Please welcome KLA EVP, CFO, and Global Operations, Bren Higgins.
Good morning. Thank you for being here for our 2026 KLA Investor Day. It's great to be back in New York. I'm gonna make a few comments. First, I'm gonna walk through the agenda overall. I'll make a few comments, and then I'll transition to our President and CEO, Rick Wallace, who will talk about compounding sustainable outperformance of the company, where we've been and where we're going, some of the dynamics that are driving the ecosystem and how that plays through to opportunities for KLA relevance. Ahmad Khan, who's the President of our Semiconductor Products and Customers business, will then stand up and talk about process control in the AI era, some of the dynamics that are driving our business, how we're collaborating and engaging with customers that drives our innovation model, and ultimately, how we execute.
Our strategy is to take advantage of what looks like a very exciting business environment moving forward. We'll take a break, and then Brian Lorig, who runs our service business, will talk about service. We're excited with some of the dynamics, despite some of the export control effects on service, about what's happening in service in terms of how our service business and how our customers are relying on it to drive higher performance out of the install base and higher levels of availability. After Brian, I will talk, come back up on stage and talk about our path to 2030. What's happening, what we've done over the last several years, and what's happening in the market that then translates into how we're operating the company, what that means in terms of financial performance as we move forward.
I'll walk through a 2030 model based on the various assumptions that get us there. At the end of the day, there's always those unknowns and questions. We're drawing some lines over time, but I think what matters is the story that we have around the ability for KLA to drive our relevancy and execute the business model over time. We have a credible history of execution, and so as we talk about that, I think our history is indicative of how we're gonna run the company moving forward. We'll then transition to a to a Q&A session, roughly 45 minutes, and then after that, we'll go to lunch, and we'll do some round robin with the executive team rotating around the various tables to to answer some additional questions.
Pretty big program here, 9:00 A.M. to roughly 2:00 P.M. today, and we hope you can all stay with us. You see the mention of the AR/VR demos. If you haven't had a chance to do those already, we also have some things we brought in terms of show and tell. As you go through the day, if you haven't had a chance, I'd encourage you to take a look at some of what we have here. The AR/VR demos you get to get up close and personal with a couple of our very high-end systems, a Gen5 wafer inspector, and our new Teron beam 8xx multi-column electron beam inspector for reticle inspection. Get a sense for the complexity scale of what we do at KLA. I'm gonna let everyone read this, so I'll take 10 minutes maybe.
Now, look, we're gonna make forward-looking statements today. Those statements are subject to risk. We have an exhaustive list of risk factors that are in our SEC filings. I would encourage you to take a look at them. You can find them at our website at kla.com. As it relates to the March quarter, and we did a press release this morning, we affirmed our guidance for the March quarter. Revenue at $3.35 billion, non-GAAP diluted EPS of $9.08. Got three weeks to go, more or less quarter's going as expected. No real update to the March quarter, and we're comfortable with the guidance that we provided at earnings last January. A couple comments about the industry.
If you look at where the industry is today, we continue to see very strong momentum across all segments, foundry logic, memory, and advanced packaging, that the wafer equipment market, including advanced packaging, is now expected to be in the range of $135 billion-$140 billion. We talked about a mid-30s profile. We think that the momentum that we're seeing, particularly as it relates to what's happening in the second half, is driving incrementally up that view. Now, the other thing that's pretty clear today as we engage with customers is that the visibility for 2027 is very high. Our customers are building new facilities, and so there's new equipment demand that is tied to those facilities and the schedules for that construction.
Our visibility into 2027, to have this level of visibility 12-24 months out is really uncommon, but gives us a lot of comfort to be able to stand up here in March 2026 and talk about 2027 and having a growth rate in 2027 that we think is at the similar level or perhaps even higher than what we're seeing in 2026. That's the growth rate, year-over-year growth rate of the industry investment in wafer equipment in 2027 versus 2026. As it relates to KLA, we expect to continue to see quarter-to-quarter sequential growth as we have for the last several quarters. Our views on 2026 are strengthening based on our visibility into the second half for semi-cap systems.
Our view for 2026 now is that we expect the total company to be up somewhere in the high teens in terms of year-over-year growth versus 2025. Given that's the total company, our semi-process control systems growth should be faster than that. You know, we'll look at, you know, couple points more or less faster than that. We'll see how the year plays out. Half-to-half dynamics.
I'm sure we'll talk about more of this in the Q&A, half to half dynamics. I don't really want to get into it at this point. Let's finish the March quarter. We'll talk about June, and then we'll give you some sense on what that looks like. At least in terms of the overall year, we think the year translates into high teens growth today. Made two announcements around capital allocation today, very consistent with our history at KLA. Our 17th consecutive annual dividend increase, which is a 21% increase to $2.30 per quarter from $1.90. An additional share repurchase authorization of $7 billion on top of the $3.9 billion that was on our remaining authorization as of the end of December.
Roughly $11 billion in authorization to support what we expect to be a pretty robust environment over the next couple of years in terms of cash flow generation. We'll generate the cash. As everyone generally knows about KLA, we're big believers in assertive capital allocation and allocating every dollar of cash. Our share repurchase programs are consistent over time, and this is in support of that and our outlook moving forward. I've never been as positive about this industry as I am right now. Let me stop. Thank you for being here today, and welcome to the KLA 2026 Investor Day. I've never been as positive about this industry as I am right now.
AI is a driver and an enabler for KLA. This is not a small inflection. It is a massive transformation. We couldn't design a better market than if we designed it ourselves. There's no question that there's this incredible demand associated with the build-out of the infrastructure to support AI. We have seen growth in logic, and we have seen massive growth in memory, and we have seen massive growth in packaging. All three segments are growing pretty significantly. Monolithic chips don't really solve AI problems.
Y ou need to integrate these AI chips with memory and other parts of chip design to ensure that the full system is working. The die size are bigger, the chips are bigger. They're more valuable. The defect challenges are more significant. All those are uniquely benefit KLA. From an enabler perspective, it's both how it supports our products, which makes our products more efficient and more economical for our customers. We started over 10 years ago looking at how we would leverage this technology in our systems. We are looking for subtle defects on the wafer. These defects are very, very hard to detect. AI, through training and inference, is able to detect these defects in ways that classic algorithms, rule-based algorithms are not able to detect.
KLA is positioned incredibly well to both support the demand and also improve all of our processes by leveraging it. Which is why our prominence with a lot of customers has increased because they recognize the unique role that we play in this incredible technology transition.
Please welcome President and CEO, Rick Wallace.
November 30th, 2022. Do you guys remember what happened on November 30th, 2022? I was in a review at KLA with our Chief AI Officer, his name Kris Bhaskar, and he said that week something had happened that had only happened twice in his 40-year career that was as profound. That was the week ChatGPT dropped. His view was that was gonna change the industry in ways that we couldn't imagine. He was right. We couldn't imagine at that time. I had not heard of ChatGPT, but I called my kids at the end of the week, college kids, and said, "Have you guys used it?" All three had. It got 1 million downloads in the first week. Let me ask, how many of you use Chat or some other chatbot on a routine basis? Okay. How many of you have created an agent?
How many? Only a couple. Three. My prediction is that by the end of the year, you all will have created an agent. It's not very hard. You can do it in half an hour. You can create an agent. What will the agent do? It'll do two major things. It'll do a lot more work for you, and it'll also drive a ton of demand for compute. ChatGPT drives one 10X what an internet search does, and reasoning and agents drive another 10X. Part of the build-out that we're seeing around AI is the fact that so many people are starting to leverage these tools, and that's what's driving consumption. What we didn't understand at the time, it was this interface to chat that was gonna drive the demand. What you're gonna hear today is how well-positioned KLA is for this inflection.
We said in the video, Bren said if we had to design a market, this would be it. We're gonna talk about our customer relationships and how that positions us, and we'll talk about our technology and the investments that we've made, some of which you can see in the room. We thought we'd bring these to you because the times where we've had investors come on-site and see the hardware, it's made quite an impression. One example over here is that's a lens in a BBP system right there. That's a lens. When I started in the industry, it was an off-the-shelf lens. That's not off the shelf. Who are we? We're founded 50 years ago. This is our 50th anniversary, and the original idea for the company was to automate what was done manually.
Ken Levy automated photomask inspection because it was being done by microscopes. We started in the industry doing that, but we've expanded, of course, to a large portfolio, and we're now the leaders in process control. We'll talk about what that means and why that's important in the age of AI. The leadership team. What you really need to know, you're gonna hear from Ahmad and Brian, as you heard, and you, Brian will come back. What you need to know as an investor about this leadership team is this is the leadership team that committed to our 2019 Investor Day and hit it. This is the leadership that committed to the 2022 Investor Day and hit it. This is the leadership team that's committing to the 2030 plan we're gonna share with you today. We've been through it.
We're very excited about the prospects that we have, and we're gonna show our work today on why that's the case. I know you know this, but I wanna add a little context to this hierarchy. You know, semiconductors is in service of overall electronics. Historically, electronics have grown GDP-ish. Not now, not with these data centers. It's significantly above that. A lot of customers or investors will ask, "What's the bottleneck now?" I'll show and we'll share, ask me in three months, ask me three months ago. It's moving. The bottleneck is moving. But one thing that is clear is there aren't enough fabs. When we talk to our customers, there are definitely not enough fabs. There's not even enough fabs contemplated in what we'll share for 2030 to meet all the stated demand out there.
If more of you start doing agents, that's gonna drive up the demand for compute radically. We know that's gonna happen. It's already happening. The agent's probably 100x the compute demand of the original search. We're pretty excited about this, and we're gonna make the case that none of this happens without process control. Without process control, ChatGPT would not have dropped, and certainly we wouldn't be where we are today. I'll talk a little bit. Of course, you know the landscape, but I wanna give you a specific perspective on this landscape. It is more about who designed the chips than the overall revenue, and why that's so profound for KLA and for the industry. Up until we started seeing mobile and smartphone, there were a couple companies that designed it. They needed their own silicon. Right?
One in Cupertino, one in Korea, that they wanted their own silicon for their phones because they could better optimize. That happened, you know, mid-early 2004, 2005, 2006, 2007 in there, and they started doing design for silicon. We thought that was interesting, but it wasn't really a profound difference. It just shifted who was doing it, who was making it. There weren't that many more designs. The real inflection for our business and for this business, if I go back, it's really what happened when things went to the cloud. More so, that was the beginning of what we're seeing now, when it went to the cloud, because it changed the economics of who wanted to drive the design of semiconductors.
We saw the first company, and AWS was one of the first ones to acquire design capability to build their own chips, and we're like, "Why are they doing that?" If you fast-forward it today, there's so many more players designing their own silicon. It went from everybody buys the same microprocessor to people start designing their own. That flipped the dynamic, and we'll show you the design starts go up. There's more players in it because the economic advantage for these players, it's now viewed as their differentiation. Yesterday, there was another hyperscaler that announced four new designs. Yesterday. That was on 3 nm. There's over 100 designs on 3 nm now. This dynamic has changed significantly because it's in the economic interest of the cloud provider and the hyperscalers to have an advantage in their own compute stack.
Originally, they're buying from the same players, and, you know, we'll show you some data about how it impacted KLA and the transition we made. At the same time comes along new ways to do compute. We've been locked in the CPU construct for a long time of how to do compute, right? There were a lot of people that said that was not the most efficient way to do compute, but there were so many barriers to make that transition. I'll share our example. This KLA system, and what we really ship is a very high-end microscope with lenses like that and sensors like that attached to a very high-performance compute network or a compute data center, if you will, that sits. Now this is an example of one of our products, an edge server, or it's really high performance.
We saw cost was one, and in this case, there are many examples like this in the company. It was literally $1 million of cost to us. That's the cost of our compute inside of a system. We're going to the next generation trying to figure out, do we stay on CPU architecture or do we transition to GPU? We had already had experience, you know, with GPUs and programming them. In fact, we had AI algorithms as early as 2018 in our products, but they were CPU-based. We hadn't made the transition. We wanted to, but what was the big barrier? It was the massive amount of rewrite of code. When we looked at staying on CPUs, we knew. Yes, we'd have minimal code change, but the cost curve was gonna go up pretty dramatically and the power consumption.
Once we bit the bullet and decided to go to the CPU-based architecture, and we did it for the algorithms. We did it for the algorithms, but look what's happened. The cost only goes up from $1 million to $1.2 million, and the power consumption actually goes down. Power goes down. If you're a hyperscaler and you're paying your own power bill, that going down, that's enough just to do that. If you're now also hosting models to run. Our view of this was like, there's no going back. It's way more efficient. By the way, GPUs, because they're parallel process, and Ahmad Khan talked to this in detail, they're a better way to do compute, but they're not the most efficient way to do compute.
There's custom silicon being designed for specific tasks that's even more efficient than GPUs, which is why there's so many more designs happening. GPU is better, custom better yet. What's interesting about this transition is once we make this transition as a company, we're never going back. We leverage this across every one of our product platforms, the savings, and now we're on a different cost curve and a different performance curve. Huge implications for us. Now, the challenges, you all know, first of all, one of the challenges is these GPUs are very large die. Why does that matter? It really matters to our customers because the yield challenges are much higher on large die. The process control requirements are much higher, and we'll share the intensity that goes with that.
The other thing is, if you've got one of these very high-powered processors, you've got to feed it data. Suddenly, the memory out of nowhere becomes a critical component. I also mentioned the shortages. Right now, it's a memory shortage. High bandwidth memory is a shortage, right? That is the thing that's a shortage. Because you gotta feed all of these processors to keep them active, 'cause you paid so much for the processor, you gotta feed it data. That's changed the device of a memory device significantly, the first change in years, high bandwidth memory. It all favors the need for more process control because of the size of the memory that's going up, the complexity, the lack of redundancy. For those of you that think KLA is not a memory player, it's not true. That was true.
Actually, if you go back far enough, it wasn't true. We grew up on memory, and then it got less true as the memory guys figured out how to do redundancy, and now it's back in a big way. We'll share why the intensity in memory manufacturers has gone up so much. The other thing that's amazing about packaging is we got into the packaging business, and Ahmad Khan show our roadmap, years ago, thinking more than Moore. Remember that? We were like. We did our first acquisition in this in 2008. Yeah, we were early. Boy, we're glad we're there now because this advanced packaging looks like a semiconductor process, which is why there's some confusion about capital investment, capital spend as it pertains to front-end and packaging, and we'll try to clear up some of that. I mentioned this. This is crazy.
This number of people doing advanced designs at the leading. This has flipped on its head. Many of you, if we'd done this Investor Day 10 years ago, you would have said that every node, there's gonna be fewer players at the leading edge, and only a few people can afford it, and therefore, the number of designs was gonna go down. A couple things happened. One, scaling resumed, and then the hyperscalers enter. Those two factors, you get scaling and you get more players who determine that silicon is strategic for them. Now we see, this is just within the first year, I just mentioned there are a bunch of new designs announced yesterday for 3 nm. Yesterday. You see the demand for 2 nm going up because 2 nm are more efficient node.
What does this mean for process control? Well, if you have a lot of designs going through a factory and you have a lot of process flows, because if I'm a hyperscaler and I want my custom silicon, I want my custom process flow too. I don't wanna buy what everybody else buys. I want the thing that I think is best. Well, if I'm providing that capability, I gotta make sure the variations are low. That's what's driven up process control. Look, this industry is all about economics. Nobody does anything if there's not an economic return, and the economic return on managing these complex fabs are now enormous, especially when there are shortages. What's the role of process control?
Well, process control is a funny term actually because, confession, I started as a controls engineer in the paper industry. Okay? It's nothing like this industry. I mean, it's super exciting and all, but nothing like this industry. But it was just a simple closed-loop control. Why is the semiconductor industry so different? Because of constant change. You can't get something under control 'cause it's constantly changing. So we talk about what do you do in process control. We originally, as a company, thought it was about finding defects and assuring quality. It is, but it's far more than that. We learned this, it was probably in the 1990s.
We learned this when there was a big study on competitiveness in semiconductor manufacturing, and the conclusion was, because every semiconductor manufacturer essentially buys the same equipment, the competitive advantage is in learning rate, how fast they learn. Their ability to learn and improve the process. Think about this, you're building a $15 billion, $20 billion fab, and you start at zero yield. KLA is about learning faster, and learning faster is the only sustainable advantage in semiconductor manufacturing. It's the only one. I could argue it's the only one in tech, is the ability to learn faster. What our systems are used for. Years ago, we had people think it was just for quality control or final check or whether we're a measurement company. No, it's about learning.
We'll show you the challenges of when people bring out this new process. They adopt more capability, and they ask for our expertise because they've got to learn how to ramp that new process up because the stakes are so high. What can go wrong in a semiconductor manufacturing process? There are defects everywhere, and they're not stable, and they change all the time. The beauty for us is nothing is stable in a fab. You could run along with a process that seems to be working, and there's a hiccup because some of the material, starting material comes in and it's not what it's supposed to be. We've had customers that have suddenly had an experience because something changed in all these processes. Then you've got to make all these measurements.
The history of process control is you put a process in statistical control, and you let it run in control. I can tell you this industry is never in control. It just isn't. That's not true. If you make analog wafers, you can be in control. Truly, they don't change that much. The other thing that's interesting, there's a couple of areas we inspect everything in, and one of them is a reticle, and I'll show you what a reticle is. This is a mask. This is a template to print, and Ahmad's going to talk about this more, to print one of the many layers in a semiconductor process. They call it lithography because you're doing a lithograph, right? When I started in the industry, I know, long time ago, $2,000 for a mask.
