All right, great. Good morning, everyone. Thank you so much for coming. I want to welcome you to the Goldman Sachs Global Semiconductor Conference. My name is Toshiya Hari. I cover the semiconductor and cap equipment space here in the U.S. Very excited to kick off the conference with Mark Murphy, EVP and CFO for Micron. I will kick off with some questions, but want to keep this fairly open. For those who'd like to ask a question, please have them ready. Mark, thank you for hosting the conference.
Thanks.
for supporting the conference.
Yeah, good morning.
I wanted to kick off near term, and then as we progress, go longer term and then more structural. In terms of the near term, I guess since your last update, what have you seen in terms of demand across DRAM and NAND? If you can kind of talk through some of the key applications, that would be super helpful. As you answer that question, talk a little bit about the dynamic in China. We saw the headlines from the CAC last week. You've had a week to potentially digest and maybe reach out to a couple customers, so any feedback from them would be super helpful.
Sure. Good morning, everyone. Thank you for joining us so early. I'll start with the safe harbor. I'll be making forward-looking statements. These statements have risks and uncertainties associated with them. I refer you to our risk factors disclosed in our public filings. As far as trends in the business, the trends in the business have been consistent with what we outlined in our third quarter or in our March earnings call for the second quarter. Volumes and prices have been generally consistent with what we expected. We are seeing bit growth and byte growth for DRAM and NAND as we expected. Now, the CAC decision came very late in the quarter here, and there was some effect from that.
Since it happened so late in the quarter, we didn't feel there was a need to update the guidance. It'll be more of an effect as we look pro forma. Yeah, we do think, as we said in March, we do think that, yeah, the market's bottoming. You know, we are seeing... You know, we do believe that, you know, we're seeing that, you know, growth in bits as we expected. We bottomed in bits in the November quarter, had an increase in February quarter, increase in the May quarter. We still see second half being stronger on bits. Customer, inventories are trending down as we had expected.
We've had broad-based, you know, reductions in supply, so every industry participant at this point has reduced CapEx and utilization. You know, in the end, we see that the market is beginning to find its footing on supply-demand balance. You know, and as a result, we're beginning to see some indications of price stability in some areas. In fact, in some isolated areas, we've had price increases, which we saw right after some utilization reduction announcements. Generally, there, you know, it's just indicative of there are some more firm areas than others on pricing. For example, DDR5 is firmer than DDR4.
On the CAC, the Cyberspace Administration of China, their review and conclusion, we talked about last week, we were notified that they concluded their review. You know, in their conclusion, they notified operators of critical information infrastructure that they were to not use Micron products. It's still very early in this. We are still, you know, customers and Micron are still working to fully understand the effects of this decision. Initially, you know, last week, we provided an indication of what we thought the impact would be, we gave that as low single-digit percent of revenue to high single-digit percent of revenue, total revenue for the company.
You know, as we've had, you know, a week of additional time, we've had some initial customer discussions. You know, as that information becomes clearer, we'd say that the better estimate is at the high end of that range, or a better estimate is at the high end of the range. Yeah, we will continue to assess it here, and provide more at earnings, but it is a, it's an uncertain period right now. It's an evolving situation, and it's very challenging. You know, over time, we would expect that we would be able to replace some of this impact.
Again, it's very early, still taking in info, you know, and the situation that's evolving and that we're working through. Back to the core business. You know, this was a very severe downturn that we've experienced. We're still in, but we do feel, again, that it's bottoming. As I mentioned earlier, we are seeing signs of supply-demand, you know, balance beginning to improve. Given the severity of the downturn, that's gonna take quite a bit of time. You know, we're working through it. Fortunately, Micron's technology leadership, the breadth of our product portfolio, our manufacturing excellence, that's helped us greatly work through this downturn.
The balance sheet has also provided us an advantage in that we can still invest in critical technologies, you know, and make the right long-term decisions. We're focused on executing well at the moment, dealing with the short-term challenges we have in the business and elsewhere, and then also just positioning ourselves for long-term opportunity. On the other side of this downturn, we're very excited about the market.
I mean, we've got artificial intelligence, we have automotive, we have a number of other high-growth areas. We just believe that with Micron's technology, our products, our manufacturing capability, that we're very well positioned to take advantage and serve those markets well, and then deliver a better financial performance more in line with our long-term financial model, through cycle model.
