So, welcome everyone. Before we start, I would also like to welcome the participants from a distance, who we don't see, who we don't hear, but who are listening in to the call. Welcome. I was told it's 46... Sorry, 70. Anyway, I don't know where the number 46 comes from, but anyway, so supposedly it's 70 additional. So welcome. I hope you had a good tour, that you got some more detailed impression of our different beautiful machines, also some thoughts behind developments, where we're going, and now we will summarize that with a presentation or a set of presentations. One about Besi in general. Number two, about, of course, the famous-- Hold on. I have to push this one? This button. Ah, the Safe Harbor statement, to repeat that also for our friends online.
This is an Investor Day concept, so this is not a market update. All information we share is information either generally known to the public, or it is assumptions we base on customer inputs. But this is not about our guidance; this is not about Besi's quarterly business. Anyway, here we go. Our agenda for today is, first, a few words about Besi. Number two is our technology trends, market trends, we call it. So our CTO office, Chris Scanlan, who you have met in the group, if you were, or if not, please, Chris, stand up. That's Chris; he was also here last year. Anyway, then we have Peter, who has the honor to present our wafer level progress, hybrid bonding, TC, but also other concepts in those applications. Then we have mainstream die attach. Christoph. Christoph, you've also met, who's standing in the back.
And then we have a Q&A session. But let me quickly start with the Besi update. So what's happening in this world? Well, not a surprise to you. AI is driving the bus, and the bus is going faster and faster, with more and more stops, with more and more development, which should, if all is right by the outside, analysts, drive this industry to significant higher levels than we have witnessed ever before. Assembly has become from a necessary evil or the flea on the dog's tail, that's how I was brought up 30 years ago in this industry, to a key enabler to further integrate on chip level and to create interconnects which are at chip level. So assembly has grown in importance, and we will share with you many of those developments, which will be critical for growth in the years to come.
Although it's still early stages, we are in a component world, but ever more on chip level, we can connect, and based on that connecting on chip level, we can create much faster circuitry, use less energy, and especially where it is about manipulating data, we need to offer solutions which generate less heat, faster circuitry, and the enabler for that is hybrid bonding, and to some extent, also TCB. But anyway, 35%-40% of our orders are coming from this new technology world. Many of you may have concluded already with the ongoing reporting quarter by quarter, the last quarter, end of April, where although the industry is in a major, let's say, downturn, Besi still performs very well in that downturn simply because our successful focus on the forefront of this advanced packaging technology.
We service the major, or the leaders, you can say, in this industry for those devices, and the adoption is spreading from logic also gradually towards the space of memory. Anticipated ramp of both combined in the most advanced applications is getting closer. 2026, 2027 is expected by outside analysts to be the kickoff of those modules using hybrid bonding technology. We've seen orders for 2.5D capacity began on a larger scale in the second half of last year... particularly for CoWoS, but also photonics, but more about that later in the presentation of Christoph in particular. But also Peter and, and, yeah, Chris, will also from a technology background, share with you what is driving this and what are the unique opportunities for Besi.
Memory customers are certainly pursuing a dual strategy in this next generation, trying to further stretch existing technologies in reflow, some in mass reflow, and then gradually towards the transition for hybrid bonding, simply because the geometries force this transition, apart from the big advantages of energy savings using hybrid technology versus a reflow process. Besi further expands its R&D, which is of course a no-brainer, because of all these wonderful challenges in AI. If we look at this chart, it tells you the forecast in the years to come. The source is Yole, where it's clearly predicting a major growth and well above the average for the industry, both for 2.5, 3D, but advanced packaging in particular, and that is where Besi's focus has been for the last 30 years.
If we look at what I already said earlier, it's amazing that in this industry, we all thought after COVID, supply chain issues, and, yeah, the virtual world, which changed society forever. So at that time, in 2021, 2022, the sky was absolutely the limit, but everyone forgot that the world always comes back to models which we have used in the past. Of course, some new models, but in general, that meant that this industry had an enormous overcapacity. And different from past cycles, where usually six to eight quarters, the industry would absorb an overcapacity, we are now faced with a bit longer, which is also not unusual. We had that in previous cases, the IT bubble, for instance, but also other moments in time. But anyway, it is what it is.
But as discussed earlier, for Besi, that offers unique opportunities to also grow significantly in new technologies. It's expected that this tide turns, although if you look at the latest independent research analysts, they all take their models somewhat down for 2024, maybe too much looking at statistics and reality is catching up. So if we look at our customers, the whole infrastructure in this industry, H2 is expected to be better than H1, but still with some caution. Anyway, general information. If we look at our performance in this downturn, I won't go over the details, but in any case, reaching gross margins of well above 65% in a situation where revenue is more than 30% down, tells you that our operating model is as unique as our technology development.
So in any case, that gives us an enormous confidence in to develop this business in the years to come. Capital allocation always has been strong. Last year, very significant, returning over EUR 430 million to shareholders. In total, in the past 10 years, I think it is now about EUR 2 billion. Anyway, technology is what counts, and we are definitely enjoying at the forefront of next generation technologies, wonderful positions. If that continues, that can offer us a much stronger growth than what we have witnessed in the past. This is to remind you that this industry, back to 2006, can you imagine?
We acquired Datacon here, where we are in 2005, and if you look at this wonderful revenue per quarter development, it definitely tells you that this is a cyclical world, and that won't change. But at the same time, you see the continuous development in downturns, preparation, upturns, demonstrating ever higher margins. That means our focus on customer developments. Pick the winners is one of our major, let's say, strategic focus is never ending. So pick the winners in the three categories: communication, data, and automotive, and at the same time, continuously improve our operating model.
Our operating model, in that sense, is unique compared to others in the industry, multiple sourcing from the start, but then also through contracts which enable us a flexibility, both in up terms, where we grow significantly faster than the average in the sector, but also in downturns to maintain our break-even cost at the lowest levels, but adapt the cost that we maintain margins which are better than we have performed in previous cycles. So that also is very important to understand about Besi's performance. If you look at this slide, it helps you to understand who is doing what. We all are, let's say, reminded every single day that this industry is preparing for a next major upturn, but that's in every downturn the case. But this time it's even more than that.
If you look at the massive investments announced by the leaders in this industry, and, especially for 2024, where we are today, that tells you that is based on a market which in the end, also needs assembly. Otherwise, these front-end fabs are not of much use. But in any case, a major investment in capacity in all technologies, and that is what is for new technologies, Besi's unique opportunities. This was too fast. Hold on. State of the industry, what is important this time, and I want to point that to you here on this slide, at the right side, that we have, for the first time, a very significant imbalance between value, revenue, and inventory. That is one of the problems.
We had, of course, a period of major inflation that has inflated pricing revenues, but the capacities still are sufficient to, on a unit basis, respond to demand. That is why this time the correction cycle takes longer than in previous cycles. Here are some more details about planned investments, and it's all in progress. There's an enormous investment going on in extending front-end, but also back-end facilities. Look at the Amkor, for instance, in the U.S., but also many others. We can expect, at some point, return to growth once again. What does that mean, in particular for assembly equipment?
We see the peak reached in 2021, where, supposedly, according to TechInsights, the market was about $6.6 billion, which is very important in itself, because the market until that time was always between $3 billion and $4.5 billion. So for the first time, partly because of, COVID, but mostly because of new advanced packaging, demands has grown above that, that high water mark. It has come down because of the, the overcapacity in the industry, and supposedly this year it should recover, but, in a single digit, number, as is expected at this moment. Although, if you also bear in mind that is added to this market size, the hybrid bonding.
So anyway, it reflects the same pattern as for the whole industry, but then expected gradual return to growth in the second half of this year in this model, and leading up to a next growth cycle, 2025, 2026. The question is, of course, do you have the right products? If we look at that question and we look at market shares, the only conclusion you can draw from this slide, which is a bit updated, but also already presented end of April. Besi has gained, once again, market share overall in the assembly market, in particular in the addressable markets. Then if you look for die attach, but also for packaging, Besi has done pretty well, but that is also not abnormal if you look at our financial performance. There must be substance behind that.
