2026.09.24 | GP | Memory | Jim Handy Memory Pricing Into 4Q26

With a Memory Expert

Primary:

MU

005930:KS

000660:KS

Associated:

NVDA

285A:JP

688825:CH

Yangtze Memory Technologies Co., Ltd.

Viewpoint:

Former Executive

Moderator: Dan Kaplan (DK)

| Sep 23, 2026 | 40 Min Read

Jim Handy (JH), Former Director - Semico Research Group

SU MMARY

Overview

The discussion focused on the memory market, specifically DRAM and NAND flash, with insights from a seasoned industry analyst. The conversation covered current pricing trends, supply-demand dynamics, and the strategic moves of major players like Micron, SK Hynix, and Samsung. The expert provided a comprehensive analysis of the market's volatility and the factors influencing pricing and supply, including long-term agreements (LTAs) and the impact of hyperscalers on demand.

Key Insights

The expert noted that memory prices are on the rise, although the pace has slowed. Spot market prices have seen significant increases, while contract prices are also elevated, causing OEMs to feel pressured. Long-term agreements are crucial for securing supply, but they also cap price increases, which can limit margin growth for manufacturers. The expert expressed concerns about the potential for an overbuilt AI market, drawing parallels to the internet bubble of the late 1990s. There is a belief that a market correction could occur if AI demand does not materialize as expected. The expert highlighted the challenges faced by DRAM manufacturers in expanding capacity due to the long lead times required for semiconductor equipment.

Competitive Landscape

Company

Market Positioning

Recent Moves and Challenges

Micron

Stable market share

Focused on maintaining a 25% market share in DRAM.

SK Hynix

Adapting strategy

Shifting focus to non-HBM products, thriving despite competitive pressures.

Samsung

Aggressive expansion

Qualified with major clients like NVIDIA, aiming to capture more market share.

Chinese Manufacturers

Ambitious growth

Facing challenges due to US-China trade tensions and IP issues, focusing on domestic market expansion.

Quantitative Insights

Metric

Current Value

Notable Trends

DRAM Price per Gigabyte

$7.50 - $8.00

Significant price increases from previous levels.

Spot Market Price Increase

7x increase

Reflects constrained supply and high demand.

Contract Price Increase

1. 5x - 3x increase

Limited by LTAs, affecting OEMs' cost structures.

Conclusion

The memory market is experiencing notable price increases driven by strong demand from hyperscalers and constrained supply. Long-term agreements are playing a critical role in stabilizing supply but also limiting potential price hikes. The expert highlighted the potential risks of an overbuilt AI market and the challenges faced by manufacturers in expanding capacity. The competitive landscape is dynamic, with companies like Samsung and SK Hynix making strategic moves to adapt to market conditions.

DK: Hello everyone, my name is Dan with Guidepoint and thank you very much for joining today's call, Micron, SK Hynix, and Samsung Memory Pricing and Demand Update into Q4 from a memory expert. If you have any questions, please email me at ask@guidepoint.com and I will work your questions into the call. It is my pleasure to introduce Jim Handy. Jim has been with a leading market research firm focusing on semiconductors, Objective Analysis, for the past 20 years, and he's a member of the Flash Memory Summit for almost 20 years as well. Jim, thank you so much for being with us today. To start, if you could please give us a few words about your background?

JH: Sure, Dan, and thank you for having me on. I'm an engineer by training with an MBA. I worked in design for semiconductor firms, Intel, National Semiconductor, and then moved into a marketing position and did that for a while for memory chips and then finally was recruited into Dataquest, which was the leading market research firm following semiconductors until it was taken over by Gartner back in the 1990s. Almost ever since then, I've been working as an industry analyst following the memory business and other semiconductors, and that's expanded out into watching over SSDs since SSDs are made mostly by memory manufacturers. There was 1 brief stint after the Internet bubble burst, or just before the Internet bubble burst, where I left Gartner, went into industry, and worked for another memory manufacturer, SST. And that lasted until SST had financial problems thanks to the bubble burst. So, I've been very memory focused for my entire career. Although I'm an engineer, I also try to look at things from an economics background, and I've used that to bring about some clarity into the incredible swings that the memory market is well known for and have been able to create some of the most consistently accurate forecasts in the memory industry as a result of that. I should also, yeah, go ahead.

DK: Would you also? Go ahead, please.

JH: Oh, no, I was going to put in a plug for my company.

DK: Sure, why not?

JH: Okay, well, my company, Objective Analysis, is ex-Gartner analysts who have gotten together after leaving Gartner. And we do focus on a broad range of semiconductors and systems. Small boutique firm and we do consulting jobs mostly.

