Micron Technology, Inc. (MU) Earnings Call Transcript & Summary

From Micron Technologies, Executive Vice President and Chief Technology and Products Officer, Scott DeBoer; Executive Vice President and Chief Financial Officer, Mark Murphy; and your moderator from RBC Capital Markets Research Analyst, Srini Pajjuri.

Good morning, everyone. Thank you all for joining the RBC TIMT Conference. Thanks, Mark, and thanks, Scott. And also, we have Satya from the IR team at Micron in the audience. We have roughly about 45 minutes or so. Mark is going to start off with a few comments, and then we'll get into Q&A. And then we should have enough time for any questions from the audience in the end. With that, let me hand it over to Mark. Mark?

Okay. Good morning, everybody, and Srini, thank you for having us here today. I'll start with the safe harbor. We'll be making forward-looking statements. Those statements have risks and uncertainties associated with them. I refer you to the risk factors disclosed in our public filings, including our most recent 10-K. So things are good at Micron right now. Business conditions have continued to improve since our September earnings call. And we're executing well on all fronts. Technology, which you'll hear about today from Scott and products from Scott as well. And operationally, we're doing very well. Data center demand, which includes a lot of our highest value products is especially strong, but also other markets -- all other markets are also healthy. We expect growing AI demand to drive a multiyear data center build-out globally. We've seen much more supply-demand tightness than we expected at the time of our earnings call comments. And that is allowing us to drive robust pricing trends, and we're doing that across markets. Now we project tightness to continue beyond 2026, and that's due to both supply and demand factors. Consequently, customers have approached us about entering into multiyear contracts. Now on contracts, as we noted in our fourth quarter earnings call, we said we expected to close HBM negotiations in the coming months. Today, we are pleased to announce that our HBM supply is fully contracted for calendar '26, and that's both for HBM3E and HBM4. On supply, in our assessment, industry supply response is limited by clean room space availability in the near term to address all the demand opportunities that we have in front of us. Inventories are very lean. DRAM, as we've talked about before, is below our target levels. NAND is improving. And by the end of the year, we expect it to be near target levels. Near term, we are growing supply through the ramp of our node transitions of 1 beta and gamma on DRAM through existing clean room capacity. We're also engaging in a lot of productivity gains and optimization of that capacity. Also, construction is ongoing for new clean room capacity for future supply requirements. Now ultimately, our success is built on a foundation of strong execution that leverages our leadership technology and product position. That's the focus today. And I'm very pleased to be joined today by Scott DeBoer, Micron's CTO, where he can provide his views on Micron's technology and product position. Scott?