That's great. Thank you so much for that. I guess a couple of follow-ups on the CAC dynamic. You talked about the impact potentially being toward the higher end of what you had provided last week. Is it both DRAM and NAND? I know the definition of critically, critical information infrastructure is still kind of vague, but what's included in that potentially high single-digit impact? Is it server, networking, and nothing else, or is it a blend, or how should we think about that?
Yeah, we're still assessing it. It's again, it's fluid and unclear and very much evolving. That high single-digit estimate would principally be networking and server.
Okay.
You know, and is that the boundaries for, you know, the principal amount? You know, we're trying to get a sense of that.
Okay.
You know, we'll learn more here as we engage with more customers, they give us feedback, and you know, we work to get clarity from the CAC, and you know, we'll just have to continue to work through it and provide an estimate. We'll, of course, have a better estimate at the earnings call.
Okay. All right, cool. I guess in terms of customer inventory, I think depending on the application or the end market, your customers and yourselves are kind of a different place in the cycle, if you will. I think PCs and smartphones are further along that dynamic, and servers may be less so. Maybe talk a little bit about that dynamic, and, you know, at what point would you expect your sell-in to match or converge toward true consumption or end consumption?
Yeah. We have, you know, of course, on our own balance sheets and across the industry, the inventories are at unprecedented levels. We can talk more later about, you know, that effect and why it's maybe an indicator of future health of the industry. I know it's counterintuitive, I actually think it's an interesting way the industry's responded to this environment versus previous decades in the space. As far as customer inventories, we do view that PC and smartphones, since they went into a downturn earlier, they've had longer to work through and rightsize their inventory. We see those inventory levels sort of at this midyear time, being healthier.
Really, in the case of PC and smartphone, it's gonna be a function of, you know, their end market growth. Less so they're worried about inventory adjustments. Unfortunately for PCs and smartphones, the growth is still elusive there. In fact, in our last call, we called down our calendar year 2023 view for smartphones and PCs. We had PCs as, you know, flat to down slightly, and we moved PCs to down mid-single digits percent year-over-year. On smartphones, we had flat to up, and we've moved smartphones to down slightly year-over-year. I think the, you know, the inventories are okay in those markets, but it's about demand.
You know, in the case of, in the case of data center, it's, you know, the demand is robust, particularly in certain parts of the server market, as you know. The inventories are still, you know, elevated, so we're still seeing, you know, either inventories being reduced or they don't have the same need with our inventory levels where they are to purchase product. We think this, you know, shakes out through the back half of the year, and by the end of the calendar year, you know, we think all markets' inventory should be in a better spot. Our inventories, hopefully, will begin to move down at that point.
Got it. Earlier, you talked a little bit about pricing in DDR5 being a little bit more resilient than in DDR4. Curious how you're thinking about the transition to DDR5 and your competitive position there. I think there were some concerns about server platforms, the new platforms being a little bit delayed, but what were your thoughts on crossover to DDR5, more so on the server space?
Yeah. We're really excited about our product position in DDR5 and, you know, our 1β technology is sort of designed for DDR5 in a way, and HBM. We've got a, you know, an outstanding technology that's ready to ramp and serve these two product lines, both of which we see increasing here at the back half of the year and then next year. You know, 1β has, you know, 1α was a good node, but 1β has 35% better bit density, 15% better power, so it's a, it's an extremely good node. As a, you know, as the server platform rollout goes, yeah, we see DDR5 being about 30% of volumes or close to 1/3 by the end of the year.
It's about 20% now. The crossover point, we expect to be mid-next year.
Okay, got it. Then HBM, I think there's obviously a lot of focus around AI, and we'll certainly come back to your exposure there and what you're doing there. Talk a little bit about what you're doing, and at what point should we expect you guys to benefit?
Sure. We're going to benefit. Unfortunately, we've got a bit of a timing issue. Yeah, there's some history here on HBM, and Micron actually had a very good technology, maybe a better technology, you know, five years ago or more. I think it was called Hybrid Memory Cube. We had worked on that product, and we're moving forward on that, and the industry standard went to HBM. We found ourselves behind, but, you know, fortunately, the volumes are still very low. We've got an HBM product coming out end of year. It's on 1β, which I talked about, is just a terrific node. We're excited that, you know, we'll be able to hit this, you know, at a decent time.