DK: Wonderful. Again, thank you for going over your background. We do have some people who are signed up for this call who haven't heard from you before. We'd love to get your thoughts. We're in late September. What are you

seeing in terms of DRAM and perhaps HBM pricing and supply demand tracking into Q4 as well as any surprises you have seen in September?

JH: Yeah, I see more of the same. The prices continue to increase, although maybe a little bit more slowly than they have in the past. You know, spot market prices we see as up about 7 times for both NAND flash and DRAM. In contract prices, they're up only about 2.5 to 3 times. But I say only, that's pretty extreme. That's bigger price raises than any OEMs have seen before. And they're feeling very victimized by this whole thing. As far as long-term agreements, everybody is signing them. They are viewing that as being their only path towards getting assured supply of memory chips. Meanwhile, the memory manufacturers are trying to add capacity, but they are aiming at a moving target because the forecasts that are being given to them continually increase. And so if they're aiming to be able to produce a certain amount of gigabytes or HBM stacks or whatever, and then their customer says, "Well, yes, we actually need 10% more than what we told you last time," then the DRAM manufacturers' plans, which they have to plan pretty much everything 2 years in advance, they're suddenly out of date and they need to be adjusted upwards. DRAM manufacturers are being very practical about this. They're looking at that and they're saying, "If we build all of this capacity and then the demand doesn't end up fully materializing the way that we expect it to, we're going to be left with very expensive depreciation costs from all the new capacity that's not being used." And so that's something that's giving them a little bit of pause as far as maybe worried a little bit about how quickly they can build out their capacity, but mostly they're limited in their ability to build out by the availability of tools from the semiconductor capital equipment companies like Applied Materials, Tokyo Electron, Lam Research, etc.

DK: Got it. So maybe if we could go a little bit more granular, what's your current view in terms of Q4 2027 price increases? Maybe what are your current thoughts in terms of 2028 by memory type?

JH: The price increases are, DK: And I understand, JH: Yeah.

DK: Sorry, how did that change over the past 1.5 months or so?

JH: And mostly it's through a better understanding about the kinds of terms that are in long-term agreements that the semiconductor manufacturers are asking from the OEMs, that they're not leaving themselves too open-ended as far as price increases go. They are setting what the maximum prices are. The products can go up and we saw in the past earnings calls that that has actually been a limiter as to how quickly their prices could rise and how high the margins can go. That's something that's been a little bit of a surprise just because I see that as something where these companies are trying to decide just exactly how much material should be sold under long-term agreements and what percent of their total output should be allowed to go with the market prices when there is a possibility the market price is going higher. What they're doing is they're trading off those caps against both the bottom prices that they can get the OEMs to agree to pay if prices soften, which there's no sign of doing that yet. And they're also trading it off against the guarantees and cash guarantees that they're getting, but also the guarantees of volume consumption.

DK: So, but in terms of... in terms of market pricing, what are your... what are your current ranges?

JH: Well, we tend to look at things in a price per gigabyte area. And so we're looking at what a year and a half ago was $2.50 per gigabyte for DRAM. Now getting up, it's approaching $10 per gigabyte. Right now they're about $7.50, $8 per gigabyte. So the prices are going up considerably for NAND flash. It's a similar kind of increase. And both of them are under very strong constraints. So the thing that really limits where the prices can go are more whether or not the manufacturers of memory chips will kill off any of their markets by raising the prices too high. And what that shows is how inelastic the market is; what's happening is the hyperscalers are saying, "Okay, we're going to spend whatever it is, $30 billion this quarter," and they don't say we have a target of a certain amount of performance that we want to

buy and once we've reached that performance goal, whether it's measured in number of GPU chips or whatever, once we reach that goal then we're not going to spend any more. They just say we're going to spend a certain amount of money and then however many processors that will buy. And so what that does is cause the prices of the processors and subsequently the prices of DRAM to go up to absorb all of the money that is being put at their disposal.

DK: Got it. So are we still on target for about at least for DRAM 10%, 15% quarterly increases through end of 2027, or are you thinking that's... That wouldn't be supported? Or perhaps could we see the opposite? Could we see actually higher price increases because of supply demand in 2027?

JH: Well, the increase is all because any increase from quarterly revenues, quarter-on-quarter revenues, that is from price increases. And so the thing that's really limiting how quickly that can go up is long-term agreements. There's not really any other mechanism that's getting in the way of that. And that's simply because of the fact of this thing that I just mentioned before, the hyperscalers saying we're going to spend this much money, but not really putting any other measures in there to determine what that means. But it's also that there is highly constrained supply. And so that money ends up getting absorbed. But I haven't tried putting a number of 10 or 15% or whatever it is per quarter to this. But I know that the amount of growth is going to be limited by the terms of the LTA, long-term agreements, rather than by any other mechanism like the market saturates or anything.