Great. Thanks, Mark, and thanks for having us here today. I'm really mostly going to focus on our position and where our products sit today. And Micron is in the strongest position in history in terms of our technology and products in the history of the company. We have product leadership in both DRAM and NAND, and that's really enabled by a core innovation engine that we have at Micron, a global technology team and a strong partnership with our manufacturing operations. that has led to both consistent -- 4 consistent nodes of technology leadership on the DRAM side and the NAND side, but also sequentially improving yield ramps on every one of those nodes through the partnership with our technology teams and our manufacturing teams. So it's really something that is a strength of Micron is our ability to generate new technology in a very efficient and quick manner and then to ramp it fast into volume manufacturing. We've enabled a number of nodes, as I mentioned. Right now, we have our proven 1-beta technology in high-volume technology and ramped already. We also have our 1-gamma node, which is now at mature yield and ramped on some products and it's continuing to ramp on more products through the rest of the next 2 years. So really strong position with those 2 nodes, both in mature yields and ramping faster. And then our next nodes, our 1-delta and our 1-epsilon nodes are both planar DRAM nodes. And those are ramping over the next several years and will be the focus both of our high-density products, our LP memory and our DDR memory as well as eventually being part of the HBM landscape. Longer term, we are focused on true 3D DRAM beyond those planar nodes. And just like we've done in the past with other technologies, our focus is on bringing that technology to market at the right time when we believe it's cost effective and brings a competitive advantage. So I think we're in a very strong position with our 3D technology on the DRAM side, and we look forward to bringing that in at the right time. We haven't picked an exact year yet because it's going to depend on how technology evolves on the planar side and when we reach both cost and performance crossovers. On the DRAM product side, we've had a very strong year with our -- in particular with our LP and our DDR products. And both of those now on our 1-beta node and our 1-gamma node are ramping, and they're a bit different in how you think of them now. Of course, both of them have significant aspects that are commodity and the fact that we have, in that sense, leading-edge technology, power and performance that differentiate themselves in terms of just the pure performance in the market. We also, though, have significant parts of those that are more differentiated. So even those products today for us, we don't think of as products without differentiation, and that's visible in the fact that our LPDDR5 DRAM was the leading data center LP in the market and sole sourced for a big part of last year. On the rest of the product front, we've focused significantly from a customer engagement point of view. So our customer partnerships over the past year or 2 years have been a strong focus because all these products, whether it's the LP that is with sole source that I mentioned or our HBM product that I'll talk more about in a couple of seconds, are really results of long-term customer partnerships to build in features that our customers want and foresee and take multiple years to build in. And that's really the story of our success on our DRAM products and our NAND products and for sure, on our HBM products. So following our HBM3E, which we've talked about publicly for several quarters and really was a ramp from 0 to a substantial market share position based on excellent execution, business strategy and the best product in the market. We're now in a position with an HBM4 product that we also think will be the standard setter for the future benchmark for HBM4 performance. And a couple of things on the HBM front, and I'm sure there'll be some questions on HBM4. We have started with a product that is internally developed and designed to run all on Micron silicon. And the strategy around that was to focus on utilizing the core technology capability that we knew we needed to have a much higher performance HBM4 than the market was calling for a year or so ago. And by keeping that internal, we were able to utilize Micron's strength that metallization technology, the right kind of CMOS for memory in an HBM to optimize the product and the ability to design a much higher performance HBM4 product than the market was asking for a year ago, and it's aligned or still above what the capability is that the customers are asking for today. So all that led to us to design a product that is going to demonstrate very high yield above 11 gigabits per second and basically was designed to operate in that kind of space. So even though the market was calling for something almost half of that, we anticipated that we needed to be in a stronger position. And that really did lead to our strategy for staying internal on the CMOS for designing with microns metallization schemes, which are much more aligned to optimization for memory and to position us with a product that is going to set the benchmark both for performance without any design revisions and also to have the leading power that we've had and enjoyed on HBM3E continuing into the HBM4 space. And then last, on the NAND side, we've had, again, multiple generations of product -- or technology leadership on NAND. And with our Gen9 NAND right now in high-volume manufacturing and ramping across new products as the leading node in the industry and really positioned ourselves with now a focus on data center SSD in particular. And as you've seen, our growth on data center SSDs has been very substantial over the past year. And we positioned ourselves with high-density data center SSDs based on this Gen9 technology and 4-bit per cell or QLC technology leading the industry to really build our product portfolio on the NAND side at a different level than we've been before. So over the past year, we've had significant success on PCIe Gen5, high-density data center drives. And now we're leading the industry with the first products coming out on PCIe Gen6. And this is the first time Micron's led a protocol change like that with PCIe Gen6 data center SSD drives, and those are getting great uptake and customer interest and really the first time that Micron has been there in that kind of position. So net, I think of my message to start with is the technology and products position is exceptional right now at Micron and really well positioned for the exciting opportunity in front of us, certainly focused on data center, but also across mobile, automotive and other spaces where our products are recognized as leading.

Great. Thanks for those comments, Mark and Scott. Scott, definitely, I do have some questions for you on the technology side. But I'm going to start off with what Mark just -- Mark's update just now. So I guess this is what, tenth quarter of the up cycle in DRAM, Mark. Historically, these cycles have lasted anywhere from 8 to 10 quarters. So I guess for an average memory investor, this looks like we are closer to the peak. But based on what you just said, you seem to have visibility that the industry tightness will probably continue through all of next year, which I think makes sense given how strong the GenAI has been. So talk to us about when you talk about allocations, especially customers asking for longer-term contracts, especially on the commodity DRAM side. So how are you approaching your allocation? And when you talk about longer-term contracts on the DDR side, are customers willing to commit to both volume and pricing? Or are we just talking volume commitments here?

