Please welcome to the stage, Dr. Scott DeBoer, Sanjay Mehrotra, and Gavin Baker.
All right. Thank you guys for being here. Man, memory, DRAM, it was a brutal industry for many long years. I went to the Micron Analyst Day, I think it was in Sun Valley, Idaho.
Not Boise, but I went to the Micron Analyst Day in Sun Valley, Idaho in the year 2000. This brings me back to my roots. But how does it feel after so long to have the sun shining on you?
Scott has been with Micron for 30 years now, and I've been with Micron for 10 years. I've been in the memory industry for 45 years. I have to say that this is absolutely very exciting time for memory. AI absolutely needs memory. AI is all about intelligence. Intelligence gets enhanced by memory. So, the value proposition of memory has just never been higher. Micron is extremely well-positioned, 48-year history of the company, 62,000 patents, very strong team in terms of driving technology and product leadership today, only U.S. company that makes memory today. And memory is absolutely at the very heart of the AI revolution.
When you look at our customer relationships, deep customer relationships in terms of driving innovation, product portfolio, look at how Micron has really transformed itself with manufacturing excellence, close partnership with the customers, and really translating into such strong financial results for the company as well. We are really well-positioned to continue to benefit and be one of the largest beneficiaries of AI revolution.
I think it's helpful, well, first, I guess, just as we talked about backstage, maybe let's just get the near term out of the way. Normally would not ask these questions, but given what has happened in the stock market over the last month, and what's happened to your stock, have you seen any change in the demand environment, over the last month, since you last reported? Anything you would care to say about the state of demand you are seeing from your customers?
Demand remains robust for our business. Supply continues to be tight. Pricing environment is constructive for our business. Actually, it's in line with the strong market environment as well. We have said before that calendar year 2027 and 2028 will be even tighter than 2026. 75% of our 2027 is sold out. We are in discussions with customers now moving to 2028. Look, I mean, AI, as it scales, the frontier models as they get larger, agentic AI usage, larger context windows, all of this just needs more memory and storage. So, our business fundamentals continue to be very, very strong. We delivered a record 2026, 2027 is going to be even stronger.
Well, memory is the most distinguishing characteristic of AI hardware. Like, when we look at startups, there's a variety of accelerator startups, and you're either trying to get more memory, DRAM per dollar of compute, or more memory bandwidth per dollar. Increasingly, we see companies, this is how they think, not about necessarily the compute they're buying, but how much memory per dollar or memory bandwidth per dollar. I just never heard GPUs and CPUs, TPUs, Trainiums assessed on this axis, which is your axis. Maybe just why is, for any AI process, whether it is training or inference, the more memory you have, the better. Maybe just talk about why memory is so central to AI.
And I am sure Scott will add some comments here, but what is important is that people have talked about memory wall for a long time. Compute performance has, if you look at over the course of last 20 years or so, compute performance has continued to advance a lot more compared to the memory bandwidth. So, a processor in an AI system is many times sitting idle. It is sitting idle because it is waiting for data to be fetched from memory and operations to be performed.
That is why you need higher performance memory, you need higher bandwidth memory, you need more memory in order to even get highest productivity out of the processor in an AI system. CPU, GPU, all of them need more memory, faster memory. And you are right to note, Gavin, that this is very different from what it used to be.
It was all about memory price and volume, whereas today, the dialogue with our customers is about performance. It is about, of course, capacity of memory, but it is also about the thermals, and I think these are the kind of things Scott can definitely elaborate on.
Well, we are sure going to dive into all of that, especially the thermals when we talk about stacked DRAM. Stacked 3D DRAM, obviously. But I do think we should talk about LTAs, SCAs, whatever we want to call them. And for the memory industry to even get SCAs is pretty wild. And to me, two things needed to happen. First, the industry really, really consolidated. When I went to Sun Valley, Idaho, for the Micron Analyst Day in the year 2000, I do not even know how many DRAM competitors there were.
Yeah.
Like, there were loads of small fabs in Taiwan. Off the top of your head, Scott, if you have been around for 30 years, how many competitors did Micron have?
In 2000, there was probably somewhere between 10 and 15.
Okay. Now we're down to really three.
Yeah.
For DRAM. Industry structure has really improved. We've really consolidated. Shares are, I would say, in a healthy and stable place. But the other thing is, memory, the engineering, the technical content, the complexity of memory had to increase to a point, combined with the change in industry structure, you could get these LTAs. Maybe just talk about DRAM today, let's save the Stacked 3D and the future stuff, all that awesome stuff. Just DRAM today, specifically HBM DRAM, from a technical complexity perspective, relative to DRAM even five years ago, where LPDDR was.
Yeah.
Like the exciting thing.