DK: Well, it was my understanding that a lot of these LTAs are open price, right? But at the same time, are you saying that while they might be open price, is there still some limitations in terms of how much prices can increase?

JH: No, they're really not open price. Former agreements in past cycles have been where they simply agree on the unit volumes, but they don't agree on the price and the price is negotiated quarterly or monthly. That's not the case this time. This time these are price agreements. And those prices do have a maximum cap. And I know at least in Micron's case that the cap is set to be the current price at the time of signing.

DK: Got it. You know, and that kind of says that the DRAM manufacturers are a little bit shaky feeling about whether or not the prices are going to continue to go up. Interesting. Because I think in our previous conversations we talked about that there is still a markable portion that is outside of LTA agreements and pricing can, of course, fluctuate on that, right? So perhaps we could still see market pricing is just if I hear you correctly, we will not see that in the performance of the companies.

JH: Yeah, that any material during a time when prices are increasing, the prices under the LTAs are not going to go up as much as the prices on the open market. And so that's a question of how does the supplier, the DRAM supplier, balance their revenues? Do they want to have more exposure to the open market or do they want to have a higher exposure to long-term agreements? And today they're preferring to get a higher exposure to long-term agreements.

DK: Got it. But in terms of memory that is outside of the LTAs, do you see any price in there?

JH: Yeah, the pricing is continuing to increase, and that was that 7 times increase that I was talking about with spot market prices for NAND and DRAM. Spot market prices, you have to be very careful about how much credence you give to them. It is basically a clearinghouse for excess inventory. And in a market like what we have today, there isn't really very much excess inventory. And so I usually tell people that the spot market represents probably less than 5% of total DRAM sales. I think that during a shortage like we have today, it's probably a far smaller proportion than that, maybe even 1% or 2%. That means that any violent movement on the spot market is not necessarily going to be reflected in the contract pricing. Okay, so that's one thing to think about. And then the other thing to think about is that spot market prices do react very strongly, very rapidly and much more extremely than contract prices do. So if there's an undersupply of something, then prices are likely to climb far higher, and that's why it's 7 times as much in the spot market as it was in early 2025, and it's only about 3 times as much in the contract market. That's one reason

for that. And so you look at that and say, "Okay, well, the spot market prices are still going up." That tells us something. It indicates that the amount of supply on the spot market is probably shrinking, but it doesn't tell us how much to expect for the contract prices to go up. It just says that they should go up.

DK: Sure. It appears to have been my understanding that the large manufacturers have been trying to price in 10%, 15% quarterly increases lately. Is that something that you think is not holding into Q4? Is it not holding into Q4, and do you think that the LTAs that they're signing do not have room for similar elevations?

JH: Yeah, the LTAs, from what I've been hearing, mostly from my clients, don't have those kinds of terms. They have firmly locked price ceilings. There would be some ASP increase from the migration from HBM3E to HBM4. So that's still going to happen. But if you look at just HBM3E prices or HBM3 prices, those under the LTAs are going to be set at whatever the current price is and then held that way for a year or 2 years. And I say that they've got a cap. The cap, I don't believe that the cap lasts more than a year, although I haven't heard that as an absolute fact. But it's in the non- LTA part of the business; those prices are likely to go up, and then it's just a question of how much are they going to go up quarter to quarter. And it could be 15%, it could be more; it just depends on what the market's capable of withstanding.

DK: Yeah, got it. But will these price increases start trickling into the LTAs a year from now?

JH: It could, yeah. And any new LTAs are going to be signed at higher prices.

DK: Got it. So do you think, again, is there some sort of cadence for these new LTAs that you would be looking at into 2027? And perhaps more importantly, what do you, or as importantly, what do you expect in 2028 versus 2027?

JH: From the things that I've heard, it seems like a new LTA might be signed by a major supplier with a major OEM. They might sign a new LTA maybe once every 2 months. So that's the kind of cadence that I'd look for with that. Now, that said, it might be that that's how it's happened in the past with the easier and more compliant customers. And it might be that the next wave of customers are going to be a little bit harder to pin down to an LTA. And if that's the case, then that cadence could slow. I certainly wouldn't expect for it to speed up, though.

DK: What do you think by 2028? We have potentially new facilities coming up, although the fabs that they may not have may not be at full capabilities, but you also have process changes and you have situations such as Samsung putting some of the less valuable production to their partners so that they could focus on HBM. Does that mean we could see meaningful relief in 2028 in supply-demand, or will some of that additional supply be easily absorbed by the demand?

