Thank you all for coming. Thank you for Hello Tomorrow and ASML for inviting me. For those of you who don't know who I am, my name is Dr. Ian Katrris. I am an industry analyst. It's my job to look at all the silicon in play in the semiconductor space. My background is physical chemistry. I became a journalist for a little while and today I now also run a YouTube channel called Tech Potato. My background has always been about programming silicon and the makeup of silicon architecture. I remember dealing with CUDA back when it was version 0.1. Those were terrible days, but it got me a thesis, so I'm quite happy. I spend a lot of my time speaking to companies like ASML, but also their clients in the foundry space and a lot of companies who build chips at those foundries. But we can't talk about chips these days without mentioning the little thing called AI. It's living, it's breathing, and I know a lot of you are working in there today. The AI boom is real, not just in hardware, but in software, and we're coming up against a few new evolutions of the technology where it matters most. Sometimes if you track the software space in this, it feels like it's moving so fast. I remember a few years ago we were talking about having convolutional neural networks. Then transformers were the thing and now it's all about agentic AI. I just came from a trade show in Taiwan last week called Computex. 150,000 people. Every question: agentic AI, agentic AI, agentic AI. And then the other conversation was about the macroeconomics around memory and of course we'd all like cheaper memory, but the thing is the companies I deal with all have a hard problem: they're trying to integrate themselves into a very complex infrastructure. Realistically, it's the infrastructure that's helping build AI today. It's not just the chips, it's the installation, it's the use case, it's the software built on top and the interconnect between them. We have thousands of companies working in each of those spaces, all of which from big companies like the Intels, the Broadcoms down to the startups, all have the same issue: how do we bring hard technology, hard science to the masses? And because everything's moving so fast, how do we make it relevant today, not in 3 years? I think a lot of you in this room are having those conversations right now about how do we bring this deep technology into the market. There's a lot about power and investment that also play into this. You can read about water usage of fabs and foundries as well. It's always a very interesting topic. But my guest today, Kristoff from ASML, he is one of the first to showcase how they implemented this sort of deep hard science into their infrastructure. Today at a conference a couple weeks ago, he was presenting how they're now using AI across their infrastructure to help optimize their production tools so that their foundry partners can maximize their revenue. If you've ever been to a semiconductor fab, you'll know these things have to operate at 80-90% utilization because every minute they're burning thousands and thousands of dollars just in opex to keep up and running. Now, I've spoken for long enough. I know Kristoff's got a few things to say, so please let me welcome on Kristoff Fuket from ASML.
Good to see you, my friend. Thank you, Ian. Well, good afternoon everyone. It's a great pleasure to join you today. I think you know that the industry in semiconductor, which now is quite easily extended to AI, is living a very, very exciting time, a time where we seem to believe that the number of opportunities when it comes to innovation is going to grow exponentially because basically of what AI will enable in the next few years. So I'm very happy to be here because you know what people forget also a bit about AI about semiconductor is that in order for that dream to become true, to realize, we are going to need major, major innovations in energy, in material, in healthcare, in all the things. I was walking a bit around in the show, all the things some of you have been working on, and all of you I think are trying very, very hard to come up with new ideas, new concepts to support basically this extraordinary dream. I'm also happy to be here because I understood that there are about 48 different countries represented, so it's not only about the Netherlands, it's not only about Europe, it's much broader than that. In ASML we strongly believe that for innovation to reach everyone, everywhere, it's extremely important to be able to collaborate and build basically some very strong ecosystem around the world. So I had a chance to see this very diverse origins around the show and I was very, very impressed by the amount of creativity, the amount of energy, passion of some of the people I've talked to around basically the topic of innovation. So this is great. And you know, of course, like in any ecosystem, there's a few big established companies: Nvidia, TSMC is one of them. But I think that we all know that without all of you, without more innovation, more creativity coming to life in the next few years, then maybe at some point the dream doesn't become fully reality. So again, very happy to be here. I know that Ian has prepared a lot of questions. I'll do my best to try to answer those. But again, it's a real pleasure to be here. Thank you very much.