Yes. So let me step back. The industry and Micron took decisive actions during the downturn to rightsize the capacity in the industry. And we're very disciplined in our CapEx investment, manage our inventories very carefully, and delayed node transitions and the wafer capacity at Micron and the industry came down. And so now we have a situation where the AI demand is very strong. And we'll talk today about reasons that AI drives higher performance product requirements. And then you have one of those in particular, high-bandwidth memory consumes a lot of silicon. So now that the industry -- or now that the demand has kicked in on AI and has broadened out to other parts of the market, the market has gotten very tight, and especially on DRAM, but is improving pretty dramatically on NAND. So we are -- in this past year, we worked our inventory levels down and supply was provided by inventories that have been built up in the downturn. And this year, we are focused on node transitions to give us the supply we need. But overall, the supply that's going to come into the market, we believe, in now and beyond '26 is going to be inadequate to supply the -- all the opportunities in the market. And then we have greenfield capacity that's needed for ourselves and the industry, which you see coming on in '27 and beyond. We are in discussions with customers. We've been approached about multiyear agreements. Supply assurance is important for them. And it's good for our visibility on investing the capital and technology that we need to support our customers. Those discussions are underway. We're not going to disclose price and volume commitments, but they are multiyear, and that is an indication of the importance of memory and the importance of us providing supply assurance. Now what I will say is we're extremely disciplined in our investments, be it what we're choosing to focus on our technology and product development. And then importantly, the rate and pace of our capital spend. And we, of course, take in demand signals, and we very regularly process those and run those through our capital planning models. And then we use external sources. We use a number of methods to get a good read on what sort of supply we need to bring on and win. Ultimately, for Micron, as we've said, we want to achieve stable bit share and then deploy our bits to the more valuable parts of the market, which Scott's team and Sumit's and the broader Micron effort around getting the right technology and the right products and executing very well in Manish's team. I would just say that given the tightness in the market, given the duration of that tightness, which we now believe will be beyond '26, and given the customer interest in long-term or multiyear agreements, our CapEx that we had provided -- our CapEx number that we had provided, which was -- we had provided a run rate of about $18 billion for the year. There's going to be pressure on that to come up this year. So I would expect at our earnings call, which is about a month from now, we would probably -- we will bring that up and update you at that time.

Got it. And then your comment about customers asking for multiyear contracts. Are you seeing that across all end markets? Or is it specific to data center where the demand seems the strongest?

We're seeing it broadly.

Okay. Got it. And then, Mark, any time pricing gets to these levels, we hear concerns about, especially in the consumer markets, the bill of materials, the de-specing, -- are you seeing any of that, any of the behavior from customers as to, okay, I have 12 gig in my phone, either I have to raise prices or going to scale back on the memory content. Do you see that -- I mean, historically, have you seen that in terms of de-specing? Or do you think it's something we should expect in this cycle given how strong the pricing is?

I can't say that we've seen that yet. I mean it's -- I can't say that it's an elastic thing. I mean, the performance of the devices and especially as we get into greater applications with AI, I think it emphasizes the importance of DRAM. And in the case of our focused data center portfolio on NAND, the high-performance aspects of NAND. But as your question on consumer devices, we see DRAM content increasing. And we think that's a trend that will continue.

Got it. Got it. And then the other question we get this time in the cycle is about peak margins. If I look at what you reported last quarter, what you guided to 51.5% gross margins, I think back in 2018 cycle, you peaked at like 60% plus. So again, it's difficult to compare cycle to cycle. But given that HBM is a bigger portion of your mix, I mean, it was 0 essentially in 2018, is there any reason artificially for us to think that margins can't go back to that previous peaks? We would -- I would argue that they should actually be better because of the HBM mix. So how should we think about as we -- given the tightness comments that you made?

