Loking to train.
So I'm Joel Webber and I'm Eric Belchernas.
Eric AI trade has been a constant for about two or three years now, but today we're going to talk about a little corner of it that has become very interesting.
Yeah, it's almost like downstream from AI, these AI data centers which they're building like crazy. This isn't news to anybody. They need memory and high bandwidth memory in particular. And there's like three companies that have almost the entire market s k Heiniz, Samsung, and Micron. Two of them are
South Korea, one US. Now, there was a push though all three of those thocks are up a ton, so people were buying the South Korea ATF and we're seeing an inordinate amount of flows into that ETF earlier in the year and last year. But you don't get Micron with that, but you still get too at pretty good way. But you also get a lot of like other South korea'stocks, so like you're going to get K pop stuff in there.
It's like a proxy, but not a one.
Then Roundhill saw this, which again heads up play and they put out a memory ETF recently, very recently, and it just blew up immediately. We're talking two hundred million dollars of trading right off the bat. Now that's normal for like the spot bitcoin ETF for or the first ever like Ether ETF, that was pretty normal. But you don't see that ever, especially for them ETF. Normally you're lucky to trade two million, not two hundred million. I'm like,
what is going on here? So I figured out the two South Korea companies aren't in ADR format yet, meaning you can't really buy them in the US, so your best way is to use the South Korea ETF, and that doesn't have micron. So this has let's give you all three in about a sixty five percent waiting and people were like going crazy for it. They must have
let the market know. And it's at this point over a billion dollars now to keep that, to put that in perspective, the top four biggest the matic ETFs, it took them all over one thousand days. In some three thousand days, they hit a billion. So d Ram pulled off in ten days. So again we have a massive
outlier smash hit in her hands. So I went and worked with the Bloomberg intelligence analyst who's a specialist to write a three Ways the Play note, which is one of our new recurring notes, and we're here to talk about it.
That ETF by Roundhill is called d ram and here to speak with us about it is Jake Silverman, who's a semi conductor analyst at Bloomberg Intelligence. He's going to cover more than that as well, this time on trillions the memory trade. Welcome to trillions.
Yeah, thanks for having me.
Okay, So what Eric just laid out for us is a wildly successful ETF launch. It's been pretty interesting times in the semiconductor business. Why did it take so long for an ETF like this to come to market? Do you think?
Yeah?
Yeah, I think what Eric pointed out it was interesting is that the lack of ADRs, so the South Korean companies specifically in this case Samsung and s ke Heiniks really haven't had that ADR format before, and so I think there was just kind of folks are trying to understand the best way to play it, and so creating this ETF format was probably just sort of the key to unlock that. And so there have been a couple ETFs, as you pointed out, based on Korea, and you know,
Korean stocks and et cetera. And so it is interesting now because there's such a memory up cycled, there's just such a renewed interest in memory.
Okay, so when we think about chips most of the time, especially with AI, we're thinking, like Nvidia, maybe the Google stuff. Memory is sort of old school. I kind of forgot about it, to be honest with you, Like, why is this taking off?
Yeah? I would say memory is usually considered the boring aspect of semiconductors, but it's gotten a lot more exciting, and a lot of it is to do with high bandwidth memory. So basically, you take these typical dram dyes that you would get off of just any wafer that would come out of these fabs, and you stack them up and you use these things called through silicon via to connect them microbumps, and you put them on a logic dye that then connects over to the GPU. It's super complicated.
It's way more interesting than can we get that in English.
It's basically just stacking DRAMs that you're maximizing the amount of space you get on a GPU board, and so this way you get super high bandwidth that you can by maximize. I don't want to get too into the weeds with it, but it's basically maximizing the bandwidth based on a limited amount of space you have on a package for a GPU.
In general, the idea is if people are building AI data centers and they want to store all this data, they clearly want to they need memory to do that, and they I want it to be in a small, smallest physical form possible. So there's just just this a demand and then be pressure to make it small. And so that's what's going on, is that right?
Yeah.