So Kristoff, I'll be honest, I'm kind of a bit of an ASML fanboy, right? As an engineer, the stuff your company puts together feels like magic. And I know you've had the arguments internally about how much magic is involved in getting this stuff. But one of the overriding things when it comes to semiconductors over the past 30 plus years has been this concept of Moore's law, right? Getting a lot of performance for free. And you're a company that's at the leading edge of building the latest generation transistors. When I speak to Intel, the home of Moore's law, they say it's still alive. If I speak to Jensen Huang at NVIDIA, he says it's dead. If I speak to Dr. Kevin Zhang, our friend over at TSMC, he said, 'I don't care.' So, where do you sit?
Well, I think what you see through the different opinions people may have around Moore's law is that there are a lot of different versions of Moore's law in fact, right? It means many, many things. The one I like the most is the law that says that every two years the number of transistors per chip should double, and that has driven the industry for many, many, many years. But to be honest, if you look at what's happening today with AI, therefore Jensen's comments: AI needs a lot more than that. So when you look at the most advanced chips, the most advanced AI product, you're not looking at a two times every two years, you're looking at a 16 times every two years. So the AI industry needs a lot more transistors, both for logic and memory, that Moore's law used to provide. So this is why Jensen says Moore's law is dead. Well, you could say also that Moore's law has been put on steroids by AI because we need to do a lot more than that. So what's really happened is, if anything, a huge acceleration of Moore's law, a huge acceleration of the appetite for transistors in semiconductors. And this is a bit what the industry is going through right now, and this is very exciting of course for all of us.
I'm sure you saw the news last week. Huawei announced their reclassification of Moore's law as to scaling to be this sort of it's not just the transistor, it's also the package, it's also the interconnect, it's also the system. Do you agree with that view?
Well, I think if you go back to the very day when Dr. Moore presented the Moore's law as we know it, he already said that density would be achieved by scaling, so putting more transistors per unit of area, but also by packaging some of the transistors together. So I think when you look at how you can achieve more density, scaling is obviously one way to do it, which is being complemented more and more by what we call in ASML 3D integration, that can be advanced packaging, but that can be also stacking wafers together or dies together. So when you move so fast on density, the truth is you're going to need both. And if you cannot scale, which is a bit what Huawei told us a few weeks ago, then you have to do more of the other things. You have to do more integration. So it's two axes, and the multiplication of the two axes gives you the density. And if you need to go 16 times more transistors every two years, you have to move very, very fast on both axes.
So changing tack a little bit, ASML has a long history at the advancement of lithography, but the way the company was founded is a little bit different to most. So I wonder if you can help explain.
Well, you know, this goes back 40 years ago, so I was still a very young kid. So I would tell you what I heard about it. But ASML is a spin-off from Philips, and the spin-off happened at a time where semiconductor was just happening. And if you look 40 years ago, in fact, a few companies doing semiconductor, I think IBM is a good example, were looking for complete integration of everything. So IBM for example was doing its own lithography machine and everything else. And the industry started to grow, and basically Philips understood that well, maybe it was an idea to provide a lithography tool to the industry. The Japanese Nikon and Canon had the same idea. So it started by saying okay, let's specialize on this. And then I think what was very important in the way it has grown, this has been a theme for the last 40 years, was to always do that in collaboration with customers, because customers were the only people who could tell us what to do, what we had to do, and with suppliers. And the second part is important because, to the contrary for example of the Japanese companies, we decided to go for system integration rather than vertical integration. And by choosing system integration, we were capable to work with the very best partner for every single part of the tool. So this is how we ended up working with Zeiss on the optics, and we've been together also for 40 years. VDL for example is a good example, very well known in the Netherlands. More recently, TR for EUV, more recently Mistral when it comes to AI. But every time we say we're going to just go and work with the best, and the job of ASML will be to understand from the customer what to do and to do it by putting all of that together. So this is really I would say the philosophy of the company, a philosophy that is still going very strong today as we speak.