Well, HBM is one of the modern technical marvels, really. One of the key pieces about it is it brings in the difficulty of the most advanced DRAM in the world, along with the most advanced packaging technology in the world. The combination of those things, and with HBM now, with the logic-based die, with the foundry-based die, you are bringing in leading-edge logic. A lot of people think when they bring in a leading-edge ASIC from TSMC and they qualify it, that they are kind of done. It is in good shape. But HBM brings in the aspect of designing and bringing in an ASIC chip is only the very beginning. It has to work thinned down with TSVs in it, in a stack, with DRAM performance optimized.
It is much, much more complex, and many times have to go modify the ASIC and everything else that you would have thought worked for other applications. So it is the most complex memory technology for sure, but it is one of the most complex semiconductor products that has ever been built.
Awesome. All that complexity creates just a barrier to entry that was simply not there before.
Right.
Coming to LTAs, my history, 27 years in semiconductor industry, is every LTA, every take-or-pay agreement, every supply chain agreement, they have either been broken or they have ended up as an albatross around the company's necks. Why is this time different? Just to be clear, these are very dangerous words for a semiconductor investor, but I do believe this time is different when it comes to these LTAs. Why do you think it is different?
In the past, LTAs would be one-year agreements. They would be based on customer forecast, and if that forecast changed, you really did not have much options. All you could do at that point within that one-year timeframe is change the mix of the wafers. That is about it. SCAs are multi-year agreements. Many of these SCAs are five-year agreements with the customers, and that is really important because five-year agreements give you that visibility into the demand. Scott's team and our manufacturing teams and business teams work with the customers, not just on next one year what is going to happen, work closely with them in terms of the roadmap and the requirements over five-year period. SCAs are not just five years. SCAs get extended multiple years in advance.
They are giving us the visibility into future demand that enables us to plan our investments, our capital investments, our supply plans to be aligned with the customer's plans. Of course, these SCAs also have take-or-pay, robust take-or-pay provisions. SCAs have cash deposits and financial commitments by our customers. There are 26 SCAs that the company has signed that we had announced at the last earnings call. They have $32 billion of cash and financial commitments upfront.
Of course, many of these also have pricing bands or fixed pricing, and some of them do not have any pricing at all. These are very different from what one-year LTAs would be. These are multi-year commitments. Other important thing is that customers wanted more commitment for supply than we were able to provide to them.
I mean, their demand, the SCA volume requirements over the course of next five years were higher than what we were able to commit to them.
But in the past, what a semiconductor skeptic or veteran would say is, "Hey, we have seen take-or-pay deals before." Let us just imagine that in 2029, for whatever reason, AI semiconductor demand is down 10% or 15%, and the smartphone and PC industry does not absorb that. Customers come to you and say, "Sanjay, you are the CEO of Micron.
We signed this take-or-pay agreement with you, but we would like to break it, and in exchange for breaking it, we will commit to continue working with you. If you hold us to this, when it expires, we are never going to buy from Micron again." In the past, the CEO of Micron would have folded like a house of cards. It just does not feel to me like that is the way it would go this time, and it goes back to the importance of memory.
Essentially, the more memory you have per unit of compute, the better the economics of your AI factory are going to be. Let's just talk about this conversation, whether it happens in 2028, 2029, or maybe it never happens. Let's talk about this hypothetical conversation. I don't think my friend Nafea would ever say this to you, from AWS, but let's just say maybe somebody from the Google TPU team comes to you and says this. Or if, God forbid, Jensen is not the CEO of NVIDIA in 2028 or 2029, the new person who does supply chain at NVIDIA comes to you this way. What do you say back?
These SCAs, of course, have had CEO-to-CEO engagement, as well as, of course, our broader teams in terms of the business teams and customers' procurement teams. I mean, these have been arrived at through very robust set of terms and discussions that have been taken into account in these commitments that are there in these SCAs. You're absolutely correct. I mean, today, it's just not about volume or price for memory, and not just today, in the foreseeable future as well.
I think we have said before that the demand overall is so strong that we were not able to commit to that demand in the SCAs as much as customers wanted, and we don't see when supply catches up with demand, despite our best efforts and large investments in bringing up new manufacturing capacity. It's also about the value of memory.
This is very different compared to anything in the past. Memory is absolutely a key enabler for the customers in terms of driving best token economics, highest efficiency for their compute platforms, getting the highest ROI on their investments. In a given compute platform, you put more memory, you put faster memory, you are able to get a higher ROI from that. These are the aspects that customers fully recognize, and I absolutely do not expect that any scenarios that you're describing from the past are going to be happening here in the future.
I have confidence that our relationships with the customers are so strong, because these relationships are, again, not just, as I said in the SCAs, about volume or pricing commitments. They also go to partnership related to technology and product roadmap that we are working together with them over this five-year-plus horizon.
