Micron Technology, Inc. ($MU)

All right. Good morning, and welcome to JPMorgan's 54th Annual Technology Media and Communications Conference. My name is Harlan Sur, semiconductor and semiconductor capital equipment analyst for the firm. Very pleased to have the team from Micron Technology here with us this morning. Manish Bhatia, Executive Vice President of global operations for Micron, and we also have Samir Patodia, Senior Director of Investor Relations, here with us as well. Manish and Samir are going to kick us off with some opening comments, and then we'll go ahead and kick off the Q&A. So gentlemen, thank you for joining us today. Samir, let me hand it off to you first.

Thanks, Harlan. So I'll start off with the safe harbor and then pass it to Manish. We'll be making some forward-looking statements today and those statements have risks and uncertainties associated with them. So we refer you to the risk factors disclosed in our filings, including our recent 10-Q and 10-K. Manish?

All right. Thanks, Samir. Standing on room in the back, thanks for your interest in Micron. Our financial outlook has strengthened since our last earnings call. We're on track for another substantial record free cash flow in fiscal Q3. Our balance sheet has never been stronger, and that's underscored by multiple credit rating agencies actually from all 3 upgrades from all 3 credit rating agencies this year. Demand continues to outpace our ability and the industry's ability to supply due to persistent structural factors. And so we expect tightness for HBM, DRAM and NAND to continue well beyond calendar year 2026.

Perfect. And so that was a great update. And so I think that we can go back and we can ground ourselves in terms of the near to midterm outlook that the team put out back in the mid-March sort of earnings call. And I think at that time, relative to the guidance put out for the May quarter, I think the view was that the strong revenue growth outlook at that time was going to be driven by predominantly pricing a little bit of volume growth. And so let's start off there. Relative to that, I mean, which of the dynamics has played out a bit stronger than expected or stronger than expected?

Yes. Well, certainly, pricing has played out. But demand is continuing to be very, very strong. And the AI world is transitioning from just human interactions to human-to-agentic interactions and even machine-to-machine interactions. And so -- and that -- those agentic workloads are driving inference and inference is becoming a bigger and bigger part of the demand workloads for our customers. And we're just seeing memory as continuing to become an even more strategic asset, adding even more value as memory can be equated with intelligence and higher performance memory, higher reliability memory, higher capacity memory, all of those help add to the accuracy of intelligence, and that's true on both the high-bandwidth memory side, it's true on the traditional DRAM and DDR and it's also true on the NAND side, where contact windows are getting longer and Sand flash is a key part of being able to not just store all the new data that's coming up, but improve the accuracy of the results of the models.

On the midpoint of the guidance for the May quarter, $33.5 billion, 81% gross margins change of earnings power. Any way to quantify the better strength that you've seen in the May quarter relative to some of the parameters -- the financial parameters that you put out back in the March?

Well, our quarters are not going to close in just a couple of weeks, and then we'll have comments on that, obviously, on our next call.

Okay. And so then maybe the way to -- maybe the better way to frame it is, let's think about it over a multi-quarter period of time. Clearly, the environment, as you mentioned, has gotten better for you. And I think that again, going back to the March call, you had talked about demand for this year being supply constrained, but you could still see the industry shipping sort of 20% sort of bit shipment growth in DRAM and low 20% sort of bit shipping growth in NAND. Has that viewpoint changed given the dynamics that you've seen over the past couple of months?