So one of the things that probably to keep in mind is when you think about some of these models like from open AI that come out and they have like trillion plus parameters, you have to store that memory. And then when you're actually interacting with a model. So let's see if you're on chat, GBT or claude or whatever and you're actually typing into it. So what it's doing is there's this thing called kb cash where it's actually storing the amount of data and then it's appending
it over and over again. But what you want to do is so that the model is don't have to run through your entire context window every time you're typing and keep doing it over and over again, it'll store that data. So basically what happens is you need memory to interact with the model, and then when you're adding to the context for the model, it needs to interact with that as well. So basically, the more the larger these parameters are, the larger these context windows, the more memory you need.
You're in equity analyst first and foremost. And we've got three companies here esk Heinex, Micron and Samsung. Distinguish them from each other or are they literally just plug and play versions of each other?
They're very different one I would say Samsung is the most differentiated because they have various business units. So when you're buying Samsung, for example, you're not buying just memory, You're buying their smartphone business, You're buying their foundry business and the various other businesses that they have. Within that, esk Heinex and Micron are considered, or at least more so in the past, we're considered interchangeable. But the difference
between Micron and eske Heinix is twofold. One. You have the capacity that eske Heinix has, that's a lot more than say Micron, and then the other aspect is that they've been the leader on a high band with memory for a while now, so when you compare it to micron A, Micron was playing catch up and has been gaining share, but sk Heinix has still had about double that share over the last I don't know, say twelve months or so roughly, So it's a question of technology leadership.
Micron is still definitely in the mix. And then there's a capacity element which we can get into in a bit.
Eric.
I'm not gonna ask Jake this because I think he knows. Do you know where Micron is based?
I would guess like Seattle, No, you'll never guess, West Virginia, No, Mississippi, No Ohio, Missouri, no.
Louisiana, all over the place. Idaho? Oh yeah, was that next?
I mean the only thing I actually drove through Idaho once on a road trip, really cool, just long stretches of road built to spill is a famous indie rock band from Idaho. That's the only thing I think of in Potatoes, Potatoes and Micron.
Like like that.
Yeah, that is wild. Yes, I mean, look this industry, obviously, these companies have been around, but the AI demand you have here that it's seventy five billion dollars this memory industry, and you think it's going to double in only two two years to one hundred and thirty five billion. That's pretty wild.
Is that bookcase or barecase?
So I think in the past, I would have just talked about memory as an entire industry, and now it's like you have two product sets. You have high bandwidth memory, which is just completely outpacing I would say at least unit growth the thing that's gotten. So it's kind of
important to understand. So because I was talking about the structure of high bandwidth memory, it actually requires three to four times the amount of space per wafer compared to the traditional products that we see that it's like you'd have in your smartphone or your laptop, anything like even traditional servers that go into like say the cloud or something. So what's happening is there's a limited amount of capacity.
They're consistently continuously upgrading that amount of capacity. We don't have to get into the specifics of how they do that. But because you shift those wafers over to high bandwidth memory, you're taking away capacity from other products. And memory is very supply and demand driven way more commodity like than the rest of the semiconductor market. I mean, obviously GPUs are the least commodity like at this point, but memory
is the most commodity like. So if you're taking off supply and demand keeps going up, pricing is going to increase at a fairly high rate, which we've already seen and this year alone is probably going to increase anywhere between one hundred and two hundred percent, if not more. That's just d ram Land is also similar.
Well, when we hear all these big tech companies, some of the mag seven are spending all this cash, even taking out debt to spend cash on AI. This is where they're This is where they're spending it, right. It's not rocket science here. They got to they got to spend it on stuff.
By the way, you said d RAM there, you were referring to the memory itself, not the ticker, but the ticker. Yes, we gotta talk about the ticker for a second. When you saw that ticker, d RAM, what did you think.
I thought, this is the best name for an ETF for memory I could come up with.
I honestly called it DRAM for a while, and then I got educated.
Well, you're still still working on that, so Idaho and Drama.
It's an ongoing education.
So when you look at what this ETF holds, by the way, Eric, we've got those three companies. What else is in this sticker?
Here are the rest of the holding. So those three stocks make up sixty seven, so two thirds of the portfolio. Then we got forgive me on my pronunciation. Okay again, I'm an ETF guy.
Jake's gonna help you.
Okay, Kyoksia, how's that pretty good?
Good?
Okay, Kia Holdings Japan, sand Disc, that's US another retro one Seagate, US Western Digital Retro, Nanya. Yeah, that's in Taiwan, Win Bond in Taiwan, and that's it. Those all the last ones I name probably collectively make up again to describe that portfolio.
