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

Okay, great. We'd like to get started. Good morning, everyone. My name is Toshiya Hari. I cover the U.S. semiconductor and semiconductor cap equipment space. Very honored, very excited to have Sumit Sadana from Micron. He's the Chief Business Officer for the company. Sumit, thank you so much for supporting the conference.

Thank you for having me, Toshiya.

Great to have you here. I have a list of questions per usual but we'd like to keep this interactive, so to the extent you do have questions, please have them ready. I'll go to you in 20 minutes or so.

So Sumit, in terms of the near term, it's been 2.5 months since you shared guidance for the May quarter. The industry is clearly in a positive stage of the cycle, both on the DRAM side and the NAND side. How has the quarter progressed relative to your expectations? And any standouts that you'd highlight?

Sure. Just wanted to start with a quick comment that I may be making some forward-looking statements today, and please visit our website and look at our SEC filings on the quarterly and annual reports for a more detailed description of the risk factors. No new update today, but just to summarize how we look at things right now, we have a consistent view of what we said the last time in terms of how some of the segments have been progressing. We continue to see the PC market unit growth in calendar '24 to be in the low single-digit range, similar range for smartphone unit volume growth as well worldwide. Of course, there are average capacity changes taking place in both of those devices. On the data center side, we had said that we do expect within the first half of calendar '24 for inventories to get better and more normal ordering patterns to resume from customers. We have continued to see that, and we do expect better second half demand compared to the first half in data center as all of that inventory gets normalized, and we have continued to see really strong demand on the AI server side and all of the products related to AI servers have been exceptionally strong. We do expect that the pricing will continue to improve over the course of calendar 2024 as the supply/demand dynamic continues to tighten. And we had spoken about somewhere in the low 30s gross margin for next quarter. That continues to be our expectation. Our CapEx is more likely to be around $8 billion, then the $7.5 billion, $8 billion range. Continue to expect free cash flow in positive free cash flow in F Q3, Q4, so second half of this fiscal year, and a record year from a revenue perspective in fiscal '25. That continues to be our expectation with much improved profitability as well. And again, all driven by what we have said, we do see Micron as one of the primary beneficiaries in the semiconductor industry of this AI trend because of the significant increase in average content of DRAM in particular, increasingly, NAND as well but particularly in DRAM and our really strong product capabilities on the high-bandwidth memory side.

That's great. Thank you so much for the overview. We'll definitely come back to AI, HBM, enterprise storage later. But at a high level, I wanted to ask you about customer inventory levels. Last year was obviously a challenging year for you all and for the industry. Based on your customer conversations and market intelligence, where are customer inventory levels relative to what you would consider to be normal? And I know you don't have perfect visibility but across end markets, where is customer inventory today?

Sure. So yes, as you said, the customer inventories are different depending on the end customer and different from a segment perspective as well and also from a geographic perspective. So it is not easy to make sweeping generalizations. But some trends that I can talk to, the data center inventories have improved quite significantly over the course of the last several quarters. They're in much better place now. Certainly, the AI server growth has driven a lot of that improvement, and there has been some improvement in the traditional compute segment as well, and that is improving, even traditional compute-related inventories. On the PC and smartphone side, generally, the inventories very much varying by customer. Some customers have been concerned about the rising trend of pricing that they see, and they have used that to build up some inventory, but they also have an expectation of improved content as AI PCs and AI smartphones start to roll out more so later this year. Our expectation on AI PCs and smartphones is more focused around 2025 growth, and that's where we are looking for some of that to happen. Not too much from a unit perspective but more in terms of the mix is what we are looking for because the mix step-up, as people buy some of these newer smartphones and PCs, in order to future-proof those devices because at some point, applications will come out that are going to be very interesting, very useful from an AI perspective. And consumers will want to have access to hardware features that can enable those applications to run. And that typically means getting to 12 gigabytes of DRAM in a smartphone, which is significantly higher than where it is today, and 16 gigabytes at a minimum on a PC, but a lot of PCs with 24 and 32 gigabytes and above. So those will be meaningful step-ups in content, and increases in average capacities on the SSD side as well in these embedded flash in smartphones. So we expect all of that to start in a more meaningful fashion in calendar '25. But I think we do expect that the supply/demand dynamic is going to be tight next year as well, and I think that's also influencing how customers think about their inventory.

That makes sense. One sort of last question before we dive into HBM, topic of the day, week, and month. So long-term bit demand growth and how you guys think about that. It's been, I guess, roughly 2 years since you hosted your Investor Day. Can you remind us how you're thinking about the through-cycle bit growth outlook for both DRAM and NAND? How has it evolved over the past 2 years? I'm sure multiple moving parts for both DRAM and NAND. And what are the key contributors as you think about the through-cycle growth profile?