JH: That ties back to something that I said before. Hyperscalers continue to adjust their requirements up. And so the amount of capacity that's coming on in 2028 actually was committed to earlier this year and late in 2025. And that was back when forecasts were lower than they are now. And so what that implies is that the DRAM manufacturers are continually going to be undersupplied, simply by the fact that they have not been getting consistent forecasts. They've been getting forecasts that have been too low in the past and then they're building to those, and there's that 2 to 3 year lag between when they commit to building a fab and when they're actually able to run it. Nobody, none of the manufacturers are looking at their fab plants and saying, "Okay, we know that we're going to sell this many gigabytes. But let's build a fab that can do 20% more gigabytes than what we know we're going to be able to sell." Nobody does that because the last thing they want to do is to have this larger capital depreciation that they're confronted with during an oversupply when parts are selling at cost.

DK: Got it. Do you expect anyone to gain or lose market share?

JH: The Chinese manufacturers are hell-bent on gaining their market share, and so they've been tooling up for a number of years to be able to do that. And I'm expecting to see them do that. They're hoping to be able to sell to outside of China, but that opens up a huge amount of intellectual property issues, plus there's the whole US-China trade war going on. So that's going to be very tricky for them. And so probably what they'll do is they'll just increase what their shipments are to China. I know that US OEMs who are dissatisfied with the amount of product that they're getting from the Korean, US, and Japanese manufacturers, they're trying to find a way to get the US government to allow them to purchase and use parts in the United States from CXMT and YMTC. But so far, that is something that's more wishful thinking that they'll have to work very hard to convince anybody in government that that's the right path to choose. And they do have a very strong argument where they can argue, "Hey, do you want US OEMs to start lagging behind OEMs from China or from other places because of the fact that they can't get the DRAM that they need?" So I'm sure that they have a number of their representatives' ears by talking that way, but then on the other hand, you've got US manufacturing, and Samsung ends up behaving like a US manufacturer because of its large facilities in Texas. And those companies are taking the exact opposite stand and trying to also convince their legislators that the Chinese manufacturers should not be given any chance whatsoever to get into the US market.

DK: What about in between the Western suppliers? Do you expect anyone to gain or lose share perhaps now that Samsung is qualified with NVIDIA and HBM? Can they take some of that share and if so, where could they get it?

JH: Okay, I missed the very first part of your question. Which suppliers are DK: Yeah, JH: you talking about?

DK: Mostly the Western suppliers, so US, Korean, perhaps Japanese.

JH: Oh, so you're talking about, DK: Are there any share shifts between those?

JH: Yeah, and there are some. Micron has reached a point where it believes that it's got... Micron's got this idea that something like 25% is what their fair share is of any of these DRAM markets. And so they're at that point. So they're happy to do it. SK Hynix has been supplying everything that Samsung has been unable to supply. And so Samsung does want to take share away from SK Hynix. And SK Hynix, there's an interesting dynamic between SK Hynix and Samsung where SK Hynix and, well, both companies view Samsung as being the natural leader and that if Samsung wants something that it will take it away from SK Hynix and SK Hynix doesn't fight hard to get it back. So we see SK Hynix actually moving towards producing more non-HBM products at a time when the non-HBM products are more profitable than HBM. So that's working very well for Hynix. It's kind of amazing how Hynix, over the course of my career, I've seen Hynix that is being hit with things that other people would see as a competitive threat and they end up thriving from it. This seems like yet one more instance of this. So I would see that, but I don't expect to see any other real share changes between Western suppliers until something is resolved between what's going on in China and the United States. What's interesting, though, and this is a little bit of a departure, is that right now the industry is extraordinarily profitable. And during times like that, that's when you usually see large industrial firms that do not have a semiconductor arm decide that they want to become memory suppliers. I've seen this happen in Japan in the 1970s, in Korea in the 1980s, and even in Taiwan, there was a short-lived push into DRAM in the 1990s. And so I wouldn't be in the slightest bit surprised to find, for example, companies in India suddenly pop up into the marketplace, as long as this undersupply lasts as long as the DRAM makers are in place as those LTAs are in place.

DK: You mentioned that DRAM will remain under supply. Is there a timeline for that or are we talking about foreseeable future? Or I think I've heard SK Hynix say 2030, may be through the end of 2030.

