Thank you very much for joining us today. Before we begin, please note that today's presentation will contain forward-looking statements based on management's current assumptions and expectations, which are subject to various risks and uncertainties. These forward-looking statements include expectations regarding our technology and product roadmaps, our new business models and multi-year customer partnerships, our business plans and performance, market trends and opportunities, and our future financial results. Please refer to our most recent annual report on Form 10-K, our quarterly reports on Form 10-Q, and our other filings with the SEC for more information on the risks and uncertainties that could cause actual results to differ materially from expectations. We will also make references to non-GAAP financial measures today.
Reconciliations between these non-GAAP measures and the most directly comparable GAAP measures are included in the earnings releases for the relevant periods and in the appendix to the presentation materials, which are posted in the investor relations section of our website. Please welcome Vice President, Investor Relations at SanDisk, Ivan Donaldson.
Thank you very much. I just want to say thank you to everyone for being here today. I've been in this industry for 22 years, and the journey with SanDisk has just been astounding. It's been an amazing ride. We have an amazing management team and board of directors, amazing employees across the globe, and we're just really excited to be here today. I'm going to talk a little bit about the agenda just really quick. We'll go over the, obviously, company overview, strategic vision, future for the company. Then we'll go through the technology roadmap, or really the way I think of it as our innovation engine in the company, which is just astounding. We'll also take a step and look at the industry transformation, what's happened, how we got here, essentially, which is pretty phenomenal.
Then we'll go into a deep dive of the AI infrastructure. Essentially, why? Why is this happening? Why are we seeing such a step change in demand, and where we see that going forward? What are some of the key variables and dynamics for that? Then followed by the financial model from Luis. Really trying to deliver that we have a tremendous opportunity to drive shareholder value well into the future, and we hope that's going to be the takeaway from today. Then at the end, Alper will come back up and talk about future innovation roadmap and where we see that going to conclude. Again, where we see some future technologies and emerging memory opportunities, followed by Q&A.
We tried to take into account a lot of your questions of the last year of what you guys are most interested in, so hopefully we'll get to all of those. Again, I appreciate everyone to be here. Thanks for your support.
SanDisk, the flash storage vendor, was added to the S&P 500.
With a look at SanDisk.
We're going to be talking about SanDisk today.
SanDisk has more than doubled.
All right. SanDisk.
Look at that chart on SanDisk.
SanDisk, the best stock in the S&P 500 this quarter.
Many on that conference call for SanDisk specifically. What do you want to ask?
I want to better understand how inferencing is going to help or impact SanDisk demand into next year. I want to know how SanDisk is going to be able to preserve margins.
Please welcome Chairman and Chief Executive Officer at SanDisk, David Goeckeler.
All right. Welcome. It's great to be back here in the room where we launched this company 18 months ago. A lot has changed in that time, and today we're going to talk about the company going forward. I can tell you, as I was talking to some of you as we were preparing or just this morning in gathering, we were talking through a lot of stories of stuff that has happened over the six and a half years, whether it's road shows or conversations we had about this franchise. I joined Western Digital, of course, in March of 2020. Actually, the same week that COVID started. People know about that a lot more than they know about me starting at Western Digital.
But we started on this conversation, I think, that's been going on for almost 6.5 years now about where we were going to take this franchise. I thought from the beginning this was just an unbelievable franchise, that we could really unlock the value. But it took some time. They're big markets. It was going to take a lot of moves about how we got there. I can tell you, as I stand here today, I feel like I've finally gotten to the starting line of where the real value creation is going to happen. Now, that may seem like a pretty big statement based on what's happened since the launch.
But I think when you walk away from here today, you'll see the conviction we have in what is really the earnings power of this franchise and why going forward, we've finally got things structured in a way we can really start to reveal that on an ongoing basis. All right. I'm going to set some context here on how I think about the business, the big picture of how all the different things we're thinking about, how we integrate it. This is a business where you can't just think about one thing. There's two, three, four variables that are in motion at all times, and it's about getting them balanced and how are we going to make changes not only to the technology. We're always changing the technology. That's something that goes on and on all the time. We're world-class at that.
And you're going to hear from the people today that are really driving that, and that's an incredible story all by itself. But then also, how are we thinking about the business model around that? How are we structuring the business? How are we changing the relationships with our customers? How are we allocating capital in the business? All these kinds of issues to create, just on an ongoing basis, relentlessly create intrinsic value in the franchise that will be revealed as we continue to execute the business. All right. There's kind of three big categories to that. The first one, when I was putting this talk together and I was thinking about this, I was going to spend some time going over what we committed to back the last time I was on this exact stage. So what was that? February of 2025.
We made a bunch of commitments of what we were going to do with the company. I think it's just fair to say it went pretty well. I think most of the things we said, we've delivered on. We talked about we wanted to win in data center. We needed to establish data center as a major growth pillar of the business. That had been, I think, an issue with the company for quite some time. I think we're getting there. I think we just delivered significant outsized growth in the last fiscal year. I think you're going to see that continue as we go through FY 2027. We talked about we're going to focus on the consumer business, and hopefully you stopped by outside and saw the products. They all look very different. It's the same products, different branding, showing up a different way.
This is just an incredible brand that I think is a bit of an unpolished gem, and we're continuing to work on it. We've gained two points of global share in that business. We'll talk a little bit about that business today, about why it's so important to the business model as well. We made some controversial statements at the time, quite frankly. We got up here and we said, "Hey, pricing is going to inflect positive in the second half of the year." That became a big talking point. No, it's not. Yes, it is. No, it's not. Yes, it is. It turns out when we got to the end of the year, things were going in the right direction, and if anything, we significantly undercalled it. Anyway, I think things went well.
Those of you that believed in the company back then and invested in us, we really sincerely appreciate that. We take that very seriously to be good stewards of your capital. We worked very hard, and you got a good return, and we're very happy about that. That's the past, right? We're not going to talk about that. We can't go back to that point in time. That was a special point in time. It's not coming back. What we can do is talk about going forward and how we're going to create value from here going forward. I can tell you, everybody you're going to see up here on stage just has unbelievable conviction that this franchise, we're finally starting to reveal the true earnings power of it. That earnings power is going to go on for a very long time.
We'll talk about that. All right. Let's start with this first. When we say we have this capital allocation strategy, number one, we're going to invest in the business. That's always the most important thing, invest in the business. What are we talking about? I'll preface this by saying there are a lot of things we committed to doing back when we launched the company. Like I said, I think we were largely successful in making progress on those. There's a lot of things over the last 18 months, and especially the last 9 months, where we were presented with some opportunities to really change the business. A lot of momentum. We took advantage of those. I think we have fundamentally restructured the business.
We kind of put a pin in the map on this day, early this year, realizing we could see what was happening in the business. We could see we were restructuring the business, and we're going to need to stand up and explain it all to you because it is very different what it was back in February. Anyway, with all that said, let's talk about all the things we did and kind of how we thought about investing in the business as we went. The first thing, the most important thing, we're a technology company. If you don't have great technology, you shouldn't be doing what we do. You always have to have unbelievable technology. This is always the most important thing we're going to invest in.
I think we're in the best position we've been in in a very long time as you look across the portfolio. Something I've been doing my whole career is investing in technology and thinking about this what I think of as a multi-horizon innovation investment plan. You can't just think about what's going to happen next or what's going to happen this year or next year, and you can't just think about what's going to happen 10 years from now. You got to think about all of it, and how do we invest across this entire horizon to make sure we have the right technology today, tomorrow, five years from now, 10 years from now. This is where I think when you look at our business, it starts with what Ivan said. The engine of the company is the BiCS roadmap, is the fundamental NAND roadmap.
If you don't get that right, it's kind of hard to make up for it at the system level. It's kind of hard to hide that. You have to have really good NAND technology. We're constantly investing in many generations of NAND technology. We don't talk about all of them all the time. We just announced BiCS 10, I think a couple of weeks ago. By the way, we announced BiCS 9 yesterday, and I'm sure everybody is confused. Why did you announce BiCS 10 three weeks ago and then you announced BiCS 9 yesterday? Because 9 is before 10, and we thought you would have announced it first. Alper will explain that I think it's a very important point to understand.
The technology strategy is changing based on the fundamental BiCS architecture, and he'll explain to you how we now have multiple ways we can move this technology. So we're investing in BiCS 9, BiCS 10. There's actually people working on BiCS 13 right now. One thing you should have confidence in, and this is with our partner, Kioxia, we have a long roadmap of really strong fundamental NAND technology. One of the big advantages of the JV, the big benefits of the JV is we invest together on R&D, and together we're a third of the market. That means we can invest as much or more than anybody else in the market in making sure we have the best technology. So that's going to be there for a very long time. Alper will go through that. We also build the systems capabilities.
The BiCS investment gets you through the wafer, through the fab. The wafer comes out of the fab, I have to do something with it. Basically, we could just go sell all the wafers, but we actually turn them into systems ourselves. So we have people working on all of the controllers, how to build SSDs, how to build all the different products in the markets we operate in across consumer, across edge, and now across data center. So you're going to see Khurram up here today. He's going to be talking about AI in the data center, and it's really his team that builds all this stuff. So just enormous systems expertise across all of these markets.
I think this is one of the big. If you look at the foundation of the company as a technology, one of the positions that we now are in, and one of the reasons I have so much conviction about the future, is we now have optionality across the entire market. We have a very unique consumer franchise. Edge, we've always been very strong. PC, smartphone, IoT, all of the things on the edge, and now we're very strong in the data center. So we've got all this optionality, and we're going to keep all that optionality, right? One of the big things of the strategy of the business is make sure that all of this technology remains very relevant, very on point, to continue to drive this innovation across all these different markets. So that's always going to get invested in.
Now we pick our head up a little bit and look a little further down the field. One of the things we talked about 18 months ago as we announced this product or this strategy around High-Bandwidth Flash. If you would ask me, a lot of stuff has happened in the last 18 months. This may be one of the things that is actually I'm the happiest about. We basically stood up here and said, "We're going to build this thing called High-Bandwidth Flash." I think the reaction was pretty much, "What are you guys talking about? Nobody has any idea what you're talking about. What is High-Bandwidth Flash?" We only knew high bandwidth memory. But we were kind of targeting this idea that, hey, when we get to inference, AI is a massive opportunity.
At that point, all the focus was on model training, and appropriately so. But I think we were looking down the field, our team, this wasn't me, this Alper and his team, a tremendous amount of insight to see that, hey, at some point, we're going to move to this inference phase. You're going to have to scale this, and we're going to have to come up with a different memory architecture or storage architecture for inference to really scale. We have an incredibly important technology that we can bring to the party. We announced it, and we said we're going to form an ecosystem, and we are going to start building this product. Alper will be here later. That's the last talk that we're going to have today. He'll be here later and give you an update where it's at.
But I think if anybody was at FMS last week, they probably saw there was a lot of activity around High-Bandwidth Flash. Even last year at FMS, High-Bandwidth Flash was awarded the most innovative technology in the industry. So, that's a horizon, and I know a big question is going to be, when is it going to ship and all this kind of stuff. We'll get to that. We're not going to get to that today, so it's a little bit of a spoiler alert, but we're getting there. I think the ecosystem is being developed. People are coming to the table. A lot of very good discussion. That is on the horizon. Then we look even further down the field. We had this idea of 3D Matrix Memory. Little longer-term project, continue to make progress.
So we're basically. This is the first priority, invest across all of these things and make sure they're all healthy and bring them to market to the extent that we're getting feedback that they resonate with our customers. All right. The second thing we talked about was get to this idea of a cash positive balance sheet, right? It's not exactly a novel idea, but it's like we need to get the debt out of the company. I think this is something we're very happy about. It happened faster than we probably thought. We got a receptive market, and we got to this position where we don't have any debt. We have significant cash reserves. One of the things Luis is going to talk about later today is this franchise.
I think one of the things you guys are all seeing, when you start to scale this franchise, it really is good at generating free cash flow. That's good, right? That's what we think our job is generate free cash flow for all of you. So we're in a position where we can generate a lot of free cash flow. What are we going to do with it? Luis will talk about that when he gets up here. But we feel like we're in a very good spot there. All right. Now, that was kind of the things we need to do every day, but once the business started turning, we started thinking about, okay, what else can we do to invest in this business? That's what we want to do first. First thing, we want to invest in the business.
So the way I think about this is we constantly go through a process of how do we systematically de-risk the business? How do we systematically make investments? We're basically de-risking the future. There are a couple things that were a big part of that we were able to get done. First one was extend the joint venture. It's one of the first things you saw us do. We invested over a billion dollars with our partner, Kioxia, which was a recognition of the scale of this joint venture. But one of the first things we spent our money on was making sure we had production of NAND from 2030- 2034. That was very important to us because we have a tremendous amount of conviction in the future of this franchise.
If you want to be in the NAND business, you need to have a NAND fab. That wasn't always the case, by the way. Part of the issue with the industry in the past is you could procure NAND very inexpensively because the people that own NAND fabs seem to be selling it at prices that were not basically the marginal cost. I think that world is gone, quite frankly. My personal view, it's not coming back, and it's going to be very difficult to be in the NAND business unless you have access to a NAND fab, which we do. We have it at scale, and we have all the R&D benefits of that and all the manufacturing benefits. The other thing I hear sometimes is, "Oh, NAND's not that hard. NAND's a commodity.
Anybody could do it." Well, if you actually believe that, I would encourage you to go to Yokkaichi and take a look at the fab that's there. If you want to duplicate that yourself, find a lot of money and I'll see you in about 10 years. It is extraordinarily difficult to be in this business, and this JV is a huge strategic asset for us, and we took the opportunity to extend it when we could. Now, one other thing we did in this kind of de-risking, we knew that, look, we want to play in the data center market. Why do we want to play in the data center market? You guys know all that. It's a very attractive market, but also it was the market that's going to help us change this dynamic with our customers.
We want to go from this kind of negotiate price every quarter, this kind of highly transactional, highly volatile business, and we want to turn this into a business where we kind of dampen that cyclicality. We have more long-term relationships. We have more long-term visibility into what demand's going to be. The customers that are most likely to do that are the data center customers. There's a lot of reasons for that. We can go into that later if we want in the Q&A. We have a willing partner that wants to go down that path with us. If you're going to be big in the data center business and you're going to grow that business, you need access to DRAM. We're no different than anybody else. The DRAM market is very tight, as they say.
We needed to make sure if we're going to go to our customers and say, "Hey, we want to strike a five-year agreement on selling enterprise SSDs, and we're going to put this huge contract together that's worth tens of billions of dollars," we need to make sure we have access to all the pieces to actually fulfill that contract. This became extraordinarily important to us. It maybe wasn't as clear to all of you at the time that we were putting all these building blocks in place that was leading to this different contractual relationship with our customers, but that's kind of what we were doing. We had the opportunity. We took a 4% equity stake in the company. It's worked out okay. Luis is smiling down there, our CFO.
I think this quarter may be one of the few times where our GAAP earnings are higher than our non-GAAP earnings because we recognized like an $800 million gain on that investment. That's been going well. But we didn't invest in it for the return. It's great that we get that. We invested in it because we need access to the technology. Look, I think the net of all this is we've taken the opportunity over the last 18 months to really make sure the foundation of the business is just incredibly solid. We have the right technology. We have the right roadmap. We have the right innovation. We have the right relationships. We have access to all the things we need for years and years into the future. On top of that, we're going to think about how do we change the business model?
How do we get this franchise where it is sustained value creation over the long term? There are a lot of questions about that. I remember when I took this job, we went through the separation. Some of you were actually pulling me aside saying, "Dave, what are you doing? Why are you taking this job? Do you realize this industry has never made any money?" I am like, "Yeah, I got it. We will figure it out. We are going to get there. We can change things and we can get a better outcome." That is one of the things we really believe in, can change things and get a better outcome. How do we think about that, what we are going to change? This is kind of how I think about it, the three imperatives for sustained value creation.