I think what's the demand is supply constrained, right, for both right now. And so we're just continuing to work very hard on enhancing supply, but I mentioned persistent structural factors earlier in my comments. What really mean on the supply side is that technology transitions for both DRAM and NAND are delivering less and less productivity, that's right. And they had before, less bit growth improvement per technology node than nodes of the past. And that -- on top of that for DRAM, we have the structural shift towards HBM, which that's -- in order to deliver the high performance of HBM, we have a trade ratio where die sizes are larger, and therefore, you end up with fewer bits per wafer. So it takes more than 3x as many bits per -- more many wafers to be able to deliver the same number of bits and that trade ratio continues to grow as you go from HBM3, HPM4, HBM, that trade ratio continues to grow. And so that all leads to needing more greenfield wafer capacity in DRAM and greenfield capacity needs more cleaning space because you're putting in full lines rather than incremental equipment to get the incremental benefit of the technology transition, you now need more cleaning space because you're putting in full lines, and that takes more time and that takes more time. So we have the strong demand drivers and continuing to strengthen because of the value memory provides to the end AI solution whether that's training or inferencing. And then you have the structural factors in terms of supply, where the whole industry is still trying to catch up, and we don't see the ability for the industry to catch up to -- supply to catch up to demand for the foreseeable future.

Yes, it is pretty amazing since the second half of last year, and we had [indiscernible]. We were just talking about -- we had Libet our keynote yesterday, and we -- we were talking about the significant increase in memory, storage, in CPU demand, literally the moment that inferencing AI inferencing workloads crossed over training workloads in the second half of the year. It's like the light switch went on and all of a sudden, memory intensity went up. Storage intensity went up. CPU intensity went up, right? And that trend has continued, right? Because as we track inferencing workloads, I mean, they're continuing to grow like at this exponential pace, and we can understand why the continued strong demand from memory. The team talked about key customers only able to secure 50% to 2/3 of their bit demand requirements in the medium term. Is that actually still the case? Or has that gap actually widened?

Well, we talked about the demand drivers getting even stronger than the supply continuing to be very, very difficult to increase. We're all working very, very hard. You had Tim before talking about how closely we work with them and we work that closely with all of our suppliers, and we're all -- we're working very hard to try and catch up, but we don't see that supply being able to catch up to demand for the foreseeable future.

Yes. So on that front, and obviously, many share Head of Global Operations, let's talk about some of the ways that you will be able to unlock the potential for more bid supply over the coming quarters, right? On your acquired, for example, Taiwan fab in Tanjong, the target was to start production bit shipments, sort of second half calendar '27. How should we think about the first few quarters of the initial ramp? We've heard 10 -- starting off 10,000, 20,000 sort of wafer starts per month. Your new Boise fab will be ramping about that same time as well. How much output should we anticipate on the first phase of that build-out right now? Most of these again are for second half of next calendar year, but give us an update on these 2 initiatives.

Sure. And Tim, again, had a great job having Tim go before me because he mentioned that we did make the acquisition of the Power Chip semiconductor ongoing earlier this year. That transaction actually closed ahead of schedule, very pleased with the support we got from the Taiwanese government for that. And we're making excellent progress towards that target of having -- converting that fab from the logic output that it had before towards being able to do leading-edge DRAM. And we expect to be able to have production -- we're making excellent progress towards production in the second half of calendar year '27. And we also announced on our last call that we're going to be building sort of a twin fab immediately next to the existing fab there as well, and we're starting construction on that this summer. So that site will become a larger leading-edge memory site. It's only 20 minutes from our existing Tanjong operation. So it's actually really going to be run like a mega cluster site. -- there. But we're doing a lot even beyond just what we're doing in Taiwan. We mentioned Idaho One is making good progress. Earlier this year, we had announced that we had pulled in our target date for wafer output to -- from second half of 2027 into mid-2027 and then we announced that Idaho 2 construction is underway. Actually, ground preparation is underway right now, and we expect to have that wafers out from that late in calendar year 2028. So a little bit more than a year after Idaho One. We broke ground earlier this year in New York, and we're actually ahead of schedule there and making good progress and expect to pour concrete later this year there. really good progress in Singapore on our high-bandwidth memory facility, which we had broken ground on a year ago at the very beginning of calendar year '25. We expect to have production impact from that in calendar year '27 to supplement the HBM operations we have existing in Taiwan. We broke ground on a new NAND fab earlier this year in Singapore as well. So we have a lot going on to try and address supply across DRAM, HBM and NAND.

