0:01 And the reason for raising capex has
0:03 always just been we need more compute,
0:05 we need more flops, we need more
0:07 intelligence. This quarter there was
0:15 Hello everyone and welcome to another
0:18 semi-doped podcast. I'm Austin Lines
0:20 with Chip Strat and with me is Vic
0:23 Shaker from Vick's newsletter. Hey Vic.
0:25 Man, a lot has changed since the last
0:27 time we recorded a podcast. What What's
0:28 up with you?
0:31 Yeah, a lot has changed. So, after 15
0:33 years of working in the semiconductor
0:35 industry across like small, medium, and
0:39 large companies, I finally decided to
0:42 hang up that corporate hat and do the
0:44 Substack, this podcast
0:50 um full-time and see where it takes us,
0:52 you know, in the podcast front. And I
0:54 hope the substract continues to grow too
0:57 because it is so much fun that uh I
0:59 figured like okay let's give this a shot
1:02 because I learn so much from like
1:04 writing and you know reading about so
1:07 many different things and uh corporate
1:09 roles tend to be a little bit more like
1:11 hyperfocused and I've spent 15 years
1:14 doing kind of the same thing. So I was
1:15 like okay instead of trying to trying to
1:17 change roles everywhere I've kind of
1:20 managed to carve out this role for
1:21 myself. So I'm like, "Okay." I said,
1:23 "Let's try it for a while. What's the
1:25 worst that can happen? Let's try it."
1:28 >> Awesome. I love it. Yeah. So, okay. So,
1:30 you'll have more time presumably for
1:32 writing, podcasting, researching. Like,
1:35 what are you most excited about?
1:36 >> Yeah. I'm just going to do the same
1:40 thing like I'm not trying to uh add too
1:41 many things because I think we already
1:44 added one other thing to the semi-doped
1:46 umbrella which we haven't really spoken
1:49 about which is semi-doped also has a
1:52 substack presence now and it is a free
1:54 for all subscription it does not have
1:57 payw walls and what Austin and I do here
2:00 is uh just we just give our daily take
2:03 so this is like a once daily newsletter
2:06 uh that's short to read like within 3 to
2:08 5 minutes just with some key highlights
2:10 of what's happening in the day because
2:13 we monitor this stuff so much every day
2:16 now that we figured okay like there's so
2:19 much that goes like unmentioned because
2:21 we can't write everything in the
2:23 newsletter and we can't talk about
2:26 everything on a limited time podcast but
2:29 not everything requires a full deep dive
2:34 substack or a full hour podcast but if
2:35 we try to stick it in a podcast, it gets
2:37 fragmented. So, the best place for this
2:39 kind of information, but that's also
2:42 very relevant because the the reason you
2:44 sign up for the semi-doped Substack is
2:46 that, you know, you get you'll get one
2:48 email a day. You open it up, you know,
2:50 while you just have your coffee, you
2:52 just like glance through it. Anything
2:54 that picks your interest, you just read
2:57 that section or or not, you know, you
2:59 you'll get another email the next day
3:02 and it's an easy read. It's unlike our
3:04 deep dives on Substack which is highly
3:05 researched and thought through and all
3:07 that. These are just like quick hits. So
3:09 that is that is the second project
3:11 that's common to us apart from the Substack.
3:12 Substack.
3:14 >> Yes. Yes. Yeah. I'm excited about it.
3:16 I'm I'm looking forward to it. It's been
3:18 fun this week. Um you know I think for
3:20 listeners definitely go sign up. It's
3:22 just basically like Vic and I having a
3:24 quick water cooler conversation and that
3:26 you're part of you know hey here's this
3:27 news from yesterday. Vic will give a
3:29 take. I'll give a quick take. Move on to
3:32 the next thing. I think it'll be fun,
3:34 informative, and you'll just be more
3:36 intelligent about semiconductors and AI
3:38 for reading it.
3:41 >> Yeah, just semi-doped.com should get you
3:42 there. Or you can just search for it on
3:45 Substack Semi-Doped. You'll see the same
3:48 logo of this sub uh this podcast and
3:49 you'll find it there.