What one of your CEOs, who's maybe less careful, might say in response to this customer trying to break a supply chain agreement is, "Sure, but if you're not a person of your word, if you're not a company of your word, I don't want to be in business with you. And without me and the two other guys like me, you don't have a business. And everybody's going to remember you broke your word the next time memory is tight, and you're not going to have a business, because without me, you don't have a business."
That might be what somebody who's not an American CEO, that might be what they might say. But maybe to come back to 3D DRAM, this is the next. We've looked at every accelerator startup. I hope a thousand flowers bloom.
You have d-Matrix, who I think is trying to do stacked 3D DRAM in the very near term, run by a great CEO, Sid Sheth. You have Fractile trying to do it. You have a top secret stealthy company that we are not supposed to talk about, that I am not invested in, but everybody knows about, that has a lot of packaging expertise and is doing 3D DRAM. I think the rest of the industry looks at some combination of those companies as kind of like, "Hey, once you have it working, everybody is going to adopt it. Every single person, once it is working, is going to adopt 3D DRAM." Can you talk about why 3D DRAM is important and what it does to your business?
Fundamentally, it is the thing that enables the most bandwidth possible and the lowest power transfer possible of the memory. It also brings into play the opportunity to put more logic directly into the memory product that is ultimately mounted on top of a GPU or whatever the accelerator chip is, and that combination is really the thing that is missing in the AI system today. Literally every big customer is asking us for partnership in developing this technology with the opportunity to co-design a base die, stack memory on top of that, and put that known good product on top of an accelerator chip. In that sense, it is really the thing that moves AI to the next level.
Already today, you look at the options that can be used, and there is no solution for this type of bandwidth and this kind of power capability in a product.
At this point, you are not a commodity at all. You are just as custom of a part as a CPU or a GPU. It is not like you are going to be able to move very easily from.
Right.
If you have your own custom base die integrating with the accelerators logic, there is very high switching costs.
Very difficult and challenging model to put this together. The technology involved with it between the kinds of specialized DRAM that has to be built, the thermal awareness, you mentioned thermals a little while ago, is a pretty intense technology area for this. As you think about different hotspots on the kinds of chips that you are mounting on, it is a big challenge.
Got some feedback. We wanted some more technical discussion. Why do you go into why the thermals are so hard and why this is such a hard problem, especially given that these dies are getting. We are going to more and more multi-die systems on a substrate, et cetera, et cetera. The thermals over a larger area.
Right.
Obviously, Cerebras with wafer scale is the ultimate.
Yeah.
Yeah.
So, if you think about the, right now, the power usage by these kind of GPUs and chips that we are talking about mounting this on, there is an extreme amount of heat that has to come out of the system. And normally, even with HBM, it is a challenge for heat because the HBM is close to the GPU, and the system has to deal with a lot of heat. In this case, you are putting the DRAM, which is a temperature-sensitive device, directly on top of this GPU, and you have to manage to know where the different hotspots are on this chip, whether they are on the edges or if they are random.
And then you have to optimize the thermal properties of the whole DRAM stack, the bonding material, and everything about the cooling system to manage to not overheat the DRAM that is sitting directly on the GPU.
It is a major technology challenge, and the winners in this game will be a major business.
What is the first year that over 10% of accelerators will have stacked 3D DRAM, first calendar year?
I do not know if I know the exact answer to that. I know that you are going to start seeing not one-layer stacks, which is fairly simple and not actually representative of what has to happen to get stacked DRAM, but to get multiple DRAM stacked on top of a processor chip, probably it will start coming into the market in 2028, 2029, and then it will grow from there because the desire to have this is massive because of the impact on AI productivity. So, certainly by 2031, 2030, 2031, you will see a big proliferation of this.
Awesome.
Gavin, HBM was called a technical marvel. This product absolutely goes even.
Oh, yeah.
To the next level. Micron, again, given the deep history that it has, really has all of the technical expertise, I think we really look forward to. These are the kind of products, kind of technologies that absolutely bring us closer to our customers. Earlier to your questions, regarding customers, looking at perhaps ways to get out of the agreements. As I told you, those agreements have take-or-pay, they have robust terms. We do not see customers being able to get out of it.
It is also these kind of products, these products are not just with data center customers. These products are also on the edge devices, Micron is leading with innovation here. These are the kind of things that build trust and relationship in the customers that we fully expect that SCAs will only bring us even closer to them.
Talking about trust and relationship with customers, Sanjay, do you want to say a few words about Jensen? It seemed like maybe you were backstage, you were comfortable talking about this. Maybe just a few words, what he was like as a partner to you and to Micron.
I mean, Jensen, I am a huge admirer of Jensen, for all his pioneering work with modern AI as well as GPU revolution. We work closely with Jensen and his team, and some of the examples of HBM and pushing the envelope with respect to performance on HBM, HBM4 being an example, HBM3E, pushing it toward higher die stack and lower power. Our two companies co-innovated very many products. Of course, our relationship with NVIDIA is across at all levels, multiple levels of engineering. These relationships have been built over time, and we look forward to continuing to work with the NVIDIA team as they navigate through their leadership transition.