No, I think you're -- generally the drivers you're thinking about the right way. We're not going to provide guidance on margins on second quarter or beyond. We have said that these trends will be positive. So today, I talked about how tight the market is and how that's driving pricing. We are also operating very well. So our cost performance is good. Our mix of products, as you point out, has really been a wonderful story for Micron. I mean if you look at -- as we said on the earnings call, if you look at the high-capacity DIMMs, the HBM and the low-power DRAM in server, that was about $10 billion in our '25. And material part of our business, which had been back in '22, '23, it was near 0. So -- and then you add on top of that, the high-performance SSDs, you're in low teens billion dollars of business for just those -- that group of 4. And it's premium products, it's better margin generally. And in -- as we look at -- look out, that will grow on an absolute basis, of course, substantially, and it will grow on a percent potentially of the mix. So these things will -- these mix effects, we believe, given our positioning on technology and products, we believe this mix effect will continue to help us. So we've said before that we think that the second quarter can be stronger than the first quarter on a margin. And even with the first quarter conditions improving, we do still believe that the second quarter will produce better margins in the first quarter.

That's great. Then switching gears to HBM, and we know that the demand is very strong, and you just told us you're pretty much sold out for next year. But there has been this persistent debate about your road map, Scott, you talked about you're having a very strong product portfolio. So maybe help us understand in HBM3E, you said you had a product leadership. Do you see that leadership, whether it's performance power -- I mean, transferring into HBM4 as well as you look out to the road map? And then maybe for Mark, as we look out to the HBM shipments over the next, I guess, year or so, when should we expect the crossover to happen between HBM4 and HBM3? And do you think Micron will follow the -- whenever the industry crossover happens, should we expect Micron to all kind of pretty much follow that same time line?

Okay. So we do expect our performance advantage to continue on HBM4. And from the very first concepts of how we created a better HBM3 product, we really look forward on HBM4 to ensuring that it was a sustainable advantage that we had. And so we have a combination of the process technology that we use to make our HBM, our advanced metallization sequence that we use on both the base die and on the DRAM chip that is different than others. And also then some very key IP on our design capability. And we do design our HBM differently than the rest of the industry. And of course, those things eventually normalize. But we have a road map of different design elements that we bring in on HBM4 and then we'll again bring in differently on HBM4E that we believe will continue to enhance our capability, both from a power and a performance point of view. And on HBM4, in particular, the challenge really was looking ahead and making a decision to design the product at a much higher capability level than the industry was calling for a year or more ago when this product was in design. And we'll do the same thing on the future generations. But for HBM4 -- for 4 specifically, we do believe we'll have the industry's best performance on that product out of the chute at high yields. And we think that the power advantage, of course, has to be proven. But like HBM3E, early indications from our customers are that we're in a very solid position.

Yes. There seems to be this concern that because you're doing your base die in-house. You're not using TSMC at least for this generation or because you're still on 1-beta that somehow, for some reason, you have a disadvantage. So how do you kind of answer -- address those concerns?

Well, we're confident in the choices we made. To me, it feels a little bit like where I was sitting 4 years ago when I was describing why it made no sense to put 5 EUV levels in a DRAM chip at that time because the technology wasn't ready, and it would be an inherent problem for anybody who did, it would slow down your technology road map. And at that time, we had a lot of public news about why Micron is in trouble because we're not putting a bunch of EUV into our process technology back then. So some of that is just living through it again. In this case, I'm even more confident that the decision on the base die was absolutely the right one. Our optimizing a logic chip for HBM for memory capability is actually quite different than optimizing a logic die for ASICs. And of course, the foundries will catch up and they will figure out the things that are most important in terms of power delivery and low-voltage operations specific to signaling for memory capability. But there's a lot that has to happen there. And we've announced we are using TSMC for HBM4E. And so we've been partnering with them over a period of time to make sure when we do introduce that, that it is optimized for building HBM and not just kind of a follow-on to the ASICs designs that have been needed for a long time. So I think that in particular, the base die is going to turn out to be a significant advantage. And yes, there's news, but a lot of that's generated in the -- a little self-interested in a specific country. But as I mentioned, absolutely the big part of the news, which is we need to redesign and we're not -- we have some technology gap. We have the best products. Right now, we've sampled the highest speeds on HBM4 in the industry. We've given our customers HBM samples over 11 gigabits per second already. We are actually the only company that can test those right now because test capability at those speeds is also important. And part of our design is a built-in self-test regime that allows us to actually test and validate our material up at that speed. So if anything, I'm more bullish on our position technology-wise, independent of all the articles that I read also, on our HBM4 position even than we were when we were looking at our HBM3E, and we felt like our product was going to be good, but we never ramped it before. So in reality, in the last 18 months, we've gone from 0 to the market share we've talked about in terms of HBM. Our customers know how good our product is on HBM3E. So they have more confidence in us now relative to our ability to execute. And when we put a technology out on HBM4 and we talk about it, we have an even greater level of confidence now that we'll be able to execute to that.