JH: Yeah. And how many people forecast accurately to that point? I certainly don't. But I look at this whole thing and

what's going on with AI and I say, "Wow, this looks a lot like the Internet bubble in the late 1990s." And in the 1990s, the Internet was a new thing. People were just learning how to use it, and everybody knew that it was going to change their lives. And there were very large investments for it, and there was an overinvestment. And suddenly, the market came to the realization that the internet was overbuilt and it threw us into the internet bubble burst in 2000. That same thing I am worried is going to happen with AI. Everybody sees AI, says it's going to change our lives. It's a good thing. There are massive investments going on now for future demand, which has not yet materialized. And I myself believe that there is going to be at some point a coming to terms with the idea that the industry has overbuilt capacity and that there needs to be a cooling off phase. And with that should come a semiconductor collapse. But I can't tell you when that's going to happen because I don't know when AI is going to be overbuilt and when hyperscalers are going to decide to decelerate their spending.

DK: Sure, and it feels at this point that as soon as someone can get a GPU working, they have a customer, right? So from that standpoint, it feels very different from dark fiber that was overbuilt in the previous... during the early 2000s bubble.

JH: Yeah, but the hyperscalers are not profiting from all of these GPUs that they're finding uses for. There are a lot of people who are using free or low-priced tokens because of the fact that they're free. And there's some companies that are even looking at the amount of time that they've got employees using AI and saying that's not time that is being profitably used and so we're going to ask people to back off on the use of AI. It's not happening in any kind of a widespread sense but the fact that it's happening at all indicates that there are things that are going to moderate AI's use. I happen to think that once people start paying what the hyperscalers' cost is for the tokens that they use, that that's going to change demand significantly.

DK: Well JH: All right.

DK: I guess we're not here to debate that, but anything specifically you're looking for as a first canary in that as well as in the memory market?

JH: It's not really easy to tell. And I wish every investor asked me this question, "What is going to tell me that the market has changed?" The first indicator that most people look at is spot market prices. And that is something that, although they continue to go up today, at some point there is going to be somebody who puts some excess inventory out on the spot market. There's going to be enough that it will cause prices to go down. And that would be an indication that somebody's clearing inventory. Why are they clearing inventory? Because they've overbuilt or because they've overcommitted to how much DRAM or NAND flash they're going to buy for this. But the spot market does jump around a lot. So I wouldn't go and sell the entire portfolio based on a downtick. I'd wait for it to show that it was really headed somewhere before doing something like that. As far as it helps to obsess over what the hyperscalers are saying and to see whether or not any of them say, "Well, we're going to slow down our spending." Because that also would be a good indicator, but I don't have any simple things to point out there.

DK: Does it mean that maybe inventory levels are a better predictor than spot pricing?

JH: Yeah, except for the fact that people don't report inventories very accurately. So, you DK: Right.

JH: know, you have a really good point.

DK: My next question was going to be what's the shadow inventory and how do you watch that?

JH: No, and you don't know. And also something about inventories is that you don't know with OEM inventories how much product they have built as a buffer because of the fact that they worry that the shortage is going to get worse than it is today. And that's the thing that really causes the extreme speed of semiconductor collapses. With a semiconductor collapse, or I should more specifically say a memory chip price collapse, which is the thing that always takes down the whole semiconductor market. So when those start to happen, then it usually only takes about 2 quarters for them to go from very high gross margins down to near 0 gross margin. It's just an amazingly fast collapse, and I've talked about it being 60% over a 2-quarter period. But where we are today with our 80% gross margins, I guess that's going to be 80%... so just a very enormous fall. And that kind of thing could happen pretty much at a moment's notice, partly because of the inventory that there will be an inventory digestion phase where the OEMs are going to look at their inventory and they're going to say, "Okay, we now have to write down our inventory a few times every quarter because the prices go down so fast." And so we're going to completely cut off buying until our inventory, which we've grown as a protection against price increases or against shortages, we're going to take that inventory down to the point where we're just barely able to satisfy the needs of our immediate production. That's something that's always a huge influence is to how hard gas prices fall.

DK: Did it feel to you that some of the price increases were driven by these OEMs that were looking to build up inventory to make sure that they can actually fit the product?

JH: Yeah. And it's not that the prices increases are driven by that, but it's that the demand was driven by that, that they know that they're going to need whatever X number of gigabytes and they order 10% more than the X number of gigabytes that they're going to need. And by doing that, they put pressure on the market and the market prices go up. And it could be that that 10% increase is going to cause market prices to go up by 1%. It could be that that 10% increase is going to cause prices to go up by 15%. It's just really hard to tell.

DK: Got it. Let's see. Have you seen any observable changes to capacity build-outs for DRAM and NAND over JH: It's not easy.

DK: the past months?