Yes. And on top of all of that, right, you're executing on all of these new technology migrations at the same time, right? And so to that point, we're world-leading technology transitions, right? Industry-leading technology transition. And so on these new process technologies, I mean, 1-gamma for DRAM, Gen9 for NAND, right? You were targeting both of these technologies to represent the majority of your bit mix by mid this year, very close, obviously. We're almost there. Is the team going to execute to this shipment mix? Can you talk about the progress on yield, manufacturability? And does the insertion of EUV actually relax a little bit of the need for multi-patterning technology. And has that also been a tailwind for yields in defect density.

Yes. Yes. So we're absolutely on track for those goals with both our 1-gamma DRAM and our Gen9 NAND to become a crossover in terms of bit output by middle of this year. And both of them are ramping very well. Both of them in terms of yields ramped faster than prior nodes. And I always like to highlight, it's not because those technologies somehow got simpler -- it's because of the tremendous engineering work, some of the technology development work, the collaboration between technology and manufacturing as well as our use of AI across which we've been doing now for many, many years, actually more than more than 10 years to be able to continually enhance our ability to get yield improvement, have productivity enhancement, improve quality, all those things in all of our manufacturing operations. And we expect 1-gamma to become the highest volume DRAM node on a total wafer out volume basis that we've ever had of any note in our history. And we did add -- we did comment on the last earnings call that we are very pleased with our EUV progress on the 1-gamma node. It's the first node, which we introduced in EUV. And we're very pleased. We're very pleased with our decision to implement at this point in time when we had more modern, more advanced tools from ASML that are actually giving us better performance, better availability. So given that, we had said on the last call that we expect to increase our deployment of EUV in 1 delta and future generations. And so to go along with that, we actually have recently concluded a multiyear EUV supply agreement with ASML that supports our technology and capacity plans over multiple generations.

Let's talk about part of the manufacturing excellence, right, that we know about Micron is new technologies, ramping new technologies very quickly, but there's the manufacturability operational efficiencies, yield side of things, right? And given the strong pricing environment, I mean, there's not been as much focused by the market on operational and manufacturing efficiencies, right? Also, given the significant mix differentials within DRAM, for example, DDR5 versus HBM versus SOCOM, right? So the cost profiles across all of these are very different. But we want to get an update from you on operational efficiencies. What operational efficiencies has the team executed over the past 12 months? And on a go-forward basis, how should we think about like-for-like cost per bit declines across your DRAM and NAND franchises?

Yes. So again, you had Tim on before. I'm in conversations with all of our core equipment vendors at the C level on every 2-week basis, focusing on trying to address the opportunities we have ahead in terms of greenfield and sure technology transitions are able to move and move forward as well as enhancing the productivity of our existing fleet. And so we are working across all those areas to be able to meet this surge in demand. You're right. Yield improvement is part of that. Productivity improvement is part of that. All of those things in this environment, accelerating technology transitions, all those together do help with our output and our cost. But even with that, we still don't see the ability to be able to catch up the demand that continues to grow for the foreseeable future. And so while cost is definitely always a focus, it's a religion for us. we're focused on trying to be able to increase supply to be able to meet our customers' growing demand, and we don't see the ability for that to align. Now we now see the tightness in supply to continue well beyond calendar 2026.

Yes. So any way for us to think about as we work forward on our models, any way to think about sort of like-for-like cost per bit declines across -- it's such a more varied product profile. Is there a back of the envelope way to think about...

It is getting more difficult. You're right. Because the product portfolio is going -- I mean, I talked about HBM as part of DRAM, right? Obviously, much larger die sizes lead to the trade ratio, much more complex packaging -- all those things lead to different challenges there. Within the DRAM segment, there's also differences even within the DRAM portfolio between LP and DDR high-capacity. And then you have -- on the NAND side, you have a big, big gradation between QLC, the being used for client applications all the way up to performance TLC that's used for data center for PCIe Gen6 we're actually having a lot of success as well. So it's getting more difficult to think about that. And I think we're just really focused on being able to provide a higher-value product. And of course, cost will always be critical, but we're trying to move our portfolio over time, we've been successful. We continue to move our portfolio towards higher value solutions.