3:52 >> Exactly. And the last thing that is like
3:55 forming still in my life is the fact
3:58 that now I am available to do consulting
4:01 and my newsletter itself actually runs
4:04 under the umbrella of semi-exponent
4:07 which I came up as a name to say like
4:09 this is the most exponential technology
4:12 I like you know that ever existed. So I
4:13 was like okay semi-exponent is a nice
4:15 it's like you know what is the exponent
4:17 the exponent is zero means it's a flat
4:19 line but it's never zero. So anything
4:20 above zero is like an exponent. So I'm
4:22 like, okay, that's a nice name. So the
4:24 consulting arm is going to be under semi-exponent.
4:26 semi-exponent.
4:29 So now I'm just starting to find, you
4:32 know, clients who want to work with me
4:34 on various different aspects of
4:37 semiconductors, have chats, pick my
4:39 brain on various aspects of what I write
4:41 and talk about. Ah, we'll see how that
4:43 goes. Should be fun. I like talking to
4:44 people anyway.
4:46 >> Yeah. Yeah. Nice. Nice. So you'll be
4:49 researching, writing, talking and
4:50 consulting. That that does sound like plenty.
4:52 plenty.
4:54 >> Yeah, that's enough, right? That's why
4:55 that's why I'm not doing anything to the
4:58 substack. Things will continue as it is.
5:01 And uh I hope there'll be more research
5:02 involved because I have more time to
5:04 think through carefully a lot more stuff.
5:04 stuff.
5:06 >> Yeah, I've always tried to maintain a
5:09 high high level of, you know, quality on
5:12 the substack. So hopefully if I've been
5:13 doing my job well enough, you won't
5:15 notice any changes.
5:16 >> There you go. There you go. Love it. All
5:18 right, let's get into it. So this past
5:21 week, you were on vacation. There was
5:24 earnings week. They didn't wait for you.
5:27 Um and but I stayed on top of it and I
5:29 know you you've caught up a lot on it.
5:31 Um so of course listeners, I know you
5:34 know all about this. Microsoft, Google,
5:36 Meta, Amazon, Samsung, Sandis, Seagate,
5:39 Western Digital, Qualcomm, Cadence, NXP,
5:42 even Rivian, which I I love to follow
5:43 even though they're kind of downstream
5:46 as like a semiconductor consumer. Um,
5:48 all those companies reported impossible
5:50 to listen to all of it. Um, but we are
5:53 going to cover it. And I thought the
5:55 angle that we could take is actually
5:57 starting with memory and storage
5:59 companies. And here's why. You know, as
6:01 everyone's been tracking, the big four
6:03 hyperscalers have now committed to
6:07 nearly 700 billion of 2026 capex, which
6:09 is up from roughly 500 billion just in
6:12 2025. So they they every quarter the
6:15 story is always like, hey, did they keep
6:16 capex or did they raise it even more?
6:18 And the reason for raising capex has
6:20 always just been we need more compute,
6:22 we need more flops, we need more
6:24 intelligence. This quarter there was
6:26 something different which was yes we're
6:30 raising capex but it's because of
6:32 needing extra dollars allocated
6:34 specifically to cover rising component
6:37 costs. Memory is more expensive. Storage
6:39 is more expensive. Even things like
6:43 fiber and optics are are more expensive.
6:45 And so I thought that was an interesting
6:47 sort of shift of like, oh yeah, we are
6:50 increasing capex, but it's just to pay,
6:52 not to buy more flops, but just to pay
6:53 for the things that we already committed
6:56 to. Um, so I thought it'd be fun to
6:58 start with the beneficiaries of this
7:01 capex raise, which would be the memory
7:03 and storage companies that reported. So
7:05 we're going to start there. How does
7:06 that sound?
7:09 >> Yeah, let's do it. Memory and storage is
7:11 something I keep ranting on on the
7:13 podcast. Like, how much more expensive
7:15 is it going to get? Like, how much more
7:17 are they going to raise prices? They
7:19 expect a price hike every time,
7:22 >> but if the price hike isn't good enough
7:23 now, they're like, "Oh, what? You're
7:25 you're only doubling it. Why not? Why is
7:26 it not five times?"
7:29 >> Yes. Investors. Totally.
7:30 Man, aren't you happy that prices are
7:32 going up? Not going up enough. Yep. Your
7:34 stock goes down.
7:35 >> That is crazy.