JH: Yeah, no, this is something that is planned years in advance and all of the tools or nearly all of the tools are custom built. And so because of that, tool manufacturers need to have very much advanced warning. And you can't really make changes that quickly with that. It isn't like tool manufacturers can go out to a farm field and recruit a bunch of farm workers to go and build their tools for them. Their tools are all built by people who have high degrees of technical specialization, probably a very skewed number of people who have science degrees who are putting these things together and so because of that, there's just some very natural limits as to how many tools can be cranked out and, if Micron, for example, goes to an Applied Materials facility and says, "Well, we thought we needed 1 of these etchers, but we actually need 2," Applied Materials is probably going to say, "We just don't have the capacity to give that to you in less than 3 years." So it's a toughie.

DK: I know. Doesn't SK sign itself some contracts at this point that gives it a bit of a control in terms of what the industry can get?

JH: I'm sorry, they have some contracts in what terms?

DK: I believe they have contracts with ASML for a large amount of the production, but ASML is increasing its capacity.

JH: Yeah, they're doing it. I'm just saying that that's something that doesn't happen at a moment's notice that there's a long path that they have to follow to be able to increase. And I know that SK Hynix has been talking to ASML, I'm sure, for the past 2 years about how hot this market is getting and encouraging ASML to devote more capacity to them.

And I'm sure that ASML is also working very hard to increase its ability to produce new tools because of how strong this market is. But it's not something where if SK Hynix says this quarter that they want to increase the number of EUV machines that they get, they'll be able to get delivery next quarter. It's years away.

DK: Nope. In terms of capacity build-outs, especially for NAND, I think Solidigm mentioned a US facility, also CXMT is thinking that they might go into NAND, but I know this is very recent. Samsung also mentioned the US facility. When would all of that come into being? When could we see that relief and would you see— is NAND easier to produce from the standpoint than DRAM?

JH: No, actually NAND is probably a little bit harder. The reason why is because a DRAM wafer takes about 2 quarters... well, I'm sorry, no, it takes about 1 quarter to produce a DRAM wafer. It takes between 1 quarter and 1.5 quarters. For a NAND flash wafer, it takes 2 quarters to 2.5. So it ties up. It's just a very slow process to make NAND flash. So I don't see that as being anything that's going to happen rapidly. DRAM and NAND flash fabs are not compatible with each other either. Your DRAM fab is going to be very lithographically intensive. More of the tools are going to be photolithography tools. With NAND flash, it's mostly deposition and etch, so you can't change these fabs back and forth or anything like that. It slows down how quickly a company can do something like that. I would imagine that any announcements... well, no, I know that any announcements that are made by a DRAM manufacturer about any new fab are made after negotiations are made on what the site is going to be and after they've planned out how they're going to tool up the fab and that sort of thing. So these are plans that probably were started over a year ago that are being discussed now, with the exception of CXMT. I think that the CXMT announcement from what I've seen on that is really an announcement at all. I think it's just kind of rumors that have escaped the company. And so it could be that CXMT is farther away from their plans than the other companies are who have a more formal rollout of what they're saying.

DK: Right. So if Solidigm already has a site that they're thinking of, they've already been planning it for some time, right?

JH: Yeah, they went through negotiations. And when you choose a site to build a semiconductor manufacturing plant, you choose it based on tax incentives. On your negotiations with water and power and the ability to service it by roads, by airports and all of these things would have been under secret negotiation for probably well over a year.

DK: What are you seeing for long-term demand? Because we have seen some changes to memory architecture, for instance, NVIDIA cut their CPU DRAM by half. Are we seeing any reason why perhaps some of this hyperscaler demand or OEM, of course, demand could slow down?

JH: One of the things that I have to do as an industry analyst is to talk to a lot of people, and I talk to people who use GPUs, who buy these services and that, and the complaint that I always hear is that there isn't enough HBM on a GPU. The way that the computer people measure it is they measure what percent of the CPU capability in the fleet of GPUs that they're using, what percent of the GPU capability is being used and what percent of the DRAM capability is being used. And they find themselves using 100% of the DRAM capability and only about 30% of the GPU capability. And so what that means is that if you were to present them with a GPU that had 3 times as much DRAM on it, they'd say that is a perfect balance. I have never heard anybody say that it was the opposite situation for them where they needed more GPU and less DRAM, less HBM. And so what that means is that these new GPUs with less HBM on them don't really match the needs of at least any of the people that I'm talking to about what their needs are for GPUs. So I'm mystified as to why NVIDIA would be bringing anything like that out.

DK: Is the issue for them bandwidths or capacity or both?

JH: It's generally

DK: Because JH: capacity.

DK: I think NVIDIA's point is that they're keeping the benefits.