Yes. That's, I think, a very detailed answer. So I just want to go back to Mark. Given your comment that you're sold out for next year and that also includes HBM4, should we assume that all the qualifications are done and we're ready to ship soon? Is that what, I guess, your comments suggest or qualifications still need to be, I guess, completed?

Yes. I don't think the systems aren't ready for HBM4 yet, right? So qualification, HBM is different in how we work with our customers. It's not just that our part works, it has to work in their system. So truly, we go through a number of milestones. We've hit all those so far in HBM4, but the next milestones do depend on the system optimization that we partner with each of our customers to make sure that, that -- and we don't call it qualified until it's actually in their system and working. And since most of those systems are not yet available, nobody can actually be qualified yet.

And maybe just to build on this point, tying it back to our capital expenditures. So this deep engagement with customers and this confidence that we have in our product obviously helps us in gaining confidence in the investments we're making for DRAM capacity because our DRAM capacity now is heavily impacted or influenced by our view on HBM. So we've got a very good view on where we stand on HBM, the increased use of HBM and systems and how that's going to determine supply requirements and our -- what sort of investments we need to make. I had mentioned earlier that based on this HBM and also a number of other products and the overall tightness in the market that our CapEx -- we're going to bring that up likely in the earnings call. But that doesn't necessarily mean that the capital intensity is increasing because as defined by as a percent of sales, that may not. So we'll work through all that. But the point is the confidence that we have to invest on the basis of our technology and product position.

Great. And then going back to my previous question about when do you anticipate, I guess, that industry crossover from HBM4 crossing over from 3E. And then when it crosses over, do you think there's still going to be some demand for 3E or is this something, I guess, just like what we saw with HBM3 today? How do you see that playing out?

Well, we start shipping in second quarter, and it will -- the systems will move and it will ramp in the second half. I don't think we've given a specific crossover point. But it is a product that unlike standard products, we're not expecting a long tail in these products. I mean the technology moves very quickly and which is important that we sustain the capability we have and the deep engagement with customers. And then we're just careful about our planning. So we would expect it to move to HBM4 and largely be that. Of course, there'll still be some HBM3E for some time, but then we're on the 4 and then the 4E.

Sure, sure. And then last quarter, you kind of -- you've been gaining share for the last several quarters, and I think you pretty much got to your target, which is your DRAM share. And as we look forward, it looks like potentially there's opportunity for you to even pick up more share. I don't know if you have the capacity or supply. So how do you think about market share as we look out to the next, I guess, several quarters? And also, do you anticipate your HBM4 market share to be somewhat similar to what you had in 3? Or do you see opportunity to even improve on that?

I would just say that HBM for us is as we've talked about in the third quarter, we believe we achieved what we said we would do and achieve a share equivalent to our DRAM share. At this point, it's a large product line that we treat like other large product lines. And we look at what our customer requirements are and the bit trade-offs amongst all our businesses, and we make those decisions. So clearly, given the trade ratio, deploying more of our capacity to HBM means fewer bits for some other markets. And these are the trades we have to make. And again, we make those all the time with our products.

