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

Good afternoon. My name is Josh, and I will be your conference facilitator today. At this time, I would like to welcome everyone to Micron's Post Earnings Analyst Call. [Operator Instructions] It is now my pleasure to turn the floor over to your host, Farhan Ahmad, Vice President of Investor Relations you may begin your conference.

Thank you, and welcome to Micron Technologies Fiscal Third Quarter 2021 Sell-Side Analyst Callback. On the call with me today are Micron, EVP of Strategy and Chief Business Officer, Sumit Sadana; Micron, EVP of Operations, Manish Bhatia; and our CFO, Dave Zinsner. Today's call will be approximately 45 minutes in length. As a reminder, the matters we will be discussing today include forward-looking statements. These forward-looking statements are subject to risks and uncertainties that may cause actual results to differ materially from the statements made today. We refer you to the documents we filed with the SEC, specifically our most recent Form 10-K and 10-Q for a discussion of the risks that may affect our future results. With that, operator, you may open the call.

Our first question comes from Harlan Sur with JPMorgan.

On the confidence on supply tightness in DRAM into next year. I know that the team has been entering into a longer-term supply agreements with some of your customers, lead times have been extending, and I'm sure customers are more open to sharing their longer-term forecast as well. So how much of your calendar year 2022 is being -- viewed as being driven by order backlog or customer forecast visibility into next year versus some sort of top-down sort of macro view? And then just a quick follow-up question on the DRAM cost front looking into fiscal '22. I know there are a lot of moving pieces. But at the highest level, does the team, at least, expect the cost profile to be within your long-term range of down kind of mid- to high single-digit percentage points?

Okay. So this is Sumit. I can take the question related to the market. So we do go through a pretty rigorous process to assess demand and supply both. Certainly, we look at the tops-down models, macroeconomic approaches, industry environment and condition. We talk extensively to our customers, to our ecosystem partners. We obviously assess technology trends that are shaping the landscape. For example, as you know, there are upcoming launches of new x86 processor platforms that have significantly more attachment of DRAM. They tend to bend some of the demand curves when these type of transitions occur. So we take all of those into account, and we determine how things are proceeding. And we have, I think, a reasonably good track record based on what we have seen over the last few years. I think the approach that we take on the supply side is also similar. We do a fairly detailed analysis. And of course, there are certain things that change from time to time in terms of supply. This time around, we tried to highlight some of them, including the headwind that comes from bigger die sizes on DDR5 compared to DDR4 because of the on-chip ECC that used to be on the module now integrated onto the die itself. And that creates headwinds for supply growth. So when we back all of that together, the other thing to keep in mind is lot of ecosystem shortages across the semiconductor ecosystem for different types of parts, could be raw material for some of these devices, could be controllers, built-in foundries in some of the legacy nodes, it could be analog parts. A lot of different things are extremely tight. And we think that some of this tightness gets resolved over the course of multiple quarters extending well into calendar year '22. And as these things get resolved, it will open up pockets of demand that today cannot be met because of these non-memory shortages. So I think overall, when we pull together all of the demand view and the supply view, we feel that the industry environment will remain healthy for a good period of time, certainly through this calendar year extending into next year. And then we feel that the supply has been pretty disciplined on the DRAM side. We have been tracking the CapEx that I know you all look at as well. And because of the disciplined CapEx in DRAM, we feel that the overall supply growth will be in fairly good shape, consistent with the demand growth and should help to extend the healthy environment in the industry. Now I'll turn it over to Manish to talk about the cost.

Thanks, Sumit, and thanks for covering a little bit of it. One of the things that is impacting the bit growth also impacts DRAM costs for next year, which is a transition to some of these newer standards with DDR5, LP5 and obviously, graphics products, all having higher performance. But -- and higher performance for our customers and their end applications, but being a mix-driven headwind to the cost as well. In terms of what we see looking forward, the 1-alpha node for DRAM is ramping well, and we feel very good about both where that is right now, how it's going to ramp over the next several quarters and what it's going to provide us in terms of like-for-like kind of cost reductions in FY '22 and bit growth for FY '22. In terms of the cost reduction range that we've talked about before of mid- to high single digits. I think on the front-end side, we feel good that we'll be towards the higher end of that range. and that when we think about the mix of the business as well as some of the more expensive back-end costs, for example, DDR5 modules are more expensive than DDR4 modules given some of the changes in the [indiscernible] aspects for those modules, as we think about some of those things, we think that the total end maybe towards the lower end of that range. But we still have, as David mentioned on the other call, we're still working on exactly that, the mix of the products in next year's plan, and we'll give more color on that as we head to the next call.

