Hello, and welcome to Western Digital's New Era of NAND webcast. I'm your host, Rob Soderbery, Executive Vice President and General Manager of the Flash Business Unit here at WD. We are super excited to be here today, sharing with you our vision for the transitions happening in the NAND market, the new era of NAND. It's a great time in the NAND market. We're seeing pricing increase, we're seeing better supply-demand balance, we're having new demand drivers, and upside in all sorts of markets and end-use applications. So it's a great time. We also think it's a good time for us to be out sharing our vision. Of course, we're headed into the separation of the two companies into a standalone flash and HDD business.
As David Goeckeler shared, we're on track for the end of the year, for second half of the year, for that separation. We believe that we have an awesome product lineup to share with you. Last week, we spent some time sharing what's happening on the product side, and really today we'll share more about what's happening on the NAND side. Let me get into the introduction here. Today I'm going to be joined by several very important persons, VIPs, here at Western Digital. The first of those is Aadharsh Kannan. Aadharsh is the head of Economics and Data Science. He joined us at WD from the Office of the Chief Economist at Microsoft.
Deeply skilled and experienced in the economics of the cloud, capital decision-making, pricing decision-making, and he really runs the decisioning engine and the data stack, whereby we make the business decisions on the NAND business for the company. I'll also be joined by Alper Ilkbahar. Alper is responsible for all NAND technology innovation. Alper works extremely closely with our colleagues at Kioxia, Miyajima-san and his team, really innovating on the future of NAND. Alper will share with you the latest and greatest in our BiCS8 technology and our NAND innovation roadmap. So today we're going to start with the business, and we'll take a look at what's happening in this transition and what we call the new era of NAND, and then we'll go on to technology.
Now, today is not an AI launch. We did that last week. We were super excited to introduce the AI Data Cycle. The AI Data Cycle is an architecture which helps us to talk about how storage is driving AI, as well as how AI is impacting storage. Along with that AI Data Cycle announcement, we announced products in four key areas. The first of those was a 32 TB HDD for capacity leadership for that deep content storage. The second of those is 30TB and 60TB class enterprise SSD optimized for building the fast data lake for the AI ingestion and preprocessing activities. Our third launch, I think, was a surprise to the market. It was a market-leading performance, PCIe Gen 5 compute drive really purpose-built for the AI compute challenge.
And then finally, we launched high-performance, high-capacity client SSD drives, really optimized for that rich media, content creator, AI edge use case. So we're super excited about that AI Data Cycle launch. All that's up on the web. You can go see the replay and see all the products associated with that. Now, as I go through the conversation today, you may have some questions. In the upper right on your screen, there should be a Q&A button. You can hit that question button, you can enter your questions, and then at the end of the session, the moderator will read those questions off, and myself and my colleagues will be able to answer those questions. With that, let's get into it.
So I'm going to start by introducing this idea of a new era in NAND. We see a new era in NAND really driving fundamental changes in the operation of the NAND market, and there are four of those fundamental changes. The first is a change in the economics of NAND. And I'm going to say something controversial here, which is we're at the end of the layers race. Now, I don't mean that we're no longer going to see NAND with more layers. That will still happen. But this idea that the NAND industry is characterized by a layers race, and that whoever could get to the most layers the fastest would somehow win, that idea has been thoroughly debunked.
We're going to share with you some data today as to why that idea was actually not only is it not true now, but it was never that good of an idea. So the new economics of NAND, and what's happening, particularly in the 3D era, we're going to do a deep dive. That leads directly to the idea of new capital dynamics. So the NAND industry, for a long time, has operated under a pretty simple investment idea, which is if we target an aspirational bit growth rate, 30% is the number that the industry is centered on, and we build to that growth rate, that the demand will be there and the applications will soak up that demand, and therefore, we should just invest the capital to achieve that growth rate.
What we're seeing in the new capital dynamics is that really no longer true. In fact, we're reversing the way that equation works, which is being much more thoughtful, thinking about what is an appropriate growth rate for the market? What is a rate that will drive revenue and margin expansion? And what does that imply for the amount of capital that we should be investing in this market? And then ultimately, that's driving that new bit growth rate. The third is new innovation vectors. There has been a lot of talk about layers, and we will talk about that today, but we think the bigger story here is a set of new innovation vectors that are really starting to differentiate NAND....
Much like has happened, already happened in the DRAM industry with LPDDR, really optimized for low power, conventional DDR for more capacity use cases, and then high bandwidth memory optimized for performance, we're seeing the same, same change in the NAND market. We're now having power optimization, capacity optimization, and performance optimization as key drivers. In fact, a big part of our launch yesterday was talking about the capacity-optimized products, really driving that 30- and 60TB class, as well as performance-optimized products, driving your PCIe Gen 5 high-performance drives. Alper will talk about these new innovation vectors. Then finally, new market dynamics. The market has been characterized in NAND by oversupply. In an oversupply case, it made the job of buying NAND a pretty straightforward job, in that you could essentially assume ample supply in all the categories.