Wish we were right there right now, but we will, we will hit it. You know, AI's got, you know, a lot of legs. It'll be a strong business for us. Again, you know, we think that the product that we have, that we're gonna deploy later in the year, we've done our own competitive assessment. We think we'll have a very competitive product entering the market.
Okay. later in the year, you'll have.
Yeah
... the product out. Okay, got it. The topic of AI is hard to go around at the moment. There's a lot of focus on it. Talk a little bit about AI as a percentage of your business. I think in your K's and Q's, you disclosed that servers account for about 20%-25%-
Mm-hmm
of your revenue. I don't know if you have perfect visibility, but, what percent of that is AI today? As you think about, you know, the implications for content growth in servers and contribution to overall DRAM industry bit growth, how do you think about that internally?
We could talk for the next 28 minutes on this topic. It, it's a very rich topic and of course, changing very quickly. What, what we can say is that, you know, right now, it's a, it's a relatively low percent, growing very quickly, but it's very difficult to, really size it. I think it's gonna be increasingly difficult to size what's, you know, determined to be AI. Because, you've not just got the, you know, the straightforward, training and inference server activity. Eventually, you've got AI in your phone now. You know, you will have it eventually in other edge devices. I think there was some news yesterday, on the client side.
It's gonna be, you know, that question of how much AI will be increasingly difficult to answer. We know it will grow, and I think that's the, that's the exciting thing for us, is, it's going to grow. I think, you know, we had already contemplated that in our long-term, you know, growth in AI, of course, to drive server activity or server volume. We had included that in our long-term growth CAGR. We had talked about it in Investor Day, and, you know, that it was a important part of the mix shift in our business, from sort of, migration of less consumer to more, you know, data center, AI-driven, and then automotive and industrial.
You know, right now, you know, with all the advancements just since our Investor Day a year ago, with all the advancements, it looks like there's potential that it's above, you know, what we had contemplated. We're, you know, we're deep into revisiting our assumptions, and we'll be able to talk more about it at the earnings call in late June. It is exciting. I mean, in the end, you know, AI is a data science that is very memory-intensive. You know, the accuracy of these AI models is a function of the number of parameters.
If you just look at the progression of, you know, the ChatGPT models, you've got, I think ChatGPT-2 had 1.5 billion parameters. My numbers may not be exactly right here, but somewhere in that range. I think ChatGPT-3 had 175 billion parameters. ChatGPT-4, there's some not real clear if the number's been disclosed. There's some mixed information on it, but it's, you know, hundreds of times plus more than ChatGPT-3 on number of parameters. These parameters and the number of parameters is highly correlated with the amount of memory that you need. I think as there's a march towards greater accuracy and utility of these models, you know, by definition, you're going to end up with more memory content.
Just, yeah, just the amount of silicon. You know, if you look at that, I mean, and, you know, the, you know, there's more silicon, you know, versus compute as you know, look to these AI servers. You know, in a, and an example is, you know, an AI server, and granted, these are training servers, but you got 8 x the, you know, memory content of a more traditional server, and for DRAM, and 3 x for NAND. You know, as a server is, designed to be optimized for, you know, for AI, it's going to, again, consume more, more memory or need more memory.
I guess in the near term, from your perspective, can you sense sort of this inflection happening in AI servers? Or because your customers are still, to your prior point, sitting on inventory, you don't necessarily see it from where you are today?
No, we see it.
You see it.
I mean, we, you know, we're working with, there's a number of different ways to, that this comes at you, right? We're working with everybody. I mean, this is not just GPUs, and there's, of course, that. That's been dramatic. I mean, if you look at, if you look at NVIDIA's generation of AI servers, four or five years ago versus the latest, you know, Hopper version, you know, memory in that has quadrupled. Of course, they're advancing and, you know, there's gonna be additional need to optimize memory for the great work they're doing. You know, there's other areas that, you know, there's CPU-related activity, there's ASICs. You know, yeah.
There's all sorts of activity that customers, they see what, you know, the advances that they're gonna need to deliver. You know, and part of, you know, this data-intensive application requires great coordination and utility of memory.