Our next question comes from Aaron Rakers with Wells Fargo.

I wanted to go to the data center business and particularly the server business. You guys talked and you just mentioned a little bit ago about the expansion of memory in memory channels in the server CPU market. I'm just curious, are you alluding to things like Ice Lake and AMD's Milan processors today? Or I thought the commentary was a bit more weighted to the back half and even into calendar '22 where, I guess, the question is, are we starting to see indications that there's further expansion of memory channels on these CPU sockets, beyond the 8 channels that we currently see? And then also any quick update on HBM, just timing or materiality of when we should start to think about that ramp materializing?

Yes, those are 2 good questions. I will just answer them in that order. So the first part, yes, it is about 6 channels going to 8 channels and the broader deployment that will be occurring with these both internal MD platforms over the course of time. Also wanted to mention that upcoming processor platform launches are going to have dramatically higher core counts. And if you just attach the same amount of memory as before to a processor with higher core count, your bandwidth per core starts to decline, which then starts to affect application performance. And so servers that were originally configured with a certain amount of DRAM memory need higher amount of DRAM memory to be attached to that same server with a more higher core count CPU socket. So that's sort of the trend that drives double-digit increases in average capacities in terms of attach rate pushover. And that's what we are seeing as a trend that will continue to do well over the course of some of these platforms being deployed in 2022 from both AMD and Intel and gathering pace there. In terms of HBM. HBM2E was the first HBM product that we introduced to the market. We have made investments in that product. The product is being qualified by our customers. Our intent with HBM, first-generation HBM, which is our HBM2E product, is not for there to be a very material ramp in that revenue. It's more a vehicle for us to prove the technology, gain the confidence, qualify it, ramp it into volume. So a pretty complex product. And then rely on subsequent generations of HBM to drive material gains in that product line in terms of revenue ramp. So that still is the plan. And we are making good progress towards that.

Your next question comes from Mehdi Al Hosseini with SIG.

Just a quick follow-up to the previous one. I want to better understand how HBM demand would trend in the context of DDR5 adoption, could HBM actually cannibalize DDR5 server DRAM? Or do you feel confident enough that HBM is driven by incremental demand creation for AI that you won't have a cannibalizing effect on DDR5? And I have a follow-up for David.

Yes, I mean, I don't think of HBM as cannibalizing DDR5 as much as a complementary way of attaching very high bandwidth memory. If you think about different applications, some applications tend to be very latency sensitive, some tend to be very bandwidth limited. Several AI applications tend to require very high memory to CPU bandwidth, then they tend to use HBM. Having said that, pretty much most server applications in the data center today are getting impacted because these very powerful CPUs with massive numbers, of course, are unable to really get the data fit to them at a fast enough rate that they can keep all the CPU cores active and busy which is what is needed to be done in order to leverage the significant investments people make in these new CPU platforms. And that is what drives the DDR4 transition to DDR5 because you're dramatically increasing the bandwidth on the DDR bus. And hence, we believe that both of those things will occur, there will be a DDR5 transition in certain applications, which are highly bandwidth limited. Like some AI applications, there will be HBM usage. Having said that, HBM is today still a relatively modest market. It's about $1 billion TAM. It is going to grow faster than the rest of the DRAM market. But the DDR5 market will become massive very, very quickly and will very early in its ramp overwhelm the size of the HBM market and then keep going. So I don't view those as being competing with each other as much as being complementary in terms of how application workload is being addressed.

Great. Thanks for the detail. And a quick follow-up for David. I understand you don't want to provide FY '23 operating margin guide. And -- but I want to better understand how effective the management team is executing on the cost structure. Obviously, we have had a couple of many downturns. You have established effect that you're kind of a trough operating margin above 0. I believe, last time, it was around 10% operating margin. In the last up cycle which was extraordinary, operating margin exceeded 50%. And the prior up cycles, prior to 2018, your operating margin peaked in the 20% to 30%, and you're already above that. So help me understand. I know you're trying to be conservative, you're trying to be prudent, especially with the guide and not letting us get carried away. But can we establish the fact that maybe your operating margin could trend higher? It may not exceed the prior peak, but at least it could trend towards the 50%.