JH: Yeah, that's nice, but the complaints that I hear are about capacity. And that's why Solidigm and SK Hynix are excited about high bandwidth flash. This high bandwidth flash, it doesn't quite have the bandwidth, the read bandwidth of DRAM. It has atrocious write bandwidth, but that's not important for the applications that they're looking at. They're looking at inference, and inference needs reads and it doesn't need writes. High bandwidth flash is something that increases the memory size that's available to the GPU without really degrading the bandwidth.

DK: So maybe that's what NVIDIA is thinking, that, okay, we will put less HBM because it's tough to get, right? And the customers will just use other memory to get the efficiencies.

JH: Yeah, the other memory that's available to them now is memory that has not got that high of a bandwidth because it's memory that's off the chip. This would be...

DK: Right JH: know, LPDDR or standard DDR or, at worst, NVMe flash. But the high bandwidth flash is not going to sample into the middle of next year. So it's definitely not available for current NVIDIA models of GPUs. And these high bandwidth memories can't be added on as a later upgrade to any system. They have to be built into the same package as the GPU at the same time that the GPU is being built.

DK: Any thoughts if it delivers on promise? Any thoughts in terms of pricing for HBF or margins for HBF, is...

JH: Yeah.

DK: that going to be higher than the current 80%?

JH: No, I don't think so. It's really hard to guess what the prices are going to be. I would just, if I had to hazard a guess, I'd say it was the same markup that we used to see for HBM, which was 5 to 6 times the price per gigabyte of standard DRAM.

DK: Is the cost to produce going to be similarly higher or is JH: Yes, DK: it going to be proportionally?

JH: There are 2 things that make HBF more expensive. 1 is that standard NAND flash is designed to be as cheap as it can be designed to be. It's got an awful lot of shortcuts are taken to make the die size very tiny. And part of that is that the read performance is really bad. Part of it is MLC and TLC is that I believe that when HBF comes out, it's going to be SLC, which means you're only storing 1 bit per cell because that's the fastest way, the fastest kind of read NAND flash that you can get. And so all of these things are going to be a problem and there are going to be many more sense amplifiers than there are in standard flash, which is something that pushes the chip size up. And that's going to make it larger, but also it's going to have through silicon vias, TSVs, which are a large part of the reason why HBM DRAM is so huge that Micron likes to say that HBM3 consumes about 3 times as much wafer area per gigabyte as a DDR DRAM. And that same thing is going to be true with the high bandwidth flash, but probably even more so.

DK: Got it. I would like to move on. I think Samsung stated that LTAs cover 50% of fiscal '27 and 2/3 of fiscal '28. Does that mean we are at the point where we should see reduced cyclicality? And also, is that similar now between NAND and DRAM now that we have some of the more similar use factors?

JH: Well, the more similar use is definitely there because both of them are being absorbed in great, new amounts by AI. And there's a lot of similarity to PCs as large NAND flash users as they are large DRAM users and cell phones are also large users. So all 3 of those applications are moving together. What I mean is the NAND and DRAM move together in AI just as they move together in cell phones or they move together in PC. And because of the fact that AI is absorbing so much NAND flash and DRAM, then that's causing shortages in all of those markets. As far as the reduced cyclicality, my experience has been that these LTAs... Any time there's a shortage that the memory manufacturers ask the OEMs to sign LTAs, and when that shortage turns into an overage, the LTAs don't really stay in place. And the DRAM and NAND flash manufacturers try very hard to find ways to make them irreversible. What kind of happens is that the OEM will say, "Okay, you can hold me to the LTA, even though the prices are lower everywhere else than what I'm paying. But if you do that, then I'm not going to be able to sell any more cell phones or I'm not going to be able to sell any more PCs because my prices will be too high. And as a result, I'll go out of business and I'm one of your best customers. What are we going to do about that?" And the answer, the only answer to that is, "Let's tear up the LTA and continue on with business without one." So that's what I'm expecting to see happen during whatever the upcoming downturn happens. Something that I don't buy is the idea that the cyclical commodity DRAM and NAND flash markets are no longer commodities. They are commodities. They're behaving just like other commodities that have huge price increases. And I'm expecting for there to be a down cycle... some people are claiming it's never going to happen again. I think that's just completely wrong that these companies are going to see another down cycle. And it's just a question of when it happens rather than if.

DK: Some of these have even third-party financial guarantors, although for relatively smaller amounts. Is that relevant?

JH: Yeah, no, that is. But, once again, that's something that's subject to negotiation. Everything can be negotiated. And if there's an amount of money that's being held in escrow or prepaid up front or something like that, in order to make the LTA work and suddenly that money is required by the customer, the purchaser, because of the fact that they need it to stay solvent, then they're most likely going to be able to convince the seller that the seller has to make a concession and give back some of that money.

