 80% margins, so that’s a lot less.” But when you actually think about ASML as making equipment for manufacturing, these margins are crazy, which of course, they have earned because it’s probably the most complex machinery in the world.
So the question is, before we get into the news, one of the things you might have noticed in the news is I talked about a memory company, Samsung, and logic companies, TSMC and Intel Foundry. And so one thing that’s worth noting is actually ASML makes their money by selling systems and by servicing these systems. And actually, the servicing of the systems was 2.8 billion euros of that 9.3 billion euros. So almost 30% of their revenue is actually recurring revenue. It’s like subscription revenue. There’s an analogy here: if you think about Apple, obviously they sell devices, but then they also want to sell software and services and get some recurring revenue. So they’re not just having this lumpy, “we sell devices during the holiday season” kind of thing. ASML’s trying to build a similar business profile where of course, they sell these $400 million machines, but then they also have this ongoing revenue of helping to maintain those machines.
But of those system sales, actually 51% go to logic companies and 49% are for memory companies, at least back in this quarter. That actually was pretty surprising to me. You tend to think more about the bleeding edge EUV for logic and not as much for memory. But was that news to you?
Vik: Yeah, to me it is because I’m always a little behind on catching up on the lithography and wafer fab equipment space. But what I wanted to mention is when for people listening to this, when you say logic and memory, fundamentally they are different process technologies, right? So when you say two nanometer transistor or three nanometer transistor, by now, we’ve explained before that this is by this point a way of saying it, that the dimensions aren’t actually two and three nanometer. It’s a gate-all-around where they stack multiple gates. It’s a different structure, but the nomenclature has stuck, right? So that is one side, that is what you refer to as logic. And memory technologies are actually quite fundamentally different because the transistors there, even the leading transistors there are more a 10 nanometer class node. So when you say EUV is required for memory technologies, it’s not to make the transistor, is it?
Right. It’s really about memory is all about getting as many bits as densely packed as possible, so you can get as high of capacity as possible and really try to drive down the cost per bit. And part of that is making capacitors and making the transistor and capacitor structures as close as you can. And so a lot of it is about the pitch, the lowest pitch that you can get to try to really cram everything in.
Vik: Right. So the memory nodes have basically this capacitor and a transistor. So what they call this is a one transistor, one capacitor or 1T1C technology, where the capacitor holds the charge. If it is holding charge, it’s a one and if it’s not holding charge, it’s a zero. And the transistor then turns on or off to charge it or recharge it or discharge it, right? So that’s how memory works. So creating this 1T1C structure is where EUV lithography plays a role in memory. And it’s amazing that it’s nearly half the utility. So EUV shouldn’t always be thought of as getting to the next greatest transistor node because wherever you read it in the news, you will hear it as, “Hey, this is for A11 or A7 nodes,” which is the Angstrom era nodes, which is basically when you say A7, it’s a 0.7 nanometer transistor equivalent, right? So that is what you usually see in the news as why we need EUV, but you actually need EUV to make memory as well. So that’s the only thing I wanted to clear up just to make sure that people understand the difference between these two regions.
Yeah, that’s helpful. And so, to that point, when you look at the geographies that ASML sold to in this most recent quarter, 43% of their system sales went to South Korea, 30% to Taiwan, 14% to China, and 9% to US. So again, you can think, “Oh, South Korea, memory companies; Taiwan, obviously logic.” China is an interesting one because at first I was like, “What, they can still sell to China?” But of course, we’re talking about older DUV equipment that you can still sell to China. For people who haven’t followed this closely, the ASML story was, if you go look at how much they sold to China, there was a period within the last few years where it really spiked almost up to 50% of their sales and then now it has come back down. And that was obviously China seeing the export controls coming and so buying up as much equipment as they could before the export controls came. And then United States, obviously, there are also, obviously Intel Foundry is buying leading edge stuff, but there’s lots of other companies that are buying older tools, you can think of the GlobalFoundries of the world. So, United States is small, but it’s not just Intel Foundry.
Vik: I wonder if it’s Micron for memory.
Oh, yes. Yeah, true. True, definitely. Okay, so let’s move forward to why there was so much news this week. Well, this week, all of these announcements were clearly timed and coordinated around the conference, SPIE Photomask Technology and EUV Lithography Conference, which was in Monterey, California. Nice place. I was like, “Dude, we should go next year.” Monterey’s cool and we could learn a lot here.