7:37 >> So, all right. Let's start with Samsung
7:40 or Samsung Electronics. Um, I the number
7:42 that stuck out to me was that their
7:46 memory revenue was up 101%
7:48 year-over-year, which obviously is
7:51 pretty crazy for memory. Um, they said
7:53 as a company their Q1 revenue and
7:56 operating profit were at all-time highs.
7:59 HBM sales are expected to triple
8:02 year-over-year in 2026. and they expect
8:07 HBM 4 to be 50% plus of HBM sales by Q3.
8:09 So there's a mix shift going from HBM 3
8:12 3E to four. Um now for people who
8:15 haven't followed Samsung as closely in
8:18 the HBM market um the the story over the
8:22 past couple years was they used to be
8:23 there there's there's basically three
8:28 big players SKHX, Samsung and Micron.
8:30 And Samsung used to be, you know, up
8:33 there with SKH Highix, uh, having a lot
8:35 of market share on the order of even
8:38 just about this time, you know, last
8:41 year it was, uh, maybe like or actually
8:43 maybe like six quarters ago that it was
8:45 like 40% market share for Samsung. And
8:48 in 2025, it dropped drastically down to
8:52 like 13%, 15%, 20%. And so the story of
8:54 Samsung is fall they've fallen behind
8:56 and now they're trying to catch back up.
8:59 Now, of course, there's always a new
9:04 technology, HBM3, HBM3, HBM4, um, and
9:07 when a new standard comes out, there's
9:10 an opportunity to try to be first to it
9:13 and and regain some market share. And
9:16 so, um, Samsung on their call, they were
9:18 really trying to position that like,
9:20 sure, uh, the past is behind us, but
9:22 we're totally ready for HBM4. And so, I
9:24 thought I'd read a couple quotes quick
9:26 and then we can get into it. Um but they
9:28 wanted to make sure to say that they are
9:29 the first and they are the best around
9:33 HBM4. So the quotes from the call Jun
9:36 Kim EVP of memory sales said after we
9:38 became the world's first to commence
9:42 commercial shipment of HBM4 in February
9:44 and then also he called out so so right
9:46 there like saying in passing like don't
9:48 forget we were the world's first um
9:52 shipping HBM4 and then um he said the
9:54 differentiated performance of our HBM4
9:58 led to concentration of demand and our
10:00 production ready capacity is fully
10:02 booked and sold out. Our outstanding
10:04 performance has been translating into
10:08 actual premium on pricing. So the read
10:10 through there is
10:12 there's always been a question around
10:14 memory around storage which is is it is
10:16 it just a commodity a bits a bits a bit
10:19 it doesn't matter who it's from and here
10:21 uh Samsung was trying to say hey not
10:23 only we're first but our performance is
10:26 the best and that's why and and we've
10:28 got the capacity ready so that's why
10:30 we're sold out customers and that's why
10:32 we're able to have a premium on pricing
10:37 is because customers prefer us. Um
10:40 any reaction to that Vic? I remember
10:42 this there was this whole discussion
10:46 about uh how many gigabits per second uh
10:50 you can get out of HBM memory by
10:55 designing the base die for HBM4 to be uh
10:58 you know on a certain node or even if it
10:59 doesn't matter what node it is because
11:01 there was some discussion that Micron
11:03 was using a memory technology to make
11:06 the base die while the other competitors
11:08 like Samsung and SK were actually using
11:10 a two logic node
11:15 Um the point is that the speed becomes a
11:17 differentiating factor. So it's no more
11:20 memory is just memory.
11:23 How how it performs is become
11:25 has become a very important factor for
11:29 like which company Nvidia or AMD will
11:31 pick for their performance. And not only
11:33 that, it comes the other way too. The
11:37 Jedex spec for HBM4 is like I think 8
11:41 Gbit per second per lane. But then
11:44 because they want supremacy on inference
11:46 performance and tokens output and tokens
11:49 per watt per dollar and all that they
11:53 are pushing the speed per lane of HBM4
11:55 faster and faster and faster. So it's
11:58 become a competition as to who can get
12:02 to like 11 10 or 11 or 12 Gbps per pin
12:07 and that is not even is way beyond the
12:10 spec but it has become the thing that
12:13 drives sales and drives lock in because
12:16 once these hyperscalers choose and qualify
12:17 qualify
12:21 a HBM vendor it is a sticky decision
12:23 because qualification if you remember
12:25 like the HBM3
12:27 Samsung couldn't really get qualified at
12:29 Nvidia for a long time. They had so many
12:32 yield issues and things like that. So
12:36 qualification and performance makes this
12:39 very sticky. Uh as so it's not funible.