Micron talked about on the last earnings call, securing their first strategic customer agreement. You guys call it an SCA, right? It's a multiyear agreement with specific commitments. I assume the parameters are duration, volume, pricing back in mid-March. You talked about securing your first SCA customer. That's 5 years in duration. Since then, how many more SCAs have you secured. And how much of your bit shipment target over the next sort of 12 months do these SCAs sort of now cover?

Yes. So -- you're right. At our last earnings call, we talked about having secured our first strategic customer agreement, and that was a 5-year agreement with a large customer. And we've made meaningful progress on SCAs with other customers who are very interested in having that kind of strategic relationship with Micron. And we'll provide an update on that in the future -- on those in the future. But I think just in terms of the parameters in the way you framed it, I think one of the things that the SCAs really enable us to do is to improve our planning, right? We're planning these. I just outlined all these different capacity expansions we have that are going to take multiple years to build out the fabs and then equip them and having these longer-term agreements with specific terms, help us to align the capacity we're putting in place in terms of both greenfield clean room space and then, of course, how we're equipping the fabs, helps us to better align supply and demand out in the future.

When we speak with some of your customers, especially the large GPU/XPU suppliers. They're telling us that the entirety of all of current gen mainstream HBM 3 and 4 is sold out for next year, right? The only thing that it remains under discussion is next-generation HBM 4E because this is the next-generation architecture, there's a certain amount of customization. E-valve call is going to take a little bit longer. But just like HBM I took a little bit longer last year. But is it fair to assume that the team's HBM3 for supply outlook for next year is already fully committed volume and pricing?

So we're not commenting on 2027 HBM at this time. But I do know that in the past, we did conclude calendar year -- prior to calendar '25. We concluded calendar year '26 in the kind of Q4 time frame of calendar year '25. And at that time, that included both HBM 3E and HBM 4, and the timing that enabled the mix to be aligned between the different technologies for our customers. And so we're not going to break that out specifically right now, but just know that that's been our history and track record in terms of how we've been able to align the mix of new technologies ahead of the year -- the last 2 years.

Yes. Okay. That makes sense. The new HVM architecture ramping this year is HBM 4, you started shipping HBM volumes into the upcoming Vararubin GPU platform starting in March of this year. How has the volume ramp progressed so far? It seems like some of your peers have had some issues on performance requirements? Any issues for the Micron team as you've continued to scale up volume shipments of HBM 4?

So we announced earlier this year that we began shipments and shipments are underway, and our Atrium 4 production ramp has actually ramped twice as fast as HBM 3E 12 high did last year, right, which was a ramp we were very, very pleased with, and we're going at to that rate in terms of production output and our yield is also improving faster there. So which we had said we expected the HBM for yields to come up faster though it's a more complex product. We said that we expected it to come up faster. And based on the learnings that we had from HBM 3E '25, we've -- so we've realized that as well. And we're just excited about our HBM 4 product. It's on our 1-beta platform, which has been really strong workhorse for us. And it has an in-house optimized base die that together is giving us excellent we think it's an excellent product.

And I think you did say in your opening commentary that part of the better outlook, better demand profile that you're experiencing this quarter was contributed by HBM. Is that correct?

I didn't give any specifics on over...

You said -- I think you said better HBM and better overall DRAM portfolio but includes HBM and non-HBM.

I didn't sort for the quarter. But in general, our portfolio has improved over the last few years as we enhance with HBM as we've enhanced with data center SSDs, I think our overall portfolio has strengthened.

Sticking with HBM. HBM 4E is the next inflection. It will ramp in calendar '27 on your next-generation 1-gamma process. And importantly, you've talked about HBM 4E customization options, offering further differentiation opportunities, deeper engagements with your customers. The base logic die offers customization. And that means that the GPU and XPU customers have to engage with Micron 12 to 18 months earlier in their design cycle, sharing IP, co-optimizing the design process flow and so on. So are you at a point where customers have decided on their memory partners to execute the base die design and, therefore, your HBM 4E architecture? And if so, like what's been the breadth of your customer design wins?

