Micron Technology, Inc. ($MU)

Ladies and gentlemen, thank you for joining us, and welcome to Micron's Post-Earnings Analyst Call. [Operator Instructions] I will now hand the conference over to Satya Kumar, Investor Relations. Satya, please go ahead.

Yes. Thank you, and welcome to Micron Technology's Fiscal Second Quarter 2026 Post-Earnings Analyst Call. On the call with me today are Sumit Sadana, Micron's Chief Business Officer; Manish Bhatia, EVP of Global Operations; and Mark Murphy, our CFO. As a reminder, the matters we're discussing today include forward-looking statements regarding market demand and supply, market trends and drivers and our expected results and guidance, and other matters. These forward-looking statements are subject to risks and uncertainties that may cause actual results to differ materially from statements made today. We refer to the documents we have filed with the SEC, including our most recent Form 10-K and upcoming 10-Q for a discussion of risks that may affect our results. Although we believe that the expectations reflected in the forward-looking statements are reasonable, we cannot guarantee future results, levels of activity, performance and achievements. We are under no duty to update any of the forward-looking statements to conform these statements to actual results. Operator, we can now open the call up for Q&A.

[Operator Instructions] Your first question comes from the line of Melissa Weathers from Deutsche Bank.

I wanted to touch on the NAND side of things. So you guys are one of the only players to openly talk about greenfield capacity adds so far. And then there are some industry participants saying that the growth in NAND bits going forward can be served by node upgrades alone without the need for capacity adds. So can you talk about the decision to add greenfield capacity there? And what kind of trends you're seeing that gives you confidence to add that capacity?

Well, hi, Melissa, it's Manish. I can -- let me talk a little bit about the decision and then maybe Sumit can talk about the NAND demand trends. So we have said that -- in the past that while DRAM required greenfield wafer capacity, meaning new wafer start capacity growth to be able to meet the long-term demand trends that we saw and NAND was able to meet that with technology transitions. And I think we still see that NAND technology transitions are able to provide strong bit growth going forward. And so our decision here while, yes, reflecting confidence in the market demand outlook, which Sumit will talk about, was also driven by our continued space consumption for those technology transitions and for the next technology transitions that we'll be having in the future as well as our decision to locate more of our NAND R&D in Singapore, where it's closer to our manufacturing. And so those 2 drivers of additional cleanroom space combined with our outlook on the market were really what was behind our -- and our confidence, frankly, in our product portfolio, which I'm sure Sumit will talk about, were behind our decision to be able to break ground for our new NAND fab.

And just on that, it's not a greenfield site, right? I mean, you mentioned greenfield in your question, and this is an existing site where we are adding this additional cleanroom space that Manish said. So in terms of the demand side of the picture, I mean, we see very robust demand for NAND driven by growth in the data center. AI servers are using a huge amount of SSDs, high-capacity SSDs as well as high-performance SSDs. And this is working really well in our favor because our portfolio is doing exceptionally well. We are the first company in the world to have a Gen6 SSD in the market. And this has been something that works well with NVIDIA systems. We have seen tremendous demand for it that we are not able to completely even come close to meeting. And then even high-capacity SSDs, we have seen tremendous growth and tremendous demand, and we have come out with a slew of new products that we are looking really excitedly to continue to grow our share in the data center SSD space from a record level that we reached at the end of last calendar year, calendar 2025. We have been growing to record shares in the data center every year for the last 4 years. So we are very excited about that, and our supply is nowhere close to being able to meet the demand that we see for the foreseeable future. And so this expansion that Manish mentioned is something we will put to good use to continue to grow our business with obviously a focus on disciplined investing when it comes to our CapEx going forward.

And I should just add that even though we're breaking ground now, that capacity -- that cleanroom isn't expected to provide new boost to our capacity until the second half of 2028. So we do think that not just for us, but across the industry in NAND, cleanroom space over the medium term is still going to be a challenge for the industry, particularly as many in the industry have redirected some NAND cleanroom space towards DRAM.