And so, SPIE is the Optics and Photonics Professional Society, and this particular event is the annual gathering of the Photomask world. So this is the people who make and use masks, whether it’s captive mask shops inside of Intel, TSMC, Samsung, merchant shops, names like Photronics, DNP, Toppan, which I don’t know a ton about these companies and so it’ll be fun to dive into them more. You’ve got mask blank suppliers like Hoya, AGC, mask writer companies, inspection folks, because not only do you have to make the mask, you have to pattern the mask, like draw your particular layouts, but then you have to inspect it to make sure that it is correct. And so you’ve got companies Lasertec, KLA, Zeiss, all in this space. And so they all come to this conference, and they are all in the supply chain of making the photomask. So, I guess before we talk about what changed, would you like to introduce the idea of what is a photomask, what are some of the dimensions, what’s the current state of affairs?
The Low NA and High NA Tradeoff
Vik: Yeah, I’ll tell you what I know and you correct me if I’m wrong because this is how I learn on the podcast live. This is good. So photomasks are essentially patterns that are on a different substrate-like thing other than the wafer itself, where there are patterns where the transistors should be or should not be. So you can have a positive photomask or a negative photomask, which means the patterns on the photomask can selectively block light from hitting the wafer underneath, or they can selectively let light through. So however you want to look at it, positive or negative. And then depending on the materials used, parts that are exposed or not exposed remain and the rest can be etched away, right? So this is how basically features on the wafer are formed.
And fundamentally, wafers have been 300 millimeters on CMOS for a long time. Those are 12-inch wafers. But the masks that pattern on these 12-inch wafers have never been 12 inches. They’ve been 6 inches. I think they’ve been 6x6 squares or something like that. Then what happens is now if you have to pattern a 12-inch wafer, you have to use the 6x6 in different spots on the bigger diameter wafer and you have to stitch them up together eventually. That takes time and it’s complicated to do. I think the benefit that you’re going to tell me next and we’ll hear more about this is that now we have gone from a 6-inch lithography mask to a 12-inch lithography mask. So we can pattern the whole thing in one go.
Yes, yes. So high level, you’re right, and we’ll get into the details. The mask is all about how you end up drawing where the transistors go. And the mask, the standard mask has always been a 6-inch square, 152 millimeters by 152 millimeters. And it is basically a quartz plate. And ultimately, the question is, how do you draw these tiny lines to allow the EUV light to either shine through or get blocked? And that’s traditionally done with a tool called an electron beam, an e-beam writer, and it’s a very slow process to draw this. But ultimately, you write on the mask.
And then like Vik said, this ends up getting used in the semiconductor manufacturing process, today, it’s 6x6 inches. And that works fine for low NA EUV, which is what the industry is using today. What happens is you have the 6x6 inch mask, but with low NA EUV, there is this demagnification step where it gets shrunk down by 4x in both directions. So the final actual reticle size that you’re patterning is 26 millimeters by 33 millimeters. And this gets you that 858 millimeter squared for the full reticle size, the postage stamp. And if you remember in the Hopper era, it was like, “Oh yeah, every Hopper is reticle sized.” We cap it to that. And the reason is because you want to just make one mask and then pattern it all over the wafer and get as many Hoppers as you could. And then if you remember, in the Grace Blackwell era, they actually said, “Well, we need even more transistors and more compute. So we can’t make that reticle size any bigger because there’s this whole industry around the photomasks.” So let’s just create the Blackwell GPU, but we’ll end up stitching two of them together to create a GPU with two dies. So that’s the way the industry has worked so far.
Vik: I want to just step back for one minute because I think there’s an analogy here that will help people understand this if that was a little difficult to follow. The reason I’m saying this is that lithography requires a certain amount of expertise to see how it actually works. But let me see if I can throw a little light on it. So when you’re saying low NA, I think we should just mention NA stands here for numerical aperture. The larger the aperture, the better. The higher the numerical aperture, the better—better in the sense that you can make smaller transistor dimensions. That’s the one thing. So we are now going from low NA to high NA, which means we can make smaller devices and patterns. The other thing is this reduction in size that you mentioned, 4x, right? The best way to think about this is think about a telescope.
Exactly. And by the way, as we talked about advanced packaging, that’s how you stitch together these GPUs. So obviously, advanced packaging, CoWoS, EMIB, very important. Okay, so back to when you talked about low NA and how we want to move to high NA.
So the numerical aperture—the question is, how do you make smaller transistors? And the simplest way is to decrease the wavelength of the light you’re using. But EUV lithography uses 13.5 nanometer light, and that’s very specific, as we’ve talked about before. It has to do with tin and these droplets and you’re hitting them twice with the laser to essentially pulverize them, and they give off light at this certain 13.5 wavelength. It’s not so simple to just be like, “Oh, well, find a different material that gives off 9 nanometer wavelength light.” That’s not so easy. So the other dial that you can turn is, okay, we can’t change the wavelength of the light, but we can change the numerical aperture. And so, if you have a bigger numerical aperture, you can have finer resolution, and so you can ultimately resolve smaller transistors.