So anyway, leading to a strong position at the beginning of a next, upcycle. Explained in a bit more detail in these circles, somebody called it donuts the other day, which I think is a nice, name for these circles. You see the addressable market basis, 35%, then for die attach over, well over 40%, advanced die placement, even close to 70%, and for packaging and plating, about, low 20s. Again, here to emphasize which parts of the market are we addressing and what is their expected growth rate, the CAGRs? For die attach is around 29%-30%. Packaging and plating, close to 22%. So in total, about 21.6%, and those are typically the segments where we are focused on. This is also critical.
R&D spent many questions over many, many years, and in several ways. First of all, it's very easy to understand that our R&D spend is increasing in line with the complexity of the end product. So every smaller device geometry requires more complex solutions. You've seen examples downstairs, where we are still able to successfully follow the ever smaller geometries, ever more critical process requirements, and you see gradually, year by year, R&D spends go up. And if revenue then doesn't go up, your total return is an issue. But luckily, that is the case. If we look at the percentage of revenue, of course, in a peak year, 2021, as a percentage of revenue, it stays slightly behind.
But then, again, where it is currently around 11% is a very acceptable number, and we can proudly say that with that spend, we can address the most successful, challenging applications for the leaders in this industry. Strong liquidity position maintained. Our balance sheet has been strong for many years. I mentioned at the beginning, that also allows us to return profits we generate, and then especially cash profits, to shareholders, which we have done consistently and will continue to do so. And as you see here, the gross cash and the net cash is in a very healthy position, even in an extended downturn. And that allocation, partly in share buybacks, dividends, is historically interesting, but that's how we try to maintain a very attractive return to our shareholders, also in a cyclical environment.
Our business model, business model shared in, in every quarter, is, is very constant. We are targeting in the next upturn to reach revenues of EUR 1 billion plus, plus, plus. And the plus, plus, plus is very much related to the advanced packaging, hybrid bonding, TC segment. 40%+ in, in share in our addressable market. If you look at our gross margins, we have reached now, consecutively above 65% gross margin, and that means in the model that we are somewhere between 62%-66% operating margins on, on a comparable basis. And, also, the other targets remain, simply intact, and that is our business model for the near-term future. And with that, I hand over to Chris. Questions are happy to be answered at the end of this presentation. Thank you.
Thank you, Richard. And good afternoon, everybody, and hello to all of you joining remotely. My name is Chris Scanlan, and I would like to share with you what we see as opportunities and, challenges for the future. First of all, Richard just got done telling you about the current business conditions, what we see in the short and medium term, but in the long term, we're still in a very, solid growing market. Our market's been growing at a pretty consistent rate over time. I started in the industry at about 1994, which isn't even on the chart anymore. But TechInsights is now projecting that by 2030, we'll finally, tip $1 trillion in semiconductor revenue. And really, what's changed over time is only the end applications that are driving the growth.
So in 1994, when I started out, yeah, smartphones didn't exist. I, I was actually pretty excited I already have a pager, because most people didn't even have that. And, you know, for the next 20 years, we developed smartphone technology. We were really focused on cost reduction, on miniaturization. That's what drove the R&D spending in the packaging industry. What's been apparent for the last two years is that we're solidly now in the AI era, so we're no longer really focused on miniaturization of things. That still is an issue for some end applications, but we're really in the AI era, and that's what's gonna drive revenue growth probably for the next 10 years. In our end markets, as Richard mentioned, we're engaged in mobile, we're engaged in computing, automotive, and industrial.
But in all these end markets, we see that AI is a growth driver going forward, even in mobile internet, even in industrial, where we see power devices that are powering data centers, for example. In computing, it's really very obvious. Most of the R&D that we're working on is directly related to devices that are powering these AI applications.... And the AI semiconductor market is poised to grow very rapidly in the coming years. By 2030, I just said that we'll be at about $1 trillion in semiconductor revenue as a total. And this data is suggesting that we'll be at about $150 billion on AI-related semiconductor revenue, so about 15% of the total. And the devices include data center devices.
So these are the devices like the CPUs and the GPUs or accelerators that are really responsible for training and inference of these AI models. Also, the memory devices that support the GPUs, but also all the edge devices, the things that utilize the models and the data, and also perform some local training. These are things like mobile phones, computers, and even cars for self-driving. The other thing that's really become an issue in the last 10 years, but as years go by, it becomes even more and more of an issue, is that simply Moore's Law is slowing. We all know this. What does it really mean, though? Today, it's possible to fit about 100 billion transistors on a chip. 100 billion. That's a lot, but we already have CPUs, we already have GPUs with 200 billion, 300 billion transistors.
There's simply no choice. Designers of these kind of systems have no choice but to split the devices up and implement them as multiple chips or chiplets. And I'll explain how this is done in the coming slides. These are a couple of examples of chiplet devices, where designers have taken some of the functions that are most critical and implement them on the most advanced process nodes that are available in the fabs. And other functions, things like SerDes or things like that, can be implemented in trailing nodes, which are cheaper, and put together in the package to form a fully integrated device. This is what we call chiplets.
The data on the left shows the projected growth of chiplets as a percentage of the total foundry market, logic market, and it's projected to grow to about 25% of the total logic market within the next few years, by 2030. Why is that? Because this chiplet approach, splitting the chip up, implementing it as optimized subsystems or subchips, has the advantage of reducing the cost, increasing the yield, and reducing the cycle time for development. For our business, what does it mean? When a customer goes from an SoC, a single chip, and then takes the next step to split the chip into a couple of parts, and then further taking the next step to really implement multiple chiplet systems with many small chiplets, simply it increases the capital intensity, and it does that in two ways.
The first one, obviously, we need more die attach steps to build the same product, but also we need much more accurate die attach steps. So the interconnect density required to put these chiplets together and interconnect them as if they were all implemented in a fab node is much higher, and therefore, we need more accurate equipment, like hybrid bonding. So I want to take a little bit of time to go through this slide. Those of you that were here on the tour have seen some of this information. For those of you on the remote link, this will be a little bit new, but this is what a typical high-performance computing product looks like in cross-section. So some of you might have heard of CoWoS. This is kind of a generic version of CoWoS, what I'm showing you. What is CoWoS?
CoWoS is taking the multiple chiplets, plus maybe memory devices and other things, integrating them onto a very high-density routing layer. We call that an interposer. This interposer is always or generally a wafer-level interposer, meaning we do die attach onto a wafer. So all these things are attached to the interposer, could be silicon, and could be made of other materials. And then we mold that, and we put it onto a substrate, so that's the on-substrate part of CoWoS. And how do we do that, and what are the steps? It starts with the most advanced bonding technology that we have available to us today, and that's hybrid bonding. So hybrid bonding is used in multiple places. The first is in the fabrication of these chiplets. We call them 3D ICs.
And what we do in the case of hybrid bonding is we take advantage of direct copper to copper bonding to create very high density interconnects between vertically stacked logic or logic plus memory devices. This is all done in the front-end fab environment. It's not really a packaging process. So you can think of hybrid bonding as the last thing that happens in the wafer fab to create a 3D IC. Once the hybrid bonded chip is fully tested and created in the front-end fab, only then it's sent to the back-end factory for assembly onto the interposer. We can also use hybrid bonding in HBM, and I'll share a little bit with you about the types of applications or the considerations in using HBM or hybrid bonding for HBM. And then we have TCB chip to wafer.
TCB chip to wafer is used, again, once we've created the 3D IC using hybrid bonding, we'll use TCB chip to wafer in order to assemble the 3D chiplet onto the interposer. We can also use it in the fabrication of the HBM stack. For many applications, flip chip mass reflow is used for the assembly of the die to the interposer. A lot of applications don't have HBM, they just have multiple chips, and for those applications, we use standard flip chip mass reflow. In certain cases, for example, the NVIDIA B100 that was recently announced, we have things like bridge die, which provide a local high density interconnect between the chiplets. This requires very high accuracy placement because subsequent lithography steps are required after the die test step. The interposer has to be attached to the package substrate.