And just to be clear, you were mentioning something about FY '23, you're not looking for FY '23, right? Is that what you were talking about?

Yes, I was just trying to make sure that you're paying attention to my question. I would just applaud.

Yes, I was definitely paying attention. So I mean, the way I kind of look at it is what the peak is and what the trough is, is less important than what the entire cycle looks like in terms of profitability. And I think I mentioned on the call in the prepared remarks that we had EBITDA margins, I think, through that entire cycle of 50%. We had gross margins, I think, through the whole cycle of 40% and then ROIC was 20%. I feel like those are very good numbers, quite honestly. And so I think if we can achieve numbers that are roughly in line with that, plus or minus a bit, I feel really good about our performance. Obviously, some of it is the -- we're at the mercy of the market. But I do think that a lot of the things that we've done structurally to drive the relative profitability of the business have been pretty spectacular, quite honestly. And so now you know that we're in a place where costs are somewhat in line with the industry. We look at our cost to drive them somewhat in line with the industry. I think we've shown, particularly this past quarter that we could do a very good job in terms of managing our pricing across the DRAM and NAND markets. We're -- as you see, we're pretty efficient on our operating spend. So I'd like to think that we could continually, through cycles, maintain very good profitability in both EBITDA margin and gross and operating margin. How that shakes out on the model year-to-year, I'll let you guys all figure that out.

If I may, just quickly, a quick follow-up. If I just go with the 50%, 40% EBITDA and gross margin, that doesn't leave me a lot of upside from here. Maybe 3 points, if I just take what you reported for the last quarter and I think the trough of 6 quarters ago. The general question becomes how long is the cycle? And we don't know. So -- and this is where it gets a little bit tricky. And on top of everything, you haven't been aggressive with buyback. So what am I missing here?

What are you missing in terms of what, about the buybacks?

Well, there's only a 3-point upside to this margin guide EBITDA 50% and gross margin of 40%, if I were just to average what we have done so far in this up cycle. And then you may push back and say, "Well, Mehdi, the cycle could be 2-year cycle, I could run a business at 50% EBITDA and 40% gross margin for 2 years, not just 2 quarters." And this is where I'm having a hard time to think about sustainability of earning growth.

Well, I mean, I think we're not -- it's hard for me to make a prediction on what the high point is and what the low point is. All I want to do is kind of drive profitability through the cycle. And that is not to say that we're making any prediction about how long the cycle is or how high the cycle will get. We're just trying to drive the profitability through the entire period to generate a good return on invested capital. The buyback, I'll tell you it has nothing to do with any view on how long the cycle is. We have an opinion that over the long term, we can drive very good profitability. We have a good product road map. We have good technology leadership, excellent cost in terms of being competitive with others. So we feel very confident about those. And when you feel confident about those and think about where the stock price is today, of course, this is a good price. And we will buy back a fair amount of stock next quarter. We have that built into the plan. So that's -- that in no way is in any relationship to what we think about the cycle. As far as we can see it, which, of course, is always a bit hazy, we feel good that things will be tight into '22, and then we'll see what '22 plays out at. I think we've done -- shown that we can do a very good job managing through upturns and managing through downturns, we'll continue to do that.

Your next question comes from Tom O'Malley with Barclays.

Dave, this one's for you. You've done a really good job of bringing down OpEx as a percent of revenue over the past really 8, 9 quarters. But it's stepping up from a dollar's perspective from May into August. Can you talk about what's driving that step up there? And then what the profile from a spend perspective should look like as we go into the next fiscal year?