You are going to have, by the end of this decade, circa 100% or maybe quite a bit more of your market cap in net cash. What are you going to do with that, man?
We have talked about, for quite some time, that we have had restrictions related to CHIPS Act. December 9th is the date when those restrictions expire. We have said excess cash, we will be returning to shareholders. Of course, company will continue to first prioritize our R&D, our manufacturing investments. We have tremendous opportunity as the only U.S. supplier, and now investing more than $250 billion here in the U.S. to bring leading-edge memory manufacturing. Of course, we are going to invest in growing the business. But yes, financial position of the company and the outlook is quite strong, and we have every intent and capacity to be providing strong shareholder return starting December 9th.
Obviously, you have a very high return buying stock back with your stock. What is it on earnings? I lose track if it is three or four times earnings or whatever it is. There is a very high return on buying stock back. I am just curious, like. I am a tiny little shareholder, but my humble thought for you and the board, and I promise I am not going to write a letter about this.
That's okay.
Is what I would do if I were you, what I would consider, not what I would do, what I would consider is saying, "Hey, we're going to keep $200 billion, $250 billion on the balance sheet." You know what? Whenever we have the next downturn, maybe Micron just goes private. Poof. At a big premium. I think that actually might be a big stabilizing factor for the stock, but you should.
Obviously, the returns on buying it right now are really high. Sanjay, we want to run through three questions with every CEO. In China, 95% of Chinese people are optimistic about AI. It is a tragedy that here in America, a lot of Americans are scared of AI. From your perspective, what is something that AI is doing today to make the lives of ordinary Americans better?
Of course, things such as summarization, translation, I think all of us are able to use that, and every American is able to use it. Think about it, that today in the world, one out of five people is using AI. Last year, it was one in 10. This is advancing fast. The opportunities are tremendous.
I think one of the biggest opportunity that people are seeing benefit, I'm seeing it, when you go to a doctor, I think they are today using AI scribe, and updating the records, giving you more time, instead of doctors, physicians staring at the screen, taking notes, spending so much time. Now they're able to spend more time with you, with the patient. I think that's one benefit that Mayo Clinic has acknowledged, Kaiser has recognized, and the studies have been done over thousands of patients.
I think this is something that ordinary Americans are already experiencing today. Just the benefits that AI brings into healthcare, into drug discovery, I think these are exciting opportunities very much aligned with Micron's vision of accelerating intelligence to enrich life for all.
What are you doing as the CEO of Micron to make sure that what you can do as CEO, that AI is positive for America and ordinary Americans?
We can spend another 25 minutes talking about that. First, we are absolutely investing in the very foundation of AI. I mentioned earlier, investing $250 billion here in U.S., creating 90,000 jobs, direct jobs as well as indirect jobs combined. That's one big piece. Investments here in America, leading-edge memory manufacturing, which is an important national security matter, as well as economic security. Then, investing in building the workforce. When we signed up to invest $250 billion here in the U.S., we actually engaged with the communities right at the start. $250 million in New York to work with the community on housing, transportation, STEM education, workforce development. Recently, we announced $70 million investment in Boise, Idaho, because as we build, we want to make sure that the community thrives with it as well.
Foundation of AI, building manufacturing infrastructure here, creating jobs, workforce investments in order to develop the workforce, and then, of course, in our technology, in our products, making them lower power, higher performance to bring the benefit to the AI ecosystem.
Three years, what's AI going to be able to do for ordinary Americans that it can't do today?
I think just think about today, there are AI drug discoveries, about 173+ are already in clinical trials. In 2023, that number was just two dozens. Imagine a parent with a child with a rare disease, that in the past 10- 15 years of wet lab experiments, hit and miss, 90% of the trials failing, these are the kind of opportunities that parent with that child, today they have hope because now AI is driving drug discovery that will solve these rare disease challenges. I do believe that three years from now, you will be seeing many of those getting addressed, getting solved, and that child will probably have a shot at life that in the past would have lost their life.
Love that. What is more valuable than the life of your child? But I do just want to say, Sanjay and Scott, and to everyone who preceded you at Micron, just as an American, thank you. Intel led the world in logic manufacturing for 50 years. They lost the crown to Taiwan. You were the American manufacturing, semiconductor manufacturing company.
National champion.
[inaudible]
We are a national champion.
That kept pace with South Korea and Taiwan, and is way ahead of China. Thank you on behalf of Americans. Thank you, and let's make sure you stay in the game for the next 30 years.
Living our vision to enrich life for all.
Yes.
Thank you.
Thank you.
Thank you.
Appreciate you. Thanks, guys.
Thank you, Gavin. Thank you.
Thanks, Scott.