This little thing, this was a Saran wrap on top of it, that was the pellicle. Why did that exist? It's 'cause if a particle fell, it would be out of focus. That's what a pellicle was about. I used to accidentally break them all the time. My first job in a fab was doing inspection on these reticles that were $2,000. This reticle today for EUV, $200,000. I need sets of them. That gets inspected a lot because if there's a defect on this, and by the way, this is what we'll call a single die reticle because that means there's only one of these, and it's printing. It reduces when it prints, but it's printing, and if there's a defect on there, it prints on everything. You can't have that.
That gets inspected 100%. Almost nothing else does. Everything else, when we talk about process control intensity, everything else is sampling. We have 1,600 apps engineers that work with our customers determining their strategy for sampling, for sampling and inspection. I was apps number seven at KLA, so we've ramped a bit. The reason is 'cause we've got to help customers figure out where to inspect and where to measure, and it's dynamic. We get a lot of questions about process control intensity and how it's changed, and it has changed. I gave you my example of analog. The way to interpret this graph is from the left to the right is the relative % of process control intensity in those different markets. You can see analog is very low.
Now, the line is the amount that was spent last year on equipment. It's low process control intensity, but they didn't spend much 'cause not much changes. Far right, almost half of a spend for mask shop is process control. Almost half. The reason it's gone up since 2018 because the cost of the mask, the number of single buy, everything has driven that number up. The three things that are in the middle of this, DRAM, WLP, advanced foundry, that's all about AI. The reason those have all gone up of process control intensity is 'cause of the challenges, and we'll share more about those with AI, and Ahmad's gonna go into more detail. I think it's important to understand what drives process control intensity.
It's a combination of the economics associated with failure, as in a mask, and the economics associated with managing the complexity. We have multiple offerings. One of our strengths as a company is we have a portfolio, and the reason this matters to our customers is because we don't go in with a single solution and say, "This is what you need." This graphic of E-beam to BBP to laser scanning. Look, the throughput difference between BBP and E-beam is like 1,000 x, so it's not a linear graph. There are specific things that you can only find with E-beam, and there are specific things that you can use laser scanning for, and our customers are always gonna look for the most cost-effective portfolio answer. On top of all these inspections, there's the measurement.
The benefit that we get as a company from having all these things is we can invest in common core technologies that get reused. Ahmad's gonna show an example of why we've been able to make such an impact in E-beam, leveraging a lot of technology we already have in the company. The other competitive advantage we have in addition to our portfolio is our huge installed base. I'm saying that because we learn a lot about what to do for our customers, and it turns out now customers buy our equipment, and Brian's gonna share our installed base support for customers. It's a great revenue stream, but it's a competitive advantage, and it creates a lot of value for our customers. The other thing that's happened, and this is kind of fascinating, is the life of an average tool has gone up dramatically.
Again, Brian's gonna share that with you. We're a measurement inspection company. How do we measure success? We have an operating model we talked about in the video. You'll hear more about it, of how we run the place. For your purposes as investors, what do we focus our team on? What do the four of us when we're in reviews look at? We spend a lot of time on intensity. What is our relevance? What is our share of wallet? And we do that because we know if we create more value for our customers, they're gonna buy more of our equipment. It's straight economics for our customers. If they get leverage out of buying KLA, they're gonna buy KLA. The other thing we look at and focus on heavily is market share. We constrain market share by gross margin.
You can have high market share, but if you're buying the market, that's not great as an investment. We spend a lot of time looking at what's our gross margin and how efficient are we and what's our operating margin. Then do we have the talent? Can we find the people to continue to turn this crank? We have many product divisions, and so we need to hire a lot of people and get them into the KLA system. How have we done on intensity? Bren will show some other cuts, 'cause I know of you who like to, you know, pick your start date, and you can get a different story. 2019 till now, process control's grown about 16%, lithography has grown similar, and we've outgrown it because we've gained share.
We'll talk more about our share and our approach to share gain, and we look at the overall totality of process control. Every one of our divisions, we look at what are they doing in terms of driving their position competitively to drive process control. This isn't, this requires a big R&D investment. If you think about what we invest as a company relative to our closest peers, typically, we're investing more than their revenue, just in R&D. We're investing in new platforms, new capabilities, and then common things that we share across. We view this as going, "This pretty much goes up every year." This isn't counting the 1,600 apps engineers that we have. This is just the R&D. What are the areas of focus that we have? We look at all these areas.
Just think about a microscope. Remember high school physics. You have a microscope, you got illumination source, you have a stage that moves things around, you have an objective, you have your eyes. Every one of those things is a subsystem. When I joined the company, one of those things was custom for KLA. The only thing that was custom for KLA back then was the image processing. Everything else was off the shelf. Everything now is custom. We have custom development in every single area. Our competitors do not. They can't afford it. We get the leverage out of those investments across here, which allows us to provide. We don't do it to win competitively. We do it to serve the market, but it sure does help competitively. This is one of our show and tell.
That thing over there, as I mentioned, is a lens that's used in the 39xx. It weighs 300 kilos, which I'm told is a lot for a lens. The lens on the 2020, as I said, would have fit in my pocket, right? One of the biggest changes we made was when we started doing custom lenses was about 1999, was the 23xx. This was we realized, and models show why we need the spectrum of capabilities to provide it, but every one of these generations, we provide more capability. There are those, I know none of you, who years ago said optics was gonna run out of gas and e-beam was gonna take over. This is, you know, Gen5. Last year, we sold more Gen4 than Gen5. Optical is not out of gas.
The reason is 'cause we keep innovating. We keep driving new capabilities, and our customers want the most cost-effective way to solve their problem. You'll get to look at that closely. It's an engineering marvel. We own the design. There's one place in the world that can make those things, and we're pretty proud of what it allows us to do. How have we done in share? We're up 2 points from 2021, and we're now 6.5 x our nearest competitor in process control. People look at the ability to defend markets. Back to my R&D, it's a pretty big spread. It was 4x seven years ago. We invest heavily, we focus heavily on market share, and we also focus on growth of service.
Brian's gonna share his story, but that's a pretty good chart for service revenue. We'll talk about why that's the case and why it's even gonna grow at a faster rate going forward. Our installed base is also a massive competitive advantage. Not just the revenue from service, but the fact that we have that to leverage. We spend a lot of time on gross margin. Bren and I and Ahmad and Brian will ask if general managers come in and say they're differentiated, we'll say, "Show us your gross margin." If you are differentiated, you'll have a better gross margin. We'll do acquisitions, and somebody will say, "We have a really great product." And we go, "Well, really? 'Cause your margin's not that good. So is it your cost too high? Are you not pricing? What is it?" And we'll drive those up over time.
Bren will show we absorb a company with lower margins, both gross margin and operational, and we've been able to improve it using the KLA operating model. We're proud of our operational excellence. Bren talked about our cash generation. Those are things we're measured on. You look at management incentive, it's on relative free cash flow. We focus pretty heavily on operating performance. Of course, we got to get the talent to make this happen. Our job got easier in the last few years, I'll be honest, because for a while, nobody cared much about the semiconductor industry. We're back. We're able to hire great people, and when we ask them how they feel about being part of KLA, they're proud. They can't explain to their family what they do, but they're proud. Where are we now? AI is clearly driving the industry.
The outsized growth that we're seeing is because of AI. There's no way to do it without process control. You're going to hear more about this. Here's a point to take away. KLA wins no matter who wins. Doesn't matter which hyperscaler wins, doesn't matter which semiconductor company wins, doesn't matter which processor wins. It doesn't happen without process control. Process control is going to outgrow the market, and we're going to outgrow process control. Here it is. The assumption for 2030. We see semi growth, and Bren is going to double click or triple click on this. 11%, because, look, if the semi industry grows at that rate, it'll still massively underserve all the data centers that have been announced, which is well. Capital intensity is going to grow faster.
We're showing $215 billion, and this includes what you think of as WFE and WLP packaging and semi, ± $20 billion, big range. Process control is going to outgrow it, and we're going to gain share. Service, we're upping our growth expectations of service based on what we've seen and what's in the pipeline. What does that get us to? $26 billion in 2030 revenue, ± 2.5, $84 a share, ±8. We've done this before. We've signed up to what looked like really big goals before. We're generally pretty conservative. If the spend. Like, people will say this, don't ever say, this time's different. I'm not saying this time is different. This time is just much bigger. This build-out is much bigger than anything we've seen before. This data center build.
There's just so many reasons to drive this build-out, and that's what drives our model. The rest of the day, we're going to spend more time going through and explaining, showing our work on how we get to this. The first person that gets to do that is Ahmad Khan, who's done an outstanding job running our product divisions and who's gonna share deep dive on his businesses. Ahmad.
Good morning, everyone. My name is Ahmad Khan. I run the systems business for KLA and also represent the customer, meaning that I engineer the products, and then I have to make sure that our customers take them. It is a humbling process. If the customer doesn't take them, I have nobody else to blame. It's a good org design. Today I'm going to talk about how the compute is changing process control intensity at our customers. That's the bulk of my presentation. I'll give you several examples of what is changing in the industry on chip design and how our customers are yielding them and why they're using more process control.
To set the goals earlier, as early as possible, inflections, share gains, and portfolio expansions, all three of them are going to drive share of WFE for KLA higher than it was today. Today, it is around 7.4%, and by 2030, we're assuming that it's going to get to around 9%. Share of WFE is going up. We work very closely with our customers, very strong customer collaborations. Why is that important? It is best for your customer to vote what we should develop, and therefore they are bought into buying that system to solve their problems. That collaboration is very critical to us, and we work very hard to maintain that customer trust. We're focusing on innovations in hardware and also in AI. I bring that topic up because in AI, you can create models easily and replicate somebody else's performance.
That is not the case for KLA systems. We build very complex systems to reach to very pure signals. That pure signal, coupled with complex models, leads to defect detection and measurements that solves our customers' problems. All of this will drive growth. Calendar 2025-2030, wafer equipment is going to grow between 11%-13%, and the systems business would be around 15%-17%. What do we do? Rick said it's really about learning rate. It is correct. We do detection of defects and measurements of parameters, critical parameters. We collect lots of data at the customer fab, but our customers have insatiable demand for data because the output of process systems is titanium nitride or tungsten or something else. The output of KLA systems is ones and zeros. It's just information.
That information, coupled with the right people, helps solve customer problems. That enables customers to yield and make more profit. Now, we participate in all segments of semiconductor manufacturing, from infrastructure, which is reticle manufacturing, wafer manufacturing, to logic, DRAM, specialty, and now advanced wafer-level packaging, and also component inspection. Finally, when the chip goes onto the PCB board, we also inspect and write the PCB board. We are in the entire supply chain of semiconductor, gives us a lot of insight on what is happening in the market because we look at each of the segments and how they're doing. We are a portfolio company. When we go to a customer, we don't have a single product that we are positioning with the customer to solve their mission-critical problems. We're a portfolio company. We go in, and Rick mentioned the 1,500 applications engineers.
We send those applications engineers to the customers to determine what is the problem in the line. Then we use the right system with the best cost of ownership and performance to solve the customer problems. It's a full portfolio company, inspection, metrology, data analytics. These boxes are only a subset of what we make. It's much larger, but the font would be very small. We use the collaboration, innovation, and execution method as our flywheel to drive performance with our customers. Collaboration is very critical because, as I said earlier, it is best for the customer to tell us what they want us to work on. Then we innovate, and execution delivers the revenue. You will see this being used across my presentation. Now let's look at the systems business in numbers. Calendar 2025, $9.8 billion in revenue.
$9 billion of the $9.8 billion came from the process control segment, $0.8 billion from the rest. 19% outperformance from calendar 2019 to 2025. The market grew 14% in the same period. Outperformance is a general thing that we work on. Every product group, every product division is focused on how to drive intensity and outperformance in the market. 62.8% overall KLA gross margin. Gross margin is an indicator that we have a differentiated product. R&D is very important. We really focus on reinvestment back in the business to ensure the problems of five to seven years from now will be solved. The one thing to remember is when a customer wants to develop a new method, let's say a vertical channel transistor, they have a pretty good understanding of what dopant materials you're gonna use.
They can go to other customers and say, other suppliers and say, "I need this dopant material, and I need this etch capability," well before it goes into production. Our customers do not design defects. They have no plan to design defects. Their plan is to have a defect-free process. It isn't. We have to think and develop systems that have wide enough capability that when the process comes to high volume manufacturing, that we are able to do that detection and solve the customer problem. We don't have four years after the process shows up in HVM. That goes into our design philosophy.
How do we predict what will the customer do and how it will lead to a failure that we need to be able to detect and design today, which will take about four years and come out in the marketplace? This is why we are a portfolio company. We have many, many instruments, and each instrument has thousands of modes to go fine-tune the system to find that defect. This is why reinvestment back into the business is a very critical aspect of our methods. There are two numbers that I look at constantly and all of my product general managers look at very, very closely. One of them is market share, and the second one is process control intensity, which is the amount of process control that our customer is going to use to yield the device.
Let's double-click on those two numbers. This slide has two lines. One of them is the annualized process control intensity. On the top, you have WFE or wafer fab equipment market. At the bottom, you have two lines. One is the annualized, and the second is the three-year rolling average. You can see the process control intensity was around about 5.3% back when WFE was around $60 billion. Fairly flat, but something happened in calendar 2021 onwards. Generally, if you back up, the main inflection that changed is customers started to print smaller and smaller devices because of EUV, high-NA EUV, many other things. This drives large variability in the fab, and I'm gonna describe that in detail later. Therefore, the amount of process control customers needed to yield the devices started to go up.
You can see a clear trend, 5.8%, 6.4%, 6.7%, 7.1%, absolute number, 7.4% intensity. We think the trend continues because I think we can all agree that complexity of semiconductor devices will go up, not down in the next five to 10 years. Large dies, high variability, all of those factors are driving process control intensity to go up. This is why we feel that now in the 2020-2030 period, this number goes from 7.4% to 9%. The second thing is, in a growing market, we want to continue to grow share. We made a commitment in the 2019 Investor Day that we would grow share by 0.5% year-over-year, and I'm happy to say that we've been doing that.
We're now at 56.5% share, up 2% since 2021. When I was here last time, I said we were 4.5x versus the nearest competitor. We are now 6.5x versus the nearest competitor, meaning the next competitor is below 10% in share. In the recent past, we have gained share in optical, in electron beam. Electron beam is a new segment for KLA. The last time I was presenting to you, the E-beam market for KLA was between $75 million-$100 million revenue. Today, we are at $400 million. I'm going to describe why we got into E-beam and why it's very important for future, and how we are leveraging it for our entire portfolio. We gained in advanced packaging.
About four years ago, we were less than 10% share in advanced packaging. We are now in number one position in advanced packaging. Mask inspection. Rick spoke about reticle inspection a few minutes ago, and we're gaining share in that segment as well. Those two numbers are very critical, but it's not just me and the senior staff that look at this number. Every product group in my organization is driving those metrics. I'll give you a few examples. Reticle inspection. One unique thing about reticle inspection, different from wafer inspection. In wafer inspection, you're generally looking for signature of the defects. There are some defects here and some defects there. If you catch 25%-30% of the defects, it is good enough for a process engineer to now make changes to improve the defectivity level.
That is not the case for mask inspection. We have to detect 100% of the defects. 99% is not good enough because that one killer defect will print on 100% of the dies, and your yield would be low, zero, if it's a killer defect. It's very critical that you detect 100% of the defects. Our method to detect 100% of the defects is to use a suite of systems for best cost of ownership and performance. If you bring a single system to do this, the chances you detect 100% is low. That drives intensity up. Unpatterned inspection, very important. As you know, about four or five years ago, a semiconductor wafer fab with equal real estate was producing about 50,000 wafer starts per month in logic.
That same real estate now only produces about 25,000 wafer starts per month. What happened? The number of process steps to make the advanced nodes has gone up significantly. In order to have more process steps, you need more process tools. Those process tools now generate just as much defects, and therefore unpatterned inspection qualifies all those process tools to ensure they are clean before you use them for high volume manufacturing. Unpatterned has grown significantly for us. Optical inspection. We used to use a single system for a single layer. That is no longer the case. We have many layers where the types of defects on a particular layer is three to four types.
Therefore, our customers will use Gen4, then they will use Gen5, then you'll use Gen5 again, and then they'll use electron beam system all on that same wafer to detect all the different types of defects. This is driving intensity. E-beam. E-beam we got into for two reasons. One, for buried defects. We were not in E-beam prior because most of the architectures were planar. Now you have a 3D architecture, so you have subtle defects buried in. We're able to detect most of them with Gen4 because it is a longer wavelength and we can penetrate, but electron beam helps as well. At the same time, because we have so many Gen4 and Gen5 systems, an electron beam system provides intelligence for the setup of our Gen4 and Gen5 system. I'm gonna go through that in detail.
This enabled us to grow from $75 million to $100 million in revenue, to now $400 million in revenue, all driving intensity and share. I'll pick on one other example, which is data analytics. All of these systems, data systems, are generating petabytes of data every single day. Our customers can use that petabytes of data for an instant or do analysis when their customers come back with reliability failures. We now have a suite of software and data analytics systems that we provide to almost all customers for advanced packaging and for front-end manufacturing, where customers can do overlay analysis and CD analysis across the board. This business now is $250 million software business that is doing quite well. So all of this is driving intensity and share. Okay.