That's rapidly changing, and this new market dynamic now is the notion of a premium node. So the NAND industry has been a very strange industry. Every 18 months or so, we've built a better product. This product has been lower power, it's been higher capacity, it's been a better product in almost every dimension, and then we've always offered those products to the market at a lower price. Well, that's a funny way to run your industry, and it's starting to change. You need to start thinking about the new and the latest, the performance node, as being a performance-optimized product, really a premium product, and a product that's going to meet premium use cases. So a very different market dynamic, and it requires a different kind of relationship between the NAND suppliers and NAND purchasers.
So those are the big four characteristics we think about in this idea of the new era in NAND, and today we're going to put a little more substance on the bones there of each of these. And just to set up that first conversation, I want to talk a little bit about the market recovery. Of course, we're in the middle of the strongest NAND recovery ever. It started in middle of 2023 with a return to growth of the edge world. The PCs, mobile devices, the consumer devices, really powered us out of the downturn. We've all been waiting for the data center to return, and in fact, we've now seen that the entry standard server market, the hyperscaler market as well, have returned.
And then, most recently, we've seen this big tailwind appear because of the AI Data Cycle and the impacts of AI materializing, particularly in the enterprise SSD world. So we're in a full-fledged market recovery. Now, when we look at the performance of the NAND market over the past 12 years, we look at the NAND industry operating margin, obviously, one number sticks out, which is we've just been through really, really by far, the worst downturn on record. But we think the story is a little bit more complicated than that, and that's not really the story of a 2023 downturn. If you look at the NAND industry from 2018 and prior, what you're going to see is about 20% operating income, so a nicely profitable industry.
Now, from 2019 forward, though, the industry's been net unprofitable. And by the way, even if you take 2023, even if you take that severe year out, it's still materially below that 2018, 2018 era. So really the question is, well, what's happening that's causing this phenomena, and what is the industry going to look like as we go forward? Are we going to see a repeat of this? Are we just back into a cycle, or is something else going to happen? Hmm. Well, that mystery starts to resolve if you look at NAND capital spending. So in the purple range here, I have the capital spending. Now, this is the capital spending of the industry in the 2D era. And what you can see here is the capital spending averaged $6 billion a year.
You know, $5-$8 billion was the range. And so that was the amount of capital the entire industry spent every year migrating from node to node in the 2D era. As we moved into the 3D era, that number changed dramatically. We moved to, on average, $17 billion, so three times over - three times as much capital. We even peaked out at $20 billion there of capital. So we put three times as much capital into the system, and so that light blue region really tells the story here. You put a ton more capital in, you actually did not change your bit growth, your cost economics suffered. We delivered too many bits to the market, and we saw this resulting impact over this 3D era.
Now, one obvious question is to the NAND vendors: Well, why'd you do that, right? Why did you overinvest in the industry during this period of time? And to be fair, there's a bunch of reasons why this happened, and they involve both the demand distortions and supply distortions. You know, on the demand side, right when this transition was happening, we saw a massive amount of COVID demand pull forward. Like $12 billion of demand pulled forward in time. We saw clients go 300-350 million units. We saw the gaming market take off with all those stay-at-home gamers. So we saw a ton of COVID demand pull forward. That encouraged investment. We also got hugely overstated demand from the hyperscalers. The hyperscalers really overstated demand by almost 50%.
So that hyperscaler demand signal was flashing green. On the supply side, there were some similar distortions. We were in a basically zero cost to capital era, so the downside of investing capital was limited. And then we're also facing a more complex market and more players, right? Solidigm was still in the market at that time, as well as you had new market entrants. And so there was this notion that I better worry about defending share. So you had really a perfect storm here of kind of flashing green supply signals, or flashing green demand signals, things on the supply side encouraging investment, and the industry invested itself right into this downturn.
So when I talk about a new era of NAND, what we're talking about is: What is NAND going to look like for the next 4-5 years as we get through this phase, and how is the industry going to emerge? And to unpack that in more detail, I'd like to invite Aadharsh to join me on stage here. Aadharsh?
All right. Thank you so much, Rob.
Thank you.
Yeah.
Now, Aadharsh, you're an economist. How do you think about this market as an economist?
Yeah. Thank you, Rob. As economists, when we try to study a market, we like to study it across two dimensions: supply or sell side, demand or buy side. So to this sophisticated crowd, the supply side dynamics of the NAND market, as it pertains to these three intertwined forces coming together, is quite obvious. Wafer fab equipment capital leads to bit growth, leads to cost reduction. Now, there is a different story when you think about it from the demand side, with these forces working in the opposite direction. Cost reduction passed through to the customers as price reduction leads to price-induced bit growth, which, when paired with revenue increase or revenue growth, leads to us as suppliers getting wafer fab equipment capital. Now, there are a lot of forces that have changed in this market from the 2D to 3D transition.
So as long as these dynamics are clear, we'd use these dynamics to unravel some of these underlying changes in the market from the 2D to 3D era.
Great. So let's maybe start with capital. So earlier I showed that, that chart, that $6 billion going to $17 billion. That was the amount of capital the industry invested each year, and of course, it was spread across the different industry players. Now I'd like to look at a different lens on capital spend, and this is the total capital bill to migrate the entire node for the industry from node N to node N+1 , and compare that total capital bill to the migration in the 2D era, N to N+1 . You can see it here, less than $10 billion. Now, in the 3D era, we saw that go up by 4x-5x, so nearly $50 billion to migrate from node N to node N+1 .