These are not, I have talked about these before, but I am going to just go through them a little bit. Number one, we have to increase profitability. I think if you go back to that time of February of 2025, when we launched the company, this was probably the big debate. I actually love all you guys because there is always a debate. No matter when we have a conversation, there is always a debate. There is always a debate about something. As soon as you get past that debate, there is a new debate. There is always a debate. The debate, I think, a year and a half ago was could you get profitability where you need it to be? I think the debate today is the second issue. How do we reduce cyclicality? Now it is like, "Oh, okay, Dave, we get it.
We understand profitability is at a level it has never been before, but it is just a matter of time." Just a matter of time. The faster things go up, the faster they come down, all these kinds of things. It is like now it is all about cyclicality, and we are going to talk about that. I think we are doing some things. Again, over the last nine months, we have been extremely intentional about the way we run the business and the way we structure our relationship with our customers to try and reduce cyclicality. We are not saying cyclicality is going away. The whole world is cyclical. The whole business is cyclical. What we are trying to do is get this wild cyclicality of this business, just dampen that down, get more sustained value creation.
If you do those two things, in every business, you got to grow. You got to get consistent revenue growth. This is hard. In most businesses, this is hard. It is hard to grow. Usually, you run out of TAM, so you have to go acquire and do all kinds of things, and we will talk about in this business, it is actually quite different. All right, let us dive into these just real quickly. One last point. What I said earlier, you have to do this across all time horizons. It is not about maximizing value for the next two weeks. What is pricing going to do the next two weeks? That is important. Got to do that. You also have to do it across the midterm, the long term.
So we are constantly thinking, how do we balance these things across all time horizons to get to this point of sustainable value creation? So it is important. You got to think in two dimensions. A lot of the questions I get, a lot of the questions we get, tend to be about one of these things independent of the other two and about a certain time horizon. Just so you know, when you ask us questions, what we are always doing is trying to translate your question into, okay, how do I think about that question across all of these variables and all time horizons and give you an answer that makes sense? Because answering point questions does not really help advance your understanding of the whole franchise of what we are trying to do. All right. Let us just go into these a little. Profitability.
Like I said, honestly, I think profitability was always the one of these three that was just hiding in plain sight. When I was managing the franchise, when they were together, and we went to separate the company, it was this issue, will this business really ever create value over the long term? To me, that seemed like I just did not have a question about that. Because I thought the intrinsic value was always there. The question was: could you get the business model right? The way the business worked, the way you engage with your customers, could you change that to actually let this intrinsic value come out? It really starts with this kind of observation. It is a very simple observation, but I think some people forget it sometimes. We own the whole stack. We do everything. This is not a fabless semiconductor company.
That is impossible. We own the NAND IP, so the fundamental IP it takes to build NAND. One thing you should take away from the fact that I just said we have people working on BiCS13, which is going to be launched sometime, I do not-- Do not come back and ask me 100 questions about I just said when BiCS13 is going to be launched. You are talking well beyond 2030. People are working on technology a decade into the future. This is just on its face, extraordinarily difficult to do. There are hundreds of engineers that have dedicated their lives to building some of the most sophisticated semiconductor technology, and that is just one part of this chain. We have all of that, and we have it at the largest scale, again, because of our collaboration with our partner.
We have it at the largest scale of anybody in the world. As somebody that has managed technology franchises for decades, very large technology franchises, your market share makes a big difference in how much you can invest in R&D. You can basically invest R&D commensurate with your market share. What that says is when you get bigger, you should get better if you are doing your job right. I think you are seeing that show up. You see that show up in our roadmap, and Alper will go through that. Once you have the NAND IP, we do not call somebody else up to build the wafers. We have our own fabs with Kioxia. Like I said, they are quite spectacular. They are some of the largest fabs in the world. The scale we operate at is just incredible. We do all the front-end manufacturing.
Wafers come out of the fab. What I said earlier, we have another engineering team, hundreds of people that are working on all the systems expertise. How do I take this wafer, cut it up into die, put it in an SSD, put it in an enterprise SSD, a client SSD, put it in something that can go into a car, whatever it happens to be. We have all those people, too. We are doing all that work, and as I said, across all markets: consumer, edge, and data center. Then we have the back-end manufacturing as well. You can go to Malaysia. We have a big factory there. We have a factory, again, with a partner in China. We do the back-end manufacturing, too, and then of course, we do the whole go-to-market piece. There is a lot of debate. There I said it.
I am adopting all of your language, there is another debate. There is this thing like, "Oh my gosh, your margins are so high. Your margins are higher than the fabless guys." We do a lot more than the fabless people. Nothing against them. They are great companies, incredible companies. When you look at this, there is just minimal profitability leakage across this whole model. That is why I say, this thing has been kind of hiding in plain sight the whole time. The issue was the business practice was wrong, and you never could see it. Not only was it there, but we have been doing this for 25 years. It is crazy. We have got 25 years of paying engineers hundreds of millions of dollars a year, right? These engineers, they are not cheap. They are expensive. They are a little bit temperamental sometimes.
They can be hard to manage. They are wonderful people. I am one of them. It is not something you just roll out of bed and do. So we have been investing that for years, decades. Quite frankly, one of the things I see today when people say, "Oh, this is just memory, it is a commodity," I am like, "Yeah, you try it." You try to do this stuff. It is incredible what people are doing. So we have been investing in this for decades. In fact, Alper will be out here later. He leads this team. A lot of the people on that team have been at SanDisk way before the Western Digital phase, right? That was just a little phase of the company. So we have got this huge amount of expertise there. Then on the manufacturing side, we have been investing billions of dollars.
Again, for 25 years, we have just been investing billions and billions of dollars of building out this incredible scale manufacturing capacity. In many ways, like I said, I think this profitability thing was just hiding in plain sight. This thing is like a coiled spring that has been fed for 25 years, and it really has not produced. Now it is producing. Now it is about how do we sustain that over long periods of time. Again, I think hopefully one of the things you take away from today is that the way we are engaging with our customers, the level of strategic engagement has fundamentally changed. That allows us to completely reveal the profitability of this franchise. Now, a little bit, what did it take to get here? How did this get unlocked?
A gain, if I go back two years ago, it was like, "Oh, Dave, can NAND ever be 50% gross margin again?" I am like, "Yeah, of course it can." Right? I think we've kind of put that to bed. But the issue was we just have to proactively manage this business from the supply side. I think 2023 showed us, the great meltdown of the industry showed us, that if you manage this from, "I have new technology, I should release it," that's going to lead to a bad economic outcome for us. And I'll go through a little bit of that later about why that's the case. A little more proactive supply management, and pretty soon things come into balance. And this next statement may be a little bit controversial.
I actually think the market is adjusting to these dynamics quite quickly, and I think that's one of the things we're here to talk about. The market is changing actually quite quickly in the way we engage with our customers around securing future supply. It's moving rapidly from this quarter- to- quarter to multi-year time frames. Let's talk about cyclicality, the current debate. How do we think about cyclicality? You may be surprised, I said earlier, the first place I think about this is the consumer business. It's one of the reasons why the consumer business is so important. Why is the consumer business important? 351,000 points of sale around the globe. This global brand equity. It's fun in this job because you can almost go anywhere in the country, anywhere in the world, and you're like, people know our brand.
Last time I was in China, there's always a photographer. I pulled him over and said, "Hey, open your camera." He opens his camera, he's got a SanDisk card in it. It was great. People know us all over the place. Billions of products sold in the last decade. The dynamic range of this company is incredible. We sell a single product to billions of people, and we sell billions of dollars worth of products to single customers. That's kind of what the estate is across the whole thing. But let's look at the financial dynamics of this. I don't think we've ever showed this chart before. But this is industry gross margin going back to the first quarter of 2017. So there's the cyclicality for you. And you can see the great 2023 washout that kind of impacted how we changed the business in a major way.
But if we plot consumer gross margin on top of this, you see that it generally tracks the business over almost all time periods. But especially when you start to see the down cycles, it insulates you from those. Because it's just a broad-based market. So it's like this shock absorber on the business that allows us to generate a consistent level of profit on the base of the business over long periods of time. And you can see, I don't expect 2023 to come back, by the way. I know some of you are waiting for that to come back. I don't expect that to come back. But you can even see, even in the worst, the darkest days, maybe in the history of the industry, this business was still producing profit.
So very important business, will always be an important business, and it's a great asset to the company. But it's not enough. The issue is it's not big enough. Not big enough to insulate the whole business. It's great. It gives you some protection. We need to make it bigger. By the way, that is the key of why we are taking a brand first approach to this business. It is about growing it. Financially, it is a great business, and there is actually some little tricks behind the scenes that make it very profitable because we can use more of the wafer in this market than we can in other markets. There are lots of little things like that that goes on, but it is about growing this business now. Like I said, since we launched the company, we have gained a couple points of share globally.
It is hard to do. We have rebranded the whole portfolio again. Janet is here. She leads the business. Stop by outside and see all the great new products. It is not enough. This is where we really went into this whole concept of new business models. We have talked a lot about this last couple of quarters. We have tried to be very transparent on what we are doing. As transparent as we can be, given these are confidential customer relationships. I am not going to go through these in great detail because Luis is going to go through these in great detail, and he is the best person to go through them in great detail because he is actually the guy that is negotiating them. That should tell you one thing about how important they are. It is literally the CFO of the company.
Now, there are a lot of people around, there is a big team, but he is the one that has got his hands on the steering wheel of what all the terms and conditions are and what we will agree to, and engaging at this most strategic level with our customers. I am not going to go through all this detail. Luis will go through it. The one thing I will tell you, I have just heard a lot of stuff about these agreements, just incredible. There seems to be people that are very informed about the way these work that have never read the contracts. It is incredible. I watch TV and people say, "Oh, there is more holes in these things than you can imagine." I am like, "Really? How would you possibly know that?"
The number of people that actually read these, you do not have to take your shoes off to count the number of people that have read these contracts. It is just both hands, right? They are just incredibly detailed, incredibly strategic, very consequential, and they are different than what has been done before. We have a lot of confidence in that. There may be a lot of reasons you choose to invest in this company or not invest in this company, but do not choose because you believe something about this just based on the history the way the industry has always worked. Take the time to understand how these things really work, and I think you will get some of the conviction that we have about why I can make a statement.
I feel like I am finally at the starting line of where the real value creation is going to happen on a sustained basis. Now, I will say this long-term customer visibility, demand visibility, and pricing stability is replacing this quarter-to-quarter world. It is a crazy industry. Everything is negotiated constantly. I only have to go back like three, four quarters. Everything is negotiated. We come into a quarter, we're still negotiating with our customers on what they're going to take and what the price is going to be. We have three months of visibility. We're making investments that are 10 or 15 years long. We have three months of visibility. Within two quarters, we've gone from three months of visibility to over four years of visibility. What I said earlier, I think the market is adapting very quickly to these changes. I think that's 100% true.
Two quarters is very quick to have customers of this scale. We can say things like 50% to two-thirds of our supply is under agreements. We have average visibility of four years. That happened in two quarters. Kind of amazing. I can tell you, even since our earnings call last week, the number of customers that are coming to us and now they're proposing the agreement, not us. We started with, we would go to them and we would propose like, "Hey, how do we put an agreement together, get more visibility?" We got some people that wanted to work with us. We got it done. Now it's turning. Not everybody, don't get me wrong, not everybody. Now I'm starting to see customers are coming in and proposing, "Hey, how about we do a three-year deal? How about we do a five-year deal? Here's what we'll offer on pricing.
Here's how we'll do the financial guarantee." This market is changing very rapidly. The third part of this is there's still high-value businesses where customers are not going to sign multi-year agreements. We're going to stay engaged in these businesses. We want to stay engaged in the whole market. I think one of the very brilliant things about the NAND business, it's kind of an evergreen business. There's always something new. There's always a new device. This is the magic of innovation. Somebody's always thinking of something brilliant. We don't know what they are, but when they think of those things, if they have to store data, they're going to use NAND. We want to make sure we have some of our supply available to make sure we can play in those markets.
There's very important customers and very important businesses that maybe don't have the scale. There'll always be a way to play in this other part of the market. Quite frankly, there's also going to be some customers who just want to do business the way it was done a year ago or two years ago. We'll continue to engage in those markets to the extent we have the supply to make it available. We're going to meet our customers on their terms. Like I said before, we want to stay engaged in the whole market. One of the things about this business, again, one of the things that encouraged me a lot about this business a couple of years ago when we were talking about coming to work here full time, was we have spectacular set of customers.
We have the who's who of technology companies in the world as our customer base. That's amazing. We want to stay engaged with all those customers to the extent we can. So that's cyclicality. I know it's the current debate. I'm sure we'll continue to debate it, but that's kind of our view of why we do think this market is changing very quickly, and this is a very intentional strategy to try and change the relationships and bridge this gap of on the supply side, having three months of visibility. But we have to make 10-year investments in fabs, and we end up in this vast middle ground where it seems like nobody is happy. Either the supply side is scrambling because we're not making enough money to invest, or the demand side is upset because they can't get everything they want.
It's hard for me to see how that strategy works for anyone, and I think we're now rapidly converging to something that is a little more sane, and we can all get enough visibility to make sure everybody can get what they want. Again, what I tell customers, "You want to buy NAND? You're in luck. We sell NAND." Not only do we sell NAND, we have the whole stack. We have all the IP, all the production from beginning to end. I think we're getting to a very different spot. I'm going to cover one more thing, and I think as an investor, this is important. This is important for you to understand, and we've talked about it before, and maybe you already understand it, but I just want to bring it out. We talked about growth.
Again, my experience in running a lot of the technology franchises, this is kind of the hard one, especially when you get to very good economics. Because when you get to good economics, you kind of start running out of TAM. You have to then figure out, how do I expand my TAM? How do I develop new products? How do I move into new markets? How do I go acquire different companies? All that stuff, I've done all that stuff. It's all expensive and it's very difficult. We have kind of the opposite issue. We said we're going to commit to grow supply, we're going to grow volume mid to high teens, and then we debate, is that too big or too low? It could be higher. It could be lower.
Last year, if we went to February 25, people would probably say, "That's too high. Pricing is going to go down." We got to January and it's like, "That's way too low. We can go higher." But we're committed to this kind of mid to high teens volume growth. When we look at the whole market, we think this is sustainable over long periods of time. When we talk about mid to high teens growth, I think one thing that's important for all of you to understand, it's become clear to me over the last couple of weeks. When we talk about mid to high teens volume growth, we talk about that as an input to a process of developing a whole fab strategy.
And we were talking yesterday, and I asked somebody on our team a question, which I am not going to tell you the answer to, but the question was, what is going to be our BiCS 10 mix at the end of the decade? They pull up a spreadsheet and they tell me. We have a plan already. We have a fab plan years and years in advance. So this investment is the input into that plan. That plan, over long periods of time, grows volume mid to high teens. Now the output on a quarter-to-quarter basis or on a year-to-year basis is going to have some variability in it because things change in the quarter you are in. If you just pick two endpoints and you say, "Oh, you are growing faster. You are growing slower." It depends what endpoints you pick. CAGRs are very sensitive to endpoints.
But again, one thing we are committed to this grow volume mid to high teens. Some quarters it is going to be less, some quarters it is going to be lower, some years it is going to be higher, some years it is going to be lower, but over long periods of time, this is what we are going to grow. And that is an amazing place. If you can get the economics right, you get the cyclicality dampened, and then you grow, that is an unbelievable franchise. And you will see it in the business model that Luis talks about. But this is important. How do we grow? This is really important from an investment point of view. How do we grow? So you go back to BiCS roadmap. Bit cost scalable roadmap. We say BiCS all the time. Sometimes people do not know what it means.
It is right there in the title. Cost scalable. How do we build this scalable technology? So I was sitting at my desk a couple of weeks ago and I just pulled out some material and I said, "Let me do some calculations. Let me look at this." And I looked at about a 10-year period. About a 10-year period from calendar year 2020 to calendar year 2030, and I looked at kind of the plan of launching nodes across that 10-year period, or let us call it nine years. Five nodes over nine years. BiCS5, BiCS6, BiCS8, BiCS 10, BiCS11. Now again, there is a BiCS 9 in there. Alper will explain that. But these are the big major nodes. The average generation to generation bit growth per wafer was 54%. It is a pretty impressive number.