Going back to collaboration, innovation, execution, I'm going to talk about first the collaboration and what is happening in logic. AI-based inflections are really driving a lot of change in logic. The logic from five years ago is very different than the logic of today. What is changing in the logic segment? Prior to going into the technical details, I will quickly explain what is happening in the market. The first initial build-out for AI was in training. That is in its late stage now, fully implemented. The second phase is inference, and I think most of our customers are experiencing the inference world now as token usage continues to go up on a weekly basis, and our customers are able to generate revenue from that.
After that, there is deep reasoning, and as Rick talked in elaborate detail about agentic AI and autonomous work, each of these phases has multiplicative effect in growth of semiconductors. We see it in our own systems, and we see it in the marketplace. This is going to continue to drive growth. Now, this growth is affecting all chip designs across the semiconductors. It is affecting logic in CPU, which CPUs are primarily now used for complex compute and data orchestration. GPU, which is very, very good at single-mode processing, and we use it across the board in our systems. High bandwidth memory. HBM is very critical because you have to feed these GPUs with millions and billions of transistors data. If you don't feed the GPUs billions of transistors the data, you will slow down the processing.
Memory went from traditional DRAM to high bandwidth memory so that you are able to load that data as fast as possible. DRAM growth, for reliability reasons and also for conversations, and we now see growth in eSSDs for conversational history reasons. All segments of semiconductor devices are growing at massive rates. If we look at and we have examples later on, you can see here is a full board of Blackwell mixed with other devices. In order to make full system, you need a couple of large GPUs surrounded by several HBM dies, and then you take all of that and integrate it on an interposer and then on a PCB board. This is advanced packaging and then PCB. This is logic, this is memory. All of that comes together to make the device.
Let's talk about leading-edge logic, then leading-edge memory, and then I'll talk about advanced packaging and how KLA plays in each of these areas. What is happening in logic? The first thing is there is a term that I'm trying to introduce, which is called variability. Most fabs deeply understand what variability is, and I'll give you a very detailed description of it later. But variability is inversely proportional to the printed feature size, and this is why I brought up that 5.6% intensity going to 9% intensity. As people started to print smaller and smaller lines and spaces with EUV, the variability in the fab goes up. If variability in the fab goes up, you need more process control to bring that variability down or detect those defects. The details of this is gonna come later on in my presentation.
The second biggest inflection is chip size. If you have 600 chips on a wafer and a 1% yield loss, and you do nothing else but increase the chip size, and you have 60 chips per wafer, that equivalent process will have a 10% yield loss. Chip size and process complexity are governed by the size of the wafer, and the size of the wafer is fixed. You deposit the same amount of material on a large chip and a small chip, but process control is governed by the size of the chip. If you print very, very, very small chips, one killer defect will kill one chip. But if you print only two chips on a wafer and you have one killer defect, your yield is 50%. As chip size goes up, process control intensity goes up.
This is true for logic and also for DRAM. What is happening in the 3D architecture? This is a 3D channel transistor, and the key thing to remember is the channel, the worst channel, is gonna determine the performance of the transistor. Foundry customers are not allowed to bend their devices, meaning you have a high-performing transistor and you can sell it to customer A and low performing. It doesn't work. You have to have the same performance. KLA comes in to look for subtle defects with our Gen4 systems, and, you know, these are many different types of defects like silicide residue and our electron beam systems to ensure that this is defect-free and will perform well.
At the same time, we have metrology solutions ensuring that the doping engineering that is done on these channel transistors is done really well. Because if your doping engineering is not doing well, you're not able to provide multiple Vts to your customer, and therefore, the value of that product is far less. We have a suite of tools that are doing this. If you summarize all the things that are happening in logic, which is variability, large die, buried defects, we see significant growth in KLA process control. 30% increase in total inspection steps in logic. If you look at this color, the first one is the surface defects, the bread and butter of KLA. We are very, very good at doing that, and we've been doing it for years.
Buried defects were introduced, followed by electrical defects, and followed by new applications like Print Check. If you add EUV, you need to detect repeater defects, and we have new systems that do that. N minus two to N minus one, we see 30% increase in n-intensity, and then N minus one to N, another 30% increase. Incremental value versus N minus two for KLA is $1 billion in logic. I will now switch to DRAM and HBM. What is happening in memory? This first thing that is happening is the customers are trying to increase overall density inside DRAM. Why? Because they need more DRAM transistors. There's a couple of things you can do. You can add EUV, and the other thing you can do is reduce the capacitor size, which takes a lot of space.
As you reduce the capacitor size, the amount of electrons that are available to you to detect if a transistor is on or off is very low. If you have less electrons, you have to make very, very good logic circuitry inside DRAM to detect if the transistor is on or off. The main thing that is happening in DRAM as it pertains to KLA, that it is more or less now becoming like a logic device. It is not like a traditional DRAM device. That is what's driving process control intensity. Rick said, "Stop thinking that KLA's DRAM intensity is low." That's exactly the point. Why? Because it's becoming more like logic. As the capacitor sizes go up, customers have to design better logic to ensure that they're able to detect it when a transistor is on or off. To drive packing density, you're introducing EUV.
Some customers are now up to five to seven layers, others are one to two layers, but the trend is in that direction. With EUV, the great advantage of EUV is that it prints smaller lines and spaces, but it also prints smaller defects because the wavelength is short. If the wavelength is short, it'll print smaller wavelengths, but it'll also print smaller defects. Therefore, high sensitivity systems from KLA are needed to detect those defects. Also, new applications like EUV Print Check and mask requalification, all of those things happen because more EUV is happening. Large chips, DRAMs. DRAM and HBM chip sizes are increasing on a regular basis. The reason they're increasing is because you have to add more transistors, and because in order to do high bandwidth, you add TSVs from top all the way to the bottom that take a lot of space.
They're not very small, they're large. That is driving chip size. Eventually, less circuit redundancy. Customers used to have immense amount of redundancy in DRAM. One of the reasons is that the iPhone, generally, I mean, we are all sitting here, most of the iPhones are idle, and when you sleep, they are idle. The DRAM performance requirements are very different than a server class DRAM, which is performing 24/7. All of those redundancy things also have reduced pretty significantly. Now the details. I explained everything already. EUV prints repeater defects and also prints smaller defects, and therefore KLA needs to come in and detect those defects. The chip sizes in DRAM are going up pretty significantly.
This is a DDR chip and HBM3, HBM3E, not HBM4, HBM4E, even bigger, is 3x the size. Then this part, this is the non-memory content and the memory content. The memory content from DDR to HBM reduces pretty significantly, and the logic portion increases pretty significantly. Therefore, you have to ensure that this part of the chip, which is the same place where this is and the same place where this is, yields very well. This drives higher process control intensity because now you're essentially making a logic chip and you're calling it a memory chip. Therefore, overall intensity in memory has gone up significantly. KLA incremental intensity gain is $450 million in calendar 2025. That does not include the fact that market went up.
If you look at the same point differently, DRAM wafer equipment market went up by 2x. KLA DRAM revenue went up by 3x. The reason for that is all the things that I just said. Now we have a logic wafer. We've spent a bunch of money making sure that wafer yields. We have a DRAM wafer, and we've spent a bunch of money making sure that yields. Now you need to package it. In packaging, you will have lots of losses because of cutting the wafers and then doing 2.5D or 3D packaging. In HBM case, you are packaging 12 dies one after the other. In CoWoS, you're doing the rest.
The total number of chips for an advanced AI chip is greater than 100 chips before a system is available, right? You can just quickly do the math, 12 DRAM chips times six, and then you have a couple of logic chips. All of that go well above 100 chips before a first system is available. These are very, very valuable dies. This is why advanced packaging became very important, very important for KLA and very important for industry. This is not wire bonding packaging, right? Though that is fan-out packaging, that is low end. This is very high end, and the intensity is high because you are gonna lose good die. In some cases, your logic customer who is doing the HBM integration will have to buy new HBM if they lose the HBM, right?
Because the HBM manufacturer is not going to say that you lost some dies, so here's some free ones. It is a complex problem, and KLA is all over this area, and let me explain how. As Rick said, we did an acquisition in 2008 for ICOS Systems, which is a component inspector. And then we also did a second acquisition, and we bought another system called CIRCL. Details not important. But what these two systems enabled us to do is to get into the specialty and packaging market, and we started to learn about everything to do with how do you handle packaging substrates. Our problem is never, do we have the next 10 systems with resolution. We have all of that.
Our KLA's packaging problem is not that. Some of the competitors may have that problem. We don't have that problem. We're doing 10 nm detection in the front end. Packaging defects are not that big. That is not an issue for us. The only issue is can we handle all these different types of substrates? Our teams started working from 2008 and 2015 on all substrate handling, glass, silicon, other materials, ABF, all of those types of things, so we can handle them and are able to detect the defects and measure the curvature and all of those things.
When advanced packaging became important, our core customers, the top four or five, you can guess who they are, said, "Please take all your front-end systems and make them packaging compatible." We were able to use that $1.44 billion that I talked about earlier and divert a large portion of it to doing full integration of packaging as fast as possible for all of our front-end systems. Today, even though Kronos is the high sensitivity, high throughput system and the workhorse of packaging, we have the next three systems already ready. Puma, and there's a bunch of versions of Puma, then Voyager, a bunch of versions of Voyager, and our Gen4 system, all ready for packaging because we have deployed that handling capability. At the same time, we are not just dealing with round things, wafers.
We are now deploying panel capability on all of our systems, so we'll be able to handle panels. Why panels? I think you all know because the chip sizes are going up. Even though everything is full field reticle, right? It's one field reticle, but it's a 4x reduction after that. But in Blackwell's case or Rubin's case, then you take four of those and tile them together. That's how you get scale because EUV systems cannot today handle bigger reticles. So all of that is driving panels. So we have a full suite of systems to do that. Now results. I'll talk about results in a minute. There's also inflections in packaging.
Today in 3D-in DRAM, almost everything is micro bumps, and we are able to do micro bump inspections. In the near future, for logic, SoIC is gonna go to hybrid bonding. For that reason, we have a whole suite of systems for hybrid bonding. In the future, DRAM goes to hybrid bonding, and you know, we don't care if they do or not because it doesn't affect us significantly. We are able to sell micro bump systems or hybrid bonding systems. We will have those systems. Now, the one difference is that the specifications are significantly tighter. In many cases, 10x tighter.
This is why the next four systems I've already developed to ensure that if customers go to hybrid bonding, we will be able to detect very, very small defects for them, with no issues. We're watching this inflection very closely and our customers, as I said, collaboration, innovation, execution, customers constantly telling us when they're gonna do this switch. We're ready for that. Results. We had 10% share in packaging in 2020, so we were number four in a two-player market. Then 4x increase in share since 2021, $950 million in advanced packaging revenue in 2025, and a number one share position in packaging for process control. I talked about technology inflections. I would now want to speak about high volume manufacturing.
I get this question from time to time from investors, with a slight doubt that KLA does really well during the initial stages of R&D, but as customers ramp up wafer starts, they slow down purchasing process control. This used to be the case. It is not the case anymore, but I will give you some data as to why that is the case. Now let's first talk about variability. Day in the life of high volume manufacturing fab manager, what is this person dealing with? First, feature sizes are becoming smaller. Second, 3D architectures are going up. Third, number of process steps are significantly increasing. I talked about the size of the fab and what the same real estate produces today. Fourth, Rick spoke about it, number of designs going significantly up.
We see this is process steps 1.2x, 1.4x, and then number of designs going up pretty significantly. Ship sizes going up. Inline process control specs coming down, right? Because you need to detect smaller defects, EUV, all of this. All of this leads to high variability in the fab. All of this. The next thing is, what do you do with this variability? Should you do inspection everywhere? That is a difficult thing to do because the cost will be very high. How do you go about determining what is sampling, right? You have 1,500 steps to make a semiconductor device. In the 1,500 steps, each step has a lot, right? It has 25 wafers going through each of those steps, each one of those steps. Inside the lot, there's 25 wafers.
Inside the wafer, you can decide to just sample this bit, or you can sample the entire wafer. Inside that wafer, you can decide to do low resolution or very high resolution. How do you do this? How do you determine what you're gonna do? That's where the 1,500 engineers of KLA come in. It's a competitive advantage for us. We understand how yield works. We deploy those 1,500 engineers across all the fabs, and then we try to determine what particular problem this one customer is having. Every customer has a different problem. We deploy those engineers, and we process map out what is happening, and then we help solve the problem. In general, this slide is going up. Sampling in general is going up.
The number of overlay points on a wafer used to be about 250 nm. Today, there's many layers, well above 3,000 points to make sure that the alignment between metal zero and metal one is perfect. If it isn't, you will have to do rework. If you have to do rework, what does rework mean? You just used an EUV scanner to print. Now you strip it and print again. That second EUV print costs you a lot of money, so people don't want to do that. Metrology is far cheaper to do, and that's what they do. How do we process map? Rick showed a slide about what's happening in the fab, right? There's deposition, etch, everything else. We go through the full loop until a wafer is etched into silicon. The pattern is etched into silicon. We do defect inspection.
If it's clean, it's great. If it's not clean, now we need to do source analysis. We then segment the line. We have a number of different systems that segment the line. We look at what happened in CMP. We looked at what happened in etch. We looked at what happened in deposition. We segment the line and then come up with the right suite of solutions. You need that many bare wafer systems. You need that many high throughput, lower resolution, dark field systems, and then you need so many bright field systems to make sure this variability is controlled. That's how we do it. 1,500-1,600 employees that are across the world doing this on a regular basis, helping our customers. Now, collaboration is very important. If a customer thinks they can do it, then the chances of success is A.
Most customers depend on KLA to do this work so that we can help dial the yield down. Yield up, sorry. There's many types of defects also, right? There is what we call systematic defects, right? Which is you kinda understand why the defect would happen. And those are easier to solve, and we help solve those. We don't want customers to have systematic defects. We just don't. They won't yield if they don't yield, and Rick says every time, all the time with this thing that it's about economics. We want them to resolve the systematic defects as fast as possible. The second is stochastic defects. These are subtle defects. They'll happen, and then they don't happen, and then they happen, and they don't happen. You have to determine where they happen, why they happen. They're related to design.
They're related to how much dose you're using, what your focus is. Then there is random defectivity, which just requires consistent monitoring. Otherwise you will have a line problem. It is about economics. I mean, the high-end logic wafers are $25,000 a wafer with, you know, 50,000 wafer starts. You can just do the math. If you have a 3,000 wafer excursion, that's a meeting that you don't wanna be in. It's a tough situation. I skip those. If you looked at process control buying behavior N minus two, the slope wasn't one. N minus one and now N. Customers are buying process control at a fairly linear rate as they add capacity.
We see this now for advanced logic, and we see this for memory as well. I hope that helps answer the question. I've been asked this question several times, so we thought that we would add this section in the presentation to answer it, how customers' buying behavior changes as they do high-volume manufacturing. All of this, again, results semi-cap relative outperformance from 2018 to 2025 delivered $8.5 billion in outperformance during that period. This is above market performance. I'm not talking about what the market was because market is we are supposed to perform at least at market. This is the outperformance number, $8.5 billion during that period for process control. First is that they do this, and second is you have the numbers.
Okay, now let's go into point product differentiation and something that I introduced last time, which was co-intelligent solutions and what does that mean. As Rick said, what do we do? We make sophisticated microscopes, optical and E-beam. We connect them with supercomputers, and then we report results. There was a customer who asked me, "How come it took you about six years, seven years to be able to do E-beam?" I said, "Well, we didn't do any of this for E-beam. We didn't do any of this for E-beam. We only just needed to change the microscope." This applies everywhere in the KLA portfolio. When we decided to do advanced packaging, we're using AI for advanced packaging today. Why would we have to use AI for advanced packaging?
You can say that the size of the defect is so large, you should be able to easily detect it. The one big difference is advanced packaging uses organic materials. Organic materials play out very differently in optical scans than what we have in silicon wafers. Now all the learning that we developed on models and detection capability in AI in electron beam systems, we are able to use that and deploy it on advanced packaging. This story goes across the board constantly, all the time. Our collaboration inside the company is very strong. We have engineering conferences that our CTO runs, where we share algorithms that were developed in group one, algorithms that were developed in group two. These are very technical conversations that enable us to disseminate our learning across the board. It is better.
We can afford doing it more than once, but it's faster and cheaper to be able to do it this way. We would make a large investment in algorithm development, let's say in E-beam, but a very small development cost of deploying that algo on a packaging tool. If a packaging company has to do that, they have to start from scratch and do it that way. The portfolio acceleration is quite large and we really like it. We do innovations in light source optics sensors. During the breaks, we have several people here, including our CTO, that can go through these innovations in detail. Optics is very important. I don't think AI can help solve that problem.
If you don't have a pure signal, then the chances of you being able to do defect detection is very low. That's why I think KLA with its systems and its algorithms helps customers solve problems. Light. DUV broadband source 2x brighter than the surface of the sun. It is, you know, we are right now one atmosphere here, which is 760 Torr. There's several atmospheres inside this bulb. It's a very tiny thing, but it generates a lot of light. We are a big light company because we have to do light for optical systems, electron beam systems, Gen4, Gen5. Now we're working on a new source for Gen6. We are a big light company.