But there’s a big tradeoff. When you go to a higher resolution, in this case going from low NA, which is 0.33 NA, to high NA, 0.55, what happens is you have to use something called anamorphic optics. You can think of it like widescreen film, and it’s a trick that they use to make that work. But ultimately, a wider aperture means that the light hits the mask over a wider range of angles. And the EUV masks, they’re actually mirrors. It’s this 40-layer molybdenum and something else reflective mirror. And that multi-layer mirror only reflects well within a narrow cone of angles. And so at 0.55 NA with the 4X projection optics, the steepest rays coming in would fall outside of that cone, and so it didn’t reflect as well. So what ASML does is they make the demagnification step anamorphic, where ultimately what this means is they keep 4X in one direction, but they actually go to 8X with their projection optics in the other direction. And so this works, but ultimately what happens is the area that it projects on ends up being half as big. Instead of your final reticle size projection, your exposure being 26 millimeters by 33 millimeters, you ultimately get 26 millimeters by 16.5 millimeters. So that means with one exposure, you get half of the area ultimately.
Vik: Okay, okay. I get it. So I’m learning this anamorphic optics, right? Okay, the term aside, I think the basic idea is that instead of the 4X reduction on X and Y, you have 4X on one side and 8X on the other side, which means you have now—Honey, I Shrunk the Kids. I don’t know if you know the movie from a long time ago.
I’m old enough. Yes.
Vik: Old enough to know that movie. So you shrunk the kids too much on one direction, which means your feature sizes get much smaller along the 8X direction compared to the 4X direction. That automatically means that your size of the reticle has dropped by half in that direction where it was 8X, where you shrunk more, the dimension drops by half, right? So that’s what this whole thing is. Remind me what it is to do with the mirrors again?
The way that as you increase the numerical aperture, as the light comes in at a steeper angle, it doesn’t travel through the mirror sandwich as well.
Vik: Okay, I see. So the numerical aperture being bigger means that the mirror can collect light from a wider angle. And therefore, a higher numerical aperture in the mirrors means that this wide angle can now come in where it could not come in in the low NA era.
Right, right. But for some reason with the 4x4 demagnification, the light coming in at the steep angles doesn’t make it through all the way. In my head, I’m thinking of it as fuzzy when it comes out.
Vik: Yeah, okay, got it. So the mirror itself allows for more of the light collection. Okay, got it. So that’s what high NA does, and now we’ve gotten more magnification on or rather, more reduction on one angle, and then the reticle size drops. Okay, now how do we get the reticle size back up?
The 12-Inch Mask Solution
Yes. So you could just accept high NA as it is and use 6x6 inch masks, but your throughput is cut in half. If you used to make a Hopper GPU in one shot, now you’re going to make half of a Hopper GPU in that shot. And then you’re going to have to make another mask that’s the other half of the Hopper GPU and have a second shot. So now the economics of the throughput of all these very expensive machines that you’re buying. You can either cut the number of wafers that you’re producing down or you can buy two $400 million machines, right? So now it’s like, “Oh, it was $250 million for low NA EUV, and now I’m going to pay $400 million for high NA EUV because it’s complicated, it’s got even bigger mirrors and stuff. And you’re telling me that my output goes by half. So actually I need to buy two of these. So I got to spend $800 million just to keep the same throughput.”
But it’s not only that, it’s not an economic problem. There’s actually a floor plan problem, which is you have to lay out half of your GPU on one mask, half of the GPU on the other, and then you have to align it and stitch it together. And no chip designer wants a seam running right down the middle of their GPU. So this is a reason why some people are like, “Dude, high NA is never going to work. Who’s going to want to do that? It seems more expensive. It makes it harder for the chip designer.”
Now, Intel is already using high NA in production because they found certain layers where they aren’t as worried about this problem. Maybe the features that they were going to build were on a sub-reticle size die. Maybe it’s okay if it’s smaller. But other people are saying, “Well, wait a minute, we need to recover that full reticle size.” How do you recover the full reticle size? Ultimately, you go from a 6x6 mask, which was projected down 4X by 4X, to a 6x12 mask, which is going to be projected down 4X and then 8X, and you recover that reticle size, the full field of 26 millimeters by 33 millimeters with no stitching.
Vik: I see. So, the masks did not go from 6x6 to 12x12. They only went in one dimension to 12.
Exactly, because only one of those dimensions on the high NA was changing in the projection optics.