So HBM 4 development is well underway, and we expect it to ramp in calendar '27. The first product that will ramp that, by the way, is going to be a genic part, right? So that will be a generic standard part that will be the first HBM.

So no customization is what you're saying...

Not -- the first product -- the first product. Yes, first prior will be a [indiscernible] part. We are working with customers on customization, but the first project will be genic. And you're right, it will have our 1-gamma DRM, which, as I mentioned before, we're really confident. And by then, it will be a terrific node for the high performance for will provide. And we're using TSMC for the logic die as well, and we are expecting that to ramp in calendar year '27.

Would that -- are you expecting to use TSMC both for the geocast die as well as the customized base sizes as well? Or is that...

Okay. That's right. The genetic-based part as well.

Given the customer-specific design of the base die for HBM 4E, like how does the team think about the potential margin profile compared to a standardized base die design, right, that is leveraged across multiple different customers, right? Will the Micron team price HBM for? I know the first 1 is [indiscernible], but beyond that, once you start going custom, will the Micron team price HBM that reflects the customization for every customer. Or will the team leverage other? And will the team leverage other design resources like some of the custom ASIC companies like our Broadcom or Marvell or some of the other guys that do have a track record of quick turnaround customization capabilities for advanced vans logic SoCs.

Yes. I mean, I think what you're seeing with the interest in HBM, the interest in ever-increasing performance and the potential for customization is just how valuable that performance is to the end accelerator company and their customers, right, who are actually driving the AI road maps. I think that's definitely -- you're seeing that the value just continued to increase. And what that really comes from is the complexity -- what that drives is the complexity in engineering, whether that's on the design side, process side -- the core process side, which is absolutely critical, whether that's on beta, 1-gamma and then advanced packaging, right? And I think the Micron team has done a tremendous job over HBM3 and now HBM 4 of being able to drive innovation and proven track or of being able to execute on all these areas. That's driving innovations in design, having really, really strong core DRAM development and then novel advanced packaging to be able to execute on these products and provide value. I think the potential for customization is the same sort of same next instantiation of that with even more value being able to be provided. And so as we think about the path that, that can take, we think that, that could end up -- our customers will be willing to pay for that.

Let's turn to your flash memory French storage franchise, right? On data center SSD, in particular, this is an area where Micron has seen significant share improvements over the past several years, right? Your data center SSD share has gone from 5% to 7% in '22 to 10% to 12% in '23, exiting last year at 15%, right? So very, very strong share gains in performance. And it comes at a time, right, where we're talking about this inferencing inflection, right, in terms of AI compute workloads. And all of a sudden, whether that's KV cash, offload and so on, right? There's all of a sudden, this very, very strong demand pool for enterprise SSD. So again, the team exited last year with global market share. You're now the third largest global market share leader in enterprise SSD. You've got your 550 PCIe Gen5 SSDs out there doing well. You were first to market with your Gen 69650 platform based on your G9 technology. And actually, that is a part of NVIDIA's STX reference platform. You've always had good base NAND process technology. But the other big differentiator, as we know, is you've got to have great controller design. And even more importantly, you have to have a great software and firmware stock. And so does the team expect -- so how did the team over the years, acquired these sort of capabilities. And more importantly, like does the team expect continued share gains going forward?

So Harlan, every year when I come, you always ask me this, I appreciate that you've been able to -- for years been seeing the progress that we've been making, and you've always called it out and highlighted it and definitely appreciate you recognizing that for us because it has been a deliberate part of our strategy over a decade or so to be able to enhance our own internal ASIC capabilities, our own internal formwork development and not just target for me-too kinds of products, but leadership in terms of our ability, like you mentioned [indiscernible] be qualified lead qualified on NVIDIA's SDX platform. And I think that, along with share gains across all the other areas, as you said, going from 5 to 10 to mid-teens really, really shows our track record of deliberate strategy as well as strong engineering execution, coupled with strong technology road map on the NAND side and manufacturing capabilities on the NAND side. But I think the more interesting thing is really what you mentioned about what NAND now means and the high-performance [indiscernible] mean to all of these AI workloads, right? And the agentic AI paradigm shift, you talked about the key value cash, right? It all leads more agenetic workloads and the importance of the key value cash is just all about extending the context window. And the context window has grown 30x a year, right? To give you a sense of how -- and the context window growing is about accuracy and accuracy is about value, right? So NAND is not just about being able to store more data that's generated by various different workloads, whatever they may be. It's also about improving the value and improving the learning rate, and that's becoming a more important part of the AI technology stack.