Every time we put a new node in the market, or every time we turn the crank on that innovation engine I talked about, we get 54% more output per wafer. It is kind of amazing. Now, the issue is we do not do a node every year. So five nodes over nine years, you can do the math on what that is per year. When I was a young executive, they sent me to PR training, and one of the things they told me in PR training is never make the statement You can do the math. Because nobody does the math, or nobody can do the math. So now all of you can do math, I am pretty sure, but I am going to do the math for you. And so when you back that out to a yearly CAGR of productivity growth, it is 27%.
So through the application of innovation, we can grow output per wafer at a rate of 27% a year. Which, by the way, tells you can't just release nodes whenever they're available. Otherwise, you're going to flood the market with supply, and then you're going to have another 2023 situation. But that's in the past. But what this tells you from an investment point of view, what it tells me from an investment point of view, is growth is primarily driven in this business by the application of intellectual capital, not financial capital. It's the paying of those engineers and those NAND designers to continue to drive that roadmap forward, is what is going to drive the growth. Now, there is more CapEx. Each node is more steps, more steps is more tools, more tools is slightly more clean room space.
But you can see from this equation why we have to reduce wafers on an ongoing basis because the technology you have is so productive. So it's like getting the business model around that really, really high-powered engine, getting that right, and it's an incredible business. And what this means is, again, from an investment perspective, the ability to take bits, turn them into revenue, and then revenue to free cash flow is quite high. And I think at scale, you have a franchise that has an extraordinary ability to generate free cash flow margins. And you'll see a little bit of that later on. All right. That's where we are. That's the big picture. I'm going to turn it over now to the people that are really driving all of this fundamental technology and financial greatness. As Ivan said, Alper's going to come up.
He's going to talk about the NAND roadmap. As I said, you're not going anywhere in this business if you don't have the right roadmap. And he is expert on this. We have a lot of people working on this. He'll tell you about BiCS 10, where that's going. We made this statement on our earnings call last week. Earnings calls are always fascinating events because something always happens you don't expect. I love earnings calls. I talk to my peers and they're like, "Oh, I have never heard somebody say they love an earnings call." We love earnings call. We get to talk about our business. But we made this statement that said we think the market's going to be $300 billion this year and $500 billion the year after that. And everybody's like, "Oh my gosh.
We never heard that number before." They started backing into thinking that was a revenue forecast for 2027. It wasn't. One thing you need to understand about that number, it includes China. So when you include China, it skews the numbers if you're trying to back into everything else in the world, and that's an exercise left to the reader to figure out what that is. But, Eric Cherrstrom, we have a team on market intelligence. This is why we were able to stand up here last year, and we said with conviction we thought pricing was going to inflect in the second half of the year. And that was very debatable. And the reason we had conviction in saying that is because of work Eric does and his team.
We thought we'd give you some visibility into how he thinks about this and how he thinks the big picture of how this market has resettled over the last two, three, four years. Very important to understand this. Not only is SanDisk changing, the playing field we're on is changing dramatically. I think when you understand the dynamics of that, it starts to unlock some of this value creation as well. Khurram's going to come up. He's going to talk about AI inference. There's obviously a massive tailwind to the business right now. There's a lot of questions, a lot of conversations about KV cache. How is NAND used in inference? He's going to try and demystify all that a little bit. How our products fit into the data center. How do we think KV cache is going to grow in the future?
Luis is going to come. He's going to wrap it all up into the business model. He's going to go through the NVMs, a little more detail. Just like last time, we're going to put HBF, we'll talk a little bit about 3D Matrix Memory, too. We're going to put it at the end. The reason we put it in, it's not in the model yet. When it's in the model, we'll tell you. Until then, what Luis is talking about is the model for all the core business, and you should see these as future innovations that we'll continue to update you on. I think Alper will give you a very good view of all the progress that's happened in the whole world around HBF in the last year and a half. It's really been really quite exciting.
Thanks for your time. Thanks for being here again. I'm going to turn it over to Alper to get into the technology. Thank you.
Please welcome Chief Technology Officer at SanDisk, Alper Ilkbahar.
Thank you. Good morning. Great to see all of you. Welcome. I am really excited to be with all of you here and talk about the memory technologies our teams at SanDisk are driving. Let me start off highlighting the main pillars of our technology strategy. Our number one priority is to keep our exponential scaling engine running. In semiconductors, especially in memories, scalability is the most critical factor we are looking for. You are going to hear us talk about scalability, the importance, and the role it plays in our business over and over today. We are going to talk about how our 19-generation strong scaling engine keeps going, and how our roadmap extends well into the next decade and beyond. Next, we are laser-focused on what our customers are looking for, which is performance, power efficiency, and density.
You are going to hear from us how we are leading in every single one of these metrics, and how we are delivering the most capital efficiency amongst our peers in the industry to deliver superior financial results. Finally, we strive to innovate to amaze. We are innovators, we are engineers, technologists, and we are really looking for ways of improving all the applications and products every single day. Not only the existing ones, but we are also innovating in creating new applications and new markets every single day. With that, let me start jumping into the next slide here. This is a slide that I shared with you last year in February here on this stage. At the time, I shared with you how since 2001, our teams have delivered 17 generations of NAND technology.
Today, only 18 months later, we have added two new generations of NAND technology, BiCS9 and BiCS10, to our roster. This data tells me two important things. The first one is the pace of our innovation is accelerating to match the demands of the markets. Number two, NAND flash, being the most scalable semiconductor technology, is actually the only technology that can match the exponential growth of AI. That is why you are seeing, and will continue to see, the increased adoption of NAND in AI architectures. Last year, I also showed you this slide to explain the vectors that are driving and fueling our scaling engine. At the time, I had talked about how we are prioritizing the more technically difficult but significantly more capital-efficient ways of scaling, which are lateral scaling, logical scaling, and architectural scaling.
These are our priorities over the easier, yet financially more challenging and costly vertical scaling, which is adding layers. Our strategy hasn't changed, and we are really pleased with the results we are getting through this strategy. It is exemplified in this slide. I shared a similar data with you last year. At the time, the data ran through 2024. This year, we added 2025 as well. What we are showing here is capital intensity of SanDisk and our JV partner, Kioxia, and compare it against the capital intensity of the industry. We are defining capital intensity as how much CapEx we have to put there to get an incremental petabyte of bit output. Of course, the higher that is, the worse you are off. So you are trying to minimize your capital intensity.
The white line here is showing the average of the industry divided by our numbers, so the ratio to that. This data, what it is telling you is in every data point, by the way, here is backward looking for three years and averaging that. In 2025, the industry spent, on average, 2.66 times more capital than we did to generate the same output. 2.66 times more. This is the capital efficiency our strategy is delivering. Looking at the same data through a different lens. Here, I am showing you the percentage output, bit output of each of the peers in the industry versus the percentage of the CapEx they spend every year. Actually, we are looking here in a period of 2021- 2025, so this is a five-year period we are looking at.
On the left, you are going to see that between us and our joint venture partners over that five-year period, we produced 29% of the industry's bits while spending only 13% of the capital. 29 versus 13. When we look at these ratios for each player one by one, I captured that data on the right side. What you are going to see is that our capital efficiency is just about 2x that of our nearest competitor. This is possible through our technology strategy that I just highlighted, through our execution of that incredible technology roadmap, our scale, our operational capabilities and excellence, as well as the intense focus we have on tool reuse, all of which make this possible. Delivering superior financial results for all of our investors and shareholders.
As we are pushing our technology forward and pushing the limits of scaling in every single generation, we actually advance the technology across two dimensions. The first dimension is what I am capturing in the X-axis here. Every generation gives us more bits. This is happening through pursuing those four vectors of scaling. Next, and you have seen earlier, every generation we get 50%- 60% more bits, and that is happening through scaling. The other frontier we pursue is what I am capturing in the Y-axis here, which is performance and power efficiency.
We get those improvements in every generation as well through device and design innovation. You have seen this roadmap before. I am mostly going to talk about BiCS 8 and 10 today on that roadmap. As you heard earlier, our teams are already working on BiCS 11 and beyond. The scaling engine continues running, and we see actually no end to the scaling limits in the foreseeable future. So we are going to run this for a decade and longer, of course. We have talked earlier about our CBA technology, which is our hybrid bonding technology where we are able to combine two different wafers to make a single wafer. By utilizing this innovation, we are actually augmenting a derivative roadmap, as I am showing here.
This CBA-enabled roadmap allows us to take an existing technology node and push its performance and power efficiency to the next level by combining the memory array technology of that node with the next generation CMOS technology and combine the two wafers together and get to the next level of performance. You may ask, why is that relevant? Why is that important? A great example for where this is needed is actually happening in the data centers right now. The storage interfaces in the data centers, and what I mean by storage interfaces, think about our enterprise SSDs. They run on a standard interface called PCIe, and there's a transition that happens industry-wide from, say, Gen 4 to Gen 5 to Gen 6. These transitions used to happen every four or five years in the past.
That time allowed us to essentially move from one technology node to the next one, ramp that technology, and maybe ramp the next one as well, so that you would have plenty of supply and transition our products gradually into the next generation of this interface. But with the advent of AI, these transitions started happening significantly faster. When the demand turns on in the data center with the volumes that we're looking at, you have to enable that transition extremely fast. You may not even have time to ramp your next-generation technology node to meet that demand. What do we do with this CBA-enabled derivative roadmap? We can take our existing technology node and very quickly move it to the next performance level and make that transition happen, move our entire portfolio very quickly to what our customers need, and customize the silicon very rapidly.
The beauty of it is that it can be done with very minimal additional capital spending because we're leveraging the existing nodes' memory array technology, which is where most of our capital sits. This is a super efficient way of moving to the next level of performance. It gives us incredible operational flexibility, it gives us great capital efficiency, and it gives us the ability to meet our customers' requirements very quickly. It's an awesome innovation, a technology competitive advantage that we can leverage and create incredible competitive advantage for our business. I'm going to talk about BiCS 9 in a quick bit, but before I get to it, I want to take a quick look at BiCS 8 because BiCS 8 today is the backbone of our current production. It is the industry's gold standard.
We introduced the CBA technology, the hybrid bonding technology, first time with BiCS 8, and it has given us tremendous competitive advantage in performance, density, as well as power efficiency. It turns out that these are exactly what our AI data center customers were looking for. When we compare BiCS 8 against some of our peers' performance and power efficiency numbers, we've seen tremendous gap where we had a huge advantage. Here I'm comparing our BiCS 8 against our peers' 2XX generation memories. Of course, our peers are moving forward. They're announcing their next-generation products, which we call 2YY or 3XX products. They're improving their performance, but when we look at the power efficiency, we're seeing that their power efficiency is really not moving a lot better. Now let me put on what's coming, BiCS10, and show you how BiCS10 is going to compare against these.
This is where BiCS10 is coming out to be. BiCS8 was amazing. BiCS10 is going to be even better, and I think it is going to be the gold standard for the AI data centers very soon. We will get to BiCS10 in a bit, but let us first talk about the latest news. We talked about BiCS9. BiCS9 is the first product where we are essentially deploying this hybrid bonding technology extension. What we have done for BiCS9 is we have taken the BiCS8 cell array, the mature cell array we have, and combined it with the next generation CMOS wafer. Through that, we achieved tremendous performance gains. We have done so with minimal incremental capital spending. We essentially are upgrading our BiCS8 deployment supply bases to the next generation performance level with minimal capital. We designed BiCS9 based on the specification from our large hyperscale customers.
They wanted to have incredible performance, they wanted to have a lot of it, and they wanted to have it yesterday. This allowed us to achieve all those objectives very quickly. I am very happy to report that this product is already as part of our NVMs, and our customers cannot wait to have this. It is going to power our next generation storage SSDs that Khurram is going to talk about. Really looking forward to seeing this powering your AI very soon. Let us go to BiCS10 real quickly. I gave you a sneak preview of this BiCS10, the first product in the BiCS10 lineup last year. This is the 1 terabit TLC die. Again, achieving tremendous improvements over BiCS8, which is the best in the industry. Since then, our teams have done a marvelous job with this technology, and it progressed ahead of our expectations.
It allowed us to start sampling this die to our customers as of this month, which we just announced. The second product in the BiCS10 lineup is the 2 terabit QLC die, QLC being 4 bits per cell technology. We are really proud of this technology. It actually is the highest density memory chip in the world. While delivering this, we have achieved more than 60% density improvements over BiCS8, more than doubled the read and write bandwidths, as well as improved the power intensity and efficiency by 75%. These are truly amazing numbers considering that we are beating world's best NAND, BiCS8. To give you a better picture of the power of scaling, I wanted to show you BiCS8 and BiCS10 2 terabit dies side by side. This is what we are able to do with the scaling engine.
Each BiCS10 2 terabit wafer has 65% more bits than the BiCS8 2 terabit. You have the pictures here, but the wafers are sitting outside. I will invite all of you to please go out and check in person and experience that scaling. The wafers are sitting out there, and you can even touch and play with it if you want. Putting this in historical perspective, going back to the data that David showed you earlier, we are looking at the generations from BiCS5 through BiCS10, and even projecting into what is coming next, BiCS11. We are delivering 27% CAGR on bit growth per wafer annually. 27%. You already heard that our production plans are based on our long-term demand forecast of about high teens.
This is delivering roughly 50% over that, which means that we have the technology productivity to meet all of our production needs and plans just by scaling the technology alone. As a matter of fact, because of this productivity, our wafer starts have come down over time. That is another way how we have created value and driven our capital efficiency. With that, I thank you all for being here. I will be back to talk to you about HBF at the end of this presentation, and I will invite Eric to be with you to share his intelligence insights in the market. Thank you very much and see you soon.
Please welcome Vice President, Market Intelligence at SanDisk, Eric Cherrstrom.
Good morning. I am excited to be here to give you a brief NAND market update. We expect the flash market to reach 1.2 zettabytes of shipments in 2026. How did we get there? As you know, this industry started with growth in the consumer and edge. The key drivers were phone scaling to 1.5 billion units of annualized shipments, PC shifting their storage needs from hard drive to flash, and the emergence of the enterprise SSD. Then, in 2022, ChatGPT was launched. AI propelled the data center segment and increased its share of TAM over time. Data center share of TAM in the early 20s was 20% of bits. Last year, it was 30%. This year, 50, and continues to outpace the market. Now when you think about the revenue overlaying the volumes from the prior slide, you can again see two distinct periods.
Period one, flash was priced as a commodity. ASPs reduced offset volume gains. The historical average during this period of time for the industry was $60 billion, and we measured cycles in $20 billion increments. Now flash is a critical component of a multi-period data center build-out, and this leads us to believe that the flash market is going to grow to over $300 billion in calendar 2026, and again grow to nearly $500 billion in 2027. What was going on with supply during the same period of time? In the late teens, the industry was targeting over 30% annualized bit growth rate. The industry had to invest, growing wafer starts all the way up to 1,800,000 wafers per month in 2022. At that peak of wafer capacity, COVID-related inventory digestion dramatically reduced flash industry demand.
Supply side had to react, underutilizing 500,000 wafers per month and structurally resetting their capacity to 30% below peak levels. In spite of that reduction in wafer capacity, industry was still able to achieve mid to high teens production growth rate via nodal migrations. Let's talk cloud. The chart on the left shows U.S. data center CapEx for select hyperscale and neo cloud customers. The chart is showing the projections over time, and as you can see, starting in 2023, we have seen 15 consecutive quarters of upward revisions to this selected CapEx. Current estimates show that $1.9 trillion of capital will be spent by these companies in 2026 and 2027, and we believe monetization is coming. The most recent Amazon earnings call, they talked about not having enough capacity to support near-term demand, and the fact that AWS could reach nearly $1 trillion of annualized revenue.