So now, I'm not an economist, but I have a feeling that has a pretty big impact on the market. What do you think, Aadharsh?
Absolutely, Rob. I think what you've outlined is the nodal economics. Now, we could take the nodal economics, pair it with fab economics, which is these nodal conversions happen, and then there are a bunch of other downstream cost implications, which creates the realized capital or realized cost implications in this industry. So we've studied that. We've looked historically at the industry-level cost reduction relative to the capital intensity, going all the way back to 2010. So each dot here is an industry-level statement, where on the X-axis you have the capital intensity, and on the Y-axis you have the year-on-year cost reduction per bit. Of course, this is a proxy, so when you look at this proxy, there's one key feature that explains this broad stroke change that's happened in the market.
In the 2D era, you required lower capital intensity to be able to deliver a high cost reduction, whereas in the 3D era, you require higher capital intensity, delivering a lower cost reduction. Of this supply side economics has implications downstream, which we can actually unravel from the demand side economics.
Now, anybody can build this at home, right? So you've taken industry available, data from the analysts, and you've just plotted this chart.
Absolutely. This is the final realized or materialized version of the cost of the industry.
It's important to remember that between the theoretical nodal investment, so the capital equipment required to migrate a different node, and the realized cost out of the fab, there are a whole bunch of different factors. There's real estate, there's labor, there's like energy cost, there's all of the fixed costs that might be associated with how efficiently a given firm runs their operation. So we're sort of flipping back and forth between this idea of a nodal economics and the idea of the realized fab economics. Is that right?
That's right. That's right. And there could be even, you know, individual factors of different companies, like there could be currency conversion charges. There are so many things at play, but this is a great proxy for the broad stroke effect that's happened in this marketplace.
Great. So let's go forward here. What does this imply? So let's start with this notion of the bit growth rate versus the memory cost. And this is a simple metric. It's actually my favorite metric. And it looks at, well, what bit growth are we achieving and what cost reduction are we achieving on an annualized basis? So what you see here in the 2D era, 27% nodal to nodal bit growth rate, a 24% cost reduction. Now, that's pretty magical, right? Because that basically says: I can offer you more bits for dramatically lower cost, and if anybody offers you a 256 gig phone for the price of 128, that's going to be a pretty compelling value prop, and that's going to drive a lot of consumer behavior.
Now, when we advance to the 3D era, we see it's quite different. That node-to-node growth rate averaging 39% on the bit growth side, but the cost reduction only 11% on that underlying sort of theoretic number of that cost per bit in the node. And so all of a sudden, that 256 gig for the price of 128 value prop doesn't really work anymore. Either I got to invite the customer to spend a bunch more money, or I'm going to have to reduce my price or some combination of the two.
Yeah, your economic intuition is spot on. So I've taken the data that you have in a tabular form and represented it in a graphical form. So on the X-axis, you have the bit growth rate, and on the Y-axis, you have the memory cost reduction per bit. Notice how the 2D line is way above the 3D line, and particularly the 2D line is very close to the 45-degree angle, and the 3D line is actually somewhat below. So this is actually again supporting the thesis that you had about demand, whereas as a supplier, when we are running our nodal calculus in sort of searching the technology for the 2D era, we could actually achieve a significant cost reduction and support that bit growth by inducing demand using price.
Whereas in the 3D era, there is a careful trade-off when it comes to selecting the cost reduction and then inducing the corresponding bit growth in the market.
So let me, let me see if I got this right. So as a NAND vendor, every year, I make a decision to invest capital, and that puts me somewhere along that blue line.
Yep.
Based on that capital decision, I'll get a certain underlying growth rate of bits, and I'll get a certain cost reduction. That's how this works.
Exactly. So the interesting thing is, I might have had a sleight of hand for the economists who will basically ask: Hey, what if you are actually able to just drive 15% cost reduction, but then with that cost reduction passed through to the customers, really achieve 80% bit growth? Turns out there is another economic variable that we would have to look at, and that economic variable is price response or demand price elasticity. So we've actually quantified the demand price elasticity. Turns out it's quantifiable, so we've used advanced econometrics and causal ML techniques to quantify this demand price elasticity or price response. Notably, we'll have to do it across different segments that we serve, and each segment varies by size and price response.
So through quantifying the demand price response and by understanding the supply-side eco- economics, we are actually ready to do something very exciting. So what we are ready to do is, with these two ingredients, which is the demand ingredient and the supply ingredient, we are ready to solve the bit growth equilibrium of this industry, and we've done that, and we've estimated that to be 18%-20%.
Okay, so let's slow down here because so I recognize that picture from high school econ. So basically, what you show me there in purple is the supply side equation.
That's right.
In blue is that demand side equation.
That's right.
Now you solve for equilibrium.
Right. Those are illustrative points. And the advantage of doing this is now we understand the incentive of the industry, and this is what we really mean by end of the layers race. Which is if we, as an industry supplier, try to overdrive the market with just layers as this myopic lever, we would be facing significant economic disincentives. For that to change, two things could change. One is something can change fundamentally in the demand structure, leading for us to be able to achieve a different bit growth target, or there could be something changing from the supply side, where there's a true technological frontier or a disruptive technology breakthrough that happens.