That's more or less a standard flip chip attach process. In some cases, it requires thermal compression bonding, in the case that the interposer is warped or very large. And finally, we have to encapsulate the entire thing. So there's a wafer-level encapsulation step, but there's also, in some extreme cases, a secondary encapsulation step required. Basically, what is Besi's strategy? It's to provide the interconnect technology and the equipment to do all these steps. But why hybrid bonding? So I mentioned hybrid bonding is the most advanced interconnect used to create 3D chiplets for logic, and simply the reason is because it has the highest performance, highest bandwidth, and lowest energy for connecting chiplets in three dimensions.
Hybrid bonding provides a direct copper-to-copper interconnect, so you can see at this layer there's no solder and there's no other interfacial material between the back end of line features on the chip, and this provides the highest density and lowest contact resistance. TCB is also required. In many cases where you don't need quite the performance of hybrid bonding, you can still use TCB, and in the case of logic chiplets, typically after you create the 3D chiplet using hybrid bonding, the next step will be to use TCB to connect the 3D chiplet to the interposer. So this is one recent example in high volume production, showing how hybrid bonding has enabled AMD's MI300. So what AMD did is they used hybrid bonding again to connect three thirteen... Oops! Yeah, sorry.
So they've used hybrid bonding to create four chiplet assemblies, with a total of 13 active tiles that are hybrid bonded onto four base tiles. Once they have completed the hybrid bonding, which is shown here, they further assemble using micro bump or TCB, those base tiles to a silicon interposer, and then finally, using standard flip chip, attach the interposer to the substrate. So you can see there in one product: hybrid bonding, TCB, and standard flip chip. Hybrid bonding has been in production now for over two years. AMD first qualified this in Q1 of 2022. They've now qualified multiple products, gaming CPUs, data center CPUs, the MI300 accelerator. So there's many different products that are being continuously introduced by AMD, and they've built a lot of experience and really proven the technology to the market.
So this is the next product that's been announced. This was announced in March. This is called Clearwater Forest, and this is a new Intel data center CPU. So this will compete, for example, with one of the EPYC processors from AMD. This is the first Intel product to use hybrid bonding. And when you look at this, what do you see? We have 12 CPU tiles that are assembled onto three base tiles, and those base tiles are also advanced silicon, but they're on a trailing node. The base tiles contain memory, they contain power management and other functions, and then they're further assembled onto a substrate and combined with another couple of die that do IO functionality. Excuse me.
So this really demonstrates that it's not just one customer; now we have multiple customers in high-volume production, and this device, as I mentioned, already is at 300 billion transistors. So again, there's no way that it would be possible to implement this as a single chip. I wanted to share a little bit about how these applications have evolved and what we expect for the future in high-performance processors. So if you look at AMD, they didn't just jump directly to hybrid bonding. They started with a single chip. They had a SoC. They started running up against cost and die size limitations. They split the die, first of all, into four identical chips, all having the same function, and put them back together in the package. Once they did that, two years later, they took the next step to a real chiplet implementation.
What does that mean? In the chiplet implementation, they broke the functionality into different nodes. So the big chip in the middle is an IO chip using a trailing node, and the small chiplets, eight of them, have the CPU cores and SRAM. After that, they took the next step to implement hybrid bonding to add more SRAM to the CPU chiplets. That was really the first implementation of hybrid bonding in the industry. The next thing that they did, they took the same technology, now that they had hybrid bonding in their toolkit, and they went through the same progression with their MI family of GPUs. They started with a single SOC, and then in 2022, they ran into the reticle size limit.
They couldn't implement a single chip anymore, so they'd broken in two, connected it with a high bandwidth interface in the package, and then two years later, they implemented chiplets using hybrid bonding. We've now seen Intel go through the same progression, from single-chip Xeon to Sapphire Rapids with four chips, all the identical four chips, and then finally, recently announcing Clearwater Forest, which will ramp next year. What do we see in the future? We now see NVIDIA, we see Apple going down the same path. NVIDIA recently announced their Blackwell family, so the B100 GPU. This B100 GPU is now implemented as two chips with well over 100 billion transistors. The Apple M1 Max CPU was introduced in 2019. In 2022, they implemented a dual-chip version, which exceeded again, the reticle limit .
We see these applications in the future transitioning to chiplet, because we see these designs kind of facing all the same challenges as the designs that came before them. We don't know exactly when or in which particular configuration, but we see this as the next step. Now turning to high bandwidth memory. Last year we talked about implementation of hybrid bonding and high bandwidth memory. We talked about the limitation on thickness, and that in order to fit 16 die into a package, we needed to eliminate the solder bumps between the chips. JEDEC, since then, has changed the thickness spec for the next generation of memory, HBM4, to increase the thickness to allow more room to extend the life of micro bump. So for that reason, we see that some of the...
Or maybe even a majority of the volume for HBM4 will remain in micro bump. But yet, all the suppliers still have both TCB and hybrid on the roadmap. So you can see the roadmap on the left from Samsung and the roadmap on the right from SK hynix. Here in the Samsung roadmap, you see TCB and hybrid copper bond, gapless, and you can see the 7-micron gap. That's the thickness that they have to deal with. In the case of SK hynix, you can see here, hybrid bonding and an advanced mass reflow, which is a kind of thermal compression assembly process. So we see our customers really pursuing dual parallel development paths. We also see TSMC supporting both hybrid bonding and TCB by developing a base die and a set of design rules for both. So why are they doing that?
We think the reason is performance. So thickness perhaps is not a constraint, but performance is a big constraint. So one of the most critical issues in these high performance, computing devices, is thermal performance. So these devices are dissipating already over 1 kilowatt. So you can imagine a very thin piece of silicon, or maybe multiple, generating 1 kW or 2 kW of power. It's gonna heat up anything that's very closely coupled to it, like the memory. And it turns out that it's very difficult to extract the heat from that stack of memory. You can see here the controller die is on the bottom, the heat sinks on the top, and in between, you have 12 or 16 layers. In the case of TCB, you have layers of epoxy, not very thermally conductive material, between every one, every single chip.
This adds thermal resistance, and as a result, the device runs hotter. It becomes a bigger issue the more devices you stack. What we've found through our own simulation, and what has been demonstrated by Samsung through their own experimentation, is that hybrid bonding will reduce the operating temperature of the memory because it has about a 20% lower thermal resistance. This is super critical because the memory has a maximum temperature limit. So we think in some applications, there won't be a choice other than to go to hybrid bonding for this reason. We don't know what the split will be, but we think there will be some implementation already for HBM4.
In addition, in the future, people are already starting to talk about stacking DRAM HBM directly on the processor, and as I just said, the processor is generating a lot of heat, and therefore, hybrid bonding is really required in order to more effectively extract the heat. You can see here from AMD, they're talking about using hybrid bonding for the stacking of DRAM on top of logic. Timing on this, we think, is about equivalent to the HBM5 timing. Now, let's talk a little bit about CoWoS. So, as I mentioned, hybrid bonded chiplets eventually go into typically a CoWoS package, and these high performance applications. CoWoS has two parts, chip on wafer and chip on substrate. We see this as a very rapidly growing market.
A lot of this market is high performance computing, but a lot of it is also mobile, PCs, and things like that. And for this application, we see two steps where BESI participates. First is in the flip chip attach process, chip to wafer. We showed you this in the tour. This is the 8800 Chameo platform, and then in the assembly of the substrate or the interposer assembly to the wafer or to the package substrate, our 2200 Evo is really the tool of record, and we see continued growth in this area. As Richard mentioned, we're also seeing big growth in photonics. So this chart on the lower left shows the growth in the photonic transceiver market. What is a photonic transceiver?
This is the pluggable device that is connected to a board in the data center that allows you to plug a fiber directly to the server, converting the light to electrons and delivering that to the processor. That's what it looks like here. This is one from Intel. Besi's participating in this market. We assemble about four of the components, four types of components in these transceiver modules using our 2200 Evo. But what's really exciting for us is we see opportunities for hybrid bonding as well in this market. So TSMC, for several years now, they've been developing a new technology called COUPE, and this is a type of photonic chiplet that is fabricated by stacking a photonic chip with an electronic chip, so an EIC and a PIC. The performance, electrical performance is super critical. They're trying to achieve very high data rates.