Shifting to the edge, the edge is going through a transition period. The chart on the right shows our expectation of unit decline year-over-year in the mid-teens for both the PC and mobile segment. As you can see, all of this reduction is being driven by the low-end devices. Our belief is that OEMs will shift their mix to more premium offerings, and ASPs will continue to grow. ASPs and revenues will continue to grow in 2026 and 2027. As you can see from the most recent earnings of major OEMs, year-over-year revenues grew between 13% and 24%. Flash is entering a new reality, a reality where data center is the majority proportion of the share, edge continues to mix to premium devices, and on the supply side, bit growth targets are met via nodal migrations.
All this put together gives us the conviction where we see the market opportunity growing over $300 billion in 2026 and can approach half a trillion dollars in 2027. With that, let me pass it over to Khurram to talk about the era of inference.
AI doesn't run on magic, it runs on data. In the era of inference, that data moves through a repeatable loop we call the AI data cycle. Each stage in this cycle is unique and demanding in its own way. Let's go on a journey through the AI data cycle. Stage 1 is the raw data archive, where information lives. Think everything an organization collects: articles, videos, social posts, business transactions, and internal knowledge. It's massive exabytes to zettabytes, and it's the source of truth. Raw data by itself isn't ready for AI yet. Stage 2 is model data preparation. Raw data is extracted, transformed, and loaded, also known as ETL, to clean it up, organize it, and make it machine ready. A key step is creating embeddings, which turn content into numeric vectors so computers can group and search by meaning, not just keywords.
And for AI training, the system has to pull data randomly from across the entire enriched data set fast and at a very high throughput. Stage 3 is model training. Massive GPU clusters process that prepared data to learn patterns and build a model. Training can run for a long time, so the system regularly saves progress in checkpoints. If something fails, you can restart from the last known good snapshot instead of losing days of work. Because training pauses during checkpointing, the faster those snapshots write and store, the more time the GPUs spend learning. Stage 4 is inference, when trained models are actually used to answer questions and generate outputs. Real services may swap models depending on the task, so keeping the right models ready matters.
Inference also often uses retrieval-augmented generation, or RAG, to pull the most up-to-date, trusted context from a vector database to ground the response. To maximize efficiency, systems use KV cache so GPUs can reuse previously calculated values instead of recalculating from scratch. As more tokens are generated with agentic workflows, the more valuable this cache becomes, and the larger it becomes. These high-capacity KV caches are tiered onto SSDs. Stage 5 is new content generation. The model's outputs are stored immediately so they can be delivered to users, and reused for things like fine-tuning and ongoing consumption. The newest content is accessed the most. Then it ages into long-term archives, becoming tomorrow's raw data. And that's the AI data cycle. Collect, prepare, train, infer, generates, and repeats.
Please welcome Chief Product Officer at SanDisk, Khurram Ismail.
Good morning. I'm Khurram, and I'm here to talk about the infrastructure outlook, specifically as it pertains to flash. The good news is that we are almost at halftime. Since we don't have any breaks for halftime, I get to be your host for the halftime. Let's get into it. David talked about engineers being at SanDisk for a long time. I'm one of those. I've been in this industry for 27 years, all in memory, and never has been a time more exciting for memory than it is now in AI. I've seen all sort of peaks and troughs. The pace of innovation that AI is bringing is tremendous, and we all can see that. We see the new frontier models being loaded, right? The system architects are changing the design every six months.
The pace of innovation is quite rapid, but the infrastructure required to deploy that innovation is also being deployed at a very unprecedented rate. What is the role of flash? In my talk for the next 20-odd minutes, I want to leave you with two things. First, like David mentioned, our conviction on the critical role that flash plays and the size of the opportunity. The second one, I hope you gain the appreciation of flash as not really a clumpy device sitting at the edge of the infrastructure, but it is actually being proliferated through all the layers of AI infrastructure. I want to start with some fun facts. As I will be going over some concepts, I think it will help us understand those concepts if I draw some analogy to the human brain.
Maybe some of you know, I was just doing ChatGPT, Gemini, and I found an interesting fact that each human brain is wired with 2.5 petabytes of memory. Now you multiply that by entire human intelligence, that is like 20 yottabytes of collective human memory. 1 yottabyte, I had to look that one up, too. I work in zettabytes and exabytes. 1 yottabyte is 1,000 zettabytes. Now you look at the right, at the cloud infrastructure, which I would characterize as being very early in the innings, is only hundreds of exabytes to maybe 10 zettabytes, all memory combined. One could argue, looking at this, as we are going to scale intelligence, the cloud infrastructure can use more storage.
The interesting thing about human brain is, it works with two types of memory, the short-term memory and long-term memory, and they both work hand in hand, utilizing each other to generate intelligence. Turns out the AI intelligence is built on very similar concept. You have a transient short-term working memory that is called ephemeral KV cache, and we will cover that. Then you have the long-term memory, which is a little bit of more persistence, which is known as persistent KV cache. There is similarity. They both work on same principle on how humans store data and process data versus AI intelligence. Eric talked about the total demand in 2026 to be 1.2 zettabytes for entire flash market. Here, I am only focused on 2030 AI data center TAM, which is equivalent to what we ship as a total output as an industry in 2026.
The opportunity is massive. I will come back to this slide again as I go through why that is the case. But it is important to note, as I mentioned, flash is not just a single device sitting at the edge of infrastructure. There are many workloads that are emerging on flash, specifically in AI data center. What are some of those workloads? Well, you have first, fast data lakes. We talked about it last year. These are the massive data lakes that require massive storage. Second is staging. These are a little bit direct attached device close to the GPU for training, checkpointing. That market is having a tremendous growth. Last is the KV cache, and that will be the focus of our conversation because that is the fastest-growing segment in AI data center.
And we look at the composition of NAND technology, we see that TLC is dominant technology in 2030, and QLC still has a very good, decent-sized share. How is the infrastructure being viewed today? The thinking around infrastructure is changing. We are moving from what used to be total cost of ownership to the total value of ownership. In the past, when you deployed the infrastructure, you prioritized cost running at very large scale. Some of those considerations are listed here. With the total value of ownership, the equation is changing. Infrastructure is no longer being viewed as a cost center, but really a driver of value generation, right? In this case, the value, and the output is intelligence, right? The race to scale the intelligence is heating up. As you can see, everybody's trying to generate more tokens.
They're trying to get more users on their systems or on their AI. But with this scale, there's a lot of challenges that come, right? There is always the challenges of power, right? Where do you store? Do you store these tokens in volatile media only? There's not enough volatile media. What role does non-volatile media plays, right? Then there's shifting architectures, right? We're moving from training to inference. Turns out flash solves a lot of these problems, and that's where I'll be taking you next. You saw the video, the era of inference. The only point I would make here is a lot of focus in the AI data as it flows through the infrastructure is on inference. I presented this 18 months ago, last February in 2025, the five-stage data cycle.
The first three we focused a lot last year, which are associated with training the model. How you store the data, how you prepare the data, and how you present the data to the GPU for training was the focus, and flash did quite well. We had our high Cap QLCs that were used in fast data lakes. We had directly attached TLC SSDs that provided active datasets for the model to train. A lot of infrastructure got built as a result of this. But now as the focus is shifting to inference, the question to ask, the infrastructure that was built for training, is the same infrastructure relevant? Can that satisfy the growing need of inference? The answer is no. You can see as we move along, there is disaggregation happening in infrastructure, right? The infrastructure for training is quite different from the infrastructure of inference.
Let's look at what's happening inside the inference. Like I said, we have our existing products that go into the fast data lake staging checkpointing, but inference something new, and that's where we'll hone in on KV cache. There are two interesting trends that are emerging in inference when it comes to NAND flash. The first is data augmentation, and the second is context remembrance, right? A lot of you, I'm sure, use the models. If you know, if you're using the models, it's becoming more persistent. It remembers who you are, right? That's the second popular use case. RAG is one of the most popular techniques that is used to provide external data so the model can provide you much more relevant and accurate responses. Second, the users are wanting richer conversations, smarter conversation, longer conversation.
What ends up happening as a result is a KV cache amplification. The KV cache amplification, the way to think about is you are having longer context length because you want longer conversations. You have longer reasoning chains because you want iterative process. You want the model to know about you. Then there are multi-modalities associated with this. All of these are driving the amplification in KV cache. The easiest way to think about KV cache is if I am having a conversation with you and you are taking notes as we are having now, so you can refer to the notes rather than listening to my conversation again, and that notebook serves as KV cache. Let's just briefly touch what is KV cache, because that's the most important part in inference. Inference has two major parts. The first is prefill, where the model is thinking.
The way to think about prefill stage in inference is where the model is thinking when you provide the input. The second is decode, where the model is responding, is giving you a response. As a user puts the input prompt, that gets tokenized, and all that input gets processed simultaneously. So one would think as that process is driving a lot of parallelism, that is compute bound. You will hear a lot more people say that prefill is really compute bound. Then when the prefill happens, a context gets generated. Now in the implementation, that gets stored in a memory, which is called KV cache. The way to think about KV cache is it's a working memory of inference or the notebook, memory notebook. It is not important for now for us to discuss where does that get stored.
I will take you through how the KV cache hierarchy works, but it's important to know that this context is growing. It continues to grow. The most interesting part of inference is the decode process. This is where you generate the response. As you may know, decode is an auto-regressive procedure where one token gets generated at a time to generate the response. Now, to generate the response, the token that gets generated has to know the context of all the previous tokens that were generated. You can imagine if you didn't have the KV cache, that would present a tremendous challenge to the infrastructure, to the power, and computational overhead, which didn't need to happen. So that's where you see when we see all the memories boards are rising is because of this KV cache, because it does make the entire AI process much, much more efficient.
Okay, so we talked about this context is growing. It gets stored in KV cache. Well, how does a KV cache look from a hierarchy perspective? We presented this last Investor Day, and we talked about the system memory hierarchy in a data center system. The way to think about this memory hierarchy is around the vectors of performance, power, and capacity. For those of us who are in love with Flash, we always made the assertion that Flash is the most scalable technology around these vectors. Well, it turns out we were right. Inference is a perfect use case. Inference is a perfect use case for Flash. Why? As we talked about the KV cache amplification and with the deployment of AI agentic workflows, it's generating a lot of tokens.
There's a need to have more pages in your memory notebook, so that all the states need to be preserved. The contexts are getting longer. The conversations are getting longer. The reasoning chains are getting longer. Last, all the data that gets generated as part of your input to the system. So we have very close partnership with our customers. The NVMs are a testament. We get to learn a lot from our customers. Now they're all hyper-focused on optimizing this KV cache because it really solves a lot of problems for them. The way they go about it is different. There is one common theme that emerges from a KV cache memory hierarchy that is generally applicable to all the AI systems that are getting deployed right now. At the top, taking you back to the human brain analogy, an ephemeral cache.
There'll be a quiz after this. Ephemeral KV cache. Those are your HBM and system DRAM. The way to think about this, all the hot context, the current context that the user needs to get the response from, those are stored here. Again, if you look at from top to down, capacities at play, HBM and DRAM are generally smaller. Next is the long-term memory, the persistent KV cache. So anything that cannot be stored in the high-tier bandwidth of DRAM and HBM gets stored in Flash. This is what we are calling persistent KV cache. Now there are multi layers. Remember when I said the Flash is proliferated throughout multiple layers of AI data centers, I just want you to remember there are multiple tiers where Flash is deployed. We'll use this later on in the presentation.
It's important to understand that you cannot scale the intelligence without having this persistent KV cache layer. As you can imagine, in inference, the model gets trained once. As the context grows, the interactions are in billions. So you need some kind of persistence in your memory hierarchy. It provides a nice extendable capacity to the AI systems without having the need to take everything through the volatile ephemeral KV cache. So now that we have covered the KV cache and we talked about KV cache is going to be 35% of the market in 2030, how do we plan? How do we size the opportunity? How big can this persistent KV cache can be? If I'm an infrastructure planner, I have to think about a few things, and this is SanDisk equation of how to think about the size opportunity for KV cache.
If I am the planner, and this we derived from talking to a lot of our customers because we have a lot of close relationship, people who are actually deploying this at scale. If I'm planning for this, the first thing that I have to think about is the number of sessions that are going to hit my infrastructure. That's number one. More importantly, as the sessions hit, how many sessions do I want to retain and for how long? We have customers who tell us, "Khurram , retain the sessions or the context for only couple of hours so my friend Luis can go have the coffee and come back and have this context." There are customers who are keeping all the context forever. They want to monetize this somehow, but that's how they're looking at it. You can think about it.
If you retain it forever, there's tremendous opportunity for this KV cache to grow. That's how they are thinking about it. How many sessions are going to hit my infrastructure, and how long do I keep them? The second part is if you are planning to build out, you obviously have existing infrastructure that has a set of KV cache pools. You want to only plan for the KV cache that your current sessions may miss, and that's represented in cache miss ratio. That's another important factor. Lastly, you have to figure out how much storage will be required in a session, and that's a function of two things. The first one being, what is going to be your session length or size. A lot of people talk about context length, but it is a series of tokens that determine.
Each token, by the way, as we talked about the decode process, generates tremendous overhead on storage needs. Each token is represented in tens of kilobytes to hundreds of kilobytes, depending on which model, which implementation you're using. As you can see, these variables are what people use to determine how big of a KV cache or persistent KV cache, rather, they need to deploy. Here is our answer. This is again, 1 zettabyte install base. When you multiply all these things up, and by the way, these are just four variables. Underneath, there are second-order calculations that come in to make sure that we arrive at the right number. The SanDisk estimated 2030 install base is 1 zettabyte. This is the fastest-growing workload. This we are saying between now till 2030, we'll have 1 zettabyte of install base.
This again, our customers are very dynamic. They're changing things. Architectures are changing. There's a lot of optimizations that are happening around flash, but this gives you a good proxy to think about that, hey, if I want to take this case up, and if I take the retention time up, the number will be quite large. We feel pretty good about this because, in general, we see the context length growing. The average session's growing. The number of users that use AI is growing. We see it in a positive place. Coming back to the 1.2 zettabyte number, and I would say this again, that the KV cache number that is represented here, because I didn't cover it first time, is 35% of the overall market. The previous chart showed the 1 zettabyte number as an install base.
In 2030 specifically, we see the size of KV cache being 35% of the market. You can do the math. This again really demonstrates that flash is present in multiple workloads of AI data center. Let's look at a little bit more physically in how flash sits in data center. This is just showing the various placements of flash in the data center. At the foundation of it, in the gray box on the right, is a sea of large capacity drives. These are your fast data lakes that contain your training data, that contain your embeddings, your vectors, your RAG repositories, all the things that enterprise needs to store to make the AI work. For this, from flash point of view, high-capacity QLC drives are perfect because they provide that enormous capacity. Now, you may ask, okay, you said KV cache.
You remember the hierarchy that I showed you, and there were multilayers, right? Think of it as very cold KV cache, the way to think about this. It is great for QLC, and we see that deployed in data lakes. Now, as you get closer to the GPU, the requirements change. This is the second one. These are our direct attached SSDs. Now, here you have actually active training data sets, right? You have checkpointing going on, and you have a lot of data orchestration that is happening from GPU, and for this, you need a very high-performance TLC SSD. Now, again, going back to that G3 Tier of KV cache, this is what I would characterize as a hot and warmer KV cache.
Lastly, what we talked about in persistent KV cache, now there is another rack scale, network scale data movement that happens between the GPU complex and something that is closer to it via network, right? This is what we call the persistent KV cache, and I would characterize this as lukewarm G3.5 that you saw in that pyramid. So this is showing the AI data center ESSD placement. A lot of people talk about AI data center, and they talk about, well, it is a GPU factory. But the way to think about it, I hope with these placements that you can see that flash is living simultaneously in many different places. So I would assert that every AI factory is ultimately a data factory, and it is true when it comes to inference. So I would be remiss if we did not talk about our products.