But unless those two forces kind of reconcile or work in a different manner, this is the industry-wide bit growth equilibrium that we estimate for us as an incentive to target.
So that, and that 18-20, that's a right-now number that could move around. You could see maybe more AI demand move that up. You could see, you know, softness in a given market move that down, but you've got an 18%-20%, essentially equilibrium point.
That's right. It's a point-in-time statement that depends on the current demand structure and supply structure, but essentially, what you've said is absolutely correct.
And now I know this is illustrative, but now I think I understand. When you say the end of the layers race, essentially, in order to overdrive the market and generate all these new bits, I have to come down that blue line, which is almost at a 45-degree line down, and that's my margin. So I have to—I lose so much margin out of this equation that the cost reduction can't keep up with it.
Right. In this illustrative figure, if you had to read it, that's how I would interpret it. But the idea is we've now characterized the aggregate price response of this market to be able to know how much of a price reduction is actually needed to be able to even induce growth in our segments of customers.
All right, Aadharsh, I am, I'm thrilled that you're on the case here. Thank you so much for, for sharing that view with us.
Thank you, Rob.
Okay, so that gives you a sense of how we think about the overall supply-demand dynamics. In a much more sophisticated way to think through the long-term dynamics in the market, what supply needs to do in order to achieve the margin goals, the revenue goals, and so on, that, you know, that we put out. Next, I want to turn to how to think about that in terms of capital dynamics. Because one of the most common questions I get is, "Hey, when are you going to start turning capital investment back on? Are you looking for a specific number? Do you need to hit a specific margin target? What's that equation for turning capital dynamics back on?" So let's unpack that. Well, the first thing that has to happen is we have to have our market really restabilize.
This chart is a simple illustration of what we all just went through over the last two years. We saw that the supply, the purple line, you know, massively overshoot the demand line. So the demand took a pause. All that capital in the system continued to produce bits for another year until we had that, one of the largest oversupply situations ever in this market. At that point, pricing came down. That spurred demand on, and vendors started bringing utilization down pretty dramatically. That brought the supply curve down. And so we saw this reverse, where for the last year, the demand curve has been way ahead of the actual production output of the fab.
Now, of course, in that, in that second hump there, that's where we've been going through inventory consumption, with all of the NAND vendors burning all the inventory down that they accumulated in that first phase. Well, we're now hitting the end of that cycle, where the market is starting to normalize again. We're back into our normal operating cycle. Our estimates is that current, demand is running higher than 100% utilized supply. And that is, we're set up for a, a balance to positive situation, from a supply-side economics. And we're getting to the end, on a vendor-by-vendor basis, a supplier-by-supplier basis, we're getting to the end of that inventory burn down. So it doesn't mean we have to invest in capital, but it means the capital decision and the thinking can, can start. So let's share what that new capital thinking is.
First of all, I think that, you know, the old era dynamics is pretty clear. It's if we build it, they'll come. If we, as a NAND industry, can produce 30% more bits every year, applications will appear, those applications will consume those bits, and the economics will take care of itself. Well, clearly, that has not worked out, and we need to reverse that. So we need to pay attention to the economics of the value created, and then what's the bit growth that that economic value can support? What that means is right-sizing nodal migration, really to drive long-term supply balance. It also means that increasingly capital investment is going to be about premium nodes for premium applications, as I was saying earlier. Building high capacity and high-performance nodes for AI, as opposed to simply driving more capacity into the marketplace.
We still have those levers of utilization and inventory, and those levers, by the way, work a lot better if you're not trying to overdrive capital. And then finally, the really, the focus and the underpinning here is on meeting new organic demand and not worrying so much on trying to induce, you know, price-induced demand. So we, you know, we think this is a much healthier capital environment. Of course, lower capital results in more cash flows, and the industry will really evolve to a much healthier state. Now, we're not changing anything from WD forecasting from an FY 2025 view here, but this is really to have a notion of what's this industry going to look like really on a long-term basis. And so the question at an industry level is: where does that WFE, that wafer fab equipment number, stabilize?
So you recognize the chart here from earlier in the session, the purple being the 2D era, cresting in blue in that $15-$20 billion area, and now, that lovely magenta being the new era. Well, the math is pretty simple here. If you're going to grow at 20% instead of 30%, you need about two-thirds of the capital that you would otherwise need. So that $17-$18 billion comes down to a low double-digit, $11-$13 billion kind of number. So our estimate is that the WFE is going to stabilize at that $11-$12 billion range.
And in fact, if you look at the bottom-up analysis and projected spend from the leading market analysts, you'll see that the industry is incrementing its way back into that range over the next couple of years as we get back into this new era and operating model. So that's what we see from a capital dynamics perspective. So with that, I think you have a better understanding of, you know, the economics of the industry. We've said this provocative thing about the end of the layers race, and the question then becomes: well, what are the innovation vectors in NAND, and what is exciting that's happening in our NAND technology stack? To talk about that topic, I'd like to invite Alper up to share his perspective with us.
Thank you, Rob. Thank you very much.
Alper.