The whole job is to transmit the high-speed signals, and the data on the lower right shows the advantage in terms of performance for data rate and energy efficiency of hybrid bonding compared to microbump. There's about a 40% benefit. TSMC has recently announced that they'll be qualifying this technology using hybrid bonding for pluggable transceivers next year and for co-packaged optics the year after that, in 2026. Co-packaged optics means these assemblies, these photonic chiplets, will be placed next to the GPU or next to the network switch on the package. In this example from Broadcom, you can see that there's eight of these chiplets. There could be more, 16, perhaps, and these will provide the high-speed communication off of the GPU and to the GPU. So another big opportunity for hybrid bonding.
We see a lot of opportunity as well in other edge computing applications, like phones, like laptops. We're working on projects in, alternative reality glasses, automotive, self-driving, many different applications. The mobile phone market. So this market's growing at a low single-digit rate, but the growth that we're seeing is really driven by AI. The S24 from Samsung has been a very, successful product in the market, really, on the back of its AI functionality that it's offering. The data on the lower left from Counterpoint is showing the growth rate and the penetration of AI-enabled phones over time. So we really see this as, one of the-...
Key drivers for the mobile market, and we do anticipate at a certain point in time, that the additional APU, MPU, combination functionality will require a conversion to chiplets, probably in order to provide additional memory capacity for the AI applications. And finally, in the automotive market, right now, we're in a very, kind of a soft phase in the automotive market, but long term, we see by 2030, the semiconductor TAM more than doubling. And this benefits all of our product groups at BESI. So in the EV drivetrain, we have applications in molding, we have applications in plating, in die attach, power die attach. But here again, AI is really central to the growth.
So is the development of self-driving capability, primarily, but other AI applications as well in the car, and then all the other components, there's thousands of components, that contribute to this growth. So that's a little bit about what's driving our future business. And with that, I think I'll turn it over to Peter to talk about the particulars for sub-micron.
Thank you, Chris. Well, welcome, everybody here. Let's move on with our third section, the wafer level section, or to be more precise, let's talk even a little bit more in detail about hybrid and thermal compression bonding than Chris did. In order to do that, I don't want to talk about the technologies itself, but the adoption, where you need it and why you need it. So we concentrate actually on the high power computing portion and on the memory, the high bandwidth memory portion of that. So that will be our leading topics for this chapter here. In order to get into the chapter, we start looking on this chart, which is showing very nicely the long-term development of interconnect technologies.
So on the very left side, you know, the starting with wire bonding decades back, and then gradually growing via flip chip to thermal compression and all kind of these technologies. And always the driver behind the next step of interconnect was always the computers, always the microprocessors, because they were on the most modern nodes with the most transistors, so they needed the most connectors. So that was always the driving force. So it's always, in the end, the amount of connectors, and that relates then to the x-axis, which is the pin count, the pitch, bump pitch, the connector pitch. And now, as we all know, today, with the real high power computing applications, we have come to the very right side of this chart, where the bump pitch is 10 micron or below. And that's well known.
I don't think we need to repeat that, the home turf of hybrid technology, because there is no other technology that can connect these small bump pitches. But that's not the only one. You know, just left of this hybrid arena, 10 micron, 10 micron +, so 10 micron, 15, 12, 20, perhaps 25. That's an arena which is also very interesting because that arena can be covered by a special methodology of thermal compression bonding, which is today normally called the fluxless thermal compression bonding. And there is other thermal compression bonding, even to the left side now of this, of this arena, you know, with bump pitches of 30, 35, 40, which is, so to speak, the home turf of thermal compression.
But really, I want you to focus on this neighboring area of hybrid, because this neighboring area is really the interesting one for this, also for future applications. Because also, if you don't go all the way down to below 10 micron, if you want to do a modern product, you still need to go from 30-micron bump pitch to 20, 20, 15, and so on. And this area here, this, what I call out here as a TC fluxless area, that is actually a TC area, and everybody's thinking, "Yeah, that's old stuff," but it is not. This is a new flavor of TC, and that is new and is coming up because it's a new version of it, and you will later see what you need special about that. Now, why I'm pointing this out on this chart?
Because these are really the areas that we want to investigate for the high power computing and the high bandwidth memory, because that's the arena where these applications will play out in the future. Now, starting with the hybrid portion of it. Well, I think it's really well known in the market that since three years, we have established ourselves as the absolute hybrid market leader with our first generation machine, which was capable of 200 nm accuracy, and which is still, until today, the only machine which is used for high volume production. And this machine, and that's the scale below, is capable of running products and manufacturing products, which have a bump pitch from 10 micron down to 7, if you push the limits, even to 6 micron. And that's all products that are in volume production today.
But as I said, it's always the computing power which is pushing for more connectors and for that reason, for less bump pitch. And the next designs, which are just around the corner, from our end customers, they will have even smaller bump pitches. They want to go down now to 4 micron, some 3.5 micron. And for that reason, we have started early enough, to launch and, bring to the market our second generation machine, which is, which is the 100 nm bonder here in the middle. Which, by the way, was not even more accurate, but, even also more productive by 10%. And that is actually now taking over, being the standard, the industry standard for hybrid bonding. That's happening actually right now as we speak.
But for sure, we are not only talking actually to the production guys of our customers, we are also heavily engaged with their R&D departments. And if you look what is happening there, actually, these gentlemen, they are already working on test vehicles with a bump pitch of 1 micron, so the very, very right part of this scale here. And they're not only talking about it, they are acting on that. And for that reason, it's very clear that also with 100 nm, which is an exceptional accuracy for a die bonder, that is not good enough for the future. And for that reason, we are already working on our third generation machine, which will then have a capability of 50 nm accuracy, and which will then cover the bump pitches all the way down to 1 nm.
So that is, in a nutshell, also the roadmap here that we are working on. But that's really more the technical stuff. Let's not only look at the, yeah, at the machine side, let's also look at the customer side. You have heard already a lot of things, either here when you were going round or also from Chris. I've put here some examples and also a lot of customer names. So we all know that AMD, together with TSMC, they were the front runners in this technology. And the very left picture, the EPYC processor, was really the first product in the world done with hybrid. And now it's the second picture here, it's the MI300, for example. And then you see, and Chris explained it already, the Intel product, which was announced.
But it's not only that these big logic manufacturers are in high volume production, it's also that, the memory guys that you see next, Samsung, Micron, and not only these two, they are also heavily working on hybrid, and they are heavily working on hybrid together with us. And not to mention, Samsung is not only a hybrid company, they also have a logic division, and we are also working with them on their next generation logic for hybrid as well. So it's really, it's really not like last year, but if you remember back, for all of you who have attended also last year, where there was more or less one high volume manufacturer in the world, and the rest was, "Well, will they adopt or will they not adopt?" It's very clear that everybody is adopting right now.
You have seen the big names here, but if you look without pictures, just on top, there are small names. Perhaps you have read recently an announcement of NHanced, which is a small scale company, which is doing special logic applications, but also very high end. But also the OSATs, like ASE, they also do not want to be left out in the race of the next technology. So they all are starting to adapt to this technology. And finally, if you look to the very right side here, even in Japan, where they have this consortia, Rapidus, which is building the first 2 nm factory in Japan, also, we are engaged with them to set up the packaging manufacturing then with the hybrid technology in Japan.
So what you really can say is, compared to last year, that it has gone from one manufacturer to really more manufacturers, and from only the big guys, also to small companies, and really, everybody is adopting to it, from starting in the U.S., all over the globe, all the way to Japan. Now, once again, back to thermal compression, as we said before, because yes, yes, there are these two areas. Well, thermal compression, as I said, there's a wide area. There's a, let's say, the standard to thermal compression, which is there for a very long time, with the bigger bump pitches, which is more the left side here on this chart. And then there is this area, which is really next to hybrid, which is 10-25 micron.
And that one, that right arena, that's really the interesting one for this high bandwidth memory and also for logic applications next to a hybrid. And exactly for this arena, we have developed our next generation machine, which is also called TC Next, which needs two very specific things in order to be successful compared to conventional. And that is really a much higher accuracy, where we are poised for with our experience from hybrid. In that case, it's not 100 nm, it's 500 nm, but that's still very challenging. And you need a methodology in the machine to reduce oxidation of the bumps. That's what is called fluxless then.