Yes, so we've been managing OpEx pretty effectively. It does -- it can be a little bit lumpy because there are the prequalification expenses that R&D incurs from time to time, depending on what activity is going around in terms of the development. The last couple of quarters were pretty light. This quarter, the fourth quarter, it's actually going to be a relatively heavy quarter of prequalification expenses. So that's part of it. I think also we had a little bit of a benefit in the third quarter from the sale of some assets. We took a gain on that. And so OpEx was a little bit artificially low in the third quarter because of that gain. It was probably about $15 million higher. So once you strip that piece out. I think longer term, when we're in the really upward part of the cycle, I think we're generally running low double digits or in -- I think in fiscal '18, we ran as low as high single digits. And when you're in the kind of lower part of the cycle, it kind of runs in the mid-teens and we'd like it to average somewhere in the low double digits to low teens over time. We want to drive, obviously, leverage as we're coming out of the cycle. But I think over time, we're not trying to drive a ton of leverage through operating expenses because the one thing that always differentiates you from competitors is your ability to invest in products. And we do want to make sure we're making the proper level of investment in terms of products that -- to obviously drive profitability and drive growth.

Helpful. Just a quick follow-up. On the NAND side, in the May quarter, you guys talked about low single-digit growth. But the last couple of quarters, you've talked about an environment in which things are stabilizing. You've sounded more positive really every single time we heard you talk on the NAND environment. Is there a reason why things slowed a bit in the May quarter? And should things pick back up there just given the fact that it seems like the demand needs are pretty robust from kind of all the end markets that you've talked about?

Yes, I can address that. So definitely, the semiconductor, the NAND industry overall gross margins got pretty challenged in that CQ1, CQ2 time frame. And part of that was driven by the fact that there was a substantial amount of CapEx that had been invested and we had articulated for a few quarters that we believe that the industry CapEx is at pretty high levels. And that's what ended up playing out in terms of the challenges in the gross margin profile. Of course, NAND has pretty high elasticity of demand. So as pricing fell in NAND, certainly, the demand picked up even more and has absorbed a lot of that supply. And things are in pretty healthy shape. We have posted ASP improvements, as we saw in our fiscal Q3 results. And certainly, we look forward to improvements in fiscal Q4 as well. There is a good deal of demand coming from multiple segments of the market. Some segments of the market remain undersupplied. And so as we look at strength in PCs, look at data center strength, look at game consoles ramping with SSDs instead of HDDs for the first time, both PlayStation and Xbox, and we look at a significant amount of strength coming from automotive and industrial markets, the environment is healthy on the demand side. We feel that as long as the CapEx remains at disciplined levels in the industry, the environment can remain healthy. Our intent has always been to grow our supply bit in line with the demand growth in the industry. And so certainly, we look forward to continuing to execute well on the NAND side. We are doing a lot more than just the supply/demand balance, of course, focusing on that. There's a lot of focus on our portfolio, the strength of our products, the mix of the portfolio. We spoke about record QLC bit mix. These QLC products have dramatically higher profitability for us compared to what TLC can drive. And TLC and QLC cost structure in NAND very, very competitive world-class cost structure. So we are feeling good about all of that. And looking ahead, we are looking forward to the introduction of our first vertically integrated platform data center NVMe SSD as well. So a lot of good things happening on that front. And we hope to certainly improve the ROI in that business going forward.

Our next question comes from Ambrish Srivastava with BMO.

Dave, I just wanted to make sure I was on the same page with everybody else, so just correct me if I'm wrong. I have heard you guys talk about cost down for fiscal '21 for DRAM mid-single digit. And I think Manish said mid to high. And then on the call, somebody said 10%, and you did not correct that. So what have you said? Could you just calibrate us on what have you publicly said about the cost down for...

Yes, I probably should have corrected. Although, I wasn't sure exactly what aspect they were talking about that was down 10%. We've said that our cost down -- originally when we came into the year, we said our cost down in DRAM would be kind of mid-single digits. I would say, over the course of the year, we've gotten a little bit more bullish about that, and it was certainly trending above that. It's still above that. But suffice it to say, some of this COVID mitigation expense, we didn't exactly have dialed in. And so that has been a little bit of a weight -- has weighed us down a little bit, but not terribly. I mean, we're still above the mid-single-digit number we thought we would be coming into it. Now we have -- the other thing we've said is that we think at least over the intermediate call the next several years, DRAM cost declines probably run in the mid- to high single digits as the industry goes. We think we would be in line with that over that time.