Wonderful. So it's great to be with all of you here today. My name is Alper Ilkbahar, and I represent the Global Strategy and Technology team in Western Digital. In my role, I'm responsible for NAND technology development and productization, advanced memory development, and corporate research functions. And, in that context, my team and I work very closely, as Rob stated, with our joint venture partners, Kioxia, in Japan. And talking about our joint venture, this partnership brings two companies, the inventors of the NAND flash memory and the multi-level cell technologies, together in a close partnership, and currently is responsible for producing about a third of global bit output in the NAND industry. Through this partnership, we get to combine our forces in technology development and share equal responsibility in that, as well as IP generation.
The joint venture allows us access to scale, not only in R&D, but also in manufacturing, as well as access to a wonderful global talent pool. Some of the guiding principles for driving our technology leadership are maximizing capital efficiency through leveraging our existing tools and investments, as well as our purchasing power. Driving features, not only what the market is anticipating and expecting, but also, anticipating what the new market opportunities and applications might require around performance, power, and capacity, as well as driving density progression while we are balancing carefully bit growth and cost reductions. So we talk about our industry working on scaling bit growth and cost reductions and how these are intertwined. As technologies, we more or less pursue four major vectors of scaling, and I'm going to talk about these four vectors today.
As I'm doing that, I'm going to introduce a sort of a figure of merit, which I'm measuring the cost reduction percentage over bit growth percentage each of these vectors produce. Of course, we want this figure of merit to be as high as possible and ideally close to the maximum number here, which is a 1. The first vector I'm going to talk about is lateral scaling. Lateral scaling talks about how many bits we get to pack in every single memory layer, as well as removing some of the redundant or actually, sort of support structures there that they were put in to deliver 3D memories. If we do our job right and essentially eliminate any additional cost by doing lateral scaling, we can achieve a figure of merit that's pretty close to 1.
That's a wonderful thing to do on paper, but technically, it also turns out that this is really, really difficult. It requires a lot of work, especially if you're trying to stay on existing tool sets. The second vector of innovation and scaling we pursue is logical scaling. Logical scaling talks about how many logical bits we store on a physical device. So we started life in old days with an SLC, single-level cell, we moved to MLC, three-level cells, TLC, and then these days, to quad-level cells, QLC. The figure of merit for logical scaling is pretty close to 1, not quite, because of some of the overhead that we have to introduce. It's 0.76 for transitioning from TLC to QLC. Still pretty good.
The challenge with logical scaling tends to be with the device physics of storing and packing more and more bits into a single device. We typically see degraded device performance and reliability as a result of that. So in order for us to deliver the desired and needed sort of user experience in the end devices like an SSD, you don't only need a great technology in the end, but you also need to have great systems-level expertise to deliver that experience to your customers. The next vector of scaling, and maybe the most famous one that people know about, is vertical scaling. Here we're talking about adding more layers of memory. Now, adding more memory layers is not all that expensive.
Most of the difficulty is in really going out and spending capital on expensive tool sets, and then you can pretty rapidly add more and more bits. You can grow your bit output fairly rapidly, but unfortunately, you're also significantly increasing your cost. As a result of that, the figure of merit is a mere 0.24 when you pursue vertical scaling. The last vector I'm going to talk about is architectural scaling. What I mean by architectural scaling is the manner in which we implement the circuits that are supporting the memory arrays. Are we putting these circuits next to the array, or tucking them underneath the array, or implementing them all together on a separate wafer and then putting them together at the end? Each of these architectural choices might drive a different optimization point, a different figure of merit.
So focusing you back on the top three vectors, you will see that vertical scaling is actually the least effective way of scaling and growing your bits and driving your cost down. Yet, maybe it's because it's the easiest thing to track and look at and count, unfortunately, many people incorrectly equate the number of layers with technology leadership. We at Western Digital pursue all these vectors of innovation and scaling together. We definitely prioritize lateral scaling and logical scaling before we add more layers and more cost. This is how we achieve superior financial results, as well as capital efficiency. So we talk a lot about capital efficiency and capital intensity. Our definition for capital intensity, the CapEx we spend as a percentage of our revenue. In this slide, I'm going to compare Western Digital's capital intensity with that of the industry.
As you will see on the left side, our market share over the past few years has been more or less stable, around 15%. So we essentially were able to grow with the industry together. Yet when you look at the capital we spent to achieve the same results with the industry, you're going to recognize that our capital spending was substantially lower as a percentage of the revenue compared to the industry. I think that this is actually a pretty objective way of evaluating a company's technology competitiveness and quality of investment decisions. So looking at this data, I actually feel pretty good about where we are. Now, of course, bit and cost scaling and some of the capital efficiencies are not the only considerations when we design a technology node. We also have to look at the features. And what features matter?
These days, the most prominent ones are performance, capacity, and power. Certainly, the new era of compute and AI have been driving storage needs and driving performance needs to drive all of data into the CPUs and GPUs in the data center. So these have been really driving performance and capacity requirements to a next level. Power, which used to be mainly a concern and domain of edge devices and mobile devices, now is increasingly becoming a major concern for data center, as AI applications are making their way into data centers and really pushing the limits of power envelope in the data center. Any device that goes into a data center these days have to really pay attention to power. So to address performance, power, and capacity requirements of the markets, last week, we introduced three wonderful SSD products.