And that's what we have come up with in a very nice package, and that's actually what we are now going to start with customers like Intel, Micron, but also all the others that you see here on the list, to get engaged with them, on their next future product that they have in the TC arena. Now, while we are talking TC, let's switch a little bit, and once again, even so Chris did mention it already, let's think a little bit once more about this high-bandwidth memory. Because you remember the last half year, I would say, or even perhaps more, there was this big debate, in the industry. Yeah, now, as the JEDEC tray standard in height was lifted a little bit, yeah, does high-bandwidth memory need hybrid anymore, or is there now anyway anything TC?
That was actually the big debate. Let's once again do a deep dive here to really understand what is driving this industry. And this chart here is from Yole, which is only summarizing very simple, the specification of the different high-bandwidth memory generations. So that means what Samsung needs to deliver, that NVIDIA is happy, or Micron, or Hynix. And I want to point out two lines here. One thing is, from generation to generation, actually, the memory capacity has to increase because you need more memory for these AI applications. And from generation one to generation four, you have a factor of 64 times bigger the memory. And the other thing what needs to increase at the same time, with the factor of 16, is the bandwidth. So meaning, how fast can you read or write data from this memory cube?
And that's the two major challenges, and that's also what the big three guys are striving for in their R&D departments. Now, let's look at that. What do you do when you want to have more memory capacity? Well, for sure, you can try to optimize your silicon, but there they are pretty much at, like, also on the other side, at the end of the opportunities. They can do a little bit, but not much. So what you easily do is say, "Well, let's stack more dies. Let's put more together." Very logic, very straightforward. And then you run into the problem, but it's getting higher then, and there is a standard, and I exceed the standard. Okay, standard is increased. You think, no problem, but that is not true.
Because there is this other problem that I don't explain the second time, because Chris did it nicely, because it's generating heat, and with all the more layers, it's more difficult to get the heat out, and you cannot overheat the part because then it's damaged. Bad enough, but it's getting worse. Because, once again, at the same time, you need to increase bandwidth. And how do you do that? Well, you have two opportunities, technically. Very simple also. Either you make more connections that you can do more read and write operations in parallel, because you have simply more electrical lines, or you run at the higher clock rate, so you mean you do more read and write operations in one second.
The bad thing is, both of those methodologies lead to even more heat, because you do more operations where you push electricity through your chip, and that means more heat. And so on one hand, because you have stacked more die, you anyway have a problem to get rid of the heat, and then you need to operate it at a higher speed, which creates even more heat. And that, perhaps, to try to put it in a little bit simple words, that's ultimately the reason there's simply no way for these guys to avoid hybrid technology. They can try to extend TC, and that's what they are doing, but there is simply no way out. Now, having that said, let's look at what the plans are.
Actually, I tried to make a more overview table, you know, from the charts that Chris has shown, from the individual roadmaps of Samsung and Hynix, which are official and everything. But I think it's perhaps a little bit a better overview here for you to follow it. So currently, with the HBM3 and 3E, which is the current generation that NVIDIA is using, they have their technologies which they have been using. So that's this, NCF thermal compression for two of the gentlemen and the mass reflow for SK Hynix. So that's what they have been doing all along. Now, if you look on their roadmaps, what they are doing for HBM4 and 4E, they all have, on one hand, an improvement of their existing technology.
So that means that thermal compression guys, they are going to fluxless thermal compression because that gives them an improvement. Or Hynix, who is using a different process, is advancing their own process, which they have developed on their own. But at the same time, they are, all of them, without any exception, they are working heavily on, on hybrid because of the problem that I have described to, to you before. And in the end, I think we should not think along the lines, is now HBM4E all hybrid or all thermal compression? In the end, I think the tipping point for the technology is not the generation, it's the, it's the stack height. And 16 will be the tipping point, where you can go up to 16 with thermal compression.
Perhaps you can even make a 16 with an advanced thermal compression version, and 16 and above will definitely need to go to hybrid. And also, Chris mentioned that we think we will see the first 16-up stacks with hybrid technology in 2026. That is our estimate, according to what we see. And then after that, if you even need more memory and more bandwidth, you're anyway bound to the latest and greatest technology. And actually to add to this more technical side, to add also a little bit, a market side.
Once again, a chart of Yole, where you see the HBM market in total, and where you also, by the way, see this nice increase, this heavy increase now about in this year and next year, which is for sure generating a lot of business for conventional interconnect technologies. But also, if you look into the future, where there is still a nice growth overall, but what is more important is that really the old technologies, HBM2, one, two, and then later on, also HBM3, that they will fade out quite fast, because the industry will only look for the most advanced stacks with the biggest bandwidth. So that means it's not only about the pure growth rate of the whole HBM market, it's really also about a very fast adoption rate, always, of the latest and greatest HBM stacks.
And that is for sure, if you look it from that angle, a really nice, really nice business opportunity. And we, as Besi, we are prepared to really pick up on this business opportunity, because as we have seen also on our tour through the company, we have, on one hand, our... on the right side, our hybrid machine prepared for that, and we are working with all of these gentlemen on developing hybrid. And on the picture below, you see, a sample, a sample build with 16 high stacks, and we have put it here in writing, because otherwise you cannot count it. Even, in the R&D lab, we are even doing 26 high stacks. So that's the one side.
And on the left side, actually, you see the thermal compression bonder, which is precisely fitting into that specific version of thermal compression that you need. And you see here a cross-section of a test vehicle of ours, of a 16-high stack. Once again, even if this is a test vehicle, we are working also directly with the manufacturers on this one as well. So for us, this is one big opportunity ahead of us, because also, if they stay on TC, they need to go to the next generation of TC. They cannot prolong their existing machines any longer. They need also new generation of machine, because it's not the same TC as they are using today. With that, I would want to come to my last chart.
You know this chart, and the gentlemen and ladies who have been here for the last years know this chart very well. This is our hybrid market modeling. So we are modeling here the amount of hybrid bonds that we think that will be used in the industry with a bottom-up model. And then we, for sure, divide it by average UPHs and availability factors, and that's how we get to machines. Now, let's have a look what has changed from last year. I've pointed it out. So overall, our outlook for 2023 has increased by 6%. So we are more confident overall than we have been last year. And why is that?
Well, first of all, and I explained that on the other slide, the adoption rate and also the companies who are adopting, has increased significantly and is much more broader today than last year. Then, what we also can say, if we look at the logic, which was the front runner and which is also today the biggest contributor, is that, besides, TSMC, well-known, also the second really real relevant, supplier for logic and computing devices has adopted to a high volume production now. I think you have seen some of our announcements of order intake recently. But also, TSMC is growing further.
They have made in an official article, it was stated that, they think that this type of technology and this type of devices in their factory will increase by 100% every year until for the next three years, until 2026. So that is a big growth rate. And actually, just I think it was yesterday, there was also, also officially available. They, they announced their next capital expenditures. And, if you read them, you know, also they need to build new factories because they are running out of space. And if you read them carefully, you'll see also Besi named as a main contributor to their next expansion phase. So that meaning it's not only a starting point where they already have a lot of machines installed, they are really also growing like hell.
Then the next one, the next point, once again, the memory. I think I tried to explain how we look at the memory business and why we think this is a big opportunity for us. And yes, we agree, there was a certain shift in there. We modeled it with one year compared to our last year's model, because there will be an extension of using thermal compression technology, but it does not, as explained, it does not change the overall plan on the long run. And there is a third thing, and I think that's not discussed heavily. Chris mentioned it a little bit, but I want to put more emphasis on that because it's really important....
Last year, I got a question here at the same event: what do you think, Peter, when will application processors adopt to hybrid, or will they adopt to hybrid at all? Yeah, the visibility last year was mediocre, let's face it. But if you have been on conferences, let's say, for the last half year, roughly, and I know some of you have—I have met on the ITF conference at IMEC in Belgium just recently. You heard everybody was talking about edge AI. And edge AI, and Chris also mentioned that in one slide, always does mean edge AI on our phones, edge AI on our computers, and if some of you remember the BMW gentleman saying, edge AI on our cars.