Thanks, David. And that's right. And what I had been talking about when I was with a longer-term kind of trend for us and the mid- to high single digits. And I think as we mentioned at the beginning of the year or maybe I mentioned on one of these callbacks last quarter, coming into this year, we made a strategic decision to invest more in 1-alpha, which would have a bigger cost benefit, really primary cost benefit in fiscal '22. And we actually reduced some of the CapEx on some of the other technology nodes, 1z in particular, and that is why our cost reduction for this year is maybe somewhat lower than what that range would indicate that we're targeting for the longer term.

Okay, okay. That's helpful because I don't miss -- I do miss a lot of stuff, but I thought I was pretty dialed into cost down. When I heard 10%, I said, huh I must have read wrong.

No, we never said we get to 10%.

Okay. I think that's important.

For the industry track kind of numbers.

And then what have you set for NAND as well for this year?

So for NAND, we said -- fairly early on, we said it would be in the kind of low double digits, which was kind of somewhat driven by mix and that's generally where it's been tracking. I would say, it's also maybe tracking a hair negative more recently as well when the dust settles for the same reasons. It was a function of this COVID medication expense.

Got it. Got it. So then my question is on the EUV side and the cap intensity going to mid-30s. What is the time frame? Does that go into -- keep staying in the mid-30s up until 2024? Or should we think about that as the peak here? What's the right way to think about it?

Yes. So as we said on the call, we're planning to introduce EUV on our 1-gamma node in 2024. And then thereafter in subsequent years 1-delta. So it will really take us until 1-delta node and beyond to be able to have a significant -- to make the whole EUV infrastructure be productive for us. So it will probably go towards that 1-delta, which would be after the 2024 time frame where we'll be making prepayments for tools, investing in the facilities, infrastructure for this equipment and getting it ramped up into production. And as I said, it will be an introduction of 1-gamma. 1-delta will be a deeper penetration, more layer count. And then so I would think it would be through that 1-delta, which would extend beyond 2024, and then we'll start to see hopefully some more improvement and reduction in some of the multipatterning that EUV will provide in 1-delta and beyond.

Got it. Manish, while I have you, what is the die size increase with DDR5 that you guys are expecting?

Yes, we're not giving that specifically. But I think one of the things, Sanjay gave a little color on is one of the drivers of that is the addition of on-chip ECC. So kind of used that as an estimate of one of the major drivers.

Your next question comes from Simon Wu with Bank of America.

Very quickly, any rough idea, the percentage of your PC [ DRAM ] at the 1-alpha node in the U.S., whether it can be more than 10% from the August end of quarter, was less than 10%. And the same question regarding the 176 layer NAND out of the total NAND. So we want to measure whether it can be more than 10% or less than 10% of the total PC NAND in DRAM and NAND for this current quarter.

I think we did say that we reached a meaningful level of production in FQ3 on both 1-alpha and 176 layer and that production, obviously, at the fab level in F Q3, Simon, would be -- would allow us to say a meaningful level of revenue generation from both of those 2 nodes in FQ1. And then as we look towards the end of the year, I think we did say in the prepared remarks that we would have 176 layer would be more than half of our NAND output by the end of the year. And then we said that between 1-alpha and 1z, combination of the 2 would be more than half of our output by the end of the year on the DRAM side as well.

And that's the end of the calendar year.

Calendar year, yes, sorry. Yes, calendar, yes.

Okay, sure. Yes, because I want to know the definition of the meaningful portion, whether it can be more than 10% or less than 10%. It's okay. Sure, sir. And then another, I think it's a little bit kind of not a concern, but I'm feeling like the 1-alpha node is very advanced technology. So that should be used to maybe 16 gigabits production rather than 8 gigabits or solid 2 gigabit. Sensing 176 layer, I think is pretty extensive, very advanced for the 256 gigabit mono chip production. I think your 176 layer should be used to maybe 1 gigabit level of the chip production. So the question is whether your 1-alpha node and the 176-layer 3D DRAM can be used effectively the larger, larger density because your OEM customers, to date prefer maybe smaller, lower density at the maybe 1z node or maybe 96 layers. So any color on this, how to maximize your new node use for the larger, larger density, which is much larger than the mainstream [ chip-a ] gigabit or 256 gigabit. So what are your OEM customers feedback or reaction to your new node base larger than the chip?