Last year when I was here, we were trying to tell you that we are going to succeed in data center, as David alluded to. Happy to report that both our TLC ESSD and QLC SSD are qualified at major hyperscalers, major customers, OEMs, and we are shipping both of them. For TLC SSD, we are shipping it in PCIe Gen 5 configuration in all the form factors, and they are great for staging and KV cache that we discussed on the previous slide. When it comes to QLC, it is also. On the compute, on the TLC SSD, we also demonstrated our next PCIe Gen 6 drive at Flash Memory Summit is going to be an industry-leading, high performance, great power drive, really going to solve a lot of inference bottlenecks. So we demonstrated that at FMS last week.
Similarly, on high capacity SSD, if you remember last year we showed you a roadmap of up to 1 petabyte. We at Flash Memory Summit demonstrated over 256 TB drive in E3 form factor, and that was very well received. You will see a lot of market shifting towards higher cap drive next year from 128- 256. So we have a great portfolio. We have great platforms that will serve all the needs of the AI data center growth now, but also for future. So we talked about a lot of close collaboration with customers, applying our thinking, understanding how the KV cache looks like, what does the market size opportunity look like. But we want to become AI practitioners ourself. We also want to be the AI practitioners.
We started an initiative at SanDisk, it is called AI Lab at SanDisk, where you can imagine we have server scale, rack scale type of systems, and we run the workloads, the models, the way our customers do. Because we want to truly understand the bottlenecks, and we want to complement it with what we learn from our customers with our own understanding. Here, I am just providing you two metrics. The way to think about this is, this data was collected in a server scale application or system with a cluster of GPUs, multiple SSDs, HBM, DRAM, all the hierarchy that I showed you. We ran hundreds of user sessions, like I explained to you in persistent KV cache equation. We assumed certain things in that equation.
What we see, that a system that has SSD versus a system that only has volatile media like HBM and DDR, consumes 75% less energy. I will extend it, this is not published data, but for us to generate 1 million token on a system with SSD versus just the volatile media or no SSD, it takes one fifth of less power. So different metrics, but you get the idea that for SSD, you consume much less power. Secondly, on the same system that had SSD versus no, we saw 75% higher throughput in tokens per second. We generated more tokens per second than you would with HBM and DDR. Simply why? Because you do not have enough capacity, you are limited, and you are having to recompute. That is an expensive process both in power and performance.
As you can see that SSD is not something that is just an afterthought, it is actually an essential, and that is where you see the explosion of KV cache workload in the market. I want to leave you with one last thought. I do not know what I was supposed to say. Let me pull it from my persistent KV cache. The persistent KV cache responded. Today, Flash represents the work that is completed. If you look at previous compute cycles, whenever the compute cycle was finished, all the intermediate states or the nodes were discarded, only to be recomputed whenever the compute needed. If you think about the millions and billions of scale of AI, that strategy is very inefficient. It will not work. From that perspective, we like to think about Flash SSD as a token battery.
It is storing the energy to be used later on. Flash truly represents the accumulated intelligence. We believe you cannot build intelligence without persistence, and Flash is great. In an era where the most valuable output is intelligence, preserving it becomes as important as creating it. In summary, we have a robust growth outlook. We have conviction in the AI data center market. We have good understanding of where the customers are headed, where things are headed, how the architectures are working. We have the right product portfolio. We have a strong product portfolio that is good for now and for future. We also acknowledge that there are going to be efficiencies when it comes to inference. You all heard quantizations, all the optimization that is happening to reduce the store in KV cache, but that is only going to fuel the Jevons paradox.
There's going to be more use cases that'll come out of it. So we feel pretty bullish that these efficiencies are welcome, and they're going to drive more utilization. Flash is not something that is an afterthought. Industry is actually innovating around Flash. Why? Because it's the most scalable technology, and like I said, you cannot build intelligence without persistence. Flash technology is a great medium to build intelligence. Thank you very much, and I would like to now invite my friend Luis to talk about financials and business planning.
Please welcome Chief Financial Officer at SanDisk, Luis Visoso.
Good morning, everyone. I thought you may want to look at some numbers. It's great to be here back after 18 months of launching the company. Frankly, this conversation is about sustainable value creation. Sustainable value creation. We're committed to do that every single year. Our journey, as I said, started on February 25, when we separated from Western Digital. Shortly thereafter, we announced our Q3 2025 results. You may remember those numbers, $1.7 billion in revenue, we lost $0.30 in non-GAAP EPS, and we generated $220 million in adjusted free cash flow. We've come a long way. Hopefully, you saw our earnings last week. We reported $9 billion in revenue, non-GAAP EPS of $39.25, and adjusted free cash flow of $5 billion. This excludes cash we received from our NVMs as prepayments and deposits. So we've come a long way.
Now, going forward, what are we going to do? We're committed to creating value for our customers. As we do that, we're confident that we can create value for our shareholders. So let's look back into the year that we just delivered. These are the metrics that matter the most. We operate in a large, fast-growing market. That market has tripled or will triple in calendar year 2026, reaching $300 billion on its way to $500 billion in calendar year 2027. So very large market. Now, what's very important is the composition of the market is changing from an edge-centric market to a data center-centric market. That brings very different dynamics, and I'll explain some of that. Our revenue for the year, $20 billion, up 175%. That's twice as high as our prior record that we delivered in 2022.
Nice growth, and importantly, that growth, that revenue improved every single quarter throughout the year. Gross margin 71.6%. That is up from 30.3% the year before. Again, our performance improved every single quarter throughout the year. We closed the year with 84.6% gross margin. That enabled our EPS to go to $39.25, up from only $0.29 the year before. Great performance on our financials. The metric that matters the most is our free cash flow. We generated $8.7 billion in free cash flow, excluding those new business model prepayments, and that also improved every single quarter. Our run rate, $20 billion. That is our run rate of generating free cash flow from this business. Very good business. Growth is there. The market is growing. We are capturing that value. You may have a few questions about the new business model. Let us go into that.
Very importantly, this is our way of strengthening our relationship with our most strategic customers. Why? Because the new business models deliver a fast-growing, profitable, and less volatile business, going back to what David just said. Fast-growing, very attractive, less volatile business. Is not that beautiful? We are building these relationships. The way this started is very custom agreements with each of our customers that center around supply and demand certainty. The conversations did not start around pricing. Obviously, we do get to pricing, but they start with supply and demand. Our customers want to make sure they can get the products they need, and we want to make sure we have somebody on the other side. That is how this conversation started. There are details by quarter, details by month in most cases. While they are custom-made, there is a framework that is consistent around all these agreements.
They start with a multi-year, in most cases. When you have a multi-year, those volumes are growing at a very fast pace, faster than we are growing as a company, and there are fixed and variable components on pricing. Very importantly, every single one of these agreements has a financial guarantee, and I will talk about that. These conversations go to the highest level of the companies. They require board approval. We are talking to CFOs, we are talking to treasurers, we are talking to CEOs. This is not like a typical conversation of the past. Let us talk about some of the details. We have eight engagements with customers. These are win-win conversations. As Khurram alluded, there is high level, very deep integration from a technical and commercial side. These eight customers, by the way, news to you, include three hyperscalers from the U.S.
Three U.S. hyperscalers are part of these eight customers. The oldest deal we signed was only in January of this year. Guess what? Two customers already came back and they said, "Hey, guess what? As I look at my models, as I do the math that Khurram was talking about, I need more." So they are already expanding, either extending the duration of their term or adding more bits to the same contract length. We feel very good about these contracts. In terms of duration, we are moving from a quarterly price negotiation to large multi-year engagement. Remember, these price negotiations lasted three months, sometimes not even three months. Over that time, when we were operating in that model, we practically generated no shareholder value. Made capital investments super difficult because they were very risky.
We did not know for how long our customers were going to take our products. We had no commitments. Go from there into an average length of our contracts of over four years. The longest contract is five years now, and we are actively in conversations with several customers to go even further. That is very important. We are allocating a significant portion of our business to these new business models. Why? Because as I said, they are fast-growing, attractive, and less volatile businesses. We like this business model. We want this to be the predominant way of doing business for our company. How does pricing work? Well, pricing will be fixed in some of these contracts. Some of them include variable portions, and very importantly, the variable portions include floors and ceilings, and our financials are very attractive even at floor pricing.
We talked about around 80% gross margin for the floor pricing. We believe there is subset to that pricing, and therefore we feel very good of financials of the new business models. The non-business model, the rest of these bits, will continue to fluctuate with the market. If the market continues to go up, obviously, we have an ability to capture that. Let us try to quantify the size of these contracts. If you look at the $93.9 billion, that is the total contract value, TCV. That is how much we expect to collect in revenue from the beginning to the end of these contracts, $93.9 billion at an average of four years. Some of that revenue has already been recognized, so the remaining performance obligation, the RPO, is $91.1 billion. A lot of the value, a lot of the revenue is still to come.
Both of these numbers reflect the floor pricing, the minimum pricing we expect from these contracts. We believe that there is upside on both of them as prices will be higher than the floors that we have. We talked a lot about financial guarantees and is there risk in these contracts. We have secured $16.5 billion in financial guarantees from these contracts. There are two big buckets of this. The biggest ones is financial guarantees held by or provided by third-party financial institutions. The other part, the smaller bucket, exactly $2.9 billion, is deposits and credits from our customers that we have either received or are about to receive. Of the $2.9 billion you will see on this slide, we already have $2.5 billion in our bank. The vast majority is financial guarantees provided by or held by third-party financial institutions.
Very importantly, our customers will pay for their products in ordinary course. This financial guarantee, other than prepayments, will stay constant throughout most of the time. That is important, and I will come to that in the next slide. How do I think about this financial guarantee? How strong of a protection is it? Well, an easy way to think about it is the ratio between your financial guarantee and your remaining performance obligation. You have the numbers. You can do the math, as David said. If you do that, and if you define that ratio at the beginning of the contract, let us call that the base ratio, as you divide the financial guarantee by the TCV, the total contract value. Some of you are questioning, how does that ratio evolve over time?
So we looked at our contracts, at multi-year contracts, and we wanted to provide you an illustrative example of how that ratio would evolve over time. For a three-year contract, two years out, in average, you should expect that ratio from beginning to two years later to be twice as high. So you guarantee your protection as percentage of the revenue to come significantly increases as the contract goes on. What are we going to do? Well, we're going to execute these NVMs with excellence. We don't want four-year deals. We don't want five-year deals. We want these NVMs to last decades. Right? Therefore, we're going to execute them with excellence. We're going to have the products with quality on time, just as we agreed with our customers. We want them to fulfill their part of the bargain. We're going to do the same.
You've seen us do some of that. We're increasing some of our safety stocks. We want to make sure that we have the agreements with our JV partners. We buy our agreements with Nanya to make sure that we have the DRAM. But we want to make sure that we can perform very well on these NVMs. Again, the goal is to make them even longer. We're going to be number two. We're going to be very patient. We're going to be patient as we continue to evaluate new deals. Just as you saw, we only have eight companies with whom we've signed an NVM. We'll be very selective going forward to make sure we choose the winners that value our products, are willing to pay for them, and want to make commitments which are longer term. How do we think about the model going forward?
Right? Going forward, our financials will be the result of a combination of both models. We will have a proportion of our business will be the NVMs. That would be the largest portion of our business going forward. Why? Sorry to repeat myself, this is a growing, profitable, and less volatile business. We like this business and it has reliable volume. So we're going to keep that NVM and we know exactly what to expect from that side of the business. We have a portion of the business which will be the non-NVM. We'll continue to support our customers. David alluded to that. Some customers are just too small to have new business models. Some of them, they're very strategic, don't get me wrong, but this business model may just not be the right solution for them.
When you aggregate all of that, for 2028 through 2030, we expect to grow revenue mid to high teens, consistent with bit growth. We talked about bit growth in the mid to high teens, well we expect revenue to grow at about that same rate. We expect non-GAAP gross margin to be around 80%. We expect non-GAAP operating margin to be 75%. How do we get there? We expect to spend about 5% in OpEx, and we do not expect significant contributions from other income and expense. So you get to that 75% and then you get to 50% adjusted free cash flow after paying for taxes, working capital spending. We have, for modeling purposes, I would assume mid-single digits capital intensity as percent of revenue. That's our gross CapEx.
Now importantly for 2027, we already talked about this as part of earnings last week, we expect bit growth to be somewhere in the mid-teens, and we expect sequential prices to be modest throughout the year. That's the model. Why do we feel confident sharing with you these numbers? Our confidence comes from our customer conversations. It comes from our new business models that we signed based on our conversations will lead us to believe that more NVMs will come. So we feel very good about our new business model, our conversations, and frankly, the growth of the business overall. So what are we going to do with the cash? We will continue to invest in the business. This is a great business to have and it requires cash and we'll continue to invest in it. What does that mean? We'll continue to invest in OpEx.
We'll properly fund the business. We'll invest in CapEx, and we'll continue to do things to strengthen us, like the Nanya type of investments, the JV extensions, those type of things that make us more robust, more sustainable as a company. That's super important for us. Number two, which we've done very well this year, we'll maintain a strong balance sheet. What does that mean? Healthy cash balance. How much? You've seen us operate in a range over the last few quarters, and we will keep to operating around that range. That doesn't mean it will be exactly the same number. There are payment terms, there are different things that happen, but within the range that you've seen us operate over the last few quarters. We have no debt. We got rid of the TLB. We intend to keep it that way.
Our revolver is unused, and we don't intend to use it either. And we'll continue to improve our credit ratings with the agencies over time. We've made progress this year. We're at BB+ overall, and we intend to continue to make progress. And the rest of the cash, the excess cash, is going to go back to you guys. That's what we're here for. Our value as a company is to create value for our shareholders. And the excess cash, not a portion, 100% will go back to you. We've done a lot of work to understand what's the best way to do it, and the current moment, we believe that the best way to do it is through our share buyback program. What are we doing? You look at last quarter, Q4 of 2026. We generated $5 billion. How much we will return to you? 4.5. Right?
We're living to whatever we're telling you is exactly what we're executing. The board authorized a $6 billion program. We executed 4.5, we have 1.5 left. The board authorized another $14 billion program. So now we have $15.5 billion authorized and not spent yet. We will give you an update as we go on, but we believe that our role is to return the cash to our shareholders. In closing, we're super excited. We're super excited not of the value we have already created. That's good, don't take me wrong. But we're very excited about the value we can create going forward. We operate in a very attractive market. It's growing, it's profitable, and frankly, SanDisk is very well positioned to capture that value. You saw our technology.
We have leading technology with NAND, leading technology with our products, and we have very close relationship with our customers. Those relationships, those NVMs are opening doors that had never been opened as wide as they are today. We feel very good about where we are in the market. What is our financial model? Super simple. Translate bits to revenue to profit, and profit to cash, and then the cash goes back to you guys. That is our model. I hope that you guys find it interesting. We do. What we are going to do next is we are going to talk about HBF. As David mentioned, HBF is not in the revenue projections. We are funding it. It is part of our OpEx, it is part of our CapEx, but we are not including the revenue projections here.
Thank you for that. I will turn it over to Alper.
Please welcome back to the stage Chief Technology Officer at SanDisk, Alper Ilkbahar.
Hello again. Wow, nobody left. That is amazing. In the second part of our technology presentation, I get to show you our innovate to amaze DNA. I will talk about two technologies that we introduced last year on this very stage. Both of these technologies address the memory wall problem. Memory wall problem is essentially simply DRAM not keeping up with the compute in AI because it just does not scale anymore as well as it used to. To solve that problem, we started working on two technologies, both of which are highly scalable and can solve this memory wall problem. The first technology I am going to cover is the 3D Matrix Memory. Let us just dive into it right away.
Quick recap first.
Oh, did I-- Okay, I'm back. Quick recap first. The 3D Matrix Memory, we started working on this technology back in 2017 in our research lab. In 2024, we moved the development to a 300 mm facility, a modern facility at our development partners, IMEC. Last year when I was here, we had just delivered a development vehicle that we could just essentially pursue the activities at IMEC with, and that's what we had shown you. Since then, we continued making steady progress. We used the development vehicle I showed you, and we started depositing memory layers on top of it, and we delivered 300 mm wafers and package parts to test and demonstrated multi-gigabit level functional memory arrays. Our devices are approaching performance levels that are getting pretty close to our product specs that we had. So steady progress, it keeps going.