It's getting more and more clear that AI will penetrate our personal lives with our edge devices that we are using day-to-day dramatically, and this is not a topic only for compute centers. For that reason, also there, it is getting more and more clear that our processors that we are using in our edge devices need a lot of compute power and also spec-specific engines for neural engines. That is driving this whole edge AI, and with that application processor types into the same direction. Because one more technical thing, remember one thing that Chris said: hybrid is also a huge enabler to provide high compute power, not alone, but in combination with the chip itself, to provide high compute power at a low power consumption electrically.
That is exactly what a lot of edge AI devices need, because the batteries in our cell phones are not endless. Also in the car, the guy said, from BMW, he needs the electrical energy for his drivetrain and not for the computer to calculate to calculate the AI. So for that reason, that one will be the third big thing coming up after the memory. That's how we modeled it, and that's what it is, what is in there. Beside all that, there are also some nice add-on and upside potentials in co-packaged optics, mentioned by Chris, and also by special CMOS applications that we are working on with some lead customers.
So all in all, I have to say, our model from last year has been confirmed, and we are running on that track and getting the business there. And with that, I want to hand over to Christoph.
Thank you, Peter. Exciting news, indeed. We come now to the last part, which is all about the mainstream die attach. I am running that group, and I will walk you through the updates there. So let me start with a quick introduction of the basic die attach product group. Just in a nutshell, it's about 80% of the overall Besi revenue, so a quite significant overall contribution to Besi. Market share has increased. Last year, in that meeting, I was talking about 40%. The latest numbers, just recently presented by TechInsights, indicated that we have increased now to 43% overall for die attach. If you look into the advanced packaging segment, there, our dominance is even much bigger. We own more than two-thirds of that market. The market itself is a very interesting one.
Also, brand-new data from TechInsights, talking about 29% CAGR in the time frame till 2026. So definitely, an interesting one. And if you look into drivers, I think it became clear today, and also in the presentations we have just seen, it's all about AI. AI translates into very advanced package formats like CoWoS. Photonics is a big thing for us as we speak, but still also the mobile is a very relevant driver for the die attach market. Automotive, at this moment, I think you have realized a bit slower, but still expected to return to a growth mode any time soon. Well, on that slide here, you see the die attach product portfolio. It's split into four different product lines. The two on the left-hand side are the more advanced packaging product lines.
The two on the right-hand side are the real mainstream product lines. I start with a prominent one, the MMA platform, our 2200 Evo, which is a very versatile multi-chip platform, highly configurable and used in a very diverse field of different applications. The most prominent one, obviously, is the mobile market, whereby we are very strong in the camera and sensor supply chain. On the flip chip side, we have a long list of different flavors of these machines. The Chameo platform for the wafer-level assembly, the Quantum platform for the substrate level assembly, and the 2100 Flip Chip for the lead frame assembly.
So really, the full portfolio of flip chip processes is being supported, and this is really the flip chip platform of choice, no matter what application is behind. Then, looking at the die bonding machine, this is the industry-leading epoxy bonder in terms of industry-leading, in terms of accuracy, but also throughput. This machine, yeah, also here we see various applications. The most important one is mobile PAs, because of its very tight bond line thickness control. And last, but certainly not least, our soft solder product line up here with a 2100 SSI, which is a industry-leading platform supporting the soft solder business, but also the diffusion soldering, which is our two technologies that are used in automotive and power applications.
So to date, the focus is more on the, on the left-hand side, and, in the, following slides, I just wanna highlight some of the recent introductions and also strong solutions, at this moment. I start with the high-end smartphone market, which is still very relevant for us and also seen as a, as a, a driver, near and midterm, for two reasons. We do see two driving vectors. One is, what Chris already explained, the tendency that these large language models are getting onto our phones, the edge AI or, Gen AI phones. And yeah, that is, Counterpoint is talking about a, a very strong introduction rate now by 2027, if I'm not mistaken.
They predicted almost half of the mobile phones out there will be edge AI phones already, and this triggers a whole replacement cycle of mobiles. The other driving vector on the bottom left is more an internal one. This is simply the fact that we as Besi are very much engaged, especially on the camera and sensor side, with leading customers to introduce more and more process steps. And as you can see in the chart here, over the last 10 years, we have been able to grow the number of supported process steps by a factor of 10. We are now at about 40 different steps we are doing, and we are working a lot to increase that number again. We are working on things like variable apertures on smartphones, so very SLR-like features, but also lens system assembly.
So very high-end stuff that will come to smartphones and will help to drive that business further. Solutions on our end, I have a couple of machines, very active in that market. As indicated already, the Evo is the platform supporting the camera and the sensor business, the flip chip machine, which is very strong in mobile, and lastly, certainly also the 2100 platform with its contribution to the mobile PAs. The strong point here is that we as Besi, working a lot on integrated lines, on fully automated lines that help to drive up the yield of the assembly process. Also, cleanliness is a key topic. We have improved our machines a lot in the last time frame in terms of cleanliness.
We introduced a lot of inspection and metrology in the integrated line, also with the end in mind to improve the yield. And lastly, and that's probably also an interesting point, we are not only supporting the manufacturing of AI cards, high-performance computers, we are also making use of AI a lot internally in our development. For instance, here, as mentioned, also, our vision algorithms are more and more relying on AI, and we make good progress there, compared to conventional systems. The second aspect I'd like to talk about is photonics. Data center growth, I think is imminent. Everybody understands that, everybody hears that. With the growing data center capacity, there is also a communication part in the data center growing like hell.
I think it was Chris explaining the transceivers that are basically translating electrical signals into optical signals and vice versa, which are needed to connect all those racks in the data center. All the short distance connections are meanwhile done optically, and this yield chart you see here is talking about that. The guys on the bottom are the key players in that field, in the transceiver field, and all of them, I can proudly say, are users of the photonics platform we are offering with our 2200 Evo. Yeah, this machine is being used to do all these optical subassemblies, the TOSA ROSA, as it's called, lasers, VCSELs, et cetera. Why the Evo, you may ask, and there is one simple explanation for that. This photonics market is still a very diversified market.
There is no standard available. Everybody has his own recipe and proprietary technology, which at the end means there is no equipment available off the shelf. You just have to develop customized equipment—customize your own equipment. And with our highly configurable and flexible Evo machine, we offer a perfect starting point to do that customization. And not only that, also once the recipe proves to be working, we are able to ramp rapidly from low volumes into HVM versions. We can deliver many machines in the same configuration and with consistent quality. So that's the reason, the main reason why this solution is used very much in that market.
We are working to extend our process capability further to even take, grab a bigger portion of that market, namely the active alignment side, which we are not supporting now, but we believe with increasing accuracy to one micron, there is a good chance that we can capture that market as well. Then coming to CoWoS. We heard it a lot. We saw that pictures many times already. On the left-hand side, you see a typical high-performance computing board, where the logic has to be placed, HBM dice have to be placed. All those components need to be placed on wafer level with high accuracy, with high throughput, with best cost of ownership. And we have been able to develop an integrated line that can do both in one line.
The logic, which typically comes from a wafer, as well as the HBM dice, which are fed in a tape and reel format. And, yeah, in this combination, we offer really an extremely competitive solution, which beats competitors really by factors. We have seen that in a recent installation in Korea, that this solution really was much faster and providing the cost of ownership that is needed when a product goes into high volume, and that's what happens right now in this market. Once this subsystem has been assembled, so the logic, the HBMs, eventually other components as well, we end up still with a chip-like structure that's at the end needs to be packaged into the final package.
This is where the on-substrate part of CoWoS is happening. And here we have another excellent solution available. Not only excellent solution, but also at the end, the tool of reference. Today, the vast majority of the CoWoS products that are out and are being assembled are done with our 2200 Evo on-substrate attach, because of the capability to handle very large dice, to handle them carefully, reliably. We have also development activities in place to further increase the capability, the die size capability, to handle also future sizes. I think with many of them, we talked about it when being at the machine. It's really amazing to see how large those dice get and which are handled automatically.