Got it. Got it. And then in terms of the profitability itself, it's been a great product for you, HBM. But you've been improving that cost structure and your margin structure also, I think just based on what you reported in your segment gross margins, we could see that there's consistent improvement in gross margins as well. As we go to HBM4, do you expect the profitability to be at this level? Or do you see that maybe getting even better? Or because it's early ramps, are there any, I guess, issues that we need to be aware of in early days as we ramp this new product?

I mean Scott can talk more about yield. It's yielding well at this state. And so we're, again, confident and positive about that. I would say that HBM is providing a lot of value to the system. And each generation has gotten more complex and is providing more value. And so -- and it's very difficult to build, and we have the highest performance product. So not only would we expect HBM to be through cycle, generally more profitable than the rest of the business. We would be expecting to get value for each generation that is providing more value to the system.

Maybe, Scott, you can talk about what drives that cost increase as we go from HBM3 to 4? What are some of the components?

Yes. So I mean, Mark talked about the trade ratio, an important piece. The trade ratio is how many wafers it takes us to build HBM cube bits versus DRAM. And one of the fundamental pieces is defined for all of us by spec is the die size on the HBM. So from that point of view, the HBM4 die size will drive some cost structure increase relative to just pure math on silicon size that we'll have to overcome in other ways. Now when we look and we've talked about this ramp on HBM, we look at kind of the evolution of Micron's capability here, we did come from not even having a manufacturing entity on HBM to the position we're at with market share today. So when you go from 0 up to that significant market share, you obviously become more efficient, you learn how to manufacture better and more cost effectively. So all through this last year, our costs have continued to get better on our HBM manufacturing because of that, because of the maturity of manufacturing and our design for manufacturing continues to get better. So when we went from HBM3E 8-high to mature yield, and then we went to 12-high. We did that -- we did the 12-high faster than we did the 8-high by a significant amount. And our expectation based on the status we're at right now in HBM4 is we'll again substantially improve the timing to mature yield for HBM4 versus what we have done previously on HBM3E. So based on that, we will keep improving our cost structure in HBM4, and that will be an offset to some of the inherent pieces around die size and other things. And we're very much focused on consistency of manufacturing flow between HBM3E and HBM4 and making sure wherever we can utilize the learning and the process improvements and the cost structure of HBM3E and translate that over to HBM4, we're absolutely taking advantage of that. So overall, I think our expectation is we continue to improve and have a very good cost structure in HBM4.

Got it. And you expect that to continue as we go into HBM4E as well, I guess?

Yes. I mean every -- HBM4E is a little more complicated because we have the mixture of more custom products and standard products. So it's, again, a new variant. Of course, we'll optimize our manufacturing costs around yield ramp and reutilization there also. But as I mentioned, we are bringing in TSMC as the base die on that. So that changes the cost structure by itself of the HBM cube. And ultimately, it will be a different kind of optimization with custom parts versus standard.

Right. Got it. Got it. So Scott, I want to switch gears a little bit to the content side of the HBM story. I mean NVIDIA gives us pretty good visibility into their road map. So we know what to expect from a content per GPU standpoint. And AMD also just recently gave us some additional details. But at the same time, as Gen AI workloads evolve from pretraining into, I guess, inferencing, and we hear new terminology like RAG and vector databases, et cetera. So I'm just curious, there seems to be a lot of interest for low-power DRAM outside of HBM as well. And in your cloud business, I think HBM is roughly half of what you reported and the remaining is probably low power and some other high-capacity DIMM revenue as well. So talk to us about how do you see that playing out as inferencing becomes a bigger and bigger portion of the workloads? Are we seeing, I guess, a new class of memory that we need? Or I just love to hear your thoughts on that.