The second is a 300 kg catadioptric objective, very difficult to make and manufacture, but it has the ability to detect very small defects, and that enables us to solve customer problems. The third is this custom sensor, right? If you look at this is the chip itself, which is able to take simultaneous 2000 images every second. That's the image generation capability of this thing. It looks very small, but to run it, you need this big thing. There's a lot of electronics, a lot of cooling. It generates a lot of power, all that stuff. This is all done custom inside KLA. We make our own sensors, own light sources. Why do we develop systems that are this capable?
As in the earlier story, I'll just repeat it quickly. Our customers don't design defects. Our customers design logic circuitry and DRAM circuitry, and then comes the defects. We have to build systems with a lot of capability before we know what defect they will generate. This is why we build these broadband systems with lots of resolution and capability, so that a customer, when generates a defect type, we'll be able to detect it. Every customer is different. Every customer integration is different. I know people say gate-all-around. No, a gate-all-around at customer A is different than gate-all-around at customer B. We have to detect all of those defects. I'll just give you one quick example of why broadband, right? This is the spectrum of the wavelength of light.
The type of image that we are generating is the same, but you would have low contrast of that image in this wavelength of light, but you will have very good contrast of that image in this wavelength of light. Now, if I designed a line tool, otherwise known as a single wavelength system, then when the defect shows up, I may not have the right wavelength. Now a customer is like, "Okay, KLA, you are supposed to be number one in the segment, and you cannot find my defect." That's another meeting I don't want to go to. We need to make sure that we can predict all of those permutations for our customer. This is why we have several systems with different wavelengths. 21xx, 28xx.
This thing goes from 240 nm to 450 nm. This goes from 190 nm to 240 nm, and then now we started to design our Gen6 system with even a shorter wavelength. We continue to increment each of these systems every couple of years to bring additional capability. Now, just to think about how much data is generated, I'm gonna make a case of why KLA wants to make their own HPC computers. Why do we make our supercomputers? Not because we want to show engineering, because it's interesting, but there's a good reason for it. An average Brightfield system takes 78 trillion pixels of images in an hour. We then arrange this into 65 million frames, then that leads to 40 gigapixels a second of image data rate. Some more details.
125 Tb an hour. That's what we generate. One petabyte in eight hours. Total GPT-4, I believe, was four petabytes of data. Total GPT-4. If you ask GPT, sometimes it says it was one petabyte. I like to use one, but I don't wanna be inaccurate. We generate that much data in a shift on a single tool, on a single wafer. That's the amount of data we're generating. That data we cannot store. If we try to store this data, we will have fabs of hard drives. We cannot store. At that data rate, you have to process everything. This is why we have a supercomputer, because if we try to store it's too much. It's literally drinking from a fire hose. I mean, if you can imagine that, right? That's the amount of water that is coming in.
There's no way you can store this. You have to process at the data rate of the system. This is why we build supercomputers, and we design them to ensure that the system has this. Now, we can build low sensitivity systems. Like, I can reduce that pixel size by half, and now the data rate goes by half. But now you don't have sensitivity. That's the reason we do it. Once we do it for one product, we can replicate it. We just need to reduce the amount of computers you need. The architecture is done. KLA also has a five-layer layer cake for AI, right? Okay, so you know the reference. I was thinking maybe you'll get it, maybe not, but no, okay, you got the reference. Physical subsystem is very critically important. That thing there, that thing over there.
High performance compute, I just talked about. The HPC is very critical. Software infrastructure to deal with all this data that is coming through. We're not a process company, so this process company software is very low, very low, right? Because you're opening actuators and things like that. Software infrastructure is very critical. AI models, we use foundational models. We use all the different types of large language models and detection models. Most important part is that we have the context of understanding what is inside the optics and what is inside the wafer, and I'll describe that in a minute. Lastly, the application layer. Our five layer ensures that we're able to provide the result to the customer. Let's double-click on a couple of them, right?
This is why do you design your own HPC? Why do you design your own HPC with CPUs and GPUs? As I said earlier, that if we go to a very large pixel inspection, now the resolution is low, right? Because the pixel size is very large. The data rate goes down, but the sensitivity goes down. You will throw away the defect. If the defect is here, you won't be able to detect it. What we try to do every two years with Gen 4 and Gen 5, we bring in more light, we make the sensor faster, and we improve the optics. Why? We can increase the resolution. We increase the resolution. The moment we increase the resolution, what happens? We take more pictures.
We take more pictures, we need to scale the data rate. This is why we custom design everything inside KLA. We don't go out and buy a rack from somebody, right? The second thing is that we want to make sure the images that we're taking are extremely accurate, right? This is the optical image. We align that optical image to the design. Our customers provide us the design. Most inspection companies don't get design, we get design. We align the optical image with the design, then we build an E-beam system, and we coordinate match the E-beam system with the optical system. Now the chances that there's a defect here and you miss it is very low because the design is telling you something, the E-beam inspection image is telling you something, and the optical image is telling you something.
All three of those come together, and then we can add other information from design from the customer. This area is more important because there is this thing next to that thing that will cause a defect. All of those enable us to reduce what the customer calls as nuisance defect. You can always find differences. It's not the difference that we care about. It is the defect we care about. Customers tell us, "Yeah, I can see that there's a difference here, but I don't care about this. I only care about that." That's the other part we have to do. We're not in the job of reporting differences. We're in the job of reporting the critical defect that is going to cause the yield loss. All of this with our systems, we're able to do.
I hope that answers the questions. Now I'm gonna talk about a concept of how we bring the full portfolio to do defect detection. Okay, mask inspection, I'm gonna start with that. What happens in a mask shop and in a wafer fab? First you take a plate. This is a glass plate, and you want to print the design on to this. You use an electron beam system to write the design. The customer has the design, and they write it. Now the design is written, you need to do a pattern inspection to ensure that there is no defects on this plate. After you've done that, you take this plate, you put it in a EUV scanner and print.
During that print, you can add defects onto the plate because the EUV uses a source that can inherently generate 10 particles, and you can get on that. What we have to do is, after this plate is printed, we have to find all the defects, and after this plate is sent to the wafer fab, if there's new defects added, we have to detect those. There's two places a mask goes through. One is a mask shop, and the second is the wafer fab. In the mask shop, you're using the electron beam writer to make the mask, and it needs to be defect-free. When you ship it to the wafer fab, you start printing wafers with it, and it needs to be defect-free. Now, how many reticles do you have in a wafer fab?
You would think maybe 10, 20, 30, 100. That's a lot of reticles. Now, why do you have so many reticles? First, number of designs. You have many designs. Number two, number of lithography layers, right? You have a lot of lithography steps, and lithography steps are going up. Number three, respins. A customer brings a design, the design doesn't work, you need to print another reticle. Number four, extra reticles, because you will have a defect on a reticle, and you need to send it for repair. All of this causes a number of reticles in the mask shop. We believe it's about 700,000 reticles per year, new reticles per year. We have to do defect inspection on all of them. Now, in advanced logic, there are 50 193, 193i immersion reticles and about 20 EUV reticles. Okay?
For inspecting those, you would think you would need systems that are 193 and EUV. That's not the case. Our 193 system is able to do all 193 reticles and are able to do most EUV reticles. Not all, but most EUV reticles. What's the difference? It is called printed pitch. If the printed pitch, which is the pitch of the design, is larger, even though it's EUV, we will be able to detect the defects with 193. If the printed pitch is smaller, very small, then we won't be able to do it with a 193 system, and you need a high sensitivity system. Okay? That's how it works. It's about three to five reticles in advanced logic that are very tight dimensions. Everything else we can do with a 193 system.
Now in advanced memory, the same thing happens. It's 25 + 6. 30 of those can be done with a 193 system and maybe one or two reticles require very high sensitivity to ensure that we can do defect detection. This is the suite of systems that we use to do defect detection. As I said earlier, it depends on printed pitch, right? Relaxed printed pitch, 17 nm, 30 nm to 40 nm, we can easily do with our optical system, and then we can easily do with an optical system with EUV capability, most of those. Then in the wafer fab, I'll go through this in a minute. In the wafer fab, you are just looking for particles, right? You're not looking for design issues. We have a 193 system that does most of the work.
We have something called Gen5 EUV Print Check, which can determine reticle repeater defects. We do that with that. With EUV and then with Gen5. All of those capture most of the defects. Now, we're talking about a few subset of defects with very small pitch. How are you able to detect those? There's two ways to do it. One is you can develop an actinic inspection system. It has the same wavelength as the EUV scanner, 13.5 nm. You could do something else, which is an E-beam system, which has very high resolution. We're doing both, and I will show you results for both. The first one, Teron beam 8xx mask inspection system, we just brought it out to the marketplace. We've been developing it for the last five years at a customer.
Very close collaboration, and that customer is TSMC, because we have published a paper recently. It's a massive technology. If you get a chance later on, you can see it. This is the twelve by twelve column array. Why twelve by twelve column array? Because we have to cover the mask completely. E-beam is slow, but if you had a lot of E-beam outlets, you'll be able to do it fairly fast. This column array, we have 24 columns, 12 of them are always running. It's a 100,000 pounds system, if you're buying by weight. We have an AR/VR model in the other room. Why this system? Because as I said earlier, the printed pitch matters quite a bit, right?
36 nm, this is TSMC paper, so I'm just gonna repeat what they said. A DUV system can cover easily 36 nm and greater, even though it's an EUV print. An e-beam system has ultimate resolution, so it can go all the way to 20 nm printed pitch. This is reticle pitch, right? 4x after that on wafer. This is very aggressive. EUV, high-NA EUV covered both. No issue. Nobody's printing 20 nm printed pitch onto reticle for wafer. They built reticles that are very difficult to test this machine. You can easily see that with actinic you will have a slight resolution problem with e-beam, you can easily do it. Now, why do people do actinic? There's two reasons.
Again, there's always a type of defect problem. Currently, we don't have a type of defect problem that DUV and E-beam cannot capture, but there could be a type of defect problem. The second is that it is easier to do because you are able to do 24 nm pitch, which is in the middle. For the next couple of years, you could continue using this system before you need this. Our customer decided to develop this system with us because then it has ultimate resolution. An actinic system will always require change in NA each time you change the NA of the scanner, right?
That's why 0.33 NA is at one system, 0.55 NA is another system, 0.77 NA will be another system, and then after that, if you change the wavelength from 13.5 to 6.5, you will change the system again. This system is more longer capability. Regardless, we're developing actinic inspection also. We have taken our first images on our prototype system in our lab, and we've done our first inspection a couple of weeks ago, and the results look quite impressive. That was the mask shop. In the wafer fab, you need multiple defenses. Now, in every case, you need multiple defenses because as I said earlier, in reticle inspection, you have no choice but to detect 100% of the defects.
Defense one, our EUV 670e XP2, it's a DPUV system. Defense two, Gen5. Now, you would do this inspection every 250 wafers. Every 250 wafers you shoot under an EUV scanner, you're going to do a DPUV inspection. The throughput of an EUV scanner is 250 wafers an hour. For one hour, you will do this. Defense two, Gen5, every four laps, so maybe like 800 wafers, you will do a second inspection. This one, maybe around 2000-5000 wafers. It's really an insurance policy inside the wafer fab because it is, you know, it's a cost of ownership on an actinic system for reticle is quite bad. Most customers use these two.
You can always add a third defense to catch all the defects. Okay, that is a co-intelligent solution for reticle inspection. Now, how are we doing on a co-intelligent solution for E-beam? Why did we get into E-beam? As I said, there are class of defects, very subtle defects, that you may not be able to detect with optical. We still think the ratio of optical and E-beam is greater than 80/20. 80/20, the actual number is 88, and whatever the other number is. It's on the fly math. That's what the ratio today is, it's 80/20 is what we loosely call it. But there's class of defects that E-beam is very good at. As 3D became important, those subtle defects can be detected with E-beam.
The question is, if you can detect with optical, will you use E-beam? No. There's no physics and no economics that'll enable you to do that. 50 nm all the way to, you decide where it starts to struggle. The second thing is, if I can do what I showed you earlier, I have the brightfield image, I have the design image, and now I'm going to provide it E-beam intelligence. The brightfield will be able to extend more. Now we're using E-beam to extend our brightfield systems more. Thirdly, I said in HVM, they're buying a lot of brightfield systems. Well, there are 1,000 modes per brightfield system, and multiplied by the number of 1,500 layers.
Well, an electron beam system can help our AI model to determine what is the best brightfield mode for the type of defect. These three reasons is why we got into it. Now, we thought we would just get into one segment of E-beam, but that was not a good plan. I mean, generally, to get into all segments of E-beam from an engineering perspective is very hard, but we decided to do that. We have a single beam E-beam inspection, ESI 50. A multi-beam E-beam inspection, 300 beam E-beam inspection, at ESV 100. Both of them in HVM today, eDRX1 for review and EM200 for metrology. All of these systems are now in HVM. They're not in development, so already moved to HVM.
I showed this slide, which is this is a Brightfield test image, and this is the difference image, meaning that you took an image, you subtracted the pattern, and this is the difference image. Where's the defect? It's very hard to tell. It could be this, it could be this, it could be this, or it could be this. The defect that the customer cares about is only this one. Nothing else they care about. If I report this, this, and that defect, guess what? You have to do a lot of e-beam review to verify. Now my relevance goes down. In order to fix that, this is what we're doing. We're connecting our e-beam inspection systems with our Brightfield systems and e-beam review systems with our Brightfield systems. This gives the example of what the defect looks like.
This is the workhorse of full wafer inspection, and then you can send the results to e-beam review. All of this is connected through a computer. Computer is a weak word for this. It's a huge machine that is updating the models continuously. This is an example of a co-intelligent portfolio delivering questions, I mean, delivering answers to our customers. Brightfield optical inspection success, $8.5 billion since 2022 to 2025. That's just the Gen 4 and Gen 5 systems that we have sold. It grew greater than 30%. The market in the same period, I believe, was around 14%-15%.
Greater than 600 Gen4 systems, and Rick pointed that out, that we're shipping more Gen4 systems and greater than 180 Gen5 systems shipped during this period. Okay, that's the technical part. I will go into the model. So I spoke about logic inflections, memory inflections, advanced packaging inflections, high-volume manufacturing. We participate in high-volume manufacturing at the same curve. Brightfield optical inspection, darkfield inspection, and then E-beam. As I said in the very beginning, right, $9.5 billion, the systems business, the process control was $9 billion. There's a bunch of other things that we do, but if I have to present on that, we'll be here for a longer time. I'll give you a little bit of flavor, right? X-ray metrology.
We developed an X-ray metrology systems for the vertical NAND and DRAM capacitor, and it's used in high-volume manufacturing today. We purchased a chemical analysis company, and it's tagged with plating systems. This metrology system is scaling as interconnect density is going up, more customers are buying this. Data analytics. This is a $250 million software business now, connecting all of our systems together to ensure that we're able to help customer do detection. Plasma dicing. As customers go to more valuable die, instead of doing saw dicing, they're gonna start doing plasma dicing. We have a suite of systems that are gonna help with plasma dicing. Component inspection I spoke about in PCB inspection.
All of this is roughly about $1 billion on top of the semi-cap business. Lastly to the model, $9.8 billion in 2025. Market baseline growth rate would be $7.6 billion. That is a 11%-13% CAGR growth. Outperformance would be $3.1 billion, reaching $20.5 billion. Overall, semi-cap CAGR, overall systems business CAGR would be 4%, and semi-cap would outgrow the overall business by 4.5%. All because of our collaboration, innovation, execution. That's all I had. Thank you.
Now it's time for a 10-minute break. Please help yourself to refreshments outside the room and join us back here in 10 minutes.
Right. Good morning. Welcome back. As you heard from Rick and Ahmad, the market dynamics set up extremely well for KLA, and we've got a strong portfolio of products to support our customers' most difficult challenges. That's good for our service business as well, because one of the bookends of our growth strategy is growing the install base and the additional value-added services that we provide on those new systems. The other bookend is the enduring value of our systems. I started at KLA 28 years ago, working on products that we now call the class of 1995. Many of those products are still running in customer fabs around the world today, generating more than $100 million in annual recurring service revenue.
It's this powerful combination of new install base growth coupled with the enduring value of our systems that makes our service business so durable. We have three key messages for today's section. First, we are a customer-focused organization. Our service excellence is aligned with customer outcomes to ensure that they succeed. We manage that through the entire lifecycle of the customer's products. Next, as you heard from Rick and Ahmad, the KLA operating model is the framework we use to work with our customers. It's rooted deeply in data, and we use that data to improve both customer experience and KLA profitability and productivity.
Finally, the headline message for our service section today, we are increasing our service growth rate from 12%-14% to 13%-15%, which means at the upper end of that range, we'll double the service revenue by 2030. Before we get into some of the details about how we're helping our customers solve some of their challenges, I wanna give it a little bit of perspective on KLA infrastructure, a little bit about why process control service is different, and close with some financial highlights. First, from an infrastructure perspective, we have more than 57,000 tools installed worldwide at more than 4,000 customer facilities. Every one of those customer facilities is driving unique requirements, and we have to invest in order to support each one of those.