So with this, only a few highlights, of course, all the products are being sold and have their unique selling points, but I only wanted to point out a few of them. With that, I'd like to point your attention to the bigger picture, where we are acting at, with die attach. You see basically here the addressable market for die attach, for mainstream die attach, I should say, which is growing from currently EUR 800 million to about EUR 1.6 billion. So a factor two in the next years until 2026 is projected, and again, referring to TechInsights, latest updates. What we also see is that this so-called advanced die placement is the fastest-growing portion within our addressable market.
Based on basic internal marketing data, we do estimate that within that advanced die placement, which should be around EUR 500 million in 2026, about one-third is in photonics, another third in this before mentioned CoWoS application. A very significant contribution coming from AI. With what I have presented with, with our positioning of this of this AI platforms in that fastest-growing market, and also knowing that there is a strong upswing in the in the overall market expected.
We are very confident that we will not only benefit from the pure market growth, but are also able to increase our market share further with the drivers AI, CoWoS, and Photonics. And that concludes my presentation and brings us to the last section, which is Q&A.
So I'm sure you understood every single slide, and it must have been a bombardment, but I see two hands sticking up.
Just for organizational, in order that our audience, which is online, can understand, please only use the microphone. Claudia will use this microphone and bring it to you.
Hi, thanks for the presentation. I just had a quick clarification on TCB within HBM. So once HBM moves to fluxless TCB, do you expect your tool to take most of the share there, or do your competitors also have a good solution? Thank you.
For these suppliers who are using TCB, because not everybody is using that, when they are using to fluxless, then we are well prepared to take a big portion of that business.
Hi, it's Alex Duval from Goldman Sachs. A couple of quick questions. Firstly, you talked about Edge AI as an opportunity. Clearly, in terms of what's in production with hybrid bonding, you have, you know, great progress that's happened on the data center side. Can you talk a bit about how you expect customer behavior to evolve in terms of working with smartphone chip players? And then secondly, you talked about advanced TCB, and clearly, that could be part of the solution, potentially, on this HBM4, as you get to higher stacks. Although they're pursuing a twin track, obviously, with hybrid bonding.
On TCB, can you talk a bit about your more advanced solutions, the customer traction you're getting there, and how we should think about the average selling prices relative to what you're getting on hybrid bonding? Thank you.
That was a lot now. Well, the average selling prices are all about that. Think of TC-- of, of these high-end TCB, you know, applications and, and hybrid from a, from a die bonding machine point of view, think of it on a, on a similar price level. So they are in the same range, UPH, also in a similar range. So the cost of ownership on our side here is not really that much different. So if you look for cost differences of these two different manufacturing paths, you really have to look in all the steps that are leading up, to, yeah, to these two steps, in the, in the preparation on that. So that's really the, the dominating part there. Then once again, what, what else?
Traction in the market and on-
Yeah, traction in the market.
Uh, solutions.
Traction in the market with TCB solutions? Yeah, well, I've shown once again, this in this new, let's say, advanced area, let's not call it new. Advanced area of the very small bump pitches for TCB. We are really now... Last year, when you were here, I don't know if you were here in person, but you saw also on the same spot, the machine, which was an R&D machine. Now, we have it finished, it's ready to sell, and we are now really engaged with a lot of customers. Once again, you have seen some big names on my presentation. Actually, we are all working with them.
On some, we are working here in our lab, on some, we are working with our machine on site with them, and so, so it's really, it's really now the time to do the development with these customers for this generation of products that, that are to come, and our traction is pretty good in the area right, right now.
Great. Just on the point about Edge AI and smartphones.
I'll take that one. So I think Peter showed the timeline that we're sort of predicting for the adoption of hybrid. I think your question is specifically related to hybrid. Is that correct? Yeah. First of all, I think we'll first see adoption in PCs and laptops, and then in smartphone. So you already see AMD, for example, using hybrid bonding in gaming, application CPUs for gaming, for PC. The next step will be higher volume applications in the PC space and laptop, and then later, much higher volume applications in smartphone. And we see not only the application processor, but also other potential applications in the smartphone. That includes the RF front-end module. There are certain companies that develop RFFEs that are doing development on using hybrid bonding for that space, motivated by different things.
And then also, there was a discussion earlier about image sensors. We also see something. Sony, for example, last week at ECTC, presented chip-on-wafer, on-wafer, meaning they have a three-layer image sensor now, in which the first two layers are bonded together using wafer-to-wafer hybrid bonding, and then the third layer being bonded using chip-to-wafer hybrid bonding. So we see more than just the application processor as potential for the smartphone.
Thank you.
Thank you. Yeah, great presentation. Just wanted to ask on the very last slide, you showed that in the mainstream advanced die-attach market, you estimate about a third of that is CoWoS in 2026. I just wanted to clarify, does that include TCB as well, and what is the size of that market in 2023?
Good that you bring that point up, because I actually forgot to make that disclaimer. That is excluding the TC and the hybrid part. This is really looking at the more conventional, die-attach product portfolio. So everything you have seen in the submicron presentation is actually adding to that business potential.
The size of that market in 2023, do you have an estimate for that?
2023 will be... And you are talking about TC and hybrid related?
No, just CoWoS.
Ah, CoWoS, CoWoS. It is—I don't have a number out of the top of my head, but it is already a quite relevant. It has been already quite relevant for us last year. We have seen, and I think we were talking about it in one of our slides, we have seen basically early 2023, mid-2023, kicking that CoWoS business really in. We have seen it on the chip-to-wafer side, but also on the substrate side. So both platforms benefited a lot already from that CoWoS introduction.
All right, thank you.
All right, thank you. Ruben Devos from Kepler Cheuvreux. I just had a first question on basically AMD, which has been, let's say, with their products, the EPYC, the Ryzen, MI300, it's been sort of the flagship products with hybrid bonding adoption. Could you maybe talk about-- I mean, last week we've heard about the MI325, MI350, and so on. It looks like the pace of innovation is really accelerating. And in terms of logic to logic stacking, is there sort of a, yeah, acceleration happening here as well? How should we think about the bonding steps for these next generation products, and how should we look at that relative to your hybrid bonding market forecast for logic, which I don't believe have really been upgraded since last year?
Yeah. I'll leave the forecast to you, Peter, but yeah, we definitely see AMD on their, on really on their data center CPUs so, and GPUs, that you mentioned, making basically employing more chiplets, which means more bonding steps. So if you look back to the first EPYC processor that used hybrid bonding, there were eight chiplets, 3D V-Cache enabled chiplets. And there's multiple bonding steps in the fabrication of each chiplet. The picture I showed you had 12, and they've now shown examples of even higher numbers of these 3D chiplets. So that will continue to evolve. And the same on the GPU.
So, I think the first implementation, the MI300, had thirteen active chiplets, but again, there are more than 13 bonding steps because there are spacer die, there's some temporary bonding steps that we also have to consider in the process. And they have announced in this last week at the event in Taiwan, a next generation. They didn't give many details, but they hinted at even higher levels of integration. And the roadmap that I showed you gave one example of that, which was the stacking of DRAM directly on, onto the accelerator. So with that, I'll let Peter talk about the forecast and how we implemented that.
Well, as I said, we in the presentation, we are doing this, we are trying to do this as a bottom-up forecast. So that means we do whatever we know, we integrate that in our forecast. But there is, you know, not everywhere on each product and each customer, only positive developments. So yes, that is considered also, the overall number in the end is going up. If you perhaps refer that perhaps for just next year or so, well, you know, a forecast model has plus and minuses, but yes, it is considered.
Oh, okay. And then just a second question on the TCB roadmap. How does that look like? Because I hear you saying that particularly as of HBM5, it's all going to be hybrid bonding for memory. So basically, if you bridge that period until HBM5, at which hybrid bonding you have the tipping point, and before you wanna meet a lot of demand with TCB, what happens with TCB after HBM5?
Very good question, because you're right. Only for HBM memory alone, you should not do a TCP, because you might simply say: I'll wait until the final technology comes along. But looking back to—if you look to the model, you know, this overall typical package model, there is much more interconnect steps in there, which are now not related to the HBM memory, but to the overall CoWoS package, which are today done in C4 flip chip. So for example, the attach of the HBM and the logic, or the already hybrid bonded logic assembly to the interposer. That can be, which once again, today is on a large scale or still C4 reflow flip chip.