Perfect. And then back to the DRAM side, I wanted to ask pretty candidly on the pricing that we're all seeing. Next year, we have line of sight to a couple of DRAM projects coming online, including you guys bringing on a couple of those new fabs. So how are you internally modeling the impact of that increase in supply going forward? Should we -- are you guys thinking pricing could come down next year or flatten out or decel? Just any color on how you guys are internally modeling the impact of all those new fabs coming online next year and the year after.

Maybe...

I'll answer the supply side or about the industry, and I think for the projects that we've talked about, which are the new Idaho Shell -- the DRAM side, Idaho Shell as well as the -- our ramp in the Powerchip Tongluo facility that we've now acquired. We've said that the supply impact from those will be towards our fiscal '28 in terms of being able to meet revenue shipments. And I think that's sort of largely the case for the industry in terms of new large cleanroom projects coming online. They will be really impacting later in '27 and really into '28 before you see meaningful supply, which is one of the reasons we have reiterated that we see the tight supply conditions to continue beyond '26.

Yes. I think we obviously don't provide any kind of price modeling or anything else type of discussion in that forward-looking way. But we have mentioned that the demand forecast that we get from our customers for 2026, for 2027 continue to escalate. And despite efforts that we are making to increase the amount of supply we can bring online to a modest extent in '26, but some more in '27 is not really making that much of a meaningful dent in the gap. There is a lot of demand growth driven by AI and also across different segments of the market, not just in the data center. And we expect tight conditions for the foreseeable future and certainly beyond 2026 and not giving any further guidance beyond that.

I would just -- the last thing I would add, Melissa, is that as Sumit mentioned, demand far exceeds supply. As Manish mentioned, these fabs talked about for us are coming on basically meaningfully in our fiscal '28. We always have the ability to modulate tool installs. So it's -- and that would be an option as well.

Your next question comes from the line of Aaron Rakers from Wells Fargo.

Yes. Can you hear me?

Yes. Yes, we can hear you, Aaron.

Yes, sorry guys. Perfect. Two questions, one and a quick follow-up. I guess, Mark, maybe just to help model-wise, I'm curious if you could maybe frame the current quarter guidance between DRAM and NAND. What may be assumptions you're making in this tight environment around your ability to ship bits? Will that grow sequentially? And any kind of separation between the 2 buckets, between DRAM and NAND?

I can provide you -- we don't break it out. I can provide you a little bit of color. I mean, let me just -- maybe you heard the 2Q report out. So both DRAM and NAND pricing was up strongly, NAND, even more than DRAM. Both also grew volume sequentially, NAND less than DRAM. In Q3, we would expect price to again be the largest factor. But we've given you enough on CY '26 industry bit shipments and that those industry bit shipments are constrained by supply. And then our supply, we expect to grow in line with the industry. If you do all that, you can sort of determine that there's -- yes, that you can assume some modest growth, volume growth in third quarter for both DRAM and NAND.

Okay. That's perfect. And then as a follow-up, maybe more architecturally in this environment, I'm curious, there's a tremendous amount of demand that you guys are seeing in servers. I think you've guided low teens unit growth this year. But one of the architecture things that I'm starting to hear a little bit more about is the idea of CXL and memory pooling and whether or not some of these hyperscalers could drive better efficiency of their memory architecture. I'm just curious at a high level what your guys' thoughts are on that. Obviously, there's a lot of other factors, MRDIMMs and just content growth in general. But any thoughts on CXL and having an effect on DRAM?