As we see new tiers of storage and memory emerge as the inferencing workloads continue to evolve, right, we are seeing new memory and storage architectures that may be required, right? One of those is something like high-bandwidth flash. We're hearing more about high-bandwidth flash. But is the team exploring new architectures like high-bandwidth flash as you work with your cloud and hyperscale customers looking at the evolution of their workloads? Are you working very closely with your customers? And how do you see this sort of evolution unfolding?

Yes. So high-bandwidth flash is an interesting technology. It has some positive attributes clearly providing larger capacity and taking advantage of advanced packaging to be able to try and address some of the AI opportunities that, of course, NAND cell by itself is it does have some limitations versus DRAM. And so those are things that engineering needs to work through. But I think the bigger picture here is that there's demand for innovation across the memory higher [indiscernible]. And I think that's just a positive across every single tier, there's more opportunity and innovation. And I think for a company like Micron that has such a strong portfolio and a history of being able to execute in terms of research and development as well as understanding all different types of memories. I think we're very well positioned to be able to be a leader in whichever trend ends up becoming an area that adds the most value to AI technology stack.

Yes. And to that point, you talked about key value cashing. And the first implementation of that was we've seen new storage architectures we like NVIDIA's STX architecture, which you guys are a part of it. But we've seen other hyperscaler companies taking advantage of it by using this new CXL technology, right, memory expansion and memory pooling technology to not leverage SSD, but to leverage more DRAM. So to create these DRAM pools, right, and then leveraging the CXL controller and pulling technology to essentially store a lot of that sort of KV cash sort of implementation. So there, you kind of benefit both ways, right? Whether your customers choose to on these higher -- like high volume, like KV cash offloading capabilities, -- it looks like it's morphed into -- yes, we can do it with DRAM offload acceleration or we can choose to do it with enterprise SSD. But is that kind of the innovation that you're seeing in the industry?

Yes. So I think what you're seeing is that there's no one standard or I mean different cloud providers have different transformer companies each have their own workload challenges that they're trying to tackle. And so they'll come up with architectures that match their -- the long-term workloads that they'll see. And so you can imagine that you're going to see innovation up and down that memory hierarchy. And again, I think we're going to be very well positioned given the breadth of our technology portfolio and the engagement we have with all of these customers, right? I mean by Micron having moved to leadership now 4 generations in DRAM, 3 generations and NAND with leadership of leadership, that's enabled us to be helping define those architectures with our customers. We're not just sitting behind JetX spec on every single one of these. We're now defining -- working with all these different customers on their workloads, helping them think through what can optimize their toconomics, what can optimize there, end applications as they see AI unfolding, whether that's in training or whether that's in inferencing. And so I think that's the exciting part is that memory has -- it's a strategic asset now to help everyone optimize those -- their end -- the intelligence that they're driving. And that's why it's never been a better time to be in the memory industry, and there's never been a better time to be at Micron, actually.

Given the tightness in NAND, I assume that the team is also working with customers on SCAs for NAND as well. Have you secured any SCAs on NAND and SSDs?

Yes. We're making meaningful progress on SCAs and we'll have more comments on the future.

Perfect. Well, we are just about out of time, Manish. Samir. Appreciate the participation, as always. And is shaping up to be another very strong growth year for the Micron team. So hope to be -- hope to continue to monitor the team's execution on that. Thank you very much.

Thanks, Harlan.

Thanks, Harlan.