But also there, you get more bumps, that means smaller bump pitches, and that once again transitions then to TC, to thermal compression. So that means, yeah, one opportunity is now big, but will go away over time. But at the same, in parallel, on, on, not on the HBM, but on the overall CoWoS package, you will get new opportunities, because there is the shift actually then from reflow flip chip to TC flip chip. So that means it's worthwhile to invest into TC.
Thank you.
Thanks. Thanks for the question. First one, last year, we spoke a bit about the gap between the competition, and you believe that it increased, given that you had the, the new generation and more machines in the field. How do you see that over the last year? I think that the competition is doing something, but, yeah, obviously you're not standing still as well. You believe how much is the gap now?
You are talking hybrid package?
Oh, yeah, Hybrid Bonding. Okay.
Well, considering that, as I have pointed out, we are still today also the only ones who are producing hybrid products in high volume production, and our competition has a R&D machine here and there. I would make a quite bold statement that they did not close on us, especially given the fact that we are already now launching our second generation, or have launched, and that our high volume customers have immediately changed from our gen one to our gen two for their high volume production. So that is also a very clear sign that we are up. And don't forget, you know, the customers, they are looking at everything in the world, so they are not blindsided, you know.
Doing this fast move on our generation switch from our generation one to generation two for high volume productions, I think that speaks for itself.
Thanks. And the other question I had, on the HBM side, it seems that you're well positioned with Samsung and Micron. What about SK hynix? I think in the roadmap you had for the three companies, you said eventually also go to hybrid bonding. Is that the moment when you can get your inroad there?
That is actually... As they are not using this, they are now using a complete different technology, since a long time, by the way, already, which we are not pursuing with our machines. You're precisely right. The inroad to Hynix is when they change their technology.
And you already have that visibility today, or?
We are working with them as well.
Thanks.
Michael Roeg, Degroof Petercam. I have a question about HBM. Ideally, you want to have your customers doing the 15 hybrid bonding bonds for the 16 layers, and then also the bond underneath. But Samsung is gonna be introducing 3D DRAM next year, and it's gonna be higher than planar DRAM and more powerful. So is there a potential threat that eventually a 16-layer planar stack will become an eight layer 3D DRAM stack, which is same height, perhaps more powerful, but fewer bonds?
I don't think it will displace HBM. I think, this is a potential for integration directly on logic as well, and this is where we also see hybrid bonding being even more critical. So we have to wait and see how that plays out.
I didn't mean that HBM would disappear-
Yeah.
but that HBM would be 3D HBM with higher individual dies.
Yeah, yeah.
Yeah.
Yeah. Well, our customers have roadmaps all the way to 24 layers of HBM, and 3D DRAM is another implementation, another type of memory, but also requiring 3D stacking. So, you know, when they make a choice of interconnect technology, they have to consider performance, they have to consider thermal and electrical performance. So we have to... And we're working on both of those, right? So, they will consider both TCB and hybrid, and we have to wait and see which one they choose, basically. Yeah.
Okay, and then the second question is wafer to wafer hybrid bonding a competitive threat to HBM for you or not?
We don't think so, because, when you consider simply the yield loss, and multiply that by 16, it's almost impossible, we think.
So 14 good die with two decent die-
Yeah, one.
Not do it.
Right. Or one bad die, even worse, right?
Okay. Yeah, well, yeah, you still have more 95% power, but with all the redundancy. But that, that would probably not do it for the customer.
We don't think so, yeah.
Okay, good. Thank you.
Thank you. You tweaked the financial model in the slide that you presented, Richard, relative to what you told us last year and had reiterated in the annual report. Can you give us an update in terms of the timing there, as well as, you know, when you're expecting to achieve that? Yes, you've said sort of the next up cycle, but just, any further detail as to what's driven the upgrade there? Thank you.
Well, you've seen the number one, the market share development over the past year, and all the, let's say, convincing evidence what has been accomplished in the past 12 months in terms of new technology, but also improving on existing technologies. So as I said, with, yeah, let's say, the best evidence of that is the margin achievements. So that should simply underpin a model which we have in place. So it's more a confirmation based on, on the one hand, the market forecast, which has explained extensively for 2025, 2026, probably 2027, will be a next growth cycle. And we, as Besi, should benefit from that. The air is a bit dry here.
Yeah. Thank you, Martin Jungfleisch from BNP . Two questions, please. First one is a bit on pricing. Can you give a bit of a detail on the pricing roadmap on hybrid bonding when you go from the 200 to 100 to 50 nm tool? And how does this tie into, your lifted or uplifted, gross margin expectation? Then the other question is more on the, this new glass interposer or substrate technology. How would this, from a technological side, impact you? Would you need a new bonding technology for that? Thank you.
Well, on pricing and margin, if you, if you look at the model, we, we are not expecting higher margins. Well, compared to, to last model, it's more refinement, as said, and the evidence is, is the margins in the past 4 quarters, how that has developed. On the pricing, we mentioned several times that the first generation hybrid bonder is between EUR 1.5 million and EUR 2.5 million, depending upon what configurations customers order. And then we said for a next generation, the range increases by, by about EUR 500,000. But that also depends on which configuration. Is it in an integrated in a line, or is it standalone? Which type of loaders, which type of software interface? But, but that's roughly the range.
It was also answered by Peter in terms of the comparison between TC bonder and the hybrid bonder generation one plus or two, however you want to call it, also points in those ranges. And you had another question as well?
I think that I give to Chris.
Sure, yeah. So Intel is really the most aggressive in terms of their glass substrate roadmap. And yes, we are aware of what they need without sharing too much detail. The motivation is really to support these very large packages that they have in their roadmap, where it becomes very challenging, you know, to handle the warpage and deformation of an organic substrate. So we will have to learn how to handle larger single unit substrates through our machines. Also, the glass core can be a little bit more fragile and so on. So the material handling has to evolve in order to make sure that we can handle these things with high yield and avoid defects and handling. The bonding process onto the substrate will not change that much.
There's some potential for hybrid in there on the roadmap, but that's very far out. The other... The one thing that is a potential, a new thing for us, is embedding components into the core on the panel, which would have to be done in a 510 mm by 515 mm format. So that is something that we're looking at and something that we have on our roadmap.
But then you can easily conclude, the more complicated it becomes, the more complex, the tighter the accuracies, process parameters, the more opportunity. So that's, that's a very consistent line in, in our history. We always benefit from those kind of developments. Next question. Ah, Madeleine. Not too difficult, please.
It might be a bit cheeky, but you had a slide that showed, NVIDIA and also Apple moving to split die and then kind of potentially moving to chiplet after that. I just wondered if you had a kind of a timeline or, you know, they're currently working on it, or is this just more of an expectation? Thanks.
It's more of an expectation. We see, you know, we can monitor their patents, that they filed that describe the use of hybrid bonding, but of course, you don't know from reading patents if people are actually gonna implement those things. So yeah, the fact is, we don't, we don't, know exactly when they're gonna implement these things.
We're at the hour. So any last question, most welcome. Nigel?
Yeah. So I've got a question on the TCB timeline. So I got confused a little bit in terms of, the tool we saw downstairs, which was R&D, but now in production, and then you introducing a next gen tool. Don't have the slide, so I might have mixed up some of the timelines, but do you expect to ship that gen tool, tool or get orders for that sort of in the next six to 12 months, or is it further out?
So sorry for confusing you. That was not our intention. No, actually, this machine downstairs that you have seen, during our walk through the building, is the latest and greatest, which has, all the features in there for the small bump pitches. So from what I said, you know, from the 10 to the, to the 15, 20 micrometers, so that's all there. And, we are expecting orders, actually, yeah, within this year, so the next months, you know, out of our R&D and, and, activity that we- that I mentioned, that we are having anyways at- since a while with, with, with these customers.
So with that, and I just took a picture to remember this wonderful event. Thank you all again for coming this way, and any further questions, don't hesitate. We're always happy to answer if we can. So all the best and have a nice summer.