12:41 So you just can't take out Samsung and
12:43 drop in Micron tomorrow. Although there
12:45 are only three companies doing this.
12:48 >> Yes. And so zooming out and and hitting
12:50 on what you said again for listeners,
12:52 the the real interesting thing is there
12:56 is a spec this Jed spec JDC
12:59 um that defines
13:02 the performance level that and you know
13:03 the various other things that a spec
13:05 define about how it's supposed to work,
13:06 how it's supposed to communicate um that
13:08 these three companies are trying to hit
13:10 and and normally if everyone just hits
13:13 that spec exactly, it would be funible.
13:16 But but what Vic is saying is that, you
13:18 know, Nvidia said, "Hey, wait a minute.
13:20 I've got, you know, I've got to compete
13:23 against AMD and against XPUs, and if I
13:26 could if you could give me um memory,
13:29 HBM memory that's even faster, then I
13:31 can get even better, you know, tokens
13:33 per second, tokens per watt, that kind
13:36 of thing." And so actually there is a
13:40 pull from the the silicon vendors to the
13:42 memory manufacturers to say I know the
13:45 spec says 8 gigabits per second per pen
13:48 or whatever. Um can you go higher?
13:49 Right? And so then there's this
13:51 interesting like someone's trying to get
13:53 to 11 and now the others have to try to
13:55 get to 11 as well. So there's this
13:57 interesting pull to go faster and now
14:00 this kind of like unwritten spec that
14:02 this performance, you know, that they're
14:04 all trying to compete on and and so as
14:06 you move away from just everyone meets
14:08 the spec, we're all fungeible to like
14:11 how fast can you go, at what yield, at
14:14 what cost, it is more of a a a true
14:17 competition on the things that do allow
14:19 you to have premium pricing over
14:21 commodity pricing, which is which is
14:22 performance, you know, and yield and
14:24 cost and that kind of thing. So that's
14:25 that's a little bit of the story the
14:29 back the backstory here for HBM4 and how
14:32 Samsung is trying to regain market share
14:35 is to to truly compete on you know
14:37 performance. Okay. So here's an
14:41 interesting thing uh a as we move out of
14:43 that looking at the the competition
14:45 there between the three companies um and
14:47 go focusing back on Samsung another
14:50 quote that they had was our demand
14:53 fulfillment rate is now at a record low
14:54 customers who are concerned about supply
14:56 shortages are actually bringing forward
15:00 their demand for 2027 already so I you
15:02 know not surprising but again it's just
15:04 crazy to hear you know if customers are
15:06 asking for this much Maybe it used to be
15:09 like we can only give you 80% of that or
15:11 50%. Maybe now it's even lower than
15:12 that. I don't know. I'm making up the
15:13 numbers, but trying to illustrate the
15:16 point that customers are asking for a
15:18 lot of memory and the memory supplier is
15:21 saying I can give you less than ever. So
15:23 that would be uh the environment that
15:24 we're in.
15:27 >> Yeah, that's crazy to me. Uh how much
15:29 longer is this going to continue?
15:32 Because people, these hyperscalers are
15:34 increasing their capex just to pay these
15:36 memory companies. Literally, that's what
15:38 their capex is going into just to pay
15:42 these players for HBM and like nan
15:44 memory and things like this. I'm not
15:46 sure how much how much higher is going
15:48 to keep going. I I keep thinking that
15:50 that's it, but I'm always wrong.
15:51 >> Yeah, you know, that's it's an
15:53 interesting point. we can get into it
15:56 more uh both here and and later which is
16:00 I am very bullish on capex continuing to
16:02 get higher when it's buying more flops
16:05 when it's buying more compute because I
16:07 like everyone else believe that the more
16:09 compute we have the more intelligence we
16:11 have the more we can do you know I I too
16:12 have experienced I'm sure everyone has
16:16 you know um chat GBT or Gemini or cloud
16:18 code just spinning or saying like oh I'm
16:21 busy right now you know or like come
16:23 back. We've reached our limits. So,
16:25 that's frustrating and that just shows
16:27 that like dude, they need more GPUs.