This definitely is a project that has a longer time horizon, and we'll keep updating you as we make more progress on this. Okay, with that, let's go on to HBF, High-Bandwidth Flash. Last year, again here, we introduced High-Bandwidth Flash for the very first time. High-Bandwidth Flash delivers the same read bandwidth as HBM, yet with 8- 16 times the capacity. We invented this device with AI inference workloads in mind that actually leverage mixture of experts type sparse models with long context lengths and large KV caches. That's what we had in mind. Today, when I look at some of the most recent developments in the world of AI and the trends, actually, these do justify the vision we had for HBF two years ago. On this table here, I have summarized some of the latest frontier models and their characteristics.
You're going to see very quickly that certain trends are emerging. First, the parameter size. The models are growing. Trillion plus 2 trillion parameter models are no longer amazing. They're just commonplace. Many of these models actually started utilizing mixture of experts, sparse models, and they're allowing their users to go up in context lengths all the way to million type of tokens. So this is creating a new paradigm. The large models as well as the long context lengths and implied KV cache sizes are driving much higher memory capacities. While the mixture of expert type sparse models are driving the compute needs down. So you're seeing memory needs going up, compute needs coming down. We call this a new paradigm called memory-centric AI. In this memory-centric AI, we think HBF is going to play a very critical role.
Before I dive into HBF further, I wanted you to hear from somebody who deals with these LLMs and AI inference on a daily basis at a massive scale. I am going to take you back to FMS, which is Future of Memory and Storage Conference in California. It was held last week with about 3,000 plus attendees, and there were several sessions dedicated to HBF during that conference. I am going to take you to a panel discussion that we had and going to share with you some of the thoughts from Dr. Xiaoyu Ma, Google DeepMind. Please roll the video real quickly.
The basic vanilla inference is already memory bandwidth hungry. Now, each of these six trends has a big memory challenge for both bandwidth and capacity. If you look at the state of art large language models or agents, congratulations, you have all of them. This creates an enormous memory challenge, and this is why we have an industry-wide crisis for large language model inference.
Okay. Dr. Ma is talking about a memory crisis. Next, let's listen to how he believes we can solve the problem.
What do I believe? I believe in three trends. First, inference specialization for transformers. The reason is because the transformer inference is fundamentally different from training due to its auto-regressive nature and the use of KV cache. Second, I believe in memory heterogeneity. That's because the HBM-only architecture has inefficiencies to scale up the capacity, so that's inefficient for inference. I also a big fan for software/hardware co-design. I believe we are in a golden age for co-design with many 10x opportunities, and HBF being one great example.
Okay. With that, Dr. Ma is one of the many researchers who are actually spending a lot of time thinking about HBF as the latest and most exciting memory technology. It is really becoming increasingly an innovation platform, and researchers are proposing new architectures showing how one could integrate HBF into AI solutions. I wanted to share with you some of the architectural proposals that have been published recently. The first one here is an XPU, where we have taken out all of the HBM stacks and chips and replaced them 100% with HBF. This is an HBF-only architecture. Very simple. The second one is where you're sort of mixing and matching depending on the workload needs and replacing part of the HBM chips with HBF. It's a hybrid architecture.
The third architecture is also a hybrid architecture, but in this case, the low-capacity HBM chips act as a caching tier in front of the high-capacity HBF. The fourth one is a disaggregated architecture. In this disaggregated architecture, excuse me, we are disaggregating the two stages of AI inference, the prefill and decode, and optimizing the solutions, the hardware solutions for these in a disaggregated fashion. The prefill XPU is compute-intensive but doesn't need a lot of memory bandwidth. What you can do is couple a performant GPU with just regular DDR DRAM. Whereas the decode stage, which is very memory capacity and bandwidth intensive, but doesn't require a lot of compute, you could take a modest GPU and couple it with HBF. You get the best of two worlds and combine to optimize the overall solution.
These are few of the ideas that are coming out, and there's many more, and results of these have been published. But I want to today double-click on the first, the simple architecture, and share with you some of the workload simulation work that we have been doing on this architecture. For this simulation work, what we have done is we've taken a GPU actually that resembles a market-leading GPU today and has 192 GB of HBM on each of them. Then we have created another version of it by replacing all of the HBM chips with HBF, and that gives it about 4 TB of HBF memory. We simulated a benchmark that essentially emulates in a multi-turn agentic workload. It simulates or emulates a code development environment. What happens is the AI agents start developing code, spawning more jobs, et cetera.
The underlying LLM model here is a 490 billion parameter Qwen 3. Let's see how the two systems are comparing, and what we're going to measure is the token output. We're going to compare the token output of these systems. First off, when we start with the HBM-only system, it turns out that the minimum viable system to run this workload requires use of eight HBM GPUs. You just cannot fit the model with less than that, so you have to use at least eight GPUs to start this job. Here's what the workload looks like on our simulator. Okay, so eight GPUs delivering pretty stable token output. Okay. Now we are going to show you how an HBF system compares. It turns out that we were able to actually run this workload with a single HBF GPU, one GPU alone.
And let's look at that. Obviously, the performance is not as high as eight GPUs, but if you just wanted to have the minimum capital spend to run this job, all you need is a single GPU. This is the result. Next, I want to show you what four HBF GPUs look like, and here is the result. With four HBF GPUs, we are able to match the performance of eight HBM GPUs. We're getting 2x the performance out of our GPUs. How is this possible? What's happening? Actually, what happens is as the workload starts running, it quickly starts more and more jobs and runs out of the KV cache capacity. It runs out of the high bandwidth memory capacity.
The moment you run out of the capacity, you spill into the system memory, and that spill, and losing that bandwidth, essentially costs you roughly half of your performance. Your GPU utilization drops by nearly 50%. That's why we're able to deliver the same performance with half the number of GPUs. Out of this work, we had two key takeaways. Number one, if you're somebody like, say, a small business or a solo software developer who doesn't need massive amounts of tokens, but you just want to run this job with the minimum CapEx, we can improve your spending by 8 times. You get 8 times CapEx efficiency using HBF. The second takeaway is that at the maximum token output, we are able to deliver you 2x the GPU efficiency.
Which means your capital will go twice as far, which means you're going to burn half of the energy and all the economics that essentially the benefits that you can gain. This is fundamentally going to change the economics of AI. This is the memory crisis Dr. Ma talked about, and this is how we intend to solve it. Obviously, we're very bullish on HBF, but we also think that it's not only for data centers. We only believe that we can dramatically change how AI is run on edge devices with HBF. We are envisioning enabling really sophisticated AI models. I'm talking about 100 billion plus parameters, sophisticated models to run on edge devices and enable an AI experience that I like to call AI that never forgets.
What I'm talking about here is an AI agent that knows everything about you, that's constantly with you, remembers everything about you on an instant. You don't need to go back and forth many times. Everything is there with you all the time. We believe that's going to significantly change the way people are experiencing AI in their lives. To that end, we've been working on a second-generation HBF device, which we call HBF for the Edge. This is what you're seeing. This has been actively in development with multiple customers. This is what's next on the roadmap that we have for HBF. Obviously, we have a pretty big vision for HBF, and having that kind of a vision, you really need to have a vibrant and diverse ecosystem to be successful to realize that vision.
We understood that from the beginning on. When I was here a year and a half ago, we told you that we intended to create an open ecosystem around HBF. True to our word, last August, we announced a partnership with SK hynix and talked about our intent to create an open standard around HBF. We followed up in February of this year, established a consortium under OCP with participation from Google and Tenstorrent. Last week, we celebrated the release of our first public specification that is going to allow XPU designers to incorporate HBF in their systems and designs. The next step is going to be to expand the membership of this consortium. I have a piece of news to share with you. Already, we have Meta joining this consortium, so we are very happy with that.
As they and other participants contribute their feedback and input, we are looking forward to incorporate those in the next revisions of the specification over the next two years. One critical element of our ecosystem we view as the advisors that we have, the technical advisory board that we have built. You may remember I talked about this, Professor David Patterson and Raja Koduri, our legendary computer architects. We are very proud to have them on board. Last week, I had the honor of introducing our next board member, Jim Keller, to our advisory board. Jim is a rockstar chip designer. Over the last four decades, he led teams in some of the most consequential processor designs at DEC, AMD, Apple, Tesla, Intel. I actually started my career as a CPU designer and competed against several of these things, and it is not fun, I tell you.
Jim brings his expertise and guidance into now leading Tenstorrent as the CEO of the company. Today, I have the great pleasure, a surprise for you, Jim is here with us, and he is going to join me on the stage for a conversation. Jim, would you please come on stage? Please welcome Jim Keller.
Thank you, sir.
Jim, thank you so much for coming. Please, and let me hand this over to you.
All right.
Great to see you here. You flew yourself?
I did.
Thank you very much.
I had the help of an airplane.
All right. Well, that's great.
Not entirely myself.
Well, thank you for being here. Jim, we talked about all of the great processors and compute projects that you led, but then now you are taking all of that into the world of AI. How has that journey been for you? Please tell us what you and your teams at Tenstorrent have been up to recently.
Yeah. A couple of years ago, well, it has been obvious for maybe five years now, that AI is going to take over most of the data center. There is going to be heterogeneous computing, so AI compute and general purpose compute, but it is built on the usual fundamentals. We built Tenstorrent around that premise. We build high-end AI processors, and high-end RISC-V processors, and we have two businesses. Business one is we license that IP for a variety of projects, so autonomous driving, robotics, a couple of supercomputer companies, and we are looking at some server projects. We build the IP, but then we put this into our high-end server design. We are in production today with Galaxy. Galaxy is a scalable AI computer.
One of the things I think it is really interesting, we are going to talk about this, is computing has always been based on the balance of memory, compute, and IO, like generally networking. What happened with AI in the last couple of years came right towards us. We built a Blackhole chip that runs AI models, and that is really good for 70 billion parameter models. We thought we would build this Galaxy server with 32 chips per server so we can scale up. In the last three years, we went 70 billion, 300 billion, 700 billion, 1.5 trillion, 2.8 trillion. The scalability of that has been amazing. We did something interesting. In our boxes, we have 56 800 Gigabit Ethernet ports per server. Then we put those together in quad servers and then hook them.
Today, 36 of our servers all hook together, and we run models on anywhere from a single chip now up to 20 servers. We are in testing with 36, and it is scaling really well. The other thing we did is, this is pretty general purpose AI. It is a combination. We have flash in the host, DRAM in the host, AI DRAM, SRAM, compute, and networking. That lets us run a wide variety of models on the same hardware. We announced in May, DeepSeek at 400 tokens a second. That is batch 32. This is very high throughput but very high token rates. We do prefill decode on the same hardware. We ran WAN 10 times faster than anybody else. It is real-time video that runs on four Galaxy servers. Recently, we just announced 900 tokens a second on Kimbu K3. Or, I guess this is 2.6.
Three is up and running in the lab. Now we have a new higher resolution video model. It is all on the same hardware. The reason we really think about this hard is AI is changing so fast. Who here heard about KV caches two years ago? Anybody? That is a part of an LLM, the fact that we can cache it. Two years from now, something different is going to happen, and everything needs to be flexible, compute, memory, and IO. What we are going to do with really large memory is amazing because memory is one of the most flexible things. You can put programs, models, weights, caches, data sets. There is so much to do with that, so we are pretty excited to be here today.
Great. Thank you. Thank you very much. I can take that if you want, or we can just leave it there. Perfect. Jim, you talk about these super scalable systems that you are building. They go all the way from smaller needs to very large scales.
Yeah.
When you look at these scalable systems, where do you see some of the bottlenecks?
Well, it depends on how the hardware is built. Like today's HBM-based models, they're limited by the local size of the DRAM. Which you talked about, and they often don't have enough network bandwidth. So one thing we did is we have a terabyte per Galaxy server. 36 Galaxies is 36 TB of DRAM. But because the network band was so high, we conserve the memory from one part of the machine to another really flexibly. I think the two biggest bottlenecks today, we're doing pretty good on compute, but memory scalability and then the network scalability, so you can serve the memory everywhere you want. Those are the big ones.
Great. When you talk about memory, there is a lot of AI architectures that people are talking about, and they're highly differentiated by the way they use memory. Like we see, obviously, the most commonplace GPUs today with HBM memories. And then we are seeing architectures like from Cerebras or Groq that are relying mostly on SRAM, and now we are talking about HBF. How does this whole thing How do you think about the variety of these memories, and how do you think about HBF in that context?
Yeah. So first of all, GPUs were built for graphics, and they read and write all their data to memory all the time. That drove them to really push hard on HBM because they do not have enough SRAM on chip. Our processor has 200 MB per chip of SRAM, and we can put a large number together. So the balance of SRAM to local DRAM, to host DRAM to flash is really important I think. The Groq Cerebras exploited essentially a gap in the GPU roadmap. But the thing that we are going to see is the ratios of compute, what we call KV cache to decode the prefill today. Those ratios were one to one to one, and then it went to seven to four to one. Now it is 100 to 10 to 1, and people are still moving that.
If you build a machine that has specific processors for different pieces and the ratios change, what is going to happen to your compute?
Yeah.
So.
How do you think about where is HBF something like a very high capacity, high bandwidth memory relevant? Can you think of some examples where it will be really useful?
Yeah, definitely. AI is a very high bandwidth problem. Memory bandwidth, network bandwidth, compute bandwidth. The limitation, flash is great because it is lower cost per bit, much bigger capacity, but it did not have the bandwidth to really play in that high bandwidth system. The really cool thing about HBF is now you have brought the bandwidth to the table.
Right.
That makes it really great. The other wild thing is, and to be honest, I did not see this coming, the fungibility of compute and memory is amazing. Who knew compute would be so expensive that we should compute it and save the results of the computation in that big format, right? People do not realize, when you send tokens in, it is pretty small stream. When you embed that and then compute the KV cache side, it is a very large footprint. With HBF, it is now effective to save that for a very long amount of time.
Right.
That unlocks the ability to balance compute memory in a new way.
Right. We have talked a lot about data center, but you do also a lot of work outside of data center in the edge. Do you see any applicability of HBF in the edge devices and edge applications?
Oh, definitely. Today, autonomous things look at the world, and they have to process everything, and they have a model that's trained. Having huge augmentation of KV cache for everything you see in flash on device is going to rapidly change how robotics work. I was joking this morning, how many people here would wish GPS worked in New York City, right? Imagine you have a device that actually knows everything around you, it knows exactly where you are. And there's going to be so many transformations, and I don't know if that one particular is going to work out. But one day I was going, "If I had enough data, this problem would be solved." And so there's a really interesting thing about robotics, is everything you already know, you don't have to compute.
As we make that memory available in robotics, autonomous driving, so many applications, it's going to be pretty transformational. Memory is way lower power than compute, and it's a good trade.
Well, thank you very much. This has been amazing. Thank you for coming and being with us. And I believe you're going to be with us available-
Yes.
-to answer questions after lunch?
Sure.
Or during lunch?
Yep, you bet.
Okay, thank you so much.
Thank you very much. Thank you.
Before I finish, I want to probably address one question that I suspect is top of mind for many of you, which is when? When are we going to see HBF? Here is the latest update. I am happy to share with you today that we actually taped out our first HBF memory die. You are seeing. Actually, you happen to be the very first people in the world outside SanDisk seeing this picture. I apologize, I had to pixelate it because we are still not quite ready to share all the magic that goes into it. Our die has taped out. You do not have to wait too long to see the whole thing. A little bit more patience, please, but we are continuing working on this super hard to deliver our first samples to our customers of inference devices with HBF next year. Next year.
Just a little bit more. With that, I thank you all very much for being here, and I am going to invite back our CEO, David Goeckeler, on stage. Thank you and have a great rest of your day.