We invest in the form of people, more than 3,500 service engineers, and in parts, more than 260,000 unique assemblies in spare parts. That all comes together, that infrastructure comes together in comprehensive service contracts, which reduce overall cost of ownership for our customers. We talked about this in the past. We have a unique signature in our install base. We're high mix, high complexity, lower redundancy, and long life. I'm gonna unpack each one of these in the coming slides. Before I do that, I wanna touch briefly on this graphic on the right-hand side, the product plus service. As Ahmad went through in his presentation, we are the market leader in process control, and that's because we deliver the highest throughput at the highest sensitivity.
The only way that our customers can maintain that highest throughput and highest sensitivity is by leveraging KLA service. We deliver our product, our service product is availability at performance. We have to meet those throughput and sensitivity on a day-to-day basis at very, very high availability rates. This concept of product plus service, one plus one is greater than two, it's a competitive advantage for KLA. But let's talk a little bit about some of the characteristics of our business. First from a mix perspective. What do we mean when we say mix? Well, each one of those pluses on the left-hand side is a KLA product division. That in and of itself would suggest a high mix. We are not a single product company.
When you double-click on one of these, and if you haven't visited the AR/VR, as we talked about, we've got our BBP system, our latest gen BBP system there, that's one of one tool model configuration that you see. If you double-click on the division, you see there are more than 30 different product models, and every one of those product models has five different configurations. It means 150 different configurations that we're managing just for a single product division. You can imagine what that looks like when you go back and look at that through each one of these divisions, the mix that we are managing across our install base. We also have. It's also a complex business. More than 1,200 unique parts and assemblies.
You can get some perspective on that. You see our objective over here, our sensor over here. These are two examples of assemblies. They're also field replaceable units. So more than 1,200, we'll talk a little bit more about that, and more than 25,000 different specifications. So again, this is for one product division. It's representative of what we deal with across the rest of the install base. This is one of our biggest challenges as a service organization, is supporting this high mix, high complexity. It's also one of our greatest advantages because we are uniquely positioned to be able to support this. We're also differentiated in a couple different ways.
Ahmad outlined, you know, ultimately, we're a portfolio company and our customers are gonna select different KLA products to put in line and use those to drive improved overall yield. If you go onto a customer fab, you're gonna see a lot of white panels and purple stripes, lots of KLA tools. When you double-click on that back to this high mix, there are lots of different tool types in there, which means there's not a lot of redundancy for any one of those given tools. Because there is low redundancy, our utilization of our systems doesn't correlate very well to overall fab output. It's inelastic, means whether you're running 10,000 wafers, 25,000 wafers, you're still gonna use your process control tools at very high availability rates.
That contributes to the durability of our business. You're not gonna idle some of that equipment. You're gonna use our equipment. You might change some of your sampling rates, but you're gonna use our equipment at very, very high availability rates, which is why, again, we see very little fluctuation in our overall service and why it's so predictable. This is differentiated from others in the semi equipment ecosystem. Two other important differences is we are. Our part content, as you see both in the AR/VR demos and as you work through looking at some of these examples that we have, these are very proprietary in nature, these FRUs. And they take. We spend a long time developing these with suppliers.
Bren will talk about their sole sourced, many of them are sole sourced, and KLA is uniquely positioned to be able to to stock that those parts and be able to service the equipment. In addition, we have very diverse labor content because of the unique nature of all those parts, the unique nature of the failure mechanisms that we have. All this, again, contributes to the durability of our service business. A little bit about how we've performed over time. If you think about our service business, the way to think about the growth equation is you've got this existing install base. Are you able to drive greater revenue, and how stable is the existing install base? Then what's happening on the new the new tools that you're shipping?
Are you attaching at a higher rate, higher contract penetration on those? Are you generating more revenue? If you do all those things, then it should lead to improved overall growth. Good news for our business is we're improving on all of those. First, from a tool lifetime, we'll touch on this in a little bit more detail later in the presentation, but median time to tool retirement for KLA systems has increased from 10 years to more than 24 years, from 2010 to 2025. Very important for our customers. Next, from attach rate. When I was here in 2019, we talked about service revenue from contracts at about 70%.
We've increased that 10 points, closing in 2025, and that's of course, driven by even higher contract penetration on the new systems that we're shipping. That's why this is increasing, because the complexity is increasing, therefore, the reliance on KLA service is increasing. Next, we're also generating more service revenue per tool. From the class of 2000 to the class of 2020, we're up about 3x. Important drivers. This is what's leading us to improve overall projected CAGR. In 2019, we said we were gonna grow the business at 9%-11%. In 2022, we increased that to 12%-14%, and today we increased that to 13%-15%. A very important revenue stream for KLA.
Bren talked about 17th consecutive annual increase on our dividend. That dividend is serviced by cash generated from our service business. It's that predictability and confidence that allows us to continue to drive increase in the dividend. Ultimately, it's about serving our customers. We use the KLA operating model framework, collaborate, innovate, and execute. We support our customers, whether they're ramping new technologies, new fabs, or if they're trying to extend the asset lifetime of their existing tools. First, from a collaboration perspective, our service engineers are very much part of the fabric of our customers' operations, and we are recognized for it.
We see recognition whether we're helping a customer ramp a new technology, ramp a new fab, ramp a new geography. We're recognized for the best supplier of the year award or in the event when our customers experience unexpected events like an earthquake, our engineers are in the fab in hours to help our customers get those tools back recovered so that they don't miss any of their production commitments. While recognition is important, ultimately, our customers vote with their checkbooks. As we mentioned, our service contract penetration has increased over time. Back in 2005, we were about 60% coming from revenue on service contract. Today, we're more than 80%. These are long-term service contract agreements.
These are more than three years, and so this is a lot of trust with our customers. This long-term commitment allows us to not just solve for today's problem, but also to solve for their roadmaps, especially in this ever-evolving, geographic expansion environment that we're operating in today. Additionally, we talk a lot about the service contract being 80%, but if you look at this pink, this is what we call KLA time and materials. This is a little bit more variable spend, but it's still spend with KLA. If you take the purple and the pink, that's basically the revenue that we generate from those 57,000 tools in the install base. Every one of those individual tools over a two-year period recurs in revenue 95% of the time.
Again, our predictability on our service revenue, our line of sight on ability to sustain the level of growth that we've talked about and our confidence to continue to deliver that is very high because of these dynamics. Next, of course, in the semi industry, we've got to continue to innovate. Our innovation in our service business, we think about it in two distinct buckets. First, how are we trying or how are we improving customer tool performance in customer fab? Next, how are we improving every part of the value delivery chain that delivers that performance in the fab? That usually comes in the form of people, parts, and knowledge. These two things come together around KLA data systems. You know, we're a process control company.
We've been collecting data for many decades. It's all about sorting signal from noise. We've done a very good job of this, and we mine this data in order to improve both customer experience as well as KLA productivity. I'm gonna spend a little bit of time giving you some examples of some of the things that we're working on, in terms of improving in-fab performance of our customers' tools. The way that we do that is we leverage our service applications portfolio. Both Rick and Ahmad talked about our applications engineers. For example, our systems are connected to recipe development servers for accelerated recipe deployment that allows them to store, modify, and deploy uniformly across the install base.
Some additional examples that I'm gonna go through are predictive tool health insights, autonomous synchronization, and then on-demand agentic expertise. First in this period where customers, you know, demand is significantly outstripping supply, any unscheduled downtime for one of our products has a material impact on our customers', manufacturing. In the historical model, we would wait for something to break, and then this period down here meant that the tool was down. You would fix it, and then you'd come back up. With our predictive tool health insights, we're now able to track assemblies like these two off to the side. What's performance of those assemblies against predefined performance specifications to understand when the assembly begins to drift. Again, KLA is a system of systems. It's a bunch of these complex assemblies that come together to create the performance that we want.
If one of those assemblies begins to drift, total tool performance begins to drift. Because we have predictive tool health insights in place, we're able to track that, and we can now pre-position material and labor in order to support that, reducing total unscheduled downtime for our customers. That's also a value unlock for KLA because now we can be much more productive with our deployment of labor as well as inventory. We do that again across those 1,200 unique assemblies on that one BBP platform. Again, you imagine how this scales when you look across the entire install base. Similarly, we have a autonomous synchronization application, and the challenge here is you heard again from Ahmad, the high-mix environment that our customers are operating in.
In that high-mix environment, you wanna make sure that the tool that you have in place for whatever application it's serving inside the customer fab is aligned to the control reference, whatever they want that, so that they have 100% of the performance 100% of the time. Our autonomous synchronization system allows us to do that and ensure that we catch those issues or deviations before we return the tool to production, which reduces overall variability inside the customer fab and improves our scalable operational infrastructure. Semiconductor manufacturing, highly automated, but still requires skilled engineers. Our customers are hiring a lot of skilled engineers, and we have to hire a lot of skilled engineers. This is important to be able to meet ramp timelines. One of the biggest challenges, accelerating time to proficiency.
The way that we accelerate time to proficiency is in addition to some of the AR/VR tools that you saw today, we also use our on-demand agentic expertise. Again, I talked about we've been collecting data off our systems and our business processes for decades, but you know, correlation does not mean causation. It's great that we have all this data. We also have the contextual information and the subject matter expertise to understand what's really happening. That allows us to synthesize the data, and put it into dynamic knowledge model, and that then converts to on-demand delivery guidance for our customer support engineers to be able to reduce mean time to repair.
Ultimately, what we're trying to do is take the tacit knowledge of a master engineer and transfer that as quickly as possible to a new engineer, whether that's a new hire or an engineer that's now learning one of the new products that Ahmad talked about we're introducing. Ultimately, we're an execution. Our service business is built for execution, and we execute in many different ways. I'm gonna talk about how we're helping customers solve against new technology inflections, how we're helping them ramp in new geos, and then also how we help them extend the lifetime of some of their assets. First, Ahmad Khan went through this in great detail. HBM inflections that we're seeing, this is driving increased device complexity, increased advanced product models from KLA, and tighter requirements.
That has a pull-through, a natural pull-through for our service business. We're able to align on new and tighter availability and performance requirements with our customers and deliver to that commitment. Because we do that, we see revenue growth increase by about 5x from 2022. This is one of the catalysts for our 2030 plan. Similarly, advanced packaging. Also a great business as Ahmad outlined, and good for our service business. We talked about pulling front-end systems, you know, configuring front-end systems for back-end application. Again, that puts tighter requirements on running those tools in the back end. That puts pressure on our service organization. We've been able to modernize our back-end support with the same rigor and intensity that we see in the front end.
We've also been able to leverage our global infrastructure to support this significant ramp that we've seen in advanced packaging. Again, a competitive advantage for KLA, and we've seen strong revenue growth of about 8x from 2022 to 2025. Again, another catalyst for our 2030 plan. Geo expansion. We know many of our customers are expanding geographically. The last thing they wanna worry about is whether or not their equipment suppliers can support them. They've got plenty to deal with on their own. What our customers want is they wanna match the culture, the intensity, the discipline that they see in the home country. We can partner very closely with our customers to do that.
Our scope and scale allows us to move resources from the home country to the new geography, ramp up, hire people, get them trained in customer processes to allow faster yield ramp. We've been very successful in doing this over the last couple of years. As we look at the number of ramps that we've got to support through 2030, we feel very good about our position. We spent the majority of the presentation here talking about the left side of this life cycle, R&D, HBM, ramp through HBM. That's about 70% of our services business, but the other 30% is supporting this legacy HBM.
This is a business, you know, many of our customers, these are profitable businesses for our customers because they have depreciated their equipment, and so they're able to run a very profitable business, and KLA plays an important role in that. The way that we play an important role in that is by extending tool lifetime. We talk about this increase. If you go back to the year 2000, the median time of tool retirement for a KLA system was about four years. That goes back to, you know, the 18-month or two-year tick-tock strategy. We'd introduce a new technology, a new node, and then two years later, we'd retire that equipment.
That began to change as you move into the 2010s and the 2020s, as you see some of these super nodes that came into play. Now we see median time to tool retirement up at 24 years, significantly longer than our customers' depreciation cycle. Of course, this does not come for free. KLA invests heavily in making this 24-year median time to retirement come true. First off, we structurally change the way we design and service tools. All those products that Ahmad Khan talked about that we're developing, when we sit in reviews, we're not asking about how to deliver those tools for two years or four years or support them for two or four or six years. We're talking about supporting them for 25 years.
This is an important value proposition for our customers. It's one of the reasons that they pick us, is because they know that we will be there to support them for the long term. We also have active obsolescence management, hundreds of obsolescence problems that we solve, and modernizing those systems on a regular basis to ensure that they can continue to get the performance that they need to support the tool. That all comes together in what we call our class of chart. If there's one chart in my deck that I think both shows KLA's 50 years of market leadership, but also the how our service business has performed over time, it's this chart. On the y-axis, this is service revenue. On the x-axis is time.
As you see on the legend over here, these are cohorts. The way that we define a cohort is, if you were a product that was introduced in the year 2020, 2021, 2022, 2023, 2024, you're part of the 2020 cohort. I mentioned at the beginning, I started in the class of 1995, and if you squint really hard, you can see a little bit of blue starting to come up for service revenue in the class of 1995. You follow it over, and it peaks in about 2008. As you follow it all the way to 2025, you see it's still there. That's the $100 million of service revenue that we're still generating on the class of 1995 products, not dissimilar to the amount of revenue that we're currently generating from the class of 2020.
This is why we're very confident as you look out into 2030. Of course, that 2020 is gonna continue to grow. We also expect that the class of 1995, 2000, 2005 will continue to deliver value for our customers and continue to deliver revenue for KLA. This is why we're confident in our growth rate projections through 2030. In closing, we use collaborate, innovate, and execute to deliver more value for our customers. We deliver more value for our customers, we deliver more value for KLA. I was here in 2019. We were $1 billion. I talked about it took us 40 years to get to $1 billion of service revenue. When we came back in 2022, we had added another $1 billion.
We went to $2 billion. 2025, we're now another billion dollars, $3 billion. Again, as I said, at the upper end of our range, we would expect to double that revenue to nearly $6 billion and at the upper end of the 13%-15% range in 2030. Never a better time to be in the semi-industry . Never a better time to be a part of KLA. Thank you.
Right, the home stretch here. As a CFO, I've been the CFO at KLA since 2013. You love that business, right? It's a business that has a lot of real positive attributes. It's unique in the industry, and those attributes are getting better. Pretty excited to have it, and it certainly is something that as you think about just how what happens in the world, having this anchor that drives growth consistently is a pretty powerful base to have as we think about operations, as we think about capital allocation and so on. In 2019, I stood up here and said, "Hey, the world is changing, and it's changing in a real positive way for KLA, for the industry and for KLA.
For the industry, we were gonna move from a period of time where you had significant capital intensity decline to all those drivers of it sort of played out, and that going forward, we would have capital intensity flat to rising. That would create structural growth in the industry. Semi equipment would grow at the rate of semi. The other good thing that was happening is we finally had a scaling roadmap after years of delay with EUV. As a result of that, you would have the 7 nm node and later on with DRAM, but the 7 nm node where you had a lot of design activity. You saw Rick's chart earlier. First time bigger than 28 nm. Because of that, you had designs, you had a big difference between intensity and logic and foundry versus memory.
As a result of that, it was gonna drive the relevancy of KLA higher. 2022, we stood up here and said, "All those things are true." There's more momentum as we saw the successive nodes driving more designs, and that the industry was broadening in demand. You had a lot of legacy demand. Semiconductor content was rising, all the digitization related to COVID was part of that. Legacy demand is really good for our customers because of their profitability. You had what was a GDP plus growth rate for semi revenue start to accelerate. Of course, as we moved into 2024 and 2025, we saw the high-performance compute drive. All of that, as you saw on Ahmad's chart, took KLA's process control intensity or share of the overall market up about 200 basis points.
Translated from 19 to 25 into best in industry growth levels and a very strong financial model. Of course, over the last 12 months or so, I get lots of questions from our investors asked in very different ways. Really, the crux of the question is, can the next five years be better than the last five ? 'Cause the last five were pretty good for KLA. Can the next five years be better than the last five? If you look at the conditions that you heard a lot about today that are driving the next five years, the next five years are gonna be better than the last five . What I'm gonna cover today is, as you've heard, we're well-positioned to drive sustainable outperformance across all different segments, and that's contributed to increasing process control intensity. We're excited about that.
The second message, I'm gonna talk a lot about the operating model and how it translates into the financial model and how we run the company, how we make decisions and stuff, what's underneath the hood, if you will, in terms of our how we engage with customers, how we have to invest, what's driving the income statement and balance sheet over time. We have to execute, and I'll talk about some of the execution drivers in the business and how that translates to a go-forward view. Then finally, I'll wrap it up with a discussion of the 2030 target model, work through some of those assumptions. That model is really based on a history of credible execution. We've done a, I think, a pretty good job against the public targets that we've established. As Rick said, we define winning inside the company.
There's no daylight between the external plans, the internal plans. This is how we measure ourselves. This is how we evaluate the company. This is how we're incentivized, and this is how we measure our success over time. The 2030 revenue plan, $26 billion ±$2.5 billion, and I'll walk through that. Non-GAAP EPS of $84. Should continue to drive strong free cash flow generation consistent with our history, and I'll share some of that. A capital returns target that was greater than 85% on a go-forward basis, expected to be greater than 90% of the free cash flow we're generating. Go back to last Investor Day, where were we and where are we today? We talked about $14 billion and $3.8 billion.