Sure. So in terms of through-cycle bit growth, we continue to see mid-teens type of growth for DRAM, and we continue to see low 20s type of CAGR for NAND. Now these numbers obviously change depending on what your base year is, and 2023 was a pretty challenging year for the industry from a bit growth perspective. So if you take it from the 2023 base, the numbers can get distorted a bit. But by and large, those are the bit growth numbers we're looking at. Generally speaking, we are looking for data center to grow faster out of the total. And some of the device segments like PCs and smartphones to be growing slower than those averages. Of course, the rate and pace at which AI PCs and smartphones get adopted, and if there is a change in the amount of time consumers are going to upgrade their phones or their PCs, that can noticeably change these numbers. But we are taking more of a conservative approach in our planning that unit volumes and growth in aggregate unit volumes of PCs and smartphones, we still expect to be mild and more of a mix change and average capacity change driven by mix is what we are expecting. So those are the numbers. I will say that the tech transitions on the NAND side are still providing more than adequate bit growth. So there isn't really any kind of wafer growth that we are foreseeing on the NAND side, but there is a very significant trade ratio when you think about HBM growth that's taking place. HBM 3:1 trade ratio for DDR5, which means it takes up for every bit of HBM we have to ship, it takes up 3 bits of DDR5 in terms of supply. So because of that trade ratio, a considerable percentage of the industry's wafers over time are moving to HBM. And because of that, the non-HBM portion of the DRAM wafers is getting squeezed. And consequently, aggregate wafer growth will be needed in DRAM to even meet the mid-teens type of aggregate bit demand in DRAM.

Got it. And just transitioning to HBM, how are you thinking about the multiyear growth rate in HBM specifically? And I think you might have mentioned this, maybe we talked about this last night at the dinner, but percentage of bits going to HBM today? And how do you see that evolving over the next couple of years?

Sure. So last year, HBM was less than 2% of the industry bits. This year, it ought to be somewhere less than 4%, in that range. We do expect the HBM bits to grow in the medium term in the 50% CAGR sort of range, again based on the 2023 base year, of course, the industry bit growth rate, say, mid- to high teens kind of a range. But I would say that the HBM growth continues to be strong. The demand continues to be extremely robust. And we are seeing a lot of that significant demand because our HBM3E product is an industry-leading product with some very breakthrough type of power consumption capability that's at least 30% below competitive offerings. And so yes, it's a good environment.

Understood. And then your competitive position within the context of the HBM market. I know in overall DRAM, you're at 20%-plus. In HBM specifically, I think you're in the single digits today. And the reason for you guys being under-indexed as of today, is that simply a function of past technology bets or is there something structural to it? And I think you've talked about approaching 20% and 20%-plus sometime in calendar '25. What gives you that conviction that you'd go from single digits to 20%?

Yes. So Micron had a similar technology called Hybrid Memory Cube, which we had commercialized and gotten to a pretty good scale. And when HBM started to gain some traction several years ago, we switched from Hybrid Memory Cube technology to HBM. And HBM2E was the first product we brought to market and commercialized. And we did put that into volume production, and we got a decent share of the HBM2E market. Because we were making that transition from a different technology, we were -- we intersected the HBM2E market towards the latter half of the life cycle of that product. And so we determined that instead of being late to HBM3 and 3E and so on, that we needed to leapfrog the industry and get to a new product and really industry-leading specs. And that's what we decided to do. So after HBM2E, we did not do HBM3 but then straight to HBM3E in order to get to the leading edge of the industry. And so we're the first company, we believe, to sample HBM3E to our customers. And not only that but we also had the best specs in the industry, the best performance headroom to even go beyond published specs, but most notably at least 30% lower power consumption compared to the next alternative. And we feel really good about that because customers have validated those numbers, and they are amazed by the capabilities and the specs of the product. And now it is about ramping the product into high volume. So we have spoken about several hundreds of millions of dollars of fiscal '24 revenue with HBM and going to fiscal '25 and calendar '25, getting to multiple billions of revenue in HBM. And that's the trajectory we are on. So now it is about, we have the space, we have the tools and slots and orders and everything all in place, and now it's the operational aspect of ramping the product with high quality, which is what we always focus on. And we feel good about our ability to deliver on those goals, which we have outlined as being getting to our HBM share on a quarterly basis sometime during calendar 2025, being similar to our worldwide DRAM supply share, which is in the low 20s.

Understood. In terms of customer traction and customer mix going forward, I think you've publicly shared that you're shipping to the H200 that NVIDIA is ramping as we speak. As you look to grow your business, to your point, from several hundred million dollars to multiple billions over the next 12 to 18 months, what are your expectations in terms of customer mix? If you can describe customer engagements outside of the big GPU company, that would be helpful.