Vik: I see. I learned something today. So, that’s the whole point because you have 8x reduction in one direction, but to make that equal so that you could make the reticle a whole reticle shot in one go and not half the throughput, you now have to make a 12-inch mask on that dimension where the reduction was 8x. And the other dimension can remain same as the low NA era of lithography. Okay. Awesome, awesome. Oh, by the way, I wanted to say, Intel is already using this on their Core Ultra Series 3 Panther Lake processors. That is in production. So, they already have adopted this and it’s kind of cool, if you end up with one of those in your devices or something. But yeah, it’s cool to think that high NA is already in products and stuff.
A Coordinated Industry Roadmap
Yeah, it is wild. There’s literally CPUs out there running right now as we speak that used high NA EUV. It’s pretty cool. So, okay, now, let’s say everyone says, “Okay, okay, okay, 6x12. That doesn’t seem so hard. Just make a 6x12-inch mask. Let’s do this.” Okay, not so simple. This is where you have to get the whole supply chain coordinated around this. And this is why you really ultimately need ASML’s biggest customers to stand up and say, “We’re going to buy this,” right? Because if ASML just says, “Okay, I buy my 6x6-inch very high quality mask substrate from someone. Oh, hello, dear sir. Would you mind making me some 6x12-inch ones?” And they’re like, “Whoa, that’s twice as wide. None of my tooling is built for 6x12-inch.” And maybe that increases a yield concern because now the area is getting a lot bigger and so they have to make this pure wafer in a lot bigger area.
So it’s the mask substrate. It’s the companies that are testing it. Now instead of building equipment that tests a 6x6-inch mask, they have to test a 6x12-inch mask. Anything that’s automated that was used to carrying around 6x6 inches has to change to 6x12 inches. And so this is why it takes the whole industry and why those people in the supply chain ultimately need to see the end customers—TSMC, Samsung, Intel—say, “Yes, I will buy this.” And they also need the customers to give them the timelines. “I want to move to this by 2030, 2031, 2028,” whatever. So that they can all plan, invest in the R&D to make it happen and then plan accordingly.
Vik: When—so, now that you mentioned timeline, I think anybody listening to this would have that question immediately. So, when is all this really going to happen? High NA EUV. I know some of it is happening now, but what’s the plan for TSMC going forward? Because they weren’t very big fans of high NA as far as I could understand. They were like, “Hey, let’s just use low NA and push the limits of that as far as possible with other techniques like multi-patterning,” which we won’t get into here. We did speak about it earlier though. But they were never for going for high NA EUV, but now that they seem to have jumped on board, I’m assuming because they don’t see any other way out. But then what’s the timeline?
Yeah, well, so, TSMC says, “Hey, we are going...” They announced that they’re aiming for high NA, high volume manufacturing by 2030. And the industry is together saying, “Yes, let’s target the 12-inch masks in 2031, pilot lines in 2031 and 12-inch high NA systems in full production in 2033.” So, the way I read that is the industry says, “Okay, it’s going to take us five to seven years for everyone in the supply chain to make this 6x12-inch thing happen.” And TSMC said, “Okay, great. I’m going to wait until you’re 80% of the way there before adopting high NA.” And so, what TSMC is going to do is they will adopt high NA just like Intel did in certain situations where only having 6x6-inch masks is not a problem. And that will let them start to run it in pilot, run it in production, learn how the tool works, make sure the tool works. And then basically, they’ll have done all their learnings just in time for the 12-inch mask. So, the way I read it, TSMC is waiting as long as possible.
Vik: Also, it’s because it does take that long to bring this monumental change. It’s a monumental change when you’re changing the mask size like that, I imagine. And as you say, the supply chain has to adapt to it over time. And not only that, all of the capacity needs to go into using those process nodes. It’s probably going to be a future process node that uses all of this high-end features. And also the question of how many layers will you use EUV for or high NA EUV for? Are you going to use it for all the steps in a transistor or in the wafer? No, right? Probably not, right? Like metal patterning and things like that, I don’t think you need EUV for. Maybe there are only a fixed number of layers on which you will use EUV for. And initially those will start to be a low number of layers and then slowly they may increase it.
Who Adopts First: Logic or Memory?
Yes, yes. You make two interesting points there. One, which is you could selectively use high NA in places where you had to multi-pattern before. And which could be where Intel’s using it today. Hey, if I had to multi-pattern with low NA, but I could do it in a single exposure with high NA, then even if my throughput is worse or I choose selectively in places where it’s not worse, then it actually could be economical for me because it’s ultimately not only less patterning steps, but then less deposition, less etch, so on and so forth. And then two, yes, you make the interesting point, which is ultimately this is trying to intercept future process nodes. So, although Intel is using it on 18A today, ultimately, it would be about making 14A as high NA ready as possible.