If you look at where we are today, you can see that particularly given the commentary that I gave you earlier around our expectations for the year, we're in excess of that. We said we wanted to improve our share of market by 100 basis points. We improved it by over 150. We said service would grow 12%-14%. Service grew right in line with that target range, and as I said earlier, had some disruption related to access to fabs. Pretty happy about that. Certainly, the strength of the systems and some of the drivers of incremental opportunity were a factor in that. We said we were gonna broaden our portfolio. We're gonna invest in reticle inspection. We're gonna invest in e-beam, and we were gonna expand our market leadership.
As Ahmad showed you, we have done that. We feel very good about how we're positioned from that point of view. Advanced packaging. We didn't talk about it before, but it is a significant market and a significant market opportunity, and I'll talk a little bit more about why we see it the way we do. We were gonna integrate acquisitions, which we did. We put them on a trajectory. They were dilutive. Put them on a trajectory to corporate level incremental margins moving forward. We're gonna invest to support our growth. We're gonna deliver our incremental operating margin model, which we did. 2025 operating margins were 43.6%. We said our target range was 41%-43%, and that's despite tariffs and some other things that we hadn't contemplated back then.
Our capital returns was in just slightly in excess of 85% since 2022. In line with our target there. Where are we going? This all translates. $26 billion translates into 15% at the midpoint in terms of revenue growth. We're gonna gain another 150 basis points of share in terms of share of the overall market. Service revenue, as Brian just talked about, 13%-15%. We're gonna continue to optimize our business through technology leadership. How do we leverage the portfolio? It's unique to a competitive advantage for KLA, but we're also driving productivity improvement across the enterprise. Certainly have a strong growth environment, but opportunities for us to continue to drive our productivity. If we do all that, we'll sustain our model.
At a higher growth rate, we should be in the higher end of the range of our 40%-50% target model, and we'll drive 45%-47% operating margins. I talked about the capital return change. It's interesting when you look at the revenue over time, and this was that time where there wasn't much happening. You had capital intensity that was down. A lot of the growth for the company came from service. You had this period, then you had 7 nm node. You had EUV into DRAM, EUV into logic. You had a down year in 2019. This was a $1 billion increase year to year in a down year. Then we saw this path to mid-teens growth. 2019, 2025 mid-teens and basically our plan is a continuation of that.
You can see a lot of the drivers on what are the characteristics of the target revenue mix and why. In an environment where we feel like, again, the conditions are strong and that it will sustain our ability to grow it in a mid-teens range. Now, if you look back to the three Investor Days, you know what's happened. As Rick mentioned, we had acquired a company, integrated the company, drove our margins up best among the leaders in the semiconductor industry. A lot of focus, the indicator of differentiation, and then driving leverage on our infrastructure as we scaled over time and incremental margins at the top end of our 60%-65%. Incremental gross margins, 60%-65%. Operating margins differentiation lets your profitability drive your operating margins.
We've, again, a superior margin profile versus peers. I'm gonna talk more about this, but this extendability of product platforms and what does that mean to the R&D intensity, how much we spend on R&D. The industry has a pace, a cadence of innovation that isn't as fast as it used to be, but is fast enough that it drives the end markets. It's really good for structural profitability for our customers because they can amortize cost of development and cost of capacity over a broader period of time and more volume, but also good for us. I'll talk more about that in an upcoming slide. Obviously, focus on productivity and how we drive the business and our performance on incremental operating margins at 51%, so slightly ahead of the top end of the model.
Earnings leverage, operating margins, less than our capital structure decisions, taxes, all those sorts of things. We improved share outstanding, so drove our shares outstanding down about 18%. We had a target model that we would deliver 1.5x the revenue growth rate in terms of EPS growth. Revenue grew 16% over this time frame, diluted EPS at 26%. Achieved that as well. Free cash flow margin. We have a history of strong free cash flow margin. It's trending above 30%. I'll share some data on that. Our business model is a unique one. Obviously, we have the profitability levels we do. We invest in working capital, spend a lot on inventory, have a lot committed there.
If there's a thing that we accept to drive our differentiation, that goes to a lot of what we talked about earlier in terms of design and capabilities, we accept that the asset velocity of inventory turns will be lower, but it positions us to enable the differentiation and to meet the dynamics in our business. I'll talk a little bit more about that. CapEx. Absolute value of CapEx, we've been spending more, but generally between 2% and 4% of revenue. As a result, a lot of that profitability drops to free cash flow margin. It's not just about semi equipment. You can see in the chart when you compare, take a long term, different drivers, different regions, other investments, different cycles, different drivers in the end markets consistently against the stocks.
Average performance over the last 10 years, ranked in the top five of all three. Only three companies are in the top five of all three, NVIDIA, KLA, and Analog Devices. It matters a lot in terms of long-term performance, lots of opportunities where things can go a different way and a continuation across the board of, we'll call best-in-class performance in overall, the overall semiconductor industry. It translates into very strong cash flow. Free cash flow as a percent of revenue in the 92nd percentile of the entire S&P. It's that model that I talked about, profitability, but also an asset-light business that enables very strong free cash flow performance, which has allowed us over the years to be very thoughtful and very explicit about capital allocation.
That capital doesn't get valued unless it's deployed productively, that businesses need to be financed with a tier of debt. All those sort of philosophy of KLA and how we thought about it has been really driven by the free cash flow sustainability of the company. You can see in this trend, first the dividend. Right? Mid-teens growth rate target. Two targets, the growth rate, how do we drive the growth rate relative to the growth rate in our free cash flow? Also is it 25%-30%, so we can maintain a consistent cadence of 25%-30% of the free cash flow we generate. That drives our dividend.
You can see on the share repurchase side, we have the additional authorization today, and our strategy is a little bit consistently, again, putting the capital to work consistently over time and using capital structure, where prudent, to augment the returns or accrete the value of KLA to our shareholders, and we have one of those actions back here. A consistent approach over time, explicit, so it can be modeled and valued. Brian talked a lot about the dividend and servicing and the dividend. I take it one step further. It's not just the dividend, it's the after-tax cost of debt for the company. It really services both those.
If you talk about the market, this was that world where I was talking about where we basically went from a 4% GDP-type business, GDP plus, to this transition over the last few years that we think has a continuation of approximately, you know, low double-digit growth. You can see a lot of the drivers from it, you know, the obvious discipline and pricing's been good. The value of what's happening in high-performance compute is also a factor. Lots of good factors that are driving growth, the diversification of demand. What's really important for us is you can see as you take 2030 and you think about an environment of $1.3 trillion-$1.5 trillion, $1.4 trillion is 11%.
That's where we put the stake in the ground as we think about growth. Some people think it'll be higher, some people think it'll be lower, but that's how we thought about it, 11%. You can see the composition goes from about 25% to about 50% that's high-performance compute, high-end enterprise server market. You get growth in these other markets, and you need them because they consume a lot of semiconductors. What's driving the growth is the area of semiconductor revenue spending that KLA is the most relevant in. We think it's a big opportunity as the market moves, which is why we have a comfort that really across the board, that we can continue to grow our share of the overall market. Our customers are doing pretty well. This is the top five customers, their revenue trend.
We looked at their semiconductor-only businesses and also their EBITDA trend. You can see that the customer profitability is at pretty high levels these days. As I talked in the opening remarks about the outlook, feel pretty good about our customers' ability to spend more, and also that's driving their profitability. The structural profitability in the industry is pretty healthy. The trajectory over the next couple years and what's happening with underlying dynamics, we feel pretty good about where it extends over the next several years. I'm not gonna go into the details of these.
What I tried to do here was just summarize all the reasons why the KLA relevancy increases, why the relevancy of process control increases in advanced logic, what's happening in the memory that's now bridging sort of the delta in terms of how our customers spend on high-bandwidth memory relative to logic as a percent of their equipment spending. Of course, what's happening in advanced packaging. High sample rates, a lot of processing, front-end-like processing that happens in packaging. In this situation, we have rising complexity on process, lots of opportunities for our customers and very high-value devices to have challenges. As that roadmap scales, we think it creates opportunities for us moving forward. It's really hard to gain share in a market. What usually drives it is that there's a technology inflection or there's something that's changing in the underlying economics of the industry.
As we've seen over time, of course, lithography, the introduction of EUV was good for share of market, but also the growth in process control. Because we've gained share overall, the KLA process control grew faster than above. Over this timeframe, I picked a couple of other points. You know, start in 2018, start in 2019, start in 2020, you generally get the same results. That the two markets that are gaining share of WFE or share of the equipment market are lithography and process control. As you think about the packaging market, you can see back here, we don't really talk a lot about packaging. You'll see in the next slide, you can see why. You can see not a lot of growth in advanced packaging, and this is wafer-level packaging, includes dicing, includes film frames.
Eventually, we'll include panels, nothing today. You can see fairly flat, not a very high level. This is our internal forecast, but also aligns with where TechInsights and Gartner's forecast the market. Of course, you see this acceleration with high-performance compute over the last few years. What you have here on the right, and this is not a statement on performance or market share, but you see that as this market has moved to more front-end-like, all the traditional front-end players are targeting these opportunities. As we talk to investors a lot about share of market and how important that is, I think it's important to kind of get the denominator right, because everybody is reporting in their earnings or their revenue numbers, they're reporting in their numerator the opportunities that exist in this part of the market.
We think there was some confusion about how to think about it, but when you look across all the major markets, front-end players are targeting lots of opportunity and expected growth that's likely faster than traditional WFE over time. You can see why we don't really think about it to really maybe even fully understand it, right? We were 50 basis points share of that overall market back in this time frame. But of course, that's changed in a very meaningful way, right? From 2022 forward, almost over 400% increase in our share of the overall packaging market. Very relevant market. We think we're well-positioned. I'll talk a little bit more about why that's the case and how it affects how we've invested in it over the years.
Something that we're focused on, and we think it, again, creates opportunities over the next several years. A version of the operating model, the same sort of framework around collaborate, innovate, and execute. How does it affect how we drive the business and how that flows through the financial statements of the company? From a collaboration point of view, we have to work closely with our customers in early process development. We have to understand their process. It takes a long time to develop the capabilities to meet their needs. We have to stay close so we understand where the problems are gonna be, so we can pick and choose to solve the right problems. We can't solve every problem. We need to fix the problems that we can differentiate, solve uniquely, and those problems scale to production.
We have to work closely with our key suppliers. The supplier who does this work, this doesn't happen in a short period of time. We have to make sure that we can engage early enough to make sure our suppliers are ready. It's an important part of our differentiation. We have to make sure that given our structure as a company, that we can work across the company, reuse applicable technology. Ahmad talked about some of the examples there. We have to innovate. We invest heavily. Rick showed the trend. I'll show a little bit more of that and a little bit more context behind it. Well ahead of the market requirements, we are investing today in capabilities that drive revenue outside the target revenue window, so beyond 2030. There's consistent investment that's happening there, N+2, N+3 investment.
Of course, we have to invest and go get talent. We have to invest around the world. We have a worldwide talent strategy. We have engineers in a number of locations. We have to go to where the talent is. It's an important part about how we execute the company. We have to execute overall. We consistently meet against the public financial targets. I showed that performance earlier. A cadence of new product introduction. New product introduction matters a lot to KLA, and I'll talk about why that is. It's never over. You can always be more efficient and disciplined in your investments, in your operations, and so there's always focus on that. Critical supply chain.
I'm gonna talk about supply chain, but we need to make investments to ensure we have the capacity, not only can develop the capability, but also the capacity to meet our needs and then continue to deliver sustainable performance over time. This is our spend. You see R&D, this is what Rick showed you. We've talked a lot about applications, and it's about 1,600 people. I count these people. Ahmad said 1,500. It's actually close to 1,600 or will be this year. Applications, advanced degree people that are out working with our customers to drive value out of KLA tools, a really big differentiator for us. Obviously, a lot of that insight comes back and influences product development. 70% of our OpEx is technical investments. It's R&D or apps.
When we talk about improving our SG&A as a % of revenue over time, you have this in there and it scales with volume, right? More tools in the field, we need more applications people. When I talk about our plan and what it means, it's offsetting an area of natural growth. Now, AI will help us here as these folks start to use that capability more assertively, particularly as they help do recipe creation and so on. It's in general something that scales with volume. We have to offset that as we think about the long-term growth of the company. We're investing a lot in big programs. What this slide shows you is the number of programs in the company that are greater than $100 million. Engineering programs that are in development that are greater than $100 million.
It used to be a fairly small number, and we've seen that increase significantly. It's expensive, takes a long time to develop the products we have. Yes, R&D intensity is favorable, but there's a lot of investment that's happening. This is an area we spend a lot of time as an executive team in program reviews to make sure that we're achieving what we need to achieve in product development, because these are big bets, and they're important to maintaining and driving the model of KLA. Then, of course, the average spend per program has also increased about 70% bigger over this 10- to 15-year timeframe. Rick showed an example when we talk about what's happened in our investments in AI or in GPUs and how long that's taken.
What I have here are a couple of examples that we certainly are enabling high-performance compute, but we're a user, and we've been doing it for a long time. We started over 10 years ago. First product, 2018. Across the whole portfolio today, and a roadmap for more capability to optimize over time. If you look at the image computer GPU benefit versus a CPU baseline. The requirement capability is represented data rate in this VDP product, about 60% improvement. In the Voyager product here, laser scanning, closer to a 100% . If we had stayed on CPUs, it would be more expensive. We use more power. Because we converted and we did the programming work and moved the portfolio to GPUs, it's changed our cost curve, and you can see in our power usage, right?
You can see savings on power down 30%-40%, savings in cost versus the CPU baseline of about 40%-50%. If you think about what's driving customer's customer's customer hyperscalers, and how they're thinking about the investments they're making, and the robustness and cost-effectiveness of this compute, and the fact that they have to pay power. This is a pretty good microcosm, given our use cases and what their use cases are. Ahmad talked a lot about what our use cases are. We think it's pretty compelling. We're a user of it, we're an enabler of it, but we have to invest and think in the long term to make these things happen, particularly across the entire portfolio. What you have here, these are our products, product families, and they're not to scale, right?
You got Gen 4, Gen 5, which is obviously pretty big. You have component, which is pretty small, right? We go down all the way through here, and in a lot of cases, there's multiple products. This is the product introduction cadence across these different product lines. We introduce products quickly. We need to do it. We think it's important. Our customers want it, but it also allows us to do some other things. It makes us a moving target for competition, and that matters. Our salespeople always have new capability to sell, so that's good in any environment. The third thing is, particularly over the last few years, where we've dealt with a lot of cost dynamics and so on, is it allows us to reset as we're introducing new capability and improving cost of ownership, taking a look at price versus cost and make adjustments.
Of course, the last few years, that's been a big focus area. We are a value sell. It's about returns that our customers get. Because my costs went up doesn't necessarily change the return profile of what our customers see. We have to be thoughtful about that. The way we do it is we introduce new products. We introduce them a lot faster, and it allows us to do two things. Introduce improvement in the tool capability, cost of ownership, and share in some of that. It's an important part of the operating model as we execute. When you look at service, you look at manufacturing as the go forward, and we're gonna talk about our gross margin model. You can see incremental margin improvement expectations.
We think versus this 22-25 baseline, we'll see about a 1.3% increase in the points, or 1.3 points of improvement in incremental gross margins. Part of that are the value differentiation things you see on the right. Also part of that is our ability to scale our manufacturing organization. We have growth scale, obviously, given the growth rate expectations of the business. We're driving leverage through factory consolidation to our bigger factories, drive more efficiency and leverage on the investments we're making. Then a lot of focus on driving productivity and how we introduce products and how we deliver those products and drive quality and so on. You have a version in service, again, increasing value at the top on the right for service.
The incremental margin expectations in service are about 1.3x what they were at the beginning of the decade. One of the big drivers for that is that you can see service and parts expense. We have invested a lot to support our customers' regional efforts. As those regional efforts come to scale, you start to get a return on those investments. We would expect, on a go-forward basis, the incremental margins of service. Service carries a lower gross margin, but given the growth rate of systems and the improvement in the incremental, it becomes less of a gross margin headwind over time. There's also a lot of productivity, and Brian talked about it, so I won't belabor it around how we're designing for product development.
How we're leveraging AI and automation, parts planning, predictive maintenance to, as contract goes up, to optimize our cost structure under those contracts. That's the best thing about our contracts, is we know what we have to deliver to. We can make sure that the costs underneath it are optimized. There's a lot of focus in this area, and I think the combination of these things will drive some incremental gross margin improvement off of a pretty high level as we go forward. Supply chain. With the growth of the next couple of years, I'm getting lots of questions from investors on supply chain and can we ramp to support the business. I thought it would be prudent to spend a little bit of time talking about, well, what, how do we manage the supply chain? How do we think about it?
You can see our supply chain. I really break down our components into three types. Our complex subsystems like this, sole sourced relationships, long-term relationships with customers, things we have a few more options, but still long-term relationships and our higher volume critical components, and then more of a traditional supply chain in this tier three. Now, one of the things you have here in the lower left is that what is our time fence? When we start to order material to the time we ship it, the natural time fence is almost 60 weeks. It takes a lot of time to procure glass and deliver volume level optics. We have to do a lot to manage against that.