Jake.
Yeah, So when you expand beyond eske Heinex, Samsung, Micron, you get into sort of more like nand specifics. So sand Disk and Kyoksa have a JV to do research and development and they share the same fabs together. But that's nand only. So the ticker for the ETF is d RAM. But within that you also have nand as well, because Micron, Samsung and Eskahannis make DIRAM and nand nand
is more like storage. So if I if you bought it, buy an iPhone, you'll see like it'll have tiers of storage, for example, five hundred and twelve gigabytes, one terabyte that.
Stuff that fills up with the gooey storage. Yeah yeah, I'm like, what's on my iPhone?
It's like when I take too many pictures of my dog and now I need to back it up on the cloud. Yeah, and then I need more storage. So someone's got to buy that storage.
Yeah yeah, Okay, So when you think about what we just rattled off, there, is there something that's missing or is this literally the perfect portfolio for this trade right now?
It's pretty good. So non Ya Tech, for example, is pretty small. They are Taiwanese based, But there aren't that many pure play memory suppliers out there, and that's because memory has been historically tough industry, a lot of bankruptcy, a lot of consolidation. That's changed over the last two decades roughly as that's kind of stabilized, But there aren't that many players out there that exist today that simply
just supply those products. Once you expand beyond that, you start to get into the suppliers of those businesses and those are a lot more diversified. So if you're looking for a specific way just to play memory, that's probably the best pure play opportunity that you're going to get that also has a little bit of diversification within there as well.
You said that memory is like almost like a commodity, right, And if we think about how commodities trade, there are these fabulous booms and bus right, and like a commodity supercycle kind of means that everything's going up and then eventually the air comes comes out of the trade. Where are we in that supercycle commodity trade cycle?
We're definitely a boom cycle. Where in that cycle is I think the million dollar question for every portfolio manager and retail investor out there, you know, the question is like are we in inning two? Are we in inning seven? Are we starting to approach the peak of the cycle? And that's where these stocks tend to You either tend to make a lot of money or you tend to
get in a lot of trouble. There's no question right now that we're in an unprecedented upcycle like we've so we've done analysis where we kind of look at the length of these cycles and they typically last somewhere between I would say seven and ten quarters. We're well beyond that now based on the reporting and the guidance from these companies. So the question is, Okay, if we're beyond the sort of ten quarter roughly peak that we've seen in terms of that upcycle, can we go another ten quarters?
The question is then becomes, okay, is AI demand enough to sustain that? And then if I'm if we are in a very commodity like industry, how to supply factor into this as well. Our base case has kind of been more like, okay, we think come maybe second half of twenty twenty seven, twenty twenty eight. There's a lot
more of these green field fabs coming online. Historically, what you've seen is a lot of these fabs have unused space, and what you can do is you can install these tools and you can get additional capacity out of them. But now these fabs are essentially full, and we can get into the history if you want a little bit, but basically these fabs filled up. They need to build new fabs, but the lead time on fab construction is at least two years, and that's if you do it
incredibly fast. So if you want to build it from the ground, if I have dirt that's just sitting there ready to go, if you've already leased that space the green field, then you want to actually construct the fab and you know, again they need to be absolutely perfect. You can't even have particles of dust in these things, so it requires a lot of really specific construction that's specialized, and then you need to get these tools like, for example,
as mel Cell's EUV lithography machines. Those cost hundreds of millions of dollars. You have to get those in install them, and that's just one piece of equipment. So this is incredibly complex and takes probably at least three years really from just digging dirt actually getting these functioning, and then that's the first wafer that comes out. So it's a long lead.
Time and we're talking lot of greenfield coming online.
A lot of greenfield. So part of that is ske Heinez has a fab coming online pretty soon. They probably already just started getting wafers out. Micron is expanding in Boise, They're also going to expand in New York. Samsung has talked a little bit about it. But another underrated aspect is probably China. So there's a couple of fabs. There are a couple of companies that have fabs in China
YMTC for nand and then CXMT for DRAM. They have been aggressively expanding their capacity and gaining market share over
the last five or so. Longer than five years or so, and so basically does that start to impact the industry come and say twenty twenty seven, twenty twenty eight, because they don't have the market share yet that if they doubled their capacity over the next year or so, could that actually really drive negative pricing as opposed to not just slowing down the price increases we've seen.