It's interesting you bring up everybody trying to build their own lithography tool, because the way I also understand it is that at the time when ASML was formed, a lot of those competing companies also co-invested into ASML. How much of an impact do you think that had?
Well, you know, it was very important because like any company, I'm sure a lot of the people in the room will recognize that you can have the technology, but sometimes you cannot scale. And when we were developing EUV, the company was healthy because we had a good business on our DUV technology, but not wealthy enough in order to put the right amount of money in R&D. And this is where co-investment became important. And what we told our customers is: well, you're going to need this product, we are willing to do it, but if we do it with our own means, it will take too long. And on the way, we also integrated, for example, a company like Cymer that was critical to the success of EUV. So it's very important when you have something very important to be able to scale, and I think one good way to do that is make sure that the people who will use your technology contribute to that scaling. I think that's in fact a pretty good model.
But with those collaborations, how do you... Can you describe what goes into those collaborations when the roadmap for example for EUV is defined by decades and involves really hard science? I mean, for those of you who don't know the story of EUV, it was theorized in the 80s. People assumed it would be ready in the 90s and it took another 20 years beyond that. So can you go into what goes into those collaborations?
Yeah. So to be able to do what you describe, you need to be willing to share information, plans, sometimes dreams you have for the next five, ten, sometimes fifteen years, without knowing of course exactly what's going to happen because who knows what's really happen in 15 years from now. But to do that, you need to have a huge amount of trust with the people you're going to work with. So you need to know that the people you share the information with are not going to use this information against you to create leverage, to create products they may sell for too much money, or to create a technology they will cut you out from. And that's been I think one of the key elements of the success of ASML and the ability to look at the long term: to build this trust basically. And that's always what we put first in the company. There's a lot of things we're willing to compromise, maybe lose sometimes, but the one thing we never want to lose is the trust from our customer, from our partner. Because if we lose that, we know that the horizon is going to shrink dramatically, and then we won't be able to develop the next great product because, as you said, those products have a 10-15 year lead time, which for a lot of people is very scary, right? But for us, it's a bit what it takes indeed to bring materials, technology, software, algorithm together and get a product going. So we're about to move our High NA system into production. Our customer, we started a program I think in 2012, 2013.
Yeah, it's great that you say that and it's great that as a company ASML is in that position where you offer such highly differentiated technology that nobody else can realistically build without having all these collaborations. Does that not open yourselves up to entertaining a lot of bad ideas about how to solve some of these hard science problems? How do you determine what accelerates, what separates, for example, what may be a brilliant idea from what ultimately ends up being the successful one?
Yeah. So I think that first you have to really decide what you really need, right? So it goes back to what is the problem you're trying to solve. Just being creative for the purpose of being creative is not enough. You need to know what is the problem you're going to solve. And that usually calls for a lot of innovation, creativity, and I would say a lot of mistakes. So you have to be willing to learn by making mistakes, to learn by failing. I always say, I mean EUV is a great story. Of course today everyone looks at EUV as a miracle, right? Or something that came from the sky. But it took us most probably 20 years to develop EUV. And maybe the first 15, maybe the first even 18 years were just a succession of very painful mistakes, right? And for a long time, we couldn't find the solution to some key problem. One of them was how to get enough power on an EUV source. Until I would say the resilience of the engineers of ASML was such that we got some breakthroughs. But if you knew the amount of bad designs, disappointments, major disasters we had on the way, it was enormous. And the people just kept going. And you have to be willing very quickly to realize that maybe whatever you thought didn't work out and you have to go to a completely different approach. So you have to be able to do that and nothing else. This requires a lot of attention, a lot of focus. So there's not too much space for just random ideas. You need to stay focused on the problem you want to solve at the end.
So flip that a little bit because I know a lot of people in the room are on the other side of that. They're the ones building the ideas or investing in the deep technology that could be a solution for a bigger business for a multi-conglomerate. How much does that change when you're on the other side and what advice would you give?