Back to my comments around inventory, where we buy long lead time materials, we build hedge kits to try to shorten that so we can action for customers early. Still takes time to do that. One of the issues in the first half of this year that we talked about at earnings is that the momentum in the industry picked up really, really fast. When we were making a lot of decisions around the first half capacity, we were inside a lead time and that momentum picked up. Now as we get in the second half, we have a lot more flexibility and capacity, and it's influenced the outlook I provided earlier. We have these decade-long relationships. We work very closely with engineering, supply chain, long-term commitments. We provide our suppliers eight quarters of visibility. We don't cancel orders.
If you go back to 2023 and 2024, the industry corrected a bit. Our inventory levels went up. I need them to be positioned to support us going forward. We've invested about $250 million over the last couple of years in supplier capacity. We invest a lot here. This drives the total volume here as well, and then we have the opportunity to manage all the usual issues that pop up in a constrained environment. This doesn't matter so much today. Rear view mirror. You all remember this. Maybe the best indicator of future performance is what happened in the past and the validation of the model, and I think we're better positioned today in terms of how we manage supply chain.
When the last time the industry was supply constrained, our ability to ship tools and get them out and recognize revenue was at a level that exceeded all our peers. I think we're in a better position today. We learned from that experience. We've leaned in. I feel pretty good about our ability to compete over the long run and to meet the market demand that's out there. I don't generally, for high-value components, we don't compete against our competitors because the relationships are mostly sole sourced. We compete within KLA, sometimes with these suppliers, but we don't compete. When industry strained, I'm not gonna have to worry about an allocation necessarily, but you know, you do have you know, the challenges of ensuring that you're well-positioned, and I think we've made the investments to do that.
This is the trajectory of the OpEx curve over time, and you can see the R&D curve, and you can see SG&A. As you look at our 2030 plan, we think the R&D stays about 11%, which is an improvement versus the last plan with 13%. SG&A basically comes down from where we were around 8.5% or so, 9%, down to about 7% moving forward. More leverage in our model in OpEx as we go forward. When you look at this R&D concept I talked about earlier, and I'll show some examples of this.
When you look about the pace of innovation, what you have here is you have, you know, different revenue levels in these five-year periods, and then what you have is the translation in the R&D as % of revenue in the middle. Then what you have over here is the ROI, so the return on investment. What this means is that we can invest over a longer period of time to sustain our products, to deliver new capability to customers. At the same time, we can improve R&D intensity, but also drive up our returns. Back to my point, as it moves fast enough, we can make investments, we can do them over a longer period of time, and it drives a structural profitability benefit.
When you look at a couple of our products, and this is the Gen 5 product, you can see we have to invest way ahead, so I'm accomplishing a few things in my messages here, but like six years. Then if you look here, you can see the revenue, then there's more work to do, right? Then it finally scales. You have to invest way ahead of the requirements. What happens today, because the pace of innovation is not every 18-24 years, this spending, if it were that tight, this spending would contract, and you'd have to do a lot more of it in a much shorter period of time. You're also investing in the next generation platform, which is the platform, meaning that it's a new architecture wavelength of a system.
You can change a lot here, but you can invest, and you can spread that investment out, you can optimize it over time, you can still deliver unique value to customers, and you can see what the revenue ramp looks like. In this particular product, you see an 8-to-1 return to KLA as we measure it. This is laser scanning, and you can see, in this case, much longer period of investment before requirements, and then we see a similar trajectory of revenue growth over time. The ability to spend over a broader timeframe all translates into lower R&D as a percent of revenue. In this case, it was a 4-to-1. Good examples. We have some that aren't as good, but in general, every product has a similar sort of investment cycle and then translates into new capability delivering customers.
Our customers benefit, and we benefit. Here's the example from packaging. Ahmad talked a lot about investing in the handling requirements of packaging, very different than 300 mm hard wafers, very consistent. This is macro inspection, which is the market that or the product that serves most of the packaging market. You see back here, this was the front-end revenue, this lighter color. What you have here in this is, this is the packaging investment that we were making over time to be able to support this market. Of course, we see this accelerate, and this represents the number one position in this part of the market. All this R&D is transferable to the rest of the portfolio.
The rest of the portfolio, which has higher capability, so when the design rules become more aggressive over time, we're very well-positioned to deliver without incremental R&D to support the requirements of that market. Very hard to compete against that. I think we're well-positioned, and ultimately, the operating margin translation because the R&D investment has been made and now ported, puts us in a very good position in terms of how it affects our R&D spending over time. Now let's talk about AI. The way we're looking at AI, not about in our products and product development, but how we're looking at it in terms of business operations across the company. We are a very data-driven company. We generate data everywhere. As you might imagine, we generate the most data in the fab, so that translates back into how we run KLA.
We have access, we have data, so how do we accelerate the access to that? All these things are little small things, but across whether manufacturing, service or corporate operations, you start doing them, picking them off, start saving time, and then it starts to fundamentally change how people work. We're big believers that we can do this. There's also process automation. Where do you have humans in the process, repetitive tasks and so on? There's actions that are happening here. We're also doing things that drive multiple vectors of return as you look at machine learning, mask correlation, pattern recognition type activities, where you're integrating a lot of heterogeneous data to drive either cost improvement, asset optimization, and so on, predictive maintenance. We're doing those efforts, bigger lifts.
We're spending money today, but we expect over time, we're gonna start to see meaningful benefit to the company in all these areas. You can see some of the things here. It's not a full list of some of the things we're doing that I thought people would recognize that are things that are in flight inside the company today. Ultimately, as I tell our employees, got to save money, got to bend the cost curve, got to drive up revenue per head count. What you have here is revenue per SG&A head count. We have a target that we think that we can drive it from 1.3x, the beginning of the decade baseline, to 1.9x.
It's a big factor in how we're thinking about how we're scaling SG&A over time and how we offset some of the dynamics related to applications and other drivers in our business that are more volume-intensive. Let's talk about the target model. When you look at the target model, what you have here, traditional WFE and WLP, and then you have the wafer equipment intensity over time. The way we thought about it was a range of 14%-16% or so, about 15% in terms of the overall capital intensity or equipment intensity in the market. We thought that this was a reasonable place to think about it with a range. Some people have a different view of growth, some people have different view of capital intensity, but plus or minus roughly 10% is how we thought about it.
Within that, the $215 billion, about 20, low 20s, advanced packaging versus core WFE. You can see a lot of the drivers of wafer equipment intensity over time. Our model is if semiconductor revenue is growing 11%, that WFE will grow, or wafer equipment will grow about 12%. If you look at what will happen, the next thing, what size of market, but then also how does it affect, how does KLA relevance change? We talked about, you know, we gained 50 basis points here, 150 basis points here. We think it goes up another 150. Part of our range includes, you know, whether it's, you know, 9.25%, so ± 25 basis points on 9%. You can see some of the assumptions in the middle.
Obviously, what's changing the most in our long-term revenue growth model is what's happening structurally in the industry. Then what are the contributions, the overall growth rate from systems, which is gonna be between 1% and 3%, the overall service, which drives an incremental 1% CAGR on our long-term growth. That translates, and you saw versions of this earlier, right? $12.7 billion in 2025, $10.6 billion for systems, $2.7 billion for service, gets us to $26 billion. If you look at semi-cap service or semi-cap systems, 16.5% CAGR against the industry at 12%. We think we're positioned to continue to outperform, and that's how that 1.5% on the 2.15% translates back in terms of growth rates relative to the market baseline and our expectation from our business.
In terms of capital allocation, given our expectations about, look, we like the businesses we're in, our focus is on driving those businesses and capitalizing on this opportunity. Most of the capital, if you're looking at the capital base and what's changed from 2022 to 2025 versus 2026 to 2030, will be share repurchases as we deploy that cash. You can also see a change in expectations around SG&A, where it was 14% coming to 11% as an indicator or driver of scaling the model. We need $4 billion-$5 billion to run the company. We think that's a reasonable and conservative place to be in how we'll think about cash on the balance sheet as we move forward. Leverage ratio target at 1.5x-2x gross leverage. We're about 1 today.
Have some capacity implicit in our ratings that we would flex that over time, but live more or less in our target range. You can see the announcements we made earlier. Here's the target model. You can see 2030 model versus 2026. $26 billion ± $2.5 billion. 63.5% gross margin ± 50 basis points. Obviously offsetting what looks like a 50 basis point headwind related to tariffs if things kind of continue the way we think. R&D at 11%, SG&A at 7% translates into 45%-47%, so 46% at the midpoint, up from 42% at the midpoint. $8.4 in earnings, $8 ± $0.8 in target, free cash flow of 90%. You can see the macro assumptions on the right.
One of the assumptions around mix of business, we expect foundry logic to be 60% plus, but kind of in that kind of lower end, you know, 60% plus a little bit, as we go forward. I think. You know, we've got our EPC business, which is not in the semi-cap calc, is, you know, grows more mid- to high-single in terms of how we're modeling it. Tax rates are about a point higher than the last plan for completeness. Before we move to the Q&A, I think the key takeaways from the Investor Day today, first, you heard Rick talk about best-in-class outperformer. We think we're uniquely positioned. We understand the market drivers. We think they favor and drive KLA relevance. We're excited about them. You heard Ahmad talk about, we think we understand the market pretty well.
We have good strategies. Our portfolio is well-positioned. We can meet customers' technical requirements and their economic ones, and those change in the different stages of maturity in a process node ramp. As Brian talks about the service business, great business that gets better in every metric as we go forward. In terms of our 2030 target model builds on a credible history of execution and operating leverage growth and consistent capital allocation. Structural semi growth is changing and raising. We like it. Industry 12% WFE, 11% semi. Intensity and share gains greater than 150 basis points. Operating leverage through the model at the high end of our operating margin incremental operating margin target. Continuation of a disciplined capital allo
cation and return strategy leads to durable EPS compounding.
I think the answer to the question as I asked earlier before is it's gonna be better in the next five years. Thank you.
Now it' s time for our Q&A session. Please welcome back to the stage Bren Higgins, Rick Wallace, Ahmad Khan, and Brian Lorig.
Okay. We have two mics, one over here, with me, one over on the other side with Ed. Raise your hands obviously for Q&A, and I'll come to you. Start over here with Atif.
Thank you for the excellent presentations. Ahmad, I have a question. You made a good point that as the HBM-based die becomes a bigger portion of the area it plays into KLA's hands in terms of higher process intensity on the logic side. My question is, if the industry were to pivot more towards SRAM-based accelerators for ultra-fast inferencing, how will that impact, obviously some opportunity will come out of HBM and play more into the logic side. How will that impact your process intensity?
Yeah. Yeah, great question. First to clarify, I didn't mean to say that just the base die is logic. What I was actually emphasizing is the DRAM die itself, which is above the base die. The logic content of that DRAM die is increasing significantly because of the sense amps, which is closer to now logic. Plus you have a base die which is all logic. Those are the two factors today. We love SRAM. In short, SRAM is very logic-ish, and it's hard to yield. And our process control intensity in SRAM is high, and that's the reason why process control intensity and logic is high. SRAM would be beneficial to process control. The trend is real.
We'll allow it.
Speaking about it. Yes.
I don't have anything to add to that one.
Stacy Rasgon with Bernstein. Bren, I wanted to ask you a little bit about the medium-term numbers that you gave. You know, as strong as calendar year 2026 WFE is, it's a constrained year. We're missing cleanrooms and everything else. They come online, you know, starting in the next year. I guess what I'm asking is like, you know, why would WFE not actually grow faster in 2027 versus 2026, given the constraints on cleanroom capacities? I guess number two, is it reasonable to assume that you should outgrow WFE in 2027 as that happens? Your lead times are longer. You know, you're maybe more dependent on those cleanroom builds, and your overall market seems to suggest you can outgrow WFE as process control gets-
No, yeah, the first question, so I did say at least as fast. I think if you look at expectations moving forward, we feel pretty good about the growth rate. So I think, you know, look, we'll see. We're in March of 2026. We'll see when we get there. Yes, there's a ton of momentum. There's a lot of sites that are being built. There will be customers who are trying to align schedules with that construction schedule, but we feel very good about how well we're positioned, what we're seeing in terms of customer engagement, what we're seeing in terms of backlog. I think as you move to greenfield, I think it also creates opportunities as new fabs are fully equipped.
I feel pretty good about our ability to continue our track record of outperformance over the next couple of years.
Just the current calendar year, you took it up to high teens, right? Versus whatever it was before. My math suggests that's something like a 20% half over half, depending on what you wanna assume for next quarter. Is that kinda what you have in mind? That seems to be where it's dialing in.
We'll see how this quarter finishes and how we guide June, but yes, I would expect the half to half to be pretty strong in the second half. The overall company high teens, the equipment business, semi-cap will be closer to 20%.
Okay, thank you.
Based on what we see today. We'll see as we move, as we go, you know, what happens in terms of, you know, customer momentum and whether that changes. We have more flexibility in the second half. We'll see how that goes. For now, that's where we are. We're moving back to this side.
Thanks for taking the question, Joe Quatrochi at Wells Fargo. Maybe one for Brian on the services side. How do you think about just kinda, you talked about some of the ramp in services you've seen in advanced packaging and in DRAM, but how do you think about, like, the service intensity across the different end markets, and how does that changing, you know, as we move to 2030?
Yeah. Thanks, Joe. Certainly there are some differences across the different segments, but we also think about servicing products. The products, as Ahmad outlined, are now being used, traditional front-end products used in advanced packaging, used with tighter requirements in HBM, and so service intensity increases in those areas. That's part of the reason why we showed that, you know, really good growth from 2022 to 2025, and those are catalysts for growth as we look out through 2030.
Joe, I think some of the billable business happened in memory segments, packaging segments in the past. As the requirements for those customers and the capacity constraints that have, you know, given, you know, sold out a lot of cases, has started to change the service mindset of a lot of those customers. As we think about growth, that's one of the elements of growth is that service mindset is gonna change because the value of the process control and the value of what they're inspecting is increasing.
Thank you. Chris Caso from Wolfe Research. The question is the mix of advanced logic versus memory within this forecast. You had talked about, you know, obviously, process control intensity increasing in memory side because of HBM and such. Does that mean that memory will become a larger part, percentage of your business as you go through this period? Obviously, that also has to do with the relative growth rates between those markets.
Yeah, I think so. If you would go out a few years ago, the industry was, you know, high 60s, mid-high 60s in terms of % of equipment. Our model here has, you know, a view that it moves from that into the low 60 range. We'll see how it plays out over time. The changes in intensity in memory mean that the mix effect has less of an effect on KLA. But we thought in terms of how we would model it, you know, low 60s made sense given the growth expectations that are happening in and around high-performance compute.
Put differently, because of that, then actually, perhaps you're more agnostic as far as mix than if memory grows a little faster.
Correct.
The logic grows a little faster, maybe that doesn't actually make as big of a difference.
Correct. Yep. No, that's true. Look, in high bandwidth memory, we are seeing in some product lines today, we're seeing intensity levels that are as high as high-end logic. As that's a bigger percentage of the total, it creates opportunities. We'll see how that flows through the rest of the portfolio. Yeah, it's the business is less sensitive to the mix of business moving forward.
Okay. Over here.
Good morning. Thanks for hosting this event. Harlan Sur, JP Morgan. You know, I caught up with one of the largest fabless AI XPU ASIC chip companies recently. They design a high volume custom XPU chip for one of the top hyperscaler titans in the world. They're currently getting about 60% yields at 3 nm. An increase of those yields by just 3 percentage points, right, drives incremental revenue per wafer that pays for the entire cost of that 3 nm wafer. Their foundry partner is already at historical targets for defect density and yields. In this tight supply environment, larger die sizes, customers are pushing their foundry customers to drive even higher than historical yield targets, and that's on their foundry partner's existing capacity footprint. That's not even on next year's greenfield fabs, right? These companies need these chips now.
Well, what is the KLA team doing to sort of help your customers on this front, improving yields beyond target historical defect densities on existing technology, existing capacity footprint, and how is the team monetizing this opportunity? I assume it's both the tools and services opportunity, but wanted to get your views.
Ahmad, why don't you?
I didn't fully cover this point earlier, but we do some estimates on what the incremental revenue generation is with improvement of yield. Rough numbers are 1% improvement in yield could lead to at least $150 million in profit because you have already spent the money buying the equipment and everything else. It's a pretty substantial number, and this is one of the reasons why intensity is going up. You keep seeing this. That's why I felt comfortable with this 7.4% number going to 9%. On how we do it is what I described. We are working very closely with all of the high-end logic customers to drive yields up.
It is a very complex thing to do. It requires time and effort, and eventually the yields do go up.
Is there a services aspect to it? In other words, helping your customers maybe implementing different yield implementation methodologies and so on?
We do have a team of people we call PCS. These are not KLA experts, but they're device physics experts. We send those people out to specific locations where they will have a yield problem on a particular module. We deep dive it, and we go work on those issues. This team is in very high demand. I don't have enough of these people, and they're able to really help customers drive yields up, yeah.
The sampling's pretty valuable, right? As it translates to not only the performance of the systems across the fleet and that they have to match, but also the availability at that performance level becomes even more critical. It is a factor in how we think about, you know, how we monetize, you know, the opportunity as we work closely to be able to enable that. There's a service element that's, "Hey, look, this information's as valuable today given the cost of what we're inspecting." The fleet needs to be up and active, and they rely on us to enable that.