Got it. You talked a little bit about sort of reassessing the long-term CAGR for the DRAM business and perhaps the DRAM market. At the Analyst Day, I think it was a mid-teens CAGR that you threw out.
Right.
Does that feel like this is still the right range as you think about your business medium to long term?
Yeah, we've not updated it. You know, I think it, for now, it seems directionally correct. You know, in that mid-teens assumption is, you know, automotive and industrial, which will be more than double that growth rate. You've got data center, which we just talked about, which will be above that growth rate, and it'll be AI-driven. You've got, as we talked about earlier, you've got lower growth, but still durable businesses, you know, PCs, smartphones. It'll be interesting to see if, you know, with greater AI activity and, you know, what sort of increased content will happen at these edge devices. That'll be interesting to follow.
Okay. All right. I guess transitioning to the NAND side of the portfolio, where I think there tends to be more price elasticity vis-a-vis DRAM. NAND pricing is down, you know, more than 50% year-to-year. Do you sense price elasticity kicking in across some of your key applications, or not so much?
You know, it's hard to say. The pricing environment in NAND is worse than DRAM. Inventory levels are higher than DRAM. You know, I think there may be some elasticity and may be helping demand. I think in some areas, like, let's say, as a hard drive replacement by SSD, I think folks realize that the pricing in NAND at the moment just isn't sustainable. You know, while there can be an economic trade-off for that, you know, use case, in the near term, you know, the question is: Is it being, you know, stalled by just a price that's not sustainable at the moment?
Yeah, that part of the market, you know, we've said that the NAND business, or the NAND industry will benefit from some consolidation. You know, I think we've said our long-term growth CAGR in that business is low 20%, so we see it being a strong growth business. It's just hard to say at this point how much elasticity is affecting demand.
Got it. Mark, that low 20% CAGR on the NAND side, I know NAND is a little more consumer-heavy relative to DRAM, does it feel right as of today, or?
We've had no update on that.
Okay.
Feels okay.
Okay. All right. Maybe one question before I take questions from the audience. On, on the supply side of the equation, you know, you guys were probably one of the earlier companies in terms of actually cutting CapEx and cutting production. I think you're cutting production by 30%, give or take?
25%.
25%, I'm sorry. 25%, both in DRAM and NAND. You're cutting WFE by more than 50%. I guess, as you think forward and look forward, what would you need to see to consider increasing production in both DRAM and NAND? What are some of the potential trigger points?
Yeah. You know, as you've said, we feel very good about how quickly we took action, and we responded. We saw demand weakness, about this time last year, like, back half of May, actually. On our June earnings call last year, we talked about that we saw demand weakness, that we were responding to that weakness and beginning to curb spend and so forth. Of course, we know it intensified through the year. We reduced utilization, we cut CapEx. We cut CapEx meaningfully then, and then, you know, again, intensified those efforts through the year. We stand here, as you said, with WFE down over 50% year-over-year, with utilization off 25% on starts.
You know, we have, you know, as to what it takes to turn that, you know, increase investment and first increase utilization, it's just gonna take the inventory levels coming down and some visibility that the trajectory of that is what's needed to, you know, come back to more normal levels. We're, you know, we're over, well over 200 days of inventories, and, you know, and we've said we'll exit the year over 150 days.
Is it fiscal or calendar?
Well, it's calendar as well now.
Okay.
So-
Okay.
I mean, it's just that's a function of, you know, in the last call, we talked about the recovery is a bit more elongated than we had expected. The inventories are, you know, they're high. As we see demand firm up, and inventories begin to trend down, and, you know, fortunately, the entire industry has taken supply actions. Eventually, we'll start to see that correct itself. You know, the industry has utilization to work into. I talked about, you know, the 1β node we have, and we actually have the 232-layer on the NAND side. Unfortunately, both those nodes are being deployed just as we entered the downturn.
We've got two nodes that actually, as we increase utilization, we'll also have two new nodes to work into that'll be very capital efficient. In fact, on both those nodes, we've achieved targeted yields in record time. Both those nodes, though they're not fully ramped, they both have yields that are associated with products that have been out longer. We'll be able to very efficiently ramp when we do, and then we'll talk about CapEx and building out capacity. We expect next year to be also depressed on WFE, and that our spend will be a bit heavier weighted on construction as we think about back half of the decade capacity.