Sure. So I think that will -- it will play out over the next few years. And I think the key thing to come back to is there is growing -- general and growing recognition of the importance of memory products in enabling all kinds of systems, AI systems in the data center, at the edge, memory is much more front and center of system architects thought pattern today than it ever has been in history. So it will evolve to be more optimized maybe for edge applications, for robotic situations, even for automotive and certainly for data center inferencing versus training. The thing that we look at is there's no question it's going to be memory focused on how you solve those problems across that space. And absolutely, LP is a key part of enabling that, just like HBM. We see growth both in terms of HBM content and LP content going forward to support this. And we need to be positioned for subtle changes in that, and that's what we're focused on. And a piece of that is making sure our LP products are differentiated just like our HBM products are. And I think that's part of what we've demonstrated over the past year. And we'll continue working with customers to make sure we have absolutely the best low-power DRAM products for them, whether it's LP6 or some modifications of that in the future and certainly LP5 in the near term, making sure those products are solid. So overall, we think we can shift around to meet the different applications, and there's going to be growth across all of them.

Got it. I think maybe also on this issue of inferencing the longer context windows, the larger models, the deep reasoning, I mean, this is all also bleeding over into the need for higher performance SSDs. So as you know, we've focused our NAND business on that. And we're seeing that drive requirements for faster performance SSDs as we need more access to the data faster. And then the higher capacity SSDs, just so that more of that data is warm and can be used in the inferencing process.

Right. Actually, that's a great segue to my next question, Mark. So if you look at the SSD market, you have done quite well in this market. But not too long ago, we were talking about excess inventory in this market. And suddenly, demand has picked up quite a bit. It looks like there are a couple of debates. One is that the AI evolution of the workloads is what's driving this demand. And then the other argument is that, it's not necessarily because of AI, but it's because cyclical reasons and also the tightness on the HDD side. So I'd love to hear your thoughts on what do you think is driving the near-term demand and how sustainable this is?

We think it's sustainable for the reasons mentioned. There are a number of factors that inferencing is going to drive access to this vast pool of data that SSDs can perform a better job of accessing that data. So we think that's the fundamental driver now. And that over time, we see right now, as we sit here today, our inventory levels in NAND, we expect to be near target by the end of our fiscal '26. So we're going to see those inventory levels continue to decline, particularly in the back half. And that market, just like DRAM, the pricing has gone up substantially. So that's -- I'd say the principal driver is use cases, and that's good because those are durable drivers. Certainly, in the short term here, there's been the hard drive market is tight. So there's been some -- probably some demand related to that. That is a factor long term. I think the continued replacement of HDD by SSD is a long-term value proposition that we have to offer. But in the near term, I think it's driven more by use case.

Got it. We have a few more minutes left. I want to see if there are any questions from the audience. We have a couple of mics...

Okay. Maybe just a quick comment here on -- while we have some time and if there are no questions. I did want to draw folks' attention to the cash flow in the business is improving. We talked about a significant increase in cash flow, and we're seeing that. We have actively worked down our debt, technology and products and that and reinvestment in the business are high priorities, but the balance sheet is also a high priority. So our debt has gone down from what was approaching $16 billion -- has gone down below $12 billion at this point. We took out some debt through the course of this quarter. And we expect to be net cash here in the near term. I also want to point out that under the -- as far as it relates to capital return, we were authorized under the CHIPS definitive agreement to do a repurchase of $300 million, and we completed that in this quarter as well.

So maybe a couple more questions, one for Scott. Scott, you kind of mentioned Moore's Law and potential 3D structures as we look out to the next few years. How do you see -- I mean, I guess, how much runway do you think does DRAM industry has in terms of kind of using the 2D structures? When do you think we need to kind of consider going to 3D structures? And then as we go through this transition, do you think it's going to be somewhat similar to what we saw with the NAND industry? Or is it going to be different? And also maybe and potentially you can comment on the impact on capital intensity as we go through the transition.