Good job. I really want to thank Jim for coming all this way to support us and more importantly, for joining the advisory board around HBF. This has really been quite, as I said earlier, if I look back over the last year and a half, a lot of really great things have happened at SanDisk. But this one, the ability to make a market and attract people as capable, saying Jim is capable is a bit of an understatement. But people that are this distinguished in the field to come help drive this technology forward is just really amazing. We are super happy about where this is, and we will keep you posted on product availability as we continue to drive these milestones forward. As Alper said, the fact that we now have a die that we have taped out, this product is real.
It is going through the fab. We are producing it, and we will get it back, and then we will put the systems together and get it in customers' hands, in our partners' hands. I think this point that was made about co-development. That is what it is all about when you are developing new technology. If you are co-developing with some of the largest customers in the world, that is a really, really fun place to be. As we make progress on that, as we get those samples in customers' hands, we will get a lot more information about what the future of this technology looks like from a market, financial, all of that perspective. Stay tuned as we move through that process next year. All right. We have a I am not going to read through all this because you just listened to all of it.
But it is a recap of where we are. I think you can tell we are. Hopefully you can tell we are extremely excited about where we are. Like I said, I have been doing this for now six and a half years, really trying to unlock the value of this franchise. I feel like we have made a lot of progress. Since we separated the company, we have seen both companies just bloom, and really start to get the valuation. But I really do believe we are entering a very different phase of where we are going to take these franchises. The ability to really recognize the true value of this technology we have been building for decades.
As the market changes, we build new customer relationships, and we really get this engine running of, again, turning bits to revenue to profit to free cash flow, and returning that cash flow back to you. All right. I am going to bring everybody back, and I do not know, somebody may have a question in this group. It has been my experience that some of you often have a question. We are going to open it up for Q&A. We will bring everybody back. Everybody is fair game, and we will do our best to answer whatever questions you have. There will be a mic runner, so if you have a question, I guess raise your hand. Okay. Right here. Go ahead, Jim.
Thanks for taking my question. Jim Schneider, Goldman Sachs. I have one business question, one technical question. First of all, on the business question, can you maybe talk a little bit about how you're thinking about the diversity of customers you want to include in the NVMs? You talked about the hyperscale component, the data center component. How do you think about the broader mix and having too much risk in any one given end market? Then maybe secondly, on the technical side, if you think about, we're hearing a lot of discussions about some GPU customers wanting to reduce the amount of HBM content in their systems today. So before HBF comes to the market, how do you think about the amount of HBF or conventional ESSD content you need to add to an existing GPU configuration to deliver the same performance?
Luis, you want to start with the customer mix?
Yeah. Jim, we've been very thoughtful in which customers we sign NVMs with. We're looking at different markets. It includes data centers, as we talked, it includes edge customers as well. So we're looking across. Even when you look at the hyperscalers, their business models are dramatically different. So being very thoughtful in driving that diversity for the reasons you mentioned. But we're betting on winning customers. We believe that they're going to be here with us for many, many years to come. The level of integration, both technical and commercial, is as strong as it has ever been. That diversity of business models is one of the criteria we look at.
Jim, on the second point, I'll say a few words and then Alper and Khurram can have a point of view on the various specifics. But I think what you're drawing out is what we're seeing in the market. It's definitely what Jim just said. This is changing at an incredible pace, right? That makes it difficult to understand what product to build and how much of it to build. Especially with the way the market used to be organized, right? Build it and we'll talk about what the price is later. What it says to me is there's just a huge premium on staying very close to your customers because this is going to change. It's going to continue to iterate over and over again. The scaling of inference is incredible.
It is the most incredible technology transition I have seen in my career by far, and I have been involved in some pretty big ones. It is going to change very rapidly. There is going to be constant innovation. As Khurram said, there is a constant focus on how do you drive the requirements down. How do I use less power? How do I use less space? How do I make this more efficient? Because the more efficient you make it, you can scale it faster, right? And more economical. Also if there is three different people scaling inference around the globe, if one of them is twice as expensive as the other, that is not going to work very well from the business model.
Our customers, the great thing about where we are from a broad technology point of view is you have these companies now that are just spectacular. They can scale on a global footprint something this complex very, very quickly. Staying very close to them and understanding where they are going is, in my opinion, extraordinarily important for where we are going to drive this franchise. That is another angle on these NVMs. As Luis said, we have NVMs with some of the largest customers in the world. We have their commitment of what products they are going to deploy quarter by quarter for the next three-five years. That gives us a lot of insight into all of this confusion of what is the product, what is the architecture, how is it going to play out?
It gives us incredibly unique insights about how that is going to play out, and what are the right products to build, and where to put our resources to make sure they are successful. Now, one of you guys want to talk about the specific question?
Yeah, sure. David is right. Our customers provide us with a lot of data, but I will give you a little bit of long-winded answer on this. Like I talked about AI lab at SanDisk in action. You are absolutely right. The question is the system DDR, how much of it you need, as you have the continuum between the volatile and non-volatile media. It was coincidental, I was sharing the data with David yesterday, from our lab, where we ran a 1.2 trillion parameter model. We ran, again, multiple user sessions, prompts. An interesting thing emerged from that. When you look at performance and power of the system to execute that workload, what we saw that after, and we ran both the HBM sweep and DRAM sweep, like to think of it as a shmoo, right? When you, and running the shmoo around capacity.
You start with, let's say, and I am making up the number, 4 gig of volatile memory, like HBM and DDR, and then go all the way up to the maximum capacity of that server scale. It was interesting to see to run that complex workload, having just the HBM SSD was good enough, and you really required a minute amount of DDR to run the batch services. That was a revelation to us, too. By the way, we heard the news that, hey, you are reducing on the Vera Rubin, I think, cutting the system DRAM by half, the SO-DIMM DRAM. I would contend with the studies that we are doing, you can go even further, right? If you have an ESSD and HBM in the system. Like David mentioned, things are going to change.
With respect to the ESSD capacity, again, it is workload dependent, how much you want to run, taking you back to the persistent KV cache pool, right? How big you want to grow it. There is a lot of factors that go in, and those are more use case dependent. But certainly, one thing is emerging that the system DDR with respect to running the inference, it can be reduced. Now, could it be reduced by half, one eighth? Depends.
Ben, you want to. Where is the mic? You guys will run the mic up.
Hey, guys. Ben Reitzes with Melius. It is great to be here, thanks. Thanks for doing this. First, I got an observation, which I hope some people find kind of amusing, and I wanted you to react to it, and then I have a question. I have been going to tech conferences and analyst days since about 1992, and I have never seen a company guide for a year, three years out, and be trading at less than 3x that number. It is pretty amazing. I just wanted you to react to that, trading at less than 3x your FY 2030 number. Second, HBF. It is not in the model, but you are only growing bits, mid to high teens. How do we put it in? What if this is a hit and we got to add it to the model?
What are you getting rid of that we just guided for to make room for it, and how do we calculate upside if bits can only grow a certain amount?
Well, so I appreciate your observation. I may come up with a different word than amazing in my reaction to that, but that's a whole different discussion. Look, this is a conversation I think we'll have as we get the product in customers' hands and we really understand what the demand is. Again, I am going to sound like a broken record. We keep coming back to these NVMs. We are going to follow our customers, right? When our customers tell us they need something, this is the way most technology businesses work, right? You work with a customer, they want to buy something, and you build it. You do not talk about what the industry supply is. Who cares? We have the ability to produce what we can produce, and it is about getting these incentives aligned.
What I would argue, if I go back a year ago, the incentives in this industry were just completely misaligned, right? That leads to all this volatility, and that is good for nobody. We are clearly walking down a path. In two quarters, we have gone from three months of visibility to four years of visibility. I think it is just a little bit of wait and see what it is like a year from now when we actually have this product in those customers' hands, what their demand is going to be for that, and we will then figure out what the production plan is behind that. Remember what I said earlier, we own the whole stack, right? From the NAND IP to the production front end, back end, the whole thing.
We are not ready to go there just yet, but imagine what could be possible in the future. All right. I will go to the back next. CJ.
Yeah. Thank you. Aaron Rakers.
Where's Aaron?
Yeah. Aaron Rakers with Wells Fargo. Appreciate the day and all the details. I want to go down the path with HBF as well because I think it's a fascinating technology. Alper, one of the evolutions in DRAM that we're seeing, specifically HBM, is this idea of customization, right? Driving towards a base die that has certain elements that compute in it. One, do you see that as a roadmap that you could explore on HBF? And then secondly, I guess back to Ben's question, if it's successful and we start to get into next year and we start to see design in, right? And maybe that means 2028 or even 2029 volume, should we expect the CapEx discussion to change? What's involved in a production of HBF? Advanced packaging, is it a different capital intensity that we should be thinking about?
Thank you. I agree. I'm also in love with HBF. Your first question, I apologize, was-
Customization.
-customization. Yes. Customization that we started seeing. It is a natural flow of actually compute moving towards memory, right? Because it is just shuffling data around is just so expensive, so energy consuming. It is something that we had been anticipating in or near memory compute is going to definitely happen, and we are entirely prepared because we do have all the capabilities. Khurram showed you all the memory management capabilities we have. We only welcome doing compute next to the memory or in the memory. It is our wheelhouse, and we love to do that. We will cherish that, and we are highly capable of doing that. It is entirely a customer-based discussion, what part of that pre-compute or the compute itself do they want to perform inside the controller that sits next to that memory?
It is that conversation you need to align, but after that, you are going to see us being entirely capable of doing that. The second question in terms of how do we think about all the investments. Certainly, one of the reasons I am in love with HBF is because it entirely leverages our knowhow, expertise in NAND flash, right? We have the best flash, and that is why we were able to create HBF to start with. In terms of what does it mean for capital and whatnot, you have seen us talk about the productivity of our current NAND roadmap generating, actually giving us the potential to easily increase our output if we choose to do so. This is a great outlet, right? The beauty of HBF is that as a business, it is entirely orthogonal to our storage business.
So we could definitely entertain that, and as Dave said, as time comes, we will see some of the other capabilities certainly are, again, within our capability range, and we are looking forward on working on all of those. So it is going to be definitely very exciting. But let us wait till next year and see how this will get into the market. I think those are excellent high-class problems that we are willing to work on.
Let us go back.
Hi, thanks so much for doing this today.
Hey.
Mark Newman from Bernstein. Actually, I wanted to ask about the NVMs. It seems like great progress you are sharing here again today. You talked about FY 2028 being 66, two- thirds of volumes on NVMs by FY 2028. Do you have any target in mind for what percentage of volume that can get to? Related to that, what is the volume in FY 2029 and beyond? Is it also similar two-thirds level of commitment on these NVMs? Then a second question, this may seem a silly question, but on the 100% excess cash to shareholders, does that mean 100% free cash flow, or do you have a different definition of excess cash? Just wanted to clarify that. Thanks very much.
Yeah, I hope I was clear on the free cash flow. Excess cash is defined very simple. The cash we are generating minus whatever we are investing in that bucket number one, investing in the business. There is no trick. It is 100% of the cash. Excess cash will go back to shareholders. If you look at what we just did in Q1, that is exactly what we did. We generated $5 billion, and we spent $4.5 billion, right? So pretty much there. I think on your question on the percentage of NVMs, we are going to be optimizing over time. We are going to be learning a lot, and we will define what that optimal number is. We like the numbers that we see for 2027 and the numbers that we see for 2028. We think we have optimized based on the data that we have so far.
But we will keep on talking to customers, we will keep on evaluating the situation, and that may evolve over time. Our goal is to create the maximum shareholder value that we can on a sustainable basis. So that is the filter, and that is what going to guide that percentage. What is the number for 2029? Yeah, I got the question before the session started. It is consistent with the numbers that you are seeing. It is consistent with the 2028 numbers so far, but that number will keep on evolving. That is why we did not put it there, because we will keep on optimizing our number. Okay?
Yeah. Hi, thanks for taking my question. It's Krish Sankar from TD Cowen. I had a question for Khurram. Thanks for your interesting presentation on KV cache. The two pushbacks I heard on NAND for KV cache is, one, the tail latency is much longer than average latency, so that impacts system throughput. And number two, it takes a long time to write KV cache onto NAND, and so lowers the lifetime of the device. So I'm kind of curious what your answer to that is. Thank you.
Yeah, thank you. I saw your post FMS commentary on this, and I wanted to send you a chart that I showed to David. Look, when we talk about latency, again, you have to look at the mix between the top tier, high bandwidth tier really suffers from capacity. We all have to acknowledge that first, right? And like Jim was talking about, the need for the space to have all that context cannot be serviced by that. You have to go to the next tiers. So far what we see is that the overall latency is actually quite good, and it's actually workload dependent. If you're looking at something that is real time, obviously you won't go to the next tier of the KV cache. But you don't need to. You can put that in the high bandwidth tier, right?
It's not just one or the other. It depends on what you are running. And what was the second part of your question? Sorry.
Endurance.
The lifetime.
Oh, the lifetime. So look, we are making continuous progress. Alper keeps giving me great technology with very high endurance metrics. Part of this, we are learning a lot, not only on our core technology, but also on HBF that is allowing us to get phenomenal endurance number, as you saw on the warm cache that we talked about is exactly that. That requires, for example, three drive write per day. So we are increasing the endurance at the same time. But again, it is also workload dependent, where you need to read more, like we use our stock standard NAND, but like for the warm KV cache, to your point, we are offering a 10X much higher drive write per day as we would on our conventional SSD. So, you are right. But we have tricks on how to do that.
Let's go back.
Hi. Joe Moore, Morgan Stanley.
Hi, Joe.
I also wanted to ask about the business mix between the segments. If you are 1/3 data center now, and by most accounts, data center is going to roughly double in the next 12 months, what does that imply for your mix a year from now? Because if it is 60% data center, you really starve the edge businesses a lot. Just how do you see that mix? It seems clear to me that there is not enough supply. How do you think you balance out that shortage?
Joe, this is what we are talking about now. There are a couple of things the way we think about this, right? Part of it goes back to Mark's question. We want to engage with customers on a longer duration, right? There are different contracts for different duration now, and that is why it is a little hard to say what 29 and 30 are because you are starting to get in a time where some deals are ending and others may start and those kinds of things. We also want to keep some flexibility in the system as well. But as Luis said, we are going to be very judicious about where we go from here. We are going to follow our customers. It is very clear how we want to operate our business. We want more visibility. We think that is better for everybody, right?
It helps us make the best decisions. It is very clear that this model started in data center. I think that is fair to say. Their interests are different. As I said earlier, we are starting to see now that change pretty dramatically, where customers are coming to us and wanting to have these kinds of deals. What I said earlier, we want to make sure the whole portfolio stays robust, and that means you have flow of products through edge, you have it through consumer, you have it through data center. So we want to keep all of those alive. I want to make sure Khurram's engineers stay very busy building the best edge products in the world, just like they are very busy building the best data center products in the world, and of course, they are very busy building the best consumer products in the world.
I think it is fair to say our data center mix is going to go up, right? We are going to continue to rebalance that to make sure that we keep all of those markets alive and we get the best return possible for all of you, again, over that multi-time horizon that we are looking at. It is a little bit of a squishy answer, but it is a little bit we know it when we see it. The deals that get presented to us, we have got a long roster now of here is the potential opportunities we have to engage in the next contract. Luis and I must talk about this four or five times a day.
Exactly.
Like, literally for the last nine months, we're constantly talking about this. First it was the structure. How are we going to structure this to make sure it works for both of us? We just iterated around that, and we came up with some ideas of how we thought it would work. We talked to our customers, and we got that landed. So we got the model landed now. Some of the first ones were some very big ones, quite frankly, and we got those locked in, and we're moving in around that. As we go through this process now, we're going to make sure we do all those things I just said. How do we optimize value creation over all of these time horizons?
We want to make sure we keep the full portfolio alive and running, and we want to make sure we mix into more data center. Where is the final target? It depends.