So thanks, Simon. So I'll start from a technology perspective and Sumit can add the customer side and how that's going. I think I'll start with 1-alpha. We have -- it's a very competitive node, and we're deploying it across multiple different densities and multiple different aspects of -- whether it's -- it will be deployed for mobile, it will be deployed for computer and it's going to have a broad portfolio, and we see very good competitiveness on multiple different densities that we're going to have in the 1-alpha node. So I think that -- I don't think we see it as hampering us for some of the lower capacity. We see good competitiveness on multiple different densities across multiple different end market applications. On the NAND side, same thing on our 176. You're right, it is a very advanced node, and we really feel very good about our technology position and our decision a couple of years ago to kind of move forward in the replacement gate as we transition there to do a small implementation of 128 layer and then jump to 176 for the more significant deployment there. One thing to keep in mind on the 3D NAND side is that because of our CMOS under the Array capability, that means that we're going to be maybe more capable and effective -- efficient on the die size floor plan because we were able to fit all of the CMOS under the Array versus maybe some of the competitors that -- and what their comments might be about what they have to do because not all of our competitors have adopted CMOS under the Array as yet.

Yes. And in terms of customers, just like Manish said, we have a pretty broad portfolio of capacity points in our die road map. And hence, we can meet most of our customer requirements. Of course, there are some low capacity points that we can't always meet, but we don't try to create a portfolio of products that can be everything to everyone. We pick the areas where we feel the portfolio can best address customer needs in the most optimized way. And some of the applications that require far smaller capacity points are often just better served through legacy nodes. And we do that, for example, in our automotive and industrial markets, some consumer markets also serve from legacy nodes. So we use the different module mix in our manufacturing footprint to meet the varied kinds of customer demand that we have. But for the volume part of the market in most of our segments like server, client, mobile, et cetera, our 176-layer NAND and our 1-alpha can really meet the broad range of capacity requirements that our customers have.

Yes, very clear, Sumit and Manish. I really appreciate. But very one quick clarification, 1-gamma delta DRAM node, newly the EUV, right? Still using the market patterning, as you previously commented, right? No EUV there, 1-gamma, 1-delta?

No. So Simon, so we had previously said that we were evaluating EUV for 1-delta. And what we're saying now is that we do see that we will have a significant introduction at 1-delta, but we will have, not necessary, significant deployment on 1-delta, but the introduction will happen in 1-gamma with fewer layer count and then more layer count on the 1-delta. And we think the 1-gamma will be in the 2024 time frame and then 1-delta will come after that.

And our last question comes from Sidney Ho with Deutsche Bank.

I actually a couple -- 2 quick questions. One is again on DDR5. What is your current expectation of ramp schedule in terms of when and how quickly maybe versus DDR4? I know you talk about the bigger die, I also heard you also doing more server modules. Do you expect a margin trajectory over a certain period of time to be similar to DDR4 with DDR5, maybe low in the beginning and maybe come back up? I'll ask the second question as well. Just on EUV. I know you talk about 1-gamma in 2024 and then 1-delta. I wonder if you have any views on your development of your R&D project on 3D DRAM because one of the equipment suppliers talk about you don't really need EUV for 3D DRAM. I'm curious if you agree with that assessment.

Yes, I mean I can take the first part of your question on DDR5 and then Manish take the other question. So in terms of DDR5, we expect that ramp will start later this calendar year and then fiscal '22 -- sorry, calendar '22 will be the time where the ramp starts to really pick up steam in a meaningful way. Later this calendar year '21 is just the start of the ramp. And then as our customers deploy these more advanced CPU platforms in greater numbers, they will obviously be increasing their consumption of DDR5 beyond '22 in a very material way. So I think that's the schedule we are on. And I'll turn it over to Manish to talk about the EUV discussion.

Sure. Right. So the -- on the question about EUV and how it helps our scaling road map, we see node, that we have multiple more nodes than EUV is going to be able to help us with. We -- as I mentioned to Simon, just before, we see a meaningful number of layers on -- or a significant number of layers in 1-delta. And then we see nodes beyond that, that EUV will allow us to continue to scale. And we feel very confident in being able to maintain our leadership in DRAM scaling as we go into the foreseeable future. So -- and we feel very good about what EUV is going to allow us to do and what our scaling -- and the combination of our expertise in multi-patterning as well to help us maintain our lead in that.

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