Yes, Aaron, CXL is likely to be experimented with -- at some of our customers. And they are -- different companies are in different phases of either assessing it or figuring out what they want to do with it, if anything. And we are certainly going to have our memory be available in those CXL configurations. Now my feeling is that the demand is in such a robust place in terms of gap between supply and demand being so significant that any and every available opportunity that can be used and deployed at scale is going to be something that customers will likely try if it is architecturally feasible and can be made to work with their software and their systems. There is a lot of work to be done to bring these solutions to a large deployment scale, and it's not easy, and it's not something that can be pervasively used. There are lots of technical limitations as well. But I'm sure those experimentations will continue and some limited deployments will be done in order to test it out and see how it performs.

Your next question comes from the line of Jim Schneider from Goldman Sachs.

Jim, can you hear us? Do you have a question? Operator, why don't we go to the next?

It seems like we might be having some technical difficulties. [Operator Instructions] In the meantime, we will move on to Atif Malik from Citi.

On the call, you mentioned that the non-HBM margins are higher than HBM. So my question is really on allocation. How are you guys thinking about allocation? And the reason I'm asking that question is because one of your Korean peers has been reported to be raising pricing by 100%. And you had pricing mid-60s for DRAM and I understand there's always delta across different memory makers. So are you leaving money on the table by not pivoting more towards non-HBM? So just your understanding on HBM versus non-HBM allocation.

Yes. I mean, we definitely go through different quarters and these relationships between different products can change over time. There is no doubt that HBM pricing was set for a large portion of the shipments in 2026 calendar year, late in calendar '25, as is generally the case with HBM, where the pricing is determined sometime before the start of the calendar year. And that kind of a model is a good model to have. It provides good stability and visibility into the business. The pricing that we negotiate has very robust ROI and profitability, and we feel good about that business. Of course, the upsides that we have been able to create from our operations team and supply chain teams on the HBM, we have been able to sell those for even more robust levels of pricing as those upside volumes have materialized. With that said, we don't view a lot of these HBM versus non-HBM allocations to be that tactically focused. These are strategic things that we have to do in terms of providing our customers with match sets of products so they can build AI systems. You cannot ship DDR5 into an AI system that doesn't have enough HBM in it and vice versa. So there is a natural level of balance that is needed for the market to be appropriately having match sets of products to be able to ship these at scale. So with that in mind, certainly, the HBM market -- margins are good and the non-HBM margins, of course, have become even higher. But then this is not just a data center issue, the non-HBM margins outside the data center, meaning the DRAM margins outside the data center are also exceptionally robust. So we don't tend to, like, just jerk around the allocations to different customers in different segments just based on where the pricing is. We have a goal of working with our customers to meet their business needs, and we do that with an intent of helping them meet their business goals as well.

And as my follow-up, NVIDIA talked about Groq chips could be, like, 25% of the ultrafast inferencing market. And my understanding is currently, these LPU chips have embedded SRAM at one of the Korean makers. And in the future, if these chips move to a Taiwanese foundry, will the SRAM be embedded on that chip? Or can the SRAM be stand-alone and somehow bonded to the processor?

There are definitely on-chip SRAM approaches that are currently in use. And there is -- I have seen some talk about bonded SRAM, et cetera. I would not want to comment on what directions our customers would go in terms of how they work with the SRAM over time. My main focus -- our main focus here is that we look at all of these systems as being very well balanced in the way they are evolving. So these SRAM-based systems complement in small users, the larger systems that are being utilized and deployed at scale like the Vera Rubin, for example, and other similar systems that are based on ASIC accelerators. And we see these as continuing to move the ball forward in terms of making these systems more balanced, more efficient. And the DRAM usage in these systems continues to grow over time and has gotten to levels which, of course, we don't have adequate supply for. But over time, we continue to focus on the growth of these average capacities, the growth of the high-performance tiers and even the growth of LP DRAM in these AI systems, all of which are really big positives for us on the DRAM side of the business.

Your next question comes from the line of Vijay Rakesh from Mizuho.