The first one is the fastest QLC client SSD product in the market, delivering performance of TLC-level performance with a QLC NAND, and this is a BiCS6 device. I'll get to this device back in a second. The second one is the fastest PCIe cloud storage product for a compute enterprise SSD device, delivering the best-in-class random read performance as well as power efficiency, again, using BiCS6 class NAND. Finally, 30TB and 60TB capacity-optimized enterprise devices for AI training, utilizing large data sets. As I promised, I want to touch on this NVMe device that is using QLC NAND. To exemplify our commitment to QLC and logical scaling, this device really shows you what we were able to achieve using BiCS6 NAND technology and vertically integrate that with our in-house controller and firmware capabilities.
With this device, we were actually able to achieve significant performance improvements from previous generations' TLC device. So delivering QLC performance and experience with a QLC device is what it will take to drive mass adoption of QLC. And looking at these numbers, I'm actually very confident that we're going to just do that, and not only in this market space, but down the road, also in the enterprise space. I want to talk a little bit about our latest BiCS, 218-layer NAND technology. We call this the eighth generation BiCS or BiCS8. We announced this product late last year, and we were able to demonstrate significant performance and density gains over our BiCS6 generation with this product. You can look at the numbers on the right side, 30%, 21%, 80% performance improvements.
You'll quickly recognize that these sort of improvements in performance are significantly higher than what we have historically delivered generation over generation. And that was possible through a breakthrough technology we implemented in BiCS8, called CMOS Bonded Array technology or CBA technology. In the CBA technology, we implement the memory cell array and the support circuits, we call CMOS, in two different wafers. And that allows us to optimize the process conditions for each of these separately and optimize performance for the cell array and the CMOS circuit separately, delivering superior performance characteristics when they are combined. Then we take these two wafers, we flip one over, and then through the magic of hybrid bonding, we attach these wafers to each other and then implement top layer metallization and deliver it as a standalone single wafer.
CBA not only gives us dramatic performance improvements, we will also deploy some new innovation that allowed us to squeeze the die size quite a bit further, as well as reduce our throughput in the fab, delivering us significant operational flexibilities and improvements. So net-net, in short, we really love the CBA technology and believe that it's giving us a significant competitive advantage for the foreseeable future. So let's look at some competitive numbers that we were able to achieve with the CBA technology in BiCS8. I'm going to show you our analysis comparing some of the results that we collected using 2xx-layer devices from our competitors. So the first one is looking at our die density. You can see here is that we are observing us to be 15%-19% better in die density compared to competition.
As a matter of fact, you will see that our TLC die is better than some of our competitors' QLC die in this regard, which is pretty amazing. Next one, I'm again showing our TLC and QLC die, and I'm looking at the I/O speeds. This is sort of giving you an indication how fast we can feed these GPUs and CPUs with data. And this I/O speed, we are 50% faster than our competitors. Again, thanks to the CBA technology. And the third one I'm looking at is around power, and you can see that, based on our measurements, we are consuming 13% less power to program per gigabyte compared to competition. So on both TLC and QLC, significant power improvements over competitors' devices.
Now, I'd be remiss if I don't talk about, again, about QLC with BiCS8, and indeed, BiCS8 is taking QLC to the new state-of-the-art definition. When we talk about QLC devices, we typically optimize these devices either for performance or for capacity/TCO. Please look at these numbers. These are some pretty amazing improvements from BiCS6, BiCS6 , which we already showed was a pretty, pretty, pretty darn good device. This is giving me really very significant confidence in where we're going to take QLC technology moving forward with BiCS8. So I have so far talked to you quite a bit about BiCS8, and a lot of these things we had covered over the past months.
But I also wanted to show you something pretty new, and I'm really, really excited to share with you a preview of the BiCS8 2-terabit QLC die. We've designed this die to meet the data center and AI storage needs. We're going to shortly announce this product, but I wanted to share it with you today. This happens to be world's highest capacity memory die, so you're seeing something pretty amazing. We typically show you a wafer, but I felt like the view of a wafer doesn't quite convey what we have achieved. So if you allow me for a quick second, I want to show you the die. I have one here. And if my helpers here can please zoom in to this die that I'm holding on my finger here.
This is the size of the die, much smaller than my fingertip, and it stores 2Tb , 2Tb on this. So this is what I think technology leadership looks like, and I'm really thrilled about this. Now, in conclusion, my last slide here, I want to give you a sense of where our future innovation is heading. Talked about the CBA. It is definitely opening up new horizons for innovation for us. It's going to allow us to explore new trade-offs around performance, functionality, and density. We're very excited and going all in on CBA. It's going to be the basis starting point for our future technologies. We've talked about our emphasis on QLC. We will continue being the leaders in QLC and drive it into many more markets.
We talked about the new era of compute and AI driving new levels of performance and power requirements. We're going to be right up there, driving to the new heights of performance and power efficiency, while we drive our journey through the cost and bit growth scaling and pursuing capital efficient design, leveraging all of the design vectors that we talked about. With that, I thank you all very much for your attention, and invite Rob back to conclude our presentation.
Super. Now, I think we could build a pretty big eSSD with that 2Tb die there.
Well, yeah, not physically big, but high capacity.