I think it never changed. I think that the key for everyone in the room is again: what is the problem you are trying to solve? That's question number one. Question number two is: can you really identify people that really deal with those problems every day? Someone that will welcome you to solve that problem for them. That's what you will call customers. And if you have a good idea of what is the problem you want to solve and a very good idea of the people you want to solve it for, then start to work with those people and put all your energy on that problem. And anyone in this room at some point of time will have to do that because if they don't, whatever great idea they had mostly will go nowhere. So what makes the difference between really a great idea and an idea that becomes a business are those few key elements.
Does the best idea always win?
I think that no, because the other element we talked a bit about is scalability. There's always a point where you have to do things more than once, right? So for example for us, of course we can make one EUV machine, but the challenge is to make hundreds of them, ship them all over the world, and make sure that they run all the time. And in order to do that, you have to operationalize at some point what you do. You have to make sure supply chain is there. People realize nowadays how important is a supply chain. But I can tell you that most companies can get in real trouble if they don't take care of the supply chain. Supply chain is always important. And then there is a stage in the life of a company where the people who had the idea, made a prototype, where scientists realize they also need to understand a bit more on how operation works so that they can really scale the idea into a business. So operation is very, very important.
So moving on to the word of the year, I think this sort of AI and machine learning. I've noticed that over the especially the last couple of months, we're now starting to talk about it not just in a pure software industrial, we're talking about it more in a physical and this sort of infrastructure challenge that we have. How does that change for the people who are working with it and underneath it? And I'm pretty sure it's a lot of people in this room trying to adapt that technology to what they're doing. How does that compare to especially sort of pre-AI era?
Yeah. Well, I think for me one thing to realize is most probably still today, most probably also tomorrow, the most important part of AI is not the model, it's the data. The reason why AI came to life a few years ago was of course because people developed some models. But the only reason they could do that is because there was a huge amount of data that had been collected for many years around the world and that could become available. So when you think about AI, especially as an industry, as a company, as a startup, the first thing to think about is: what are the data you generate that are unique to yourself that you could use to feed potentially some of those powerful algorithms in order to develop things faster, in order to improve the performance of your product, in order in some way to boost everything you do today? And AI can really help, but you need to have the data, the information to feed it. If not, it's very difficult. And if you have that, you have to develop what I would call customized AI models that will be more adapted to what you do. And this is why I think LLMs were a first step, but they are mostly addressing what I will call the internet data set. People are talking more and more about agentic AI, about physical models, because this comes a lot closer to what companies, to what the industry is doing. So I think that transition is really a huge value. I still believe that the biggest value of AI will be for the industry, will be the ability to accelerate innovation, ability to improve performance, etc., on many, many different places. It's true for semiconductor, it's true for healthcare, it's true for education, it's true for energy, it's true for pretty much any vertical axis of technology.
So recently you've announced this partnership with Mistral. And I was lucky to attend a conference a few weeks ago where you went into some of the details on that. Now, Mistral is known for having open models and you just spoken about having the importance of having essentially a data set specific for the interface. So, can you go into a little bit of detail about how as a company ASML you're integrating those machine learning tools and what exactly is it improving?
Yeah. So you remember I talked about our model of integrating the best possible partner. I think Mistral fits that model very well. ASML is not an AI company. You know, we have good engineers who understand litho, optics, mechanics. But we don't have a DNA in developing models. We don't have even a DNA in understanding exactly how to make the best possible use of our data. Even as a company, we generate terabytes of data per minute if you look at what's happening around the world. So the idea of the partnership with Mistral was to bring a company that will be willing to work with us very deeply in order to use our data, help us creating an architecture around that, and then develop what I will call very customized models for our product, for our operations, for our R&D, etc. So that's a bit the model we put in place. And I think Mistral is indeed working, or even I will say supporting open source, because they also believe that the real value of AI is not just the model, it's the application you will drive out of the model. And the belief is if you keep open source, most people get whatever exists when it comes to the core of the model and can also spend more time on the customization.