One interesting thing about this trade drool is that if you look at the other ETFs, you know, we look at these three companies. Obviously, DRAAM has sixty seven percenters of the three companies, and EWY has forty eight percent, but no Micron, but it's heavy, and there are other two, but then there's a huge drop off even the AI ETFs. There's one AIS that has all three but only eighteen
percent total waiting. But here's what really shocked me, the semiconductor one like SMH which everybody uses, it only has micron and it's only at like a five percent waiting, so you miss a lot. And then here's why I really wanted to write about this theme and really give it attention, is that Spy has almost nothing like you own none of this trade and Spy and as big of a fan as I am of the Voo and should trade, the one thing that I would argue that you don't get with a voo and Chill is in
these trends. Eventually, if this grows and grows and grows, Micron will be a bigger part of Spy, no doubt, and you'll get at part of it. But you if you want to front run that, yeah, you can get these these runs that are earlier than when they're fully grown adults. This is why theming ETFs exists, in my opinion. So I think that's part of what we did. The other thing we looked at DROEL was, you know, Jake just talked about these memory companies are like under the
gun to build all this stuff. Well they need stuff too, So we looked at semiconductor equipment makers and that's a whole other thing. That's a whole nother trade. Now there is no semiconducting equipment ETF yet, so we try to look for one of those before we go over the ETF that we chose for this. Do you want to talk about that supply chain?
Yeah?
Yeah, So I mentioned ASML before, so they're one of the keys. Now, in the past, DRAM hadn't utilized as much EUV, so some of these older technology nodes that were used for these DRAM waves didn't require EUV. That's changing, so you need more EUV tools per fab. But then expanding beyond that, LAM Research has really been one of the leaders in different deposition and etch machines that specific
tools that specifically get used for these tabs. And then you have applied materials and then you have a very long tail of equipment companies that are In addition to that, you have KLA. I couldn't name them all right here, I would bore everybody, But the point is you have all these players. And the thing about them though, to keep in mind, memory tends to be a smaller percentage of their overall revenue, so their biggest customers are almost always TSMC. Then on top of that, you'll have company
like Intel, and you'll have Samsung. And again, if you're talking about Samsung, a lot of that is for their foundry for logics and not memory, and Intel they don't make memory anymore, so if you're selling to Intel, you're not getting memory exposure via that revenue. So again, and there's a long tail also fabs that have smaller markets like global foundaries for example, or tower semiconductor some of the others that you're just not going to be selling into.
So I think the exposure from these semiconductor capital equipment companies has been probably growing for memory, but also declining from China. So there's a couple of different moving pieces to keep in mind.
The ETF we found for this one's roll was the Pacer Blue Star Engineering the Future ETF good name, but doesn't really have a ton of assets, only five million charges forty nine basis points. But all the firms that Jake just mentioned are in the top like ten holdings
of this one. And I've been doing this a long time and sometimes when you find an analyst who's like, here's like these interesting stocks that are probably the most relevant, you have to go digging, and a lot of times you do find these these like new ones or ones that were overlooked in favor of the big popular one.
So this is part of why we try to get people to explore the toolbox more fully and something we sh do more often his analysts, which is unearthed some of these really you know, cool, what's interesting there.
Is like this one ETF becomes like this new sensation and then it makes you realize, don't look at the sun directly, look off to the side slightly, and that helps you see the rest of the eCos totally.
And you know, honestly, if I went to talk to this issuer or the at least the index maker, they'd be like, oh, yeah, we've known this for three years, dude, you know, welcome to the club. Like that's how many things are going on. And some of these ETF fishers are honestly really on the bleeding edge. But the four of them come out a day, so you just miss them and it's just tough to keep up with all this. But man, it's it's wild.
What's the ticker for that? B U l D be you wealthy?
There we go and I can give you the performance since it came out. It's not that old. It's up sixty four percent since launching four years ago.
Okay, not bad, Jake, thanks so much for joining us in Trillions.
Thank you, Thank you for having me here.
Thanks for listening to trillions and telling them Next time. You can find us on the Bloomberg Terminal, Bloomberg dot com, Apple Podcasts, Spotify, or wherever else.
You want to listen.
We'd love to hear from you. Put us up on social. Trillions is produced by Magnus Hendrickson and Kishav Pundia. Bye.