Just a quick question on the NAND side. I know you mentioned SSD NAND seeing a pretty strong uptick. Just wondering how to size it with the KV cache demand. What's the mix of SSD -- AI data center SSD, I should say AI data center NAND if we look out for calendar '26 versus last year, if you could -- if there's a way to kind of size that pickup in demand from KV cache? And I have a follow-up.

When we made our prior forecast of the extent of growth we are likely to see in the data center space, it did include a view that KV cache would be a meaningful driver. And as that has become more in focus in terms of how the -- how our customers have been wanting to deploy their SSDs, definitely, it has increased our view of the extent of demand coming from KV cache-related applications. And that has continued to cause our view of the total market opportunity to continue to grow. I'll just remind us that something we have said the last time also, which is that we have also seen a significant uptick in demand for data center SSDs coming from shortages in HDDs. And we continue to see those shortages for the foreseeable future. That has been another driver. So when we put all of these together, the NAND market is significantly undersupplied to the demand in the data center, and that demand continues to escalate in part driven by KV cache, but also driven by just the insatiable appetite that these AI servers have to have fast storage capability available as these systems get deployed more and more. And so the outlook is really strong. And as we have mentioned earlier, our portfolio is incredibly well positioned to continue to gain share in that space, including the KV cache applications, by the way, yes.

Got it. And just a follow-up on that. As you look at the CapEx, obviously, last couple of years, CapEx to NAND has been lighter. But as you look at the mix of CapEx for '26, '27 with the CapEx numbers that went up, how would you look at the mix of DRAM to NAND CapEx given both DRAM demand is up, but you're also seeing a spike in NAND?

Yes. Maybe I'll start. So Vijay, the CapEx is still going to be dominated by DRAM and HBM additions. Yes, this FY '26, we increased our outlook on CapEx to over $25 billion, which was up from the $20 billion we did on the last earnings call. The updated investment reflects this investment in the Tongluo fab, which we communicated in February conference and increases in U.S. expansion. Again, it's DRAM and HBM driven, including the increase. Today, we also provided more detail on construction, which we've said in the past was becoming a more material part of the build-out, obviously, because we need greenfield capacity. On the December call, we said expect that FY '25, '26 construction would double. And so we expect FY '26 construction to be mid- to high single-digit billions, net. When I talk CapEx, we're talking net CapEx. Now as we look out to '27, we did say that we project approximately a $10 billion incremental construction cost and also for equipment spend to increase. Now that '27 spend, NAND will begin to increase, but it will still be a much smaller portion of the spend compared to DRAM.

Your next question comes from the line of Karl Ackerman from BNP Paribas.

Can you hear me?

Yes.

This is Sam Feldman on for Karl Ackerman. So we've seen continuous HBM content uplifts with each new generation of XPUs. With the growing trade ratio of HBM, keeping the DRAM market tight and increasing AI memory requirements in the form of server DRAM, LPDDR and SRAM, do you think HBM will continue to see such large content uplifts with each new generation of processors? Or do you expect the content per XPU to eventually plateau?

Well, we are not going to make long-term projections of where these average capacities will go because those are things that our customers are going to decide as their architectures evolve. What I will say is that if you look at the direction in which AI is trending and the types of things that are creating value in the AI domain for customers, they go towards more reasoning capability and more longer context windows and the ability to do more with agents and multi-agent orchestration, all of these things are really requiring more DRAM capacity and more DRAM bandwidth. And when it comes to delivering on that kind of bandwidth and being able to really optimize the system for all of the different stages of prefill and decode and different aspects of training as well as inference, you really come to the conclusion that these accelerators, whether it's GPUs or ASICs, do require increasing amount of HBM and increasing amount of DDR5 or LP5 capacity with time. And that's the trend that we have seen thus far, and you're also seeing in the announced architectures that have been publicized. And so we feel certainly that the trend is -- has been clear. And not only has the trend been clear, but that the way our customers' customers, meaning the end customers derive value out of the AI system and AI applications is very much connected and consistent with the trend of needing more DRAM capacity and bandwidth, which HBM delivers so effectively at very good efficient power consumption levels. And so that's the reason we have said in the past that memory is becoming a strategic asset in the AI era. It is precisely -- one of the important drivers of that is precisely this trend that you really can't have a high-performing AI architecture or hardware infrastructure without all of those capabilities that DRAM and HBM bring to the table. So that's something we feel pretty good about in terms of where the market is heading.