All right, that's great. Okay, so let me wrap up with new market dynamics. Again, we do have questions coming in. I just saw the screen there, so please send those questions in. We'll have a chance to get the team up here to answer those. So we've spent a lot of time on the old market dynamics, the layers race drove oversupply. What's going to happen in this new era dynamics? Well, as Alper talked about, we're very focused on delivering these new vectors of innovation, power, performance, capacity, and ultimately delivering premium nodes. BiCS8 is going to be an amazing node. It's going to be a long lifetime node, and it's going to be a node, I think it's going to find its way into these premium use cases.
From a market dynamic perspective, we'll expect these will end up being priced at a premium and will be used in premium applications. It's a pretty significant change for the market. What that also implies, both looking at the rate of capital going in the systems, the rate of nodal dynamics, is that we're headed to more of a long-tail, multi-node world. Today in the fab, we're running BiCS 3, BiCS 4, BiCS 5, so running long a number of generations of devices. And those devices have found their way, some into various long-tail use cases, into high-capacity use cases, into a variety of different use cases. We expect that trend to continue.
So from a customer perspective, customers now are going to need to be more thoughtful about, not just, what is the latest node, but what's the node that's optimized for their particular application and for their particular use case. All this leads to the need for really greater supplier alignment and customer alignment and greater thoughtfulness in terms of the overall nodal planning. We believe there are plenty of bits out there for the premium use cases, and a key part of our strategy is to really rotate our bits to the highest and best use in the marketplace. New demand is wonderful. We don't always need to solve that new demand necessarily by putting more capital in the system. We can solve that demand by really rotating our bits to the best and highest use case.
So pretty fundamental different dynamics in the marketplace going forward. With that, let me go ahead and wrap it up before we go to Q&A. So we've talked about a lot here. We started with the new economics of NAND and this sort of provocative statement of the end of the layers race. And we've hopefully shown that that indeed is what's happening and why that's happening. You know, ultimately, what that drives, that drives this idea of nodes must be long-lasting. They've got to be future-proofed, because we're not going to be on this hamster wheel of nodal migration. Second, new capital dynamics, low double digits in terms of WFE number, really supporting a right-sized nodal migration. I'm often asked about: Are you worried about missing the market? No, we're not worried about missing the market.
It's quite easy to catch up to the market, but it's very difficult if you get ahead of the market. So conservative stance, we have a view, a forecast on what we think the market will do, and we're going to inch our way towards that, towards that number.... Ultimately, if we do our jobs right, those capital decisions are not blindly driving an exabyte growth rate, but they're specifically delivering revenue growth and margin expansion. Alper shared with you how, the new innovation vectors, I won't repeat that. We're super excited that we're back on offense here with our nodal strategy, and with both BiCS6 and BiCS8, which are really powering this new generation of amazing products. And then finally, the new market dynamics.
We look forward to working with all of the market stakeholders and finding our way through this new era of NAND. With that, I want to thank you for listening to our presentation, and invite my colleagues back up on the stage, and we'll do some Q&A. Come on up, guys. So, Eric, what do we have first?
All right, Rob, first one here, there's a few questions on BiCS8 and our capital planning. So basically, what's the status on BiCS8? And then how are you thinking about making capital decisions into that node, as you look to the future?
Yeah. So, BiCS8 has been a wonderful node. As when Alper shared this crazy idea of jumping to CBA and leapfrogging the market, I know there was some skepticism, but I believe it's been the fastest ramping, fastest yielding node we've had in our history. Is that right?
Yes, from what I can remember, that's absolutely the case.
So, great job to the NAND engineering team, a fantastic result. So we have BiCS8 capital in place. We're using that for going into all of our qualification cycles, will be in volume production with BiCS8 in this calendar year with leading products. Now, incremental BiCS8 capital will be a function of what I talked about, which is our ability to see and confidently forecast sufficient demand to support our revenue plan and our margin expansion. So we'll ease our way in. We are actively ramping up BiCS6, which is delivering market-leading performance, and so we have a great device today to really drive the opportunity, and we have a lot of scale in front of us with BiCS8.
Thanks, Rob. Related to that, is there a risk that you see right now that you could potentially undergrow the market? Or, and how do you think about your capital, our capital planning on that line of thought?
Yeah, well, I guess our thinking in terms of the market growth rate is the numbers that Aadharsh showed are really a point in time, and there's a lot of things that can happen, technology changes can happen, market changes can happen, so we'll stay adaptive as we think about the marketplace. It's much easier to catch up if we happen to undershoot the market than it is to put ourselves back in the situation in which we've overshot the market. So you can look forward to us being conservative, but absolutely, we don't intend to cede any share necessarily as part of that strategy.
Thanks, Rob. Now, when we talk about the 18%-20% bit growth target, how do we think about the TLC versus QLC mix in that number?
Yeah. So when we look at, when we look at those nodal numbers I showed you, that figure of merit, which what Alper showed, what you saw was the figure of merit for that, multi-bits per cell was, was much, much, much, much better than layers. And you see that in terms of nodal migration, where these nodes produce almost no benefit other than QLC. So we're seeing very, very little cost reduction with TLC on a like-for-like basis, say, going from a 512Gb die to a 512Gb die. Almost all of the cost reduction is actually coming from QLC. So QLC, estimate of the QLC is built into this. I wouldn't expect to see that 18-20 go way up.
But that 18%-20% is gonna have a pretty strong mix of QLC, helping us get to those numbers in a more capital-efficient fashion.