You know, but again, it's a, it's an inventory, you know, related, focus at the moment.
Got it. Got it. I think you made this comment earlier, but in terms of what your competitors are doing from a supply perspective, I think 6 months ago, 9 months ago, you guys had thought specific competitors were not cutting enough. At this point, would you say that all parties in DRAM and NAND are taking appropriate steps, or are you still waiting for select competitors to take action?
Well, we, I mean, we can only control what we control. We're doing what we can do. We, we think we have ample inventories. We think it's appropriate to reduce utilization based on what we see. We believe that, you know, incremental capacity is not helpful to the industry at the moment. The pricing just is not, it's not sustainable. It's not sustainable for investment in the technology and the business overall. That has to be fixed. I think based on, you know, what we see on others, you know, reducing utilization and CapEx, it appears that, you know, there's, just, you know, companies are not healthy at the moment, and they're dealing with it. Yeah, we'll.
We'll keep an eye on inventories, and our own business, we will manage, as we see how we're projecting those inventories, over time.
Got it. That's helpful. I'll pause here and see if you have any questions from the audience. Yeah, if you can use the mic, please.
I just wanted to follow up on that.
Sorry, if you can... I think there's a button.
Sorry about that.
Yeah, yeah.
Just wanted to follow up on the comment, going back to the 1α node. I think at the time Micron first started talking about that node, the comment was it would provide 40% greater bit density, and I think you didn't really say what you've achieved there. It was just a productive node. Could you talk about what you really did see at 1α? If I combine that with 1β, am I right that you're getting, like, 70% improvement in bit density over two nodes?
Yeah. The 1α node was also a very good node. The bit density improvement was in the zip code that you're talking about. Yeah, it's over, you know, from 1α node, 1β node, you've got 40-ish%, and you've got another 35% bit density improvement in 1β, so.
Thank you.
Any other questions? You made a comment about how high inventories in the industry might be a positive sign going forward. Could you elaborate more on that?
That's a... I'm glad you brought that up. It's obviously, on one level, it's not good. It's, you know, we obviously don't want inventories at this level, and they should be, you know, roughly half of what we've got. To me, what's interesting is it's an effect that you would not have seen, let's say, a decade ago or more. You know, the cost downs have become so challenging in the space, and they've become shallower. The, you know, one effect of that is you can then hold inventories longer than you used to.
In the past, when you were getting cost downs that were significantly higher, you would be compelled to liquidate your inventories, try and get rid of it as quickly as you can, of course, depressing the pricing all the more. And then you would, in parallel, you'd be investing heavily for the next node, and because you needed that cost down in order to survive. And again, you'd liquidate inventories because very quickly those were not gonna be price competitive. What you see now with shallower cost downs, very hard to drive incremental improvements in the business. You know, these inventories are good longer.
You know, when we were, you know, last summer and through the early fall, you know, we knew we were building inventories and, you know, and didn't want to, but, you know, but we have some comfort in building them, and it's carrying costs, but it's, you know, it's not as severe as it was in the past. We've got, you know 1α product, for example, is gonna be good product for many years, and competitive in the marketplace. I think that's a difference, and I think it speaks to the, what we believe are going to be some positive factors in the structure of the space going forward. You know, if you think about supply-demand balance, you've got positive things happening on both supply and demand.
On demand, it's, you know, contrasted with, you know, 15 years ago, it's not largely consumer-driven anymore. Consumer is still a big, important market in memory and storage, but it's not the dominant feature in the market. You've got, you know, it's less than half. It's, and the, and the drivers of growth, the, you know, what we know will be much bigger going forward will be data center, automotive, industrial. The demand is broader, it's more durable, and that's better. On the supply side, we just talked about the cost downs are more challenging and shallower. As a result, inventories can be used to modulate capacity, like they couldn't be used before.
I think the combination of more levers on the supply side and, you know, more discipline on the supply side, and a broader, more durable demand equation, yeah, I believe you'll end up with better capital allocation, and then, thus, a more stable pricing environment and better industry going forward.
Hi, just on the CAC, if it really happens, and, as you expect it up to a high single digit kind of impact on your revenue, is that correct?
Well, what I said is that we gave a range last week, fresh off of the notice that we received. We provide a range of low single-digit percent of revenue to high single-digit percent of revenue.
Okay.