Harlan, it's not just in the case of the one you're talking about. Even with leading-edge customers that if they get increased demand at legacy nodes and they're yielding in the 80s. They recognize more yield is worth a lot, especially when their customers feel supply constrained. When we talk to our customers, they're all frustrated that they're supply constrained, like for sure.
Thank you.
Yeah, hi, it's Krish Sankar from TD Cowen. I have a question for Ahmad . If I heard it right, you mentioned E-beam revenue was about $450 million last year. Seems like it has grown pretty nicely. Is it mainly coming from the multi-column product for reticle, or are you actually gaining share on the wafer die size also?
First, I said $400 million and not $450 million. Second, the mask part is not included in the E-beam revenue number. That is a separate segment. Therefore, I didn't mix it, even though the mask system is E-beam related. Those are very expensive systems, and the numbers would go up pretty significantly if I add those. This is a wafer part.
Yeah.
It's in its e-beam. It's e-beam inspection.
It's E-beam inspection, which is a single beam E-beam inspection, very high resolution, between 1 nm-4 nm. It's voltage contrast, also E-beam inspection, where we look at electrical defects. It is multi-beam. We are very proud of introducing a 300 beam, multi-beam system for primary voltage contrast applications, and then eventually it'll go to physical as well. It includes E-beam review, which is after you do a brightfield inspection, you review your results, and that's the E-beam review part, small portion of E-beam metrology. Those are the four segments that equate to that rough number of $400 million.
Coming back over here.
Thank you. Vivek Arya from Bank of America Securities. Thanks for hosting a very informative Analyst Day. Actually, first, I just wanted to clarify. I think we kind of know the answer, but still, any impact disruption from whatever is going on in the Middle East, whether it's to you, whether it's to your customers, right, in terms of any supply disruption? I hope you know the answer, but still would love, you know, anything that you might offer.
Not currently any disruption.
Okay, very good. My main question is, if I go back to your Analyst Day slides, right from the 2022 Analyst Day, your assumption at that time was foundry logic was gonna be 55% of the mix, right? It ended up being much higher. How much of the share gain that you had was just because, you know, the foundry mix turned out to be a much bigger part of the industry. If let's say over the next few years, if, let's say memory becomes a bigger part of the industry, I don't know, 45%, 50%, whatever, just because there is so much faster growth expected in memory, what does that do to your share gain potential?
Because I think the assumption is that if there is greater share of wallet that goes into more greenfield, more memory, more capacity rather than technology, that it naturally favors, right, some of your competitors. Would love to hear your perspective on that. Thanks.
Well, as we said in the earlier question, right, we're a lot less sensitive going forward than than we were historically, so I'm less concerned about that. The other thing is our long-term model assumes a pretty high level of efficiency in the industry, right? If you think about the last five or six years, the leading edge was incredibly efficient. If anything going forward, multiple players that are investing in leading and near leading edge. That creates kind of a market inefficiency that I think that would create an opportunity for us. The other thing is that I put up those R&D charts where I talk about R&D efficiency, and I showed Gen5 and I showed Voyager, and you see the revenue ramp, and then it dips down and then goes back up again.
There was a period of time of a lot of legacy investment, and we think the legacy investment, which has a lower process control than advanced logic, will be a smaller percentage of the business going forward. I think the legacy business will be 25%-30% of overall WFE, and all that is generally logic investment in terms of just what's accessible to us. I think for a combination of those factors, we feel pretty good about the plan we laid out.
The only thing I would add is when we were making the slides for 2022, the chips were very different than we were making the slides for 2026. These chips are very different. Large die, much more logic. It's a very different packaging, the way it comes together. It's a very different world.
No difference greenfield versus upgrades in memory, NAND versus DRAM?
Greenfield requires a lot more process control in the beginning because it is difficult to ramp a greenfield. I think to your earlier question on 2027, there's gonna be more greenfield and that is also very beneficial.
Not to confuse things, but there's this blending happening too. As you know, the foundries are making memory, and the memory guys are making a lot of logic, right? I think the bigger than that mix, the bigger change that we're experiencing is the move to the hyperscalers as customers and away from consumer. The consumer tolerated a lot more than the... And just a GPU versus CPU example, the guys that made CPUs used to bin them based on defectivity, right? You can't do that.
You can't do that.
... in high performance compute, right? You can't bin them. All the stresses in the system have gone up because these things are going into data centers, and all the stresses in the fab have gone up because more people are doing designs. I think that's the tension that we're seeing with our customers in the fab, is that they've got tremendous pressure from so many more people driving designs. We really didn't anticipate that in 2022, how dramatically that was gonna change. I think my comment earlier about you know, how much ChatGPT is being used or how much these reasoning models, that's just driving a tremendous demand for compute.
I mean, that's part of why I think this industry is gonna underserve quite a bit the demand that's been out there for the next few years, and the bottlenecks will move. Memory is a different game for us now, for sure. With high bandwidth memory, it's a different game.
Packaging, right? How that influences, you know, the share of the total equipment. We, you know, as we add that in and look at that over time, those numbers are based on the total market opportunity. That's a factor that wasn't something we considered in the 2022 plan for 2026 and is a more meaningful factor going forward.
Yeah. CJ Muse with Cantor Fitzgerald. Thank you very much for today. Great presentations. Wanted to focus on advanced packaging. When you think about process control share gains of 150 basis points over the next four years, how much of that do you think will come from advanced packaging? Moreover, when you think about how many more kind of dies are going into these packages, are you seeing you know the vision for more process control insertions, and are customers, as the back end looks more like the front end, willing to pay higher and higher premium, you know, versus, you know, five years ago where, you know, the ASPs were significantly lower than, you know, what you guys do typically in the front end? Thanks so much.
Maybe Ahmad will tell the story, and then you can fill it in. The conversation we had with our customer a few years ago on packaging.
Which one?
The one in Taiwan.
Yeah.
So-
Well, I'll start then you start.
Yeah. I got pulled into the meeting with Ahmad, and usually these guys talk about, well, it depends where we are in the cycle. You know, your stuff is too expensive versus we have these missing defects. But what we were getting was, you have to help us with EUV adoption, and their second topic was packaging. This was out of the guys out of the front end. We're like. Literally, Ahmad and I were like, "Are you sure? Because we're not gonna discount our equipment for packaging." They said, "No, we definitely need you to take your portfolio and modify it so it can work in packaging. And we're in Taiwan." They said, "We'll have a team in California next week.
Next week.
to go over your plans." With this customer, you respond.
It was December tenth, and next week they were gonna show up, so we had to figure out how to get them to come on January 5th. But that's the intensity, not the process control intensity, but that's the intensity we deal with. I will leave the share comment to Bren. But about insertion of process control, I'll take that one. As you know, in HBM, there is two different processes that are being used to stack dies. There's one company that has a process, and everybody else uses a second process. And there's more latitude in the first process, but almost no latitude in the second one. When you go with the second process, as the die stack goes up, your failure rates are on a log scale.
It's not very easy unless you do heavy process control. This is part of the reason why people are thinking of doing hybrid bonding, and they may do wafer-to-wafer bonding and hybrid bond, and they might do that. In that case, process control intensity goes up. Now, why does process control intensity go up? We participate very heavily in CoWoS, right? And we do a great job for our customers in CoWoS. Those same customers are trying to do 2.5D SoIC. All of that is hybrid bonding. The specs I showed you for hybrid bonding were 10x tighter. 10x tighter. Our Kronos systems are not going to be sufficient. We believe Puma will be successful, Voyager, and most likely, you will need to use brightfield systems for doing that type of inspection.
Brightfield, we need to ship the system with that thing that is on that corner and this thing. The cost is just. It is. The value of the package is quite high. We believe that the future of packaging, the specifications are gonna get tighter. Now, I don't personally like to get into the when which roadmap will happen. We are gonna grow intensity no matter what because hybrid bonding will drive intensity, but even the TC-NCF bonding drives intensity up. We're good either way. We're not choosing, but the trends are all positive. We think it continues to grow because of that. That's why we feel comfortable with signing up to a higher intensity number.
Look, we're gonna be greater than 50% or in that ballpark in 2025 of share of process control. Most of our share today is more concentrated in leading-edge logic. There are opportunities in memory as we move forward. Some of the things that are happening in logic will drive higher value systems. There's OSAT engagement as you start to see some offloading to OSAT. We feel. I'm not gonna put the stake in the ground of a specific share number, but we feel very good about the trajectory of our share, and we think we're pretty well positioned with existing platforms that customers know how to use that have demonstrated value in the front end that are very applicable to the back end or the packaging requirements.
Okay. We're here over on the far, far left. Your far left.
Thank you for letting me ask a question. I'm Shane Brett from Morgan Stanley. One of your leading-edge logic customers has more openly talked about opening up their foundry data externally and how external vendors, including yourself, has helped drive up their yield. To the extent that you can you talk about how you contribute to that yield improvement and how this broadening of leading-edge logic spend contributes to your growth? Thank you.
You wanna take it?
I spoke about earlier. I will give you another preview of what we did in front-end and then now we're doing in packaging. All of the systems that KLA makes and all the systems that our competitors make generate petabytes of data. We had this vision about 15, 18 years ago that the complexity is gonna go up, and we need to provide a solution by which customers can organize the data, and then after that run models. We have a system that enables us to collect all the data that comes out of KLA systems. We're agnostic in this area as to what customers buy, so we will take competitors' data also. Organize it.
We have KLA models that enable you to find correlations, meaning that if you have overlay performance problem in a particular layer, does that affect defect? And we can correlate those two. And we can now do, like, six to eight parameters of different types of things that are causing a problem. The system is also open. Customer can write their own algorithms and import this on. We have now changed that from wafer to tiles, and we're now providing that same system capability for advanced packaging. Every die can be traced back to every wafer. Every stacked die can be traced to wafers, chambers, everything. We are developing the advanced packaging module. I hope that answers the question. I see. You're talking about that part.
We're fully engaged with them, as well. It's hard for me to give specific customer details, but as I said earlier, one of the first things we are doing is using our process control team, the PCS team, to help segment their line and figure out where the defect sources are. Our engagement is pretty open. I mean, their CEO has made comments about KLA engagement with them, so we're very engaged in driving and helping improve yields.
Collaboration levels are high, and we talked about the broader base of investment we're seeing at leading and near leading edge investment. We're optimistic. Again, the collaboration's high.
Yeah. I think once you have to go to foundry model, you cannot bin devices. That's one issue, and the second is variability goes up because as designs come in. At 10 nm, which is the last successful node for some customers, you know, the margins were pretty large. If you go to 2 nm, that variability is very tight. I mean, you just cannot use that book. You gotta let go of that book, all the things you have learned. 2 nm is a very different world. 3 nm, 5 nm, very different world.
All right. This is Charles Shi from Needham. I have a question about the reticle inspection side of the business. For what it's worth, there's always some debate about KLA, the inspection metrology technology. I mean, probably more than 10 years is optical versus E-beam, and I mean, 80-20, I think that debate is probably more or less settled. But more recent one, I think it feels like it's about the reticle inspection. It's 193, Gen4, Gen5, Print Check versus actinic. Now I think you talk a little bit more about E-beam. So going back a couple years, I think you kinda tied the actinic inspection adoption to, at least on the wafer fab side, to the pelliclization.
I didn't hear much you talk about pelliclization, but we kinda feel like pelliclization rate seems to be going up. Does that change how you feel about that mix of the market between 193, Print Check, E-beam and actinic, and is pelliclization still a factor there?
Yeah, let me start, and then you take it. Okay, this idea that it's settled, it's just never settled with you guys. I mean, I remember hearing 20 years ago when I became CEO that optical was gonna run out of gas and e-beam was gonna take over. Look, the thing you ought to remember is we have a portfolio. We're typically attacked with products and with messaging with people that have a single product. What they're gonna say is that product is gonna take over, right? Because that's kind of gotta be their pitch. In general, it's not really how our customers respond or behave. Our customers keep trying to optimize the cost of inspection, the efficiency of it, and to get it done, so they use this portfolio, which is what we tried to explain today.
Specifically, as it gets to this question around actinic, and Ahmad talked it, I'll let him take it, but I think in general, it's easy to take a shot at us and say we're gonna win in this certain segment, but it's not actually how the market ends up behaving. If you see the products that are being used for reticle, it would be nice if it was such a simple thing that one system did everything. It's never actually been that way. There's multiple systems that provide the capability, which is why in aggregate, if you look at KLA's share over time, and if you plotted that against all the announcements of people taking share, right? It would be hard to reconcile the two.
Right. I'll try to be concise, but it's a... You asked a very detailed question. Just for everybody else, this is the mask, and there is a plastic film on it. And that's called a pellicle. This is a 193 reticle, not an EUV reticle. If there's a plastic thing on it, then how can a particle drop and affect the design? That's where the debate is. EUV reticles are not sealed pellicles, number one. Even if you put a pellicle, the particle can get in from the side, number one. Number two, if you add pellicles, it reduces the throughput of an EUV scanner, and therefore, now you are spending more money doing that. The pellicles do burst, and because of that, you have to clean EUV scanners. There's costs associated to it.
There's two religions here. One, people say, do not do pelliclization. Do very, very few for very few layers and do heavy metrology, and that is one religion. There was another company that has a different religion. Now, they're changing, which is you pelliclize everything, and then you don't have to worry about it. But the problem is pellicle is not sealed. So that's where the debate is. Now, if in the future, there's gonna be CNT pellicles that are very clear, and then it wouldn't matter if you pelliclize or don't pelliclize because a DUV system will work, an EUV system would work, E-beam system would work, because E-beam is done before pelliclization. So I don't think there is a big technical debate. We truly believe that this is a portfolio solution. Why? Because of the requirement of 100%.
If the requirement was to just catch 70% of the defects, I'll build one tool, and it'll work. The problem is that they want 100% of the defects caught. That's why it's a combination of DPUV, E-beam. We are developing an actinic inspection system and Print Check, and Print Check doesn't care about pellicles. I hope that is clear, but this is a very detailed subject, one I'd be glad to spend more time on. It's not lack of confusion in our mind.
We will have more time for that at lunch. I'm looking at the clock. We're gonna have time for one last question here in a moment. It'll come from the front. Just some final announcements just before we end, 'cause I know then otherwise everyone will be moving and up. Wanna thank everyone again for, obviously, your attendance and your attention, and we hope it was both informative and educational. We will be moving outside after this last question, where we'll get the tables will be set up, and it will be a round robin set up. Get your food, go sit down. The executives will fill in the tables, and we'll be rotating every 15 minutes for the next hour. Of course, the demos will remain open until 2 o'clock.
Some of you folks got here early and saw them. That's great. If not, I highly encourage you to. Last but not least, if you didn't get a chance, we're gonna have our CTO, Ben Tsai, here during the lunch hour and afterwards to talk through a lot of the technology showcase items that are here in this room. I encourage you also, if you have time to come back in and explore that a little bit further. With that, we'll have the last question coming here.
Hi. Thank you, and sorry to keep everybody from their lunch. I'm Melissa Weathers from Deutsche Bank. You talked about how you're users of AI in a lot of your products and how that's helping to improve your own performance. I assume your customers are also trying to use AI as much as they can to improve their yields, and even some of your semi-cap peers are probably using it as well. Can you talk about what are you seeing from them in their efforts to improve yields on their own through AI processes?
Yeah. In AI, something very important is context, right? Our machines provide us immense context for problems that we can zero in. That context comes from design, it comes from optical architecture, it comes from the patch we're looking at. That problem is what I would call a high variability problem. That's what we do. Slow variability problem, temperature of the fab changing, you have some other issues, humidity changing. You have all sorts of others, overlay over there changing, and then CD over here changing. Those types of things you can do on general data. This is a collaboration. They do that type of work, we do this type of work. I don't think this is in conflict for process control.
Hybrid metrology ten years ago was gonna just take over everything, but nothing. It doesn't work. You need sensors on the problem. I think we are working on two different things. We are actually also working on that based on that thing that I talked about, the entire fab and data, all of those things, and the customers can plug in their algorithms. I don't think we're in conflict here.
I think in terms of in the process world, people using it, we have a slight glimpse into that 'cause we have a division that makes process tools. There's just so much less data. I mean, most of what we are is data processing and image processing as a company, so we're naturally more inclined, and we've been working on algorithms literally for the whole 50 years of the company. I think that for you know, ASML has talked about having AI inside of their systems. They have a lot of parameters inside to try to optimize. I think this notion that we're all gonna use it to be more productive is true, but there are certain things that it's you have to have access, as Ahmad Khan says, to the data, and then you have to have.
The challenge in our image computing is it's such a high data rate that we have to process it locally in order to be able to provide information to our customers. It's been quite a driver for us, and we think that it's even inside the algorithms we've seen developed in the last six months are giving us more capability of the existing platforms than what we thought six months ago.
Six months ago.
That's gonna be really important for our customers because it'll enable us to provide even more of an upgrade path as we go forward for these investments that they've made, that we've made.
That will conclude the Q&A session. Again, thank you, everyone. It'll also conclude the webcast. I wanna thank everyone who was remote, who joined us as well. Thank you very much.
Thank you all.
That concludes the formal part of our Investor Day and webcast. Now, please join us for lunch with management, and make sure you check out the AR/VR models of KLA products before you leave.