Sure. So yes, we have a goal of having a broadly diversified customer base. We have worked hard to do that across our base business. So today, you won't see any 10% customer that we really disclose having in the past. So certainly, when it comes to HBM, our goal is to have that diversification there as well. So definitely very strong engagement in partnership with the leading GPU company. We also have really good engagements and really good work and partnerships with all of the HBM, large HBM customers around the world. And consequently, our focus is to get the qualifications done and ramp with all of those customers as soon as they are able to ramp. And given the specs of the product, we see extremely strong interest from our customers to bring our product to market. Because when they ship their processors with our HBM, it actually has a lower operating cost for their end customers due to the power consumption benefit that we bring. So we do expect that we will have a diversified slate of customers going into next calendar year. And with each succeeding quarter, we'll have a better diversification in our revenue mix. So long-term engagements for future road map-related products even beyond HBM3E, so HBM4, HBM4E, really strong engagement, a very exciting place to be.

That's great. HBM pricing and profitability longer term. I know as of today, you and your peers get a nice premium on a price per bit basis in HBM vis-a-vis conventional DRAM. The fear that I think some investors have is over many years, HBM just becomes another DRAM SKU and it's commoditized. How do you think about that? Does the foundry-like nature of the business and the complexity, right, that comes with HBM production, does that enable you and the industry to maintain premium margins?

Yes. Indeed, I mean, HBM is a super complex product. It's the most complex product that the memory industry has ever built in its history. And it's not becoming any less complex with time. It's going to continue to have very significant level of complexity compared to the rest of the portfolio. And so we continue to feel like the pricing is going to be high because for starters, the cost per bit to produce HBM is much higher than the rest of the portfolio. So the pricing by definition then has to be higher. But we, even if you look at it from a gross margin perspective, our expectation on HBM on a long-term basis is through the cycle for HBM gross margin to be accretive to the core business. Again, because of the product capability, because of the value it brings to our customers, because of the complexity of the product. But even more so as we think about going from the current version of HBM3E, which is an 8-high stack in 2025, as we go through the calendar year, it will transition more and more of the mix to 12-high. And that is even more complex product to do than the 8-high. Future generations of HBM may go to like 16-high, so again, that will be a further increase in complexity. And as we look past HBM4 to like HBM4E type of products, there is a desire amongst multiple customers to figure out ways because every customer has a slightly different approach in terms of how they are creating value. And when they look at their subsystem, a significant amount of value is in how the processor and the memory work together. So it's not just about, let me just build the best processor I can. It's about how do I build the best subsystem that I can. And what that means is a lot of these workloads in AI tend to be memory-gated, memory-bound workloads, which means that to really improve the performance of these workloads, you really have to do two important things. One is have more DRAM available for the processor, so the amount of DRAM that can be connected needs to be maximized. And you also have to significantly increase the bandwidth, memory bandwidth between the processor and the memory so that the data movement can be done with that benefit of that higher bandwidth. And of course, all of that has to happen with more and more energy efficiency. So I think those are the things that have to be done, but more DRAM and more bandwidth is the critical aspect. And different customers have different ideas about how to create some level of differentiation in HBM usage going forward. And as long as it is the same type of HBM that everyone is using, for example, we built HBM3E product. Of course, it has to be qualified with different processes, but ultimately, it's the same product. But customers are thinking about HBM4E and beyond generations, that they need to find some way to create that differentiation and how HBM works with their processor. And one approach is to put some logic into the base die of the HBM. And that then makes the HBM product more of a custom product to a customer. And when you make an HBM product into a custom product, then it becomes more of an ASIC-like engagement model because now you're codesigning the base die with the customer, with customers' logic embedded in it with your own logic that you have as a DRAM company. And then it changes the whole engagement model, the core development aspect of it, much better visibility and coordination, much better ownership on part of the customer for the orders that they place because it's a custom product. It's not like you can sell it to somebody else. And so a lot of foundational changes can happen in the industry driven by that. And that's a really exciting trend, I think, for the HBM business. And of course, then it also changes the dynamic versus how you view the business, its margin profile, its stability versus the rest of the portfolio.

That's a great segue into my next question on LTAs. I think you guys have talked extensively about both volumes and pricing being pretty much locked in, if you will, for '24 and for the most part into '25. The other question/pushback that I often get from investors is LTAs historically have been sort of a one-way commitment. You're committed to supply, customers can flexibly push out, cancel, et cetera. How would you characterize the LTAs around HBM that are placed today vis-a-vis historically LTAs? How are they fundamentally different from past LTAs?