Yes. So there's a lot there. I think 2 things. DRAM is a very 3D-oriented structure to start with. So when you -- we don't refer to it as Moore's Law on the DRAM side. Moore's Law is dead on the logic side. Scaling on the DRAM side is very much alive and has a really solid path. It's just a question of when we do certain things over the next decade, but there's no question that DRAM technology scales through an extended period of time and provides cost structure benefit as well as performance benefit. So the only question is when do we make changes in architecture like 3D or like other architectural changes to enable what we call planar, but it's very much a 3D type of DRAM. And in fact, lots of times in the press some of the planar architectures that we call planar are referred to as 3D also. There will be a number of transitions over the next several years. And most likely different companies will do them at different times. They'll largely be invisible, except in terms of the performance of the product and ultimately will be done when they provide a certain cost benefit. So relative to my true 3D DRAM comment, which is a drastic change in the memory architecture, I think different companies may approach this in different ways, but our approach is going to be when it makes sense versus what we can do next with planar. And when can we bring in a significant cost advantage. You're not going to see a drastic bit increase per wafer like you did on 3D NAND when we switched. I mean that was foundationally different and that they just totally changed the economics of how you built it. 3D DRAM, the way we see it will provide the same kind of cost down reductions node to node as the prior planar nodes. And it will be focused on matching array performance and increasing speed and power, reducing power, improving the performance on those. And I think those will be more incremental or more like in line with what we've done over the past several generations than anything like 3D NAND was. So actually, our 3D DRAM looks really good right now, and we will be in a position to implement it at the right time. It's just right now is not the right time.

Got it. And then Mark, on capital intensity, should we anticipate any changes as we go from, I guess, 2D to 3D?

I don't know.

It's another piece. It is a more complicated flow. Our objective is to keep it consistent on tooling more so than 3D NAND was when we switched. And I think we can do that largely because the DRAM flow is already very 3-dimensional and really complex. So we can -- we'll have much better utilization of tools and space on a 3D conversion than we did in NAND, where 3D NAND conversion really was a tear up of the whole fab starting over. So I think DRAM will be, again, more incremental, and we'll bring it in at a time that is not as disruptive. I also think it will come in not all at once, like a whole -- we'll have a planar path. Some of our products most likely will stay on planar and we'll bring in certain products on 3D that benefit from maybe a higher performance capability. So the capital transition also will be more muted in terms of we won't be changing out our entire supply base.

Makes sense. And then this is going to be my last question for both of you. You are at the forefront of enabling GenAI. I'm sure you're also implementing internally within your company. And so I just want to hear your thoughts on what you're doing at Micron as far as AI is concerned. What sort of productivity improvements are you seeing and general views on AI at the enterprise level?

I'll go quick because Mark has more to say on this one, but because he is our corporate champion on this. From a product development point of view, this is an area of intense focus for us and big opportunity in terms of the design space, how fast we can get new designs to market, how thoroughly we can verify those designs and make sure that they're right the first time, so we have more first silicon success on those designs. And we can shift our engineering exercises to higher-value tasks driving innovation as opposed to some of the more incremental things. So I think it's going to be a massive productivity improvement, in particular, on the design side, but also on the fab process side, the technology is very complex and the yield ramps are 100% dependent on how fast we can reduce variation. And that's one of the things that AI can really help us with is identifying sources of variation, finding the needle in the haystack that lets us reduce the process variation and get those yields up faster. So from a memory development and manufacturing point of view, it's massive.

And it's a really exciting time for Micron. Not only are we on a great trajectory for the business and AI's impact to our business, and we heard about that today. It's really exciting to be inside Micron. And there is great enthusiasm for GenAI and broader AI within Micron. Now fortunately, we stood up a smart manufacturing group, which does a lot of AI work, and that's been in place since 2019, and work was done over a decade ago, even before that group was formally established. So there's a history of getting -- applying AI technology in the company. And then with the emergence of GenAI, we've -- it's a priority internally to rework our workflows. We believe, have, if we look around the tech space and do our benchmarking, some of the highest adoption rates for Gen AI in our population of any company, we are applying it in coding and seeing 30%, 40% productivity there, and we're measuring many, many other areas within the company. We've added it to our company compensation goals. We have multiple targets related to Gen AI adoption. So we think this is a great technology for us to help manage the growth and complexity we have in our business. We believe, given our history of innovation and diverse culture and that we can very quickly adopt this and turn it into a competitive advantage for Micron.

Great. Thank you, Mark. Thank you, Scott. That's all the time we have. Thanks, everyone, for joining.