You know, today, I said, based on, you know, additional customer feedback, and it's still early, and our own assessment of things, that the high end of that range is a better estimate. Now, that estimate is still evolving. That estimate, as Toshi brought up earlier, that estimate is largely a function of networking and data center. Last week, I mentioned it does not include smartphones. You know, we've got to assess, you know, we're still in the process of assessing where all this falls out. Again, it's evolving. There's a lot of uncertainty. It is challenging, but we're of course, working, you know, best we can to clarify it, to mitigate it, and we'll provide more detail, you know, at the earnings call in June.
Yeah. Yeah. I guess in the short term, it's a disruption to maybe to the whole industry. Help us understand how this dynamic is gonna play out. If you say, if you lose that bunch of revenue, how are the guys gonna pick up, from Taiwan, Nanya, or from SK hynix, Korea, or some domestic Chinese player gonna fit in that space?
Well, I think in the near term, it will have a negative effect on our revenues versus our prior expectations. I mean, in the near term, I think we can say that at this point. However, as you point out, you know, the bits will move around, and we'll have some, you know, certainly hope and are working towards recovering that lost volume elsewhere. If it were a tight market, and bits were scarce, then this would be an easier issue to deal with. The market's loose, and so it's going to be, you know, take more time to deal with it. You know, we've got very advanced technology.
We're on the leading node in both NAND and DRAM. We've got deep technical relationships with customers. You know, we're amongst the highest quality customers that are delivering the highest quality products of any suppliers for most customers. I think that, you know, 2/3 or 80% of our customers rank us number one or two on quality. We've been reliable. I think, you know, to the extent, you know, there's a way to work with us, I think customers will work with us, of course, within the confines of the rules and regulations of local jurisdictions. You know, again, we'll continue to evaluate this and provide more as, you know, certainly on the earnings call.
Maybe, Mark, in the last couple of minutes that we have, I wanted to hit on two topics. One is EUV, and then one is how you're thinking about your global manufacturing footprint. On the EUV side of things, I think you've performed really well in DRAM in terms of your roadmap. You will be inserting EUV at the 1γ node. I guess the concern that investors have is, you know, you're late to EUV. Are you gonna experience similar hiccups to what your peers experienced over the past couple of years? That's number one.
Number two, in terms of the global manufacturing footprint, you recently announced projects, I believe in Boise, in upstate New York, and I think Sanjay was in Japan recently, talking about expansion there. In the current geopolitical environment with, you know, government incentives from, you know, various governments, how do you think about the balance across countries, both in terms of DRAM and NAND?
Yeah.
In two minutes.
Yeah.
I'm sorry.
That's okay. On EUV, we feel very good about the decisions we've made around EUV. I think with what we know now, it's very much the right decision to delay it. I mean, fundamentally, we didn't need it to advance the technology. I mean, others needed EUV, you know, to hit nodes that we've already completed and are being ramped. You know, our multi-patterning technology is the best in the world, and that's allowed us to advance the technology without, you know, the huge cost of EUV. There's been this question of when we do need it, which we've said we need it, we believe in 1γ, what is the... You know, are we gonna be able to deploy it in yield?
I think we've demonstrated already, we actually deploy it. We have it in Boise, in our R&D setting, and then we have it actually in a production setting in Taiwan. We've actually inserted, in a pilot line, EUV in a layer in 1α, and it's yielding well already, so we already know that it works. Then by the time we deploy it in 1γ, the supply chain will be much more mature than it, than it is. I think it's lower risk. We've already demonstrated it. We're in good shape. As it relates to, you know, the footprint, you know, there are government programs and incentives and so forth, I think it's important to keep in mind that, you know, these incentives determine or help determine geography.
They don't determine capacity. We add capacity, in line with the industry growth rates. Then we, you know, then we, you know, we'll choose a geography based on competitiveness. We're in a globally competitive marketplace. If we proceed with, you know, assuming we get the support we need for investment in the U.S., we would build out Boise. Construction is starting, and, you know, in 2023. Construction in New York would be 2024. Both of those, would, you know, be sort of primarily end-of-decade, demand or end-of-decade supply or demand at end of decade.
Okay, great. with that, we're out of time. thank you so much for coming, Mark. Really appreciate it.
Thank you, Toshiya.
Thank you.