Your next question comes from the line of Chris Caso from Wolfe Research.

I'm wondering if you care to update your view of long-term bit growth for both DRAM and NAND. And I know you mentioned in the call, low 20s in DRAM, 20% for NAND, but those are obviously supply-constrained numbers. And maybe can you speak to the extent to which you think long-term bit growth is increasing as a result of all the things that we've been talking about.

Yes. I mean, we haven't provided a new long-term bit growth number. I think your -- you have seen that in the past, we have spoken about high teens, mid- to high teens type of ranges for DRAM. And yet you have seen that last year, this year, our forecasts have been more robust than those levels. And we continue to feel like we are in an extended space of robust industry demand that obviously, due to HBM being part of these numbers with its trade ratio is just stressing the entire industries and certainly our capabilities to be able to meet those demand numbers. So you're right. I mean, these numbers, at least in the foreseeable future are all supply limited numbers rather than the actual level -- true level of demand. So yes, I mean, that's sort of the environment we are in. We do expect that next year, again, we will have a fairly robust level of growth in calendar '27. But yes, we are not providing a long-term number beyond that commentary.

Got it. As a follow-up, obviously, there's been a lot of discussion about cleanroom constraints. And as you folks have pointed out, you're starting greenfield in Singapore, and it's not available till the end of '28. I guess, you probably can't speak for the industry, but for at least for Micron, at what point do you think you can get caught up with having enough cleanroom capacity that gives you some headroom for what the customers need? And then obviously, then it's going to take time to move the tools into that. But it goes to the sustainability of what's going on right now and the extent to which the cleanroom constraints are going to drive that sustainability.

Well, I think, Chris, the first part of the question is really around what Sumit had described in terms of the long-term demand, and we're continuing to evaluate all the different demand drivers and signals, including what all the announcements from GTC this week. In terms of the availability of cleanroom space for us and the industry, certainly, we have -- we see this relative constraint for all the major DRAM players to be there through this year and into next year with major meaningful improvement to cleanroom space availability only out into '28. But of course, as we think through our projects that we've announced go beyond that into '28, '29, '30, as we think -- as we talk about those projects, the time line for us to get to the point, as you mentioned, will be a function of the demand. And we will be nimble to be able to adjust equipment orders and equipment installations in order to stay aligned with whatever the demand is as it plays out over that time horizon.

I'll just add to what Manish said, which is that as we look at the demand side of the picture, and we have been engaged with our customers, several of them to be talking about these 5-year, multiyear SCA agreements. And in the context of that, of course, we are assessing their longer-term demand, their 5-year demand, for example. And we are assessing those against our own supply capabilities. And within that time frame, we are also seeing the emergence and growth of some really exciting new demand vectors, including things like robotics, which we expect to become a very major demand driver. And so when we put all of those things in the equation, we don't have a high confidence view yet as to when the supply will be able to catch up with demand because the escalation of demand from these various vectors is just very phenomenal. And so to answer the question as to when do we have a high confidence view that the supply will be able to catch up with demand, we don't really have a high confidence view yet as to when that would happen.

Your next question comes from the line of Srini Pajjuri from Royal Bank of Canada.

Mark, I want to ask the previous question slightly differently on CapEx. You talked about CapEx being -- for the construction being high single digits in '25, growing by $10 billion. So that suggests that roughly next year, half and half between construction and equipment. For the projects that you already announced, do you think '27 is the peak year for construction CapEx? If so, what is a normalized, I guess, mix between construction equipment through the cycles that you kind of think about?