Thanks. Vijay at Mizuho. Just a quick question back on HBF. I think one of the things we saw on DRAM side is when HBM came along, it made conventional DRAM very tight. It took out capacity. Do you see the same thing happening with HBF, where conventional NAND becomes very tight? Is your first gen HBF on BiCS8, BiCS9? If you can give us some idea on what the pricing dynamics would be, and what is the capital intensity, I guess, on HBF versus conventional NAND? Thanks.
You ready to give out pricing yet, Alper?
I feel it's priceless. At the same time, I'll answer your question that I know something about, which is HBF. We base it on. Again, we are using our ability to mix and match different technologies, but the array technology is mostly based on BiCS8. So that's where the capital and install base is based on. We're going to use the most convenient CMOS technology that will drive all of the features and capabilities and performance we need. That's going to be something a little bit more, more than a little bit, a higher performance than what we use in BiCS8 itself. I gave an example of that in BiCS9, like whatever we did, and this is not going to be too different than that. As to how we think about its impact, as I said, our technology right now is giving us annually 27% growth capability per wafer.
We are using only two-thirds of that capability in how much we push into the production. I think we have a lot of headroom to grow through the innovation that we have, right? Allow HBF to squeeze into that window when we need it. If HBF becomes even bigger than that, as I said, high-class problems, I'm sure we'll deal with that because we know how to manufacture things at mega scale.
Yeah, I think, Danit, all you guys are asking is getting back to the same thing. Like where are you going to get capacity for it? Can it scale? One of the things that I hope you walk away from this with is NAND is a very scalable technology. To Alper's point, we have the fundamental engine for at least another decade of line of sight to scaling, right? I think that's an important issue, right? I think that in the end, Alper has taught me a lot of this, quite frankly, the most scalable technology wins, right? If you get the economics right. Now our problem is not our ability to scale. The issue is the chart that Eric put up.
We have this giant discontinuous period where we end up taking hundreds or thousands and thousands of wafers out of the system because we did not get supply and we do not understand the dynamics of the market because we do not have enough visibility. In my opinion, the answer to all of these questions, that is where the magic is. That is where the answer is. Getting more visibility into what demand is going to be. There is nothing we can do about demand this year or supply this year. The fab plan is cooked, right? When you start talking about the turn of the century, yeah, we can start turning the dials. The other thing you should take away, like Alper. Alper. Khurram put up that four variables, you multiply them all together and you get KV cache size. Again, I will go back to do the math.
When you have four variables and you are multiplying them together, when you start changing any one of them, the outcome can just be wildly different, right? You are essentially answering this question, what should I build for 5 or 10 years from now? The people that have the answer to that question are the ones that are investing the billions of dollars to build the infrastructure. The most important thing is to get as close to those people as possible, understand their roadmap. You help them understand what is possible, and I think that is really the magic of HBF. You cannot really expect somebody that is not a world-class memory expert to solve the memory wall problem. I think that would be an unlikely outcome.
The most likely outcome is the people that are world-class experts in memory and storage design, which these guys have been doing for decades, are going to be the ones that help unlock this scaling issue. Again, who do you want to stay close to? The video. Co-development is the answer. This is not new for us. We have been doing co-development with hyperscalers for a generation of technology now. Because you are going to get this, that is where you are going to find out how everything fits together. I think the answers to all these questions are we need to extend this visibility into what true demand is really going to be. If we do that, I think everybody wins.
Like I said, we are two quarters down this path of, I do not want to say inventing a new business model, but again, this is not the contracts of the past. This is not, "Well, just sign this NCNR." That is not what this is at all. It is a completely handcrafted structure of how we are going to work together. Two quarters, we have gone from three months to four years. Let us see where we are next year, and I think it will help answer all of these questions.
And if I may add real quickly. Please take a look at the wafers outside. Look at eight and 10 next to each other, and you are going to see 65% growth from one to the next one.
Pretty easy.
How many generations will it take DRAM to scale 65%? They are scaling single digit percentage every generation, every year and a half too. So you are looking at 10 years of, or more than 10 years of DRAM scaling in one wafer. So you ask, what are you going to do if there is an explosion? I will move to BiCS 10. I will get 65% more of it. Not too difficult.
DRAM's a fantastic technology. There's an easy way to get more of it. Spend a lot of money, right? Let's go in the back.
Yeah. Thanks for taking the question. CJ Muse with Cantor.
CJ.
I guess, Luis, first question for you, and then one for Alper. As you think about the NVMs and you contemplate your annual cost downs, is that something that you will 100% benefit from, or is that something you share inside the contracts? Then for Alper, you highlighted four potential solutions with HBF. Curious what your customers are responding to. Is decode disaggregation really where they're focused or are one of the other three solutions where you're getting the best feedback? Thanks so much.
Thank you.
Yeah, CJ, I love the level of details we give you on the NVMs, and you guys always want more. I am not going to get into pricing at more detail. It just gets very sensitive given the conversations we have with our customers. I think the key message you should have is even at the lowest price, at the floor pricing, at any point in time, we do not expect these NVMs to be below 80%. We expect that to be around 80%. So you can model that, and we have confidence of that over time. So every single year over the period, we should be able to be even at floor pricing around 80%.
In answering your question, CJ, when you looked at some of these architectures that I showed, the first two are almost similar architectures. You just decide based on your specific needs. The disaggregated one, you could also think of that from an HBF perspective, one being very similar to an HBF-only architecture except for at a system deployment level, you are disaggregating. So it turns out that all those three architectures can be served through a singular architecture, and that is where we are seeing the concentration of the discussion right now. Of course, the mobile space is very different, and that is what I was sort of showing you differently. There is a little bit difference. But in the data center, I think it is all going to be around those three architectures.
And there might be some other proprietary ones as well that I am not able to talk right now. But we have not quite seen a customer that shows this tiered architecture, if you will, the cached architecture quite yet. But it is also very interesting. The published data on that from the university research was very robust.
Okay. We have lunch waiting, and we have kept you in here a long time. We will all be around to answer questions. We will take a couple more here if people have them, and then we will break.
Hi, this is Sam Feldman on for Karl Ackerman at BNP Paribas. David, you indicated that your technology roadmap can support a 27% bit CAGR through the duration of your NVM contracts, well above your projection for mid-teens growth. One of the ways to meet that projection is supplying CapEx discipline through the elongated nodal transitions from historical rates of every 18 months. First, how does SanDisk think about balancing the share if not every NAND provider slows down nodal transitions? Second, should NAND cost improvement of BiCS 9 and future BiCS nodes slow down to 10% or less given reuse of NAND cell arrays and elongated nodal transitions? Thanks.
I do not know if I got that second question. I think that is always a problem with two questions. I always forget the first question as I am listening to the second question. I think you are asking about can everybody just speed up nodal transition to increase supply, right? I think the answer to that question is these are R&D productivity numbers, but then you have to translate it into an actual fab plan, and that takes years of planning. It is not a dial you just turn up and down, right? It is not that simple. You have to plan years in advance. There are people in our business, and I am sure all of our peers have all the same people, anybody in the semiconductor business.
Here is our fab plan for years in advance, and month by month, what node is running where, in what fab, what is it yielding? Then you add all that up, and you get a bit output. You can definitely start to change that. The point we are making on the productivity is we have ways to scale that are very CapEx efficient. Again, I do not want to keep talking about it over and over. We go back to the company split. I remember when we split Western Digital. It was like, oh, HDD, and it is a great business. It is 4%-6% capital intensity. Here we are in NAND. Luis, you just committed to what capital intensity?
Yeah, low single digits, right?
Yeah. This scalability is incredible in multiple dimensions. This idea that 19 generations of innovation by this incredible team has led to this technology, and there is still a very long roadmap. Understanding that scalability is important to understand the business and how it shows up in different ways. Also there is a flip side to it. Like everything in life, things can be double-edged sword. If you get it wrong, you can overproduce very quickly. I think that is a little bit of the basis of your question. That was the pre-2023 world. If you go to the pre-2023 world, it was 30 and 15, man. I was saying the same thing, 30% growth, 15% cost downs. It was not true. If you invest to that, it is just not going to work. You have to throttle that productivity, which is fine.
You just put a mix plan together that delivers it, make sure you stay focused on delivering it, and realize that when new businesses come in, when new innovation comes, it is likely going to come. Another way I think about it is the scalability is going to attract TAM. It is the kind of way I think in my head. The more scalable you are, people are going to start using your technology more and more because it is efficient, it can be produced, all these kinds of issues. As you are seeing that happen, there is enough horsepower in the engine that we can go at the level we need to go. We just need to not overrun the market. Again, it is back to the same thing I keep saying. The answer to all these questions is aligning the investment horizons in this business.
Where we were, and I have said this publicly a couple times, it was just ridiculous. This idea that you are going to make 10-year investments and then you are going to hold a quarterly auction to see who buys your stuff and at what price, it just makes no sense. I think we are rapidly transitioning to something that is very different. Believe it or not, there are the die-hard cycle people and like, "Let me tell you how semiconductors work. This is the way it has been for 20 years." I happen to have a belief you can change the future. You can change things. If you methodically make changes and they compound over time, things can be different in the future. Sitting around waiting for the past to come back, in my view, is a total waste of time.
Let us invent something new that works for everybody, and we got a whole bunch of people that want to go there with us. It is unbelievable. It is incredible. Unbelievable. I forgot what the second question was. Probably too long of an answer to the first question.
Yeah, it was just on cost downs and the reuse of NAND cell arrays for BiCS 9.
Cost downs for BiCS 9?
Yeah, like cost downs going forward.
One thing I said last time, I am going to go back. I said we are going to stop talking about cost downs. I actually said that, and we stopped talking about cost downs. I have to give all of you guys a lot of credit. Every earnings call was like, "Dave, what is the cost downs?" We are not talking about cost downs anymore, and everybody stopped asking. Mostly. So look, that is our job. Our job is to drive efficiency. Our cost downs are ours. That is part of the productivity of the company. When we talk about them all the time, everybody else thinks the cost downs belong to them. And they do not. They are ours.
They are because of all the great work these guys do and the market will determine what the price of our product should be, not how much it costs to build. And it is just as simple as that. Although it is cheaper in the future. That is what- You can count on that. Again, this is why this is such an unbelievable franchise. Am I going to tell you what they all are? Of course not. But are they there? Of course, they are. That is the magic of what we are doing, right? That is the magic of these guys building 19 generations of NAND. What I said before, if you think it is easy, give it a try. Right?
Awesome.
Right there.
Thanks a lot. Wamsi Mohan, Bank of America. Thanks for doing this. Thanks for all the details around the NVMs. I think a lot of people who have gone through a lot of cycles covering different industries have looked at these NVMs through a more skeptical lens in the sort of-- Look, we have gone through many eras where when you are in an upcycle, these NVMs obviously stick. A skeptic would say in a down cycle, these NVMs typically have not held up very well across many cyclical industries, not just in NAND. I would love to get your response on why is this kind of different, in the sense of, we went through hyperscalers back during COVID who had a demand forecast that they over forecasted, and we had an inventory correction. What makes the forecast today better than their ability to forecast out to 2030 any better?
Because I think some of what you are forecasting is dependent on what they are forecasting effectively, so-
Yeah.
-would love to get your thoughts around that.
I'll try to be brief. This may be difficult. First of all-
You are.
-in everything in life, the easiest thing is to be skeptical. Why don't we just say no at the starting line and never do anything, right? We'll all just be skeptical all the time, right? Why do anything? Why even try? It's all going to fail, right? Look, I get it. It's good to be skeptical. I'm skeptical. But you have to have a belief that you can change things and get a better outcome. This is very important. I look very deeply at this in people we hire. Are you willing to change things and have the confidence you can get a better outcome? It's somewhat unusual because most people won't change things because they think they're going to break something. We don't have that problem. It was already broken. Like hint.
My good friend Luis, who we've worked together before, I was lucky enough to have him join me in this, and he comes at it with a different set of eyes and says, "Why are we doing this?" It's like, there's a different way to do this. Other industries, to your point, do this differently and they get a different outcome. We can do things differently and get a different outcome. You have to have the confidence to do that and then methodically do it and let that compound over time, and you can get a different outcome. So is it going to be different? Maybe not. I'm fully willing to admit that. You guys are all in the investment business. I think the line is investments are subject to risks and uncertainties.
But where the most uncertainty is and where the most skepticism is also where the biggest returns are and where the biggest opportunities are. That should be the message of SanDisk over the last 18 months. The returns we've driven had more to do with the massive skepticism that existed of 18 months ago of where we started. What I said before, compounded growth rates. Very sensitive to the endpoints. When you start with the endpoint extremely low, like way less than even the replacement cost to your franchise, arguably some small percentage of it, of course you can have very good outcomes. Doesn't mean I'm not proud of what we've done over the last year. I'm very proud of what we've done.
But that's why I say you have to put that behind you and look at where we are today, and get all that out of your head, and then make a clear-eyed view of what is the most likely outcome? What is most likely going to happen? Not what happened in the past, and we're just going to go back to the past. We've changed things, and you have to make your own calculus of if you think that's going to be successful. We've made that calculus. We're very comfortable with it. We're trying to make that case to you. But if you don't believe it, and you need to see it, then we just have to wait for it to come.
But we're pretty clear, we just put a capital return policy in place that tells you what we think, and we're going to continue to execute that very dramatically. But the future is an uncertain place. Luis touched on this. You can't imagine the level. I think I've said this to some of you before. It's very difficult from your perspective to understand what we're going through. Because unless you're in this business, you don't understand the just radical fundamental change that's going on. When I started, I came into this business in 2000, and I had a meeting with one of my peers at one of our customers. A colleague that happened to work for the CEO of a hyperscaler.
I got a call the next day saying the supply chain team and that customer was unhappy that I met with my colleague, and that I needed to clear it with them first before I had a conversation with my former peer. Now we're talking to the CEO. Now we're talking to the CFO. Now it's incredible strategic. It's gone from this transactional, you are somebody where we're just going to negotiate with all the time. Your technology's good. It's respectful. There's any issues, but it's just like this is your place in the world, and we will talk about price constantly, to we're having strategic conversations with the most consequential people in these companies. Those people could be just pulling the wool over our eyes. They could be just a big head fake. They could head for the exits later.
That's not my calculus of what's the most likely thing to happen. Those people are officers of public companies, just like I am. They have responsibilities. Board of directors. I'm on the board of director of other companies. You don't sit in the boardroom thinking about, "Hey, I'm going to enter a contract so I can exit as soon as possible." You don't think about things like that. If you think that way, you don't get the job. So it's very different. Is it a guarantee? There's no guarantees. But I think it's the most likely outcome, and we're going to work really hard to make sure it is the outcome.
Great.
I said I was going to be brief, and I wasn't brief. We'll take one more.
Yeah.
Thanks a lot, and I won't ask about cost downs. Although I've used it-
Thank you.
-in estimating what the floor could be like in 2029, 2030. But anyways, just from the physical AI and maybe some of these non-data center demand drivers that are out there, assuming there is still maybe 20%, 30% of the bits that are still going to be consumed as we look beyond 2028, fiscal 2028, what is the feedback? We just keep hearing a lot about de-speccing and some of these edge client devices, a lot of them suffering from a lot of pain because of where memory and storage price points are at this point. Just help us understand the conversations that are going in. It seems like some of them are in your NVMs, some of these client edge customers. How are they sort of thinking about their demand outlook beyond just, let us say, calendar 2026? Thank you.
Luis, you want to talk about that?
Yeah. It is a little bit of a tricky question because I do not want to get into customer by customer discussions. We are engaged with physical AI customers at a deep level. We understand their needs, and we are working with them very closely. Yeah, we have those relationships, and they are great, and I just cannot get into too much details on what they are. But it is a market we are committed to. We see it growing, and there is a huge potential there for sure.
Yep. All right. Look, I want to thank all you guys for hanging in there for over 3 hours now. I see the clock is at zero, which means we're afternoon. Lunch is here. Thanks for spending time with us and well, I'm sure we'll be engaging a very detailed level going-