DK: Right. We'd love to get your thoughts on China. So what do you think of CXMT that has started LPDDR6 production as well as is reported to have started initial production of HBM3E? Have they materially closed the capability gap and how does that look forward for them?

JH: Yeah. CXMT, first of all, let me say that CXMT was founded by the members of Qimonda's old China Design Center. That Qimonda, the German DRAM arm of Infineon, was actually a premier DRAM manufacturer. They had excellent technologies. They just didn't have a good cost structure. And so during downturn, I don't remember which one it was now, they ended up going insolvent. The designers, though, are extraordinarily capable people. And the only thing really holding CXMT back is the availability of equipment. China is trying very hard to create its own internal equipment industry, which is an extraordinarily difficult business to be in. They are doing it by copying tools that are already in the country from leading manufacturers. But there's only so much you can do with the copy before you run into a place where you don't understand how they did this 1 little trick. So they've got that working against them, the fact that they can't get tools exactly right. Now, what does that do to them? That pushes their cost up a little bit, but it doesn't really prevent them from making much of anything. With HBM, I think it probably does because I think that perhaps their HBM die size won't fit into the package if they were to make an HBM using the processes that they have now. And when I say HBM, they're probably quite capable of making HBM2 using processes that everybody else would have been using 3 years ago. And that's fine. But as far as competing on the LPDDR marketplace, if they're using a technology that everybody else was using 3 years ago, that means that everybody else's cost is going to be lower than theirs. Now everybody else is making 80% gross margins. Let's say CXMT is only making 60% or 70% gross margins. Today, that's still just a wonderful thing. Sometime during the next downturn, everybody else starts making close to 0% gross margins. CXMT is going to be making negative margins, and it's going to be very difficult for them.

So the fact that the market is in this huge shortage today is actually benefiting CXMT and YMTC enormously because of the fact that they don't have to worry as much about the costs, which would have otherwise been a big hindrance to them.

DK: I guess with China being able to pull a rabbit out of the hat? Of course. Is there room for them to have similar margins while using older technology, but perhaps re-optimizing or?

JH: Not really, no. Something about the semiconductor industry is that there aren't really an awful lot of secrets as far as process technology goes. They all buy tools from the same manufacturers, and those manufacturers are motivated to share any learning that they've done with one chip manufacturer to share it with other chip manufacturers. So for example, 3D NAND, when 3D NAND came out, SK Hynix had a different process than Toshiba had. This is Toshiba before it was Kioxia. And Toshiba invented the process, and that was different than the Samsung process. And the Samsung process got put into volume production before the others, and that caused everybody else to follow the Samsung process because suddenly the tools became cheaper for the Samsung process than they were for the other two processes. So now everybody uses something that's similar to the Samsung process. You see the same thing now with, you know, CXMT is doubtlessly using a process very similar to everybody else's 1Z process. And everybody else's 1Z processes all look very much like each other. So there typically isn't very much difference.

DK: Gotcha.

JH: Then, if CXMT has a 1Z, whatever cost structure. And everybody else is playing by 1B or 1 gamma, then everybody else is going to have a cheaper die. But, like I say, when they're making 80% gross margins on their cheaper die, CXMT has more expensive die that are only going to be making a 60% or 70% gross margin, that's still okay.

DK: Right. So these things, as they continue gaining share, if I understand correctly, they bought tooling to increase the capacity by about 60% over the next few years per annum.

JH: That would be tooling that they put in orders for a couple of years ago, or a year ago, let's say. So it could be that they did that. I don't really track the restrictions that are put on companies like CXMT as far as what kind of tools that they've got, and so I can't really weigh in on whether or not they are capable of getting delivery of the tools that they want for that. But I guess the best that we can do is just to listen to them and say, "I guess we believe it." As they're a public company now, you'd expect for them to be very forthcoming with the things that they disclose.

DK: I think we are at our time. So I really appreciate your time. And so thank you so much for joining today's call. Thank you everyone who was able to listen in and hope everybody has a great day. So thank you so much.

JH: Yeah, thank you, Dan. I hope it's helpful to your clients.

DK: Take care. Cheers. Thank you.

DISCLAIMER: The statements or opinions expressed today are those of the Advisor and not Guidepoint, who disclaims all liability for the content provided. The Advisor may not disclose material nonpublic or confidential information or any information that would cause the Advisor to breach any duty or obligations. Guidepoint is not a registered investment advisor and the information provided is not intended to constitute investment advice.