Thank you. Now, a question for Alper.
Mm-hmm.
So as you think about the limitation to lateral and logical scaling, so how do you kind of think about that going forward? And isn't that basically why layers are being added? So if you can kind of give your thoughts on both the dimensions of lateral and logic going forward.
No, this is an excellent question, but we have been consistently able to actually do lateral shrink at every generation while we're progressing in the layers as well. We have never stopped thinking about that and found always new ways of doing that. Of course, it's very difficult. I said it was very difficult. It's the most difficult thing to do, yet our engineers have been super innovative, totally coming up with new ideas that they dream about, and we challenge them to do more. We try more and take more risks if necessary, but we have been able to deliver.
Of course, it gets more challenging every time we try it, and we rarely know how we're going to do two more of those things, but we always have ideas for the next one, and by the time we get to that one, we'll have ideas for the one after that as well. So it's going to be a regular story about doing that, but... and we'll continue pushing the difficult, challenging angle of that, because the returns are so powerful.
Thanks, Alper. Now back to you, Rob, a question on the data center and our participation. So how do you think about our Western Digital Flash data center strategy and outlook as you think about our portfolio of products?
Well, we've been working at improving our posture in the data center for a number of years, really, since David and myself joined the company. We've made tremendous progress against that. And really, the proof points were in the deliveries of last week, really delivering, you know, market competitive and market-leading products in the key necessary swim lanes. We're relatively low share in the data center today, so that means it's all upside from here, and our goal is to mix into the most favorable distribution of bits in the marketplace. We don't want to be overweight any given sector, and we don't want to be underweight. So we progressively intend to increase our bit allocation to data center. We'll mix more towards the data center from here, and we're feeling really comfortable where we are in the marketplace today.
I'm sure, looking forward to the Investor Day, we'll update new guidance then.
Thank you. A lot of appreciation on the story between the 2D and 3D era nodal economics. But I guess the most upfront question is, if we knew about this going into it, why did we kinda continue to drive in-
Well, well-
hit that bad 3D era of economics?
We didn't know about it going into it. There are two types of economics, positive and normative. Positive tends to look back and say, what happened? And so, as Rob shared in his slides, there were demand distortions and supply distortions that actually made it hard for the capital allocation to be made in a economically optimal manner in the past.
Thanks. Aadharsh? All right, so on the data center follow-up here, Rob, with regards to AI demand and things like that, how much content do you see as, like, going into these general purpose servers, and how does that scale over time?
The question, how much content in the AI swim lanes?
The AI server swim lane.
Yeah.
Yeah.
Yeah. So, as we talked about in our AI Data Cycle, the big news is that, in the sort of second stage of the AI pipeline, which is data's been pulled out of your mass storage device, out of your HDD content repository, and then goes through a whole set of, data preparation and data wrangling tasks. Back in the database days, we would have called this ETL, for extract, transform, and load. Now, we use fancy words like vectorization. So that is what is really driving the exabyte demand. And that, we forecast that exabyte demand as a significant, significant, tailwind to, you know, data center demand. Now, data center demand is about 20% of the market today, and so you're seeing that number move up, with those, high-cap enterprise SSDs.
There's a separate phenomenon in terms of the uptake of the PCIe Gen 5. Those tend to be lower capacity. They're not moving the exabyte totals as much, but they're higher priced, and they do better on the revenue side.
Thank you. Question on build up on top of our platform strategy. So what is the team focusing on from a controller firmware and systems perspective across the portfolio, and how do you think about the outlook forward?
Well, maybe I'll start with that, and then I'll ask Alper to provide some comments about why this is so important. So, our strategy is super clear. Over the past couple of years, we have consolidated our entire portfolio onto a single platform architecture. So previously, we had multiple architectures for each segment in which we played in. Our UFS architecture was different than client, which was different than data center, which was different in consumer. We're now on a single architecture, and that gives us tremendous scale, and it also gives us the flexibility to serve different swim lanes. In fact, you've seen us move a very significant part of our portfolio between these swim lanes, really aiming to serve the markets that are coming to us.
So that that platformization of our technology architecture is well underway, and each of these new platforms that we release is a further proof to that. Something I think is a little bit underrepresented, though, is just how important that platform innovation is in the integration of the controller with the underlying NAND. And maybe, Alper, you could provide a little more color on that.
Yeah, I think it's exactly right, and we talked about the QLC drive that we showed, right? Our teams were able to work very closely together to extract the utmost performance and experience that is required in a PCIe client drive from a QLC NAND, that otherwise it was not possible with TLC. So that collaboration, that understanding and planning and architecting the devices and the know-how is super critical getting that level of performance.
Yeah, and not all ASICs are equal. A given NAND and a given ASIC really come together to deliver these performance stats and these capacity stats and these power stats.
Okay, I think... Let's see. I think that we're hitting kind of at the end of the questions here.
Okay
... that are new and breaking. So any last comments or?
Well, again, thank you for joining us for the New Era of NAND webcast. We're really excited about the view of the market, and we think what the market has in store for our new public flash company, as well as the posture of our technology and our platforms, the tremendous success with BiCS6, now with BiCS8, and all the things that we have in front of us. So again, thank you for joining. Thanks to my colleagues and the folks in the room for attending today. Well done!