Vik: Or or even beyond, right?
Or even exactly, 10A or whatever. Exactly, because we’re talking five, six, seven years. Totally.
Vik: But this is good. So, I think the point is that although the news is very exciting to people like us about how this is a change in the way lithography is working and it’s a big step for TSMC, Intel—I believe Samsung is also on high NA EUV. So, they are also on this train. It’s a big step forward. But the adoption is going to be over a long period of time. Although I think that adoption in memory will be sooner than the adoption in logic.
Mm, say more. Why is that?
Vik: I think that logic—again, this is a little bit out of my depth per se, but I’m going to take a shot at it. I think memory structures are a little more simple than logic transistors are. The 1T1C cell that we mentioned requires this high aspect ratio capacitor and you want to pack as many of them as possible together. So, the EUV has that problem to solve. But if you’re going to use EUV to make a gate-all-around FET that has four layers and each gate is atoms thick and then you want to stack them—that’s a much harder lithography problem fundamentally because the structure of the transistor is much more complex. So, I think it’ll hit DRAM faster.
Yes, that makes sense. The transistor is very complex in three dimensions and these capacitors that they’re trying to just cram together as closely as possible. It’s ultimately like you said, just a big long tall rod or something. I think of it like digging a big long well or something. It is structurally very simple—I shouldn’t say very simple to my friends at Micron and Samsung, where they’re like, “Dude, this is not simple.”
Vik: Yeah, and if I’m wrong about the whole “memory will use EUV before logic” thing, leave it in the comments below because we continuously learn through a lot of the people who listen to us because expertise is very vast in this industry and we only can cover this much. So, please feel free to let us know.
Totally. We try to go very broad and we go as deep as we can, but we can’t go impossibly broad and impossibly deep, right? So, that’s where everyone else comes in. Now, to maybe make your point, Samsung’s announcement was that they’re going to use high NA EUV for DRAM in 2028. So two years earlier than what was announced by TSMC. And then Intel came in and said, “We’re already doing high volume high NA today,” like we talked about in Panther Lake already. And one of the new pieces of news that they announced was that they’ve processed 1 million-plus wafers to date.
Vik: Oh my God.
Now, maybe logic does come first. Well, yeah. So our dear friend, Lithos Graphene on X—really good, check him out, follow him. This is all right in his wheelhouse. He pointed out, “Okay, they probably haven’t used a million wafers to make a million production parts, but they’ve burned through a million wafers in the practice of learning, testing, qualifying, production,” that kind of thing.
The point is still valid, which is Intel is a million wafers deep in their learning process. And if TSMC is saying that they’re going to adopt this later, you can make the argument that Intel is much further along in understanding and using high NA in production.
So if you think that’s ultimately going to be a strategic benefit for them, then it’s just worth noting that both companies have been very clear that this is the direction they’re both going and Intel is saying, “We’re a lot further along.”
Learning from the Past
Vik: Actually, it’s funny how history, if you look back, the tables have kind of turned because Intel was kind of adamant against adopting EUV in the 10 nanometer era, which is what set them back a whole lot for the many, many years to come. But I think this time they’re not making that mistake again.
Yes, yes. Intel was late to EUV and it burned them and now they’re like, “We’re never going to let that happen again.”
So, okay, there was more news from ASML. We’ll have to talk more about them. Actually, a fun fact, I’m trying to learn Dutch on Duolingo.
Vik: Why? Is this to listen to ASML earnings calls in Dutch?
No, no, no. It’s brain exercise and I was like, “What language should I learn?” And then my wife’s learning Danish, and so I was like, “Well, you already know Danish, so if we go to Denmark, go to Copenhagen, you could be the one to translate.” And then I was like, “Why don’t I learn Dutch? There’s a lot of semiconductor companies there.” So, quiz me sometime on my letter loans.
Vik: I don’t believe you. You’re just doing it to get ASML Alpha at this point.
Yeah, like we’ll talk in Dutch and you can tell me your secrets and I won’t repeat it in English. Yes. Okay. So yes, maybe someday we’ll do a Semi Doped tour of the Netherlands. But with that, folks, thank you for listening. Hey, Vik and I have newsletters. You should check them out. chipstrat.com, viksnewsletter.com. I was actually talking to a CEO this week who was referencing Vik’s recent posts, so that was pretty sweet. Check out semidoped.com. We’ve got a daily newsletter and leave comments, send us your ideas and thanks for listening.