Yes, Srini, we're not going to give any more breakdown than we've provided. It will be lumpy. Just to be clear, the $10 billion was in reference to '26 to '27. I wasn't quite -- you may have said that, but I just want to make sure it's clear. And we're, of course, going to modulate spend as we see demand and to maintain stable bit share. As Sumit mentioned, it's not clear if -- we're investing as we can, including this recent acquisition of the Tongluo fab to put capacity in, but it takes time and a lot of effort and we can't get meaningful bits until our fiscal '28. So beyond '27, of course, we're going to be very disciplined and '27 could be a -- CapEx could go down after '27, but we're not making that call at this point. We're just investing through this next several year period that we can try and get the supply we need for our customers.

Okay. Then maybe a follow-up. For the next few quarters, there were questions on gross margins. Obviously, you don't want to comment on pricing, but could you give us an idea how to think about any puts and takes on the cost side and also any mix dynamics that we should be aware of? And also, given the higher CapEx, how should we kind of think about depreciation over the next few quarters and any start-up costs that you anticipate?

Yes. So good questions. Maybe tackle first just on costs. We continue to execute really well on cost reductions, and Manish can comment some more. But as we've talked about FY '26, we have the benefit of the node transitions, 1-gamma in particular, on DRAM and G9 on NAND that are driving our bit growth. So we get a lot of efficiency there and cost downs and the 1-gamma, for example, replacing 1-alpha capacity. So that's good. And the spend control has been outstanding. So as busy as we are, as good as the numbers are, the discipline in the business is very good, including managing, by the way, all this geopolitical and the team has really been on top of that. There's no impact to our operations. So that's good. I think to your question on start-up costs as we bring on ID1. and then there's going to be some with Tongluo. And I made some comments, maybe it was 1 year, 1.5 years ago talking about start-up costs. And those comments are still applicable. And at the time, I said I think between 1 point and 2 points of cost. Now revenue was much lower at the time. So if we dollarize that, it's probably $100 million to $200 million per quarter starting in the next quarter or so and then continuing on through '27 and then it would come down off of that. So again, at these revenue margin levels, it's a much smaller impact of 50 basis points or less. And then I think was there -- what was the third question?

Depreciation cost.

Yes, Depreciation just depends on when production wafers are out and then useful life of those assets when they're put in service. And keep in mind, it's -- because this is greenfield, this is very long depreciation life. So that's important to keep in mind.

Just a couple of things I'll add. Mark actually had some comments on costs. I think you covered a lot of them. 1-gamma is going really well. Gen9 are both going well, and we had positive comments on those. The only other one I'll add is HBM. I think our HBM3E 12-high has continued to execute well as we've gone through the last -- gotten to high volume over the last couple of quarters. And we actually see HBM4, even though we're in the early stages of the ramp, having an even faster yield ramp than HBM3E 12-high. So our HBM cost structure, we feel very good about, both HBM3E as well as HBM4, and then providing improvements for us over previous periods. And then the other comment I'll just have, I think Mark referred to geopolitical, managing, I think, costs relative to geopolitical issues. I think he meant about the Middle East and some of the media reports regarding various input disruptions, but we don't see any supply risks and very, very minimal impact to cost at this time.

Yes. Maybe -- Srini, just maybe one housekeeping issue since we're covering sort of those issues now. On OpEx, we indicated that our OpEx would be ticking up on -- due to R&D costs. So in the fourth quarter, we would expect OpEx to be close to $1.6 billion, just given that part of it is the extra week, but part of it is just, again, this increase in R&D spend, which we think is completely appropriate to drive the technology, the increased value of memory, Micron's position and then just customer demand on specific projects. And then we would expect in '27 for that OpEx number to be over $1.6 billion, probably kind of a $1.7 billion run rate number and stabilize from there.

This concludes today's call. Thank you for attending. You may now disconnect.