16:29 They need more XPUs. But is this is a
16:32 different conversation when it's like
16:34 how high can you convince your CFO to
16:37 let capex go when you're not buying any
16:39 additional compute, but you're just
16:42 paying increased memory prices? Is that
16:44 going to be the straw that breaks the
16:45 camel's back?
16:48 >> Exactly. Because if you if like you say
16:50 if all of this money went into expanding
16:55 compute then a lot more users of AI
16:57 tools or whatever get actually something
16:59 out of it. They get better tools. They
17:01 use it to build better projects that
17:03 drives revenue that drives an economy.
17:06 It it completes the circle in some way.
17:08 What is happening now is if all the
17:10 capex is going to memory players it's
17:11 like you're siphoning all this money out
17:14 into somebody's pocket. it's not going
17:15 into the, you know, positive
17:18 reinforcement loop that we want to see.
17:19 So, this is a bit concerning for me.
17:21 >> No, you're totally right. I mean, put
17:23 succinctly, there's no ROI on that
17:25 additional capex
17:26 >> except for a few companies.
17:28 >> Yes. Yes. But for the hyperscalers, yes,
17:32 there's no ROI, no extra ROIC, return on
17:34 invested capital. It's just a tax
17:35 really. It's
17:37 >> it's a tax. It's a memory tag.
17:40 >> It's a memory tax. So, okay, let's move
17:42 on. Um, so we're going to talk SanDisk.
17:44 So, we've talked when we talked Samsung,
17:46 we were focused on HBM, focused on
17:49 memory. So, let's talk um o other memory
17:51 and storage. So, we're going to talk
17:52 SanDisk. But really quick, I thought I'd
17:54 give a quick history sidebar because we
17:55 haven't talked about SanDisk on the
17:59 podcast yet. So, um SanDisk, Western
18:01 Digital really quick, there are two
18:05 companies. Um back in 2016, Western
18:07 Digital actually acquired SanDisk for
18:10 around $20 billion. And the thesis I
18:12 think back then was having a full
18:14 storage portfolio. So can we own both
18:17 hard disk drives and nan flash and that
18:19 therefore we can sell hyperscalers a
18:21 complete stack. Makes a ton of sense
18:24 right now in the era we're in. But
18:28 prior to AI um that thesis didn't really
18:30 age well. I think they're they're very
18:32 different businesses. They have
18:34 different cycle dynamics, capital
18:38 intensity, customer mix. Um, so running
18:40 them under one roof wasn't as simple as
18:42 like, oh great, um, we share customers
18:44 and now we can more easily cross-ell
18:46 into them. It was actually sort of like,
18:48 ah, these are two different businesses
18:50 and there's some different customers at
18:52 play and we're not getting the quote
18:53 unquote synergies that we thought we
18:58 would. Um, so back in October of 2023,
18:59 Western Digital announced that they were
19:03 going to split SanDisk back out. And by
19:05 February 2025, that spin out happened.
19:07 SanDisk reemerged as a standalone
19:09 publicly traded company. And so when now
19:11 when you think about SanDisk, you can
19:13 think about NAND flash, SSDs, embedded
19:17 flash. Um, they have a joint venture on
19:21 some fabs in Japan with Kioia, Kioxia,
19:23 not sure how to pronounce that. Um and
19:25 then so SanDisk you can think of flash
19:28 SSDs etc. And then Western Digital is
19:32 eight hard disk drives only. Um and
19:33 >> they are working when I was doing a
19:35 little bit of research. Western Digital
19:37 is working on this interesting next
19:41 generation tech HMR hammer. It stands
19:44 for heat assisted magnetic recording.
19:46 And it I'd never heard of it, but uh the
19:50 goal is to achieve 100 terabyte plus
19:53 hard disk drive capacities for AI scale
19:55 data and they're planning volume
19:57 production in 2027. Have you heard of this?
19:58 this?
20:00 >> Yes. So, HMR has been there for a while
20:02 actually. It's not that new. >> Okay.
20:02 >> Okay.
20:06 >> HR has been cited for I've heard this at
20:08 least for like 5 years now. Uh I can't
20:10 pinpoint exactly how long it's been
20:13 around. It's always been the next uh you
20:15 know the next greatest thing in hard
20:18 disk drive technology. But then what has
20:20 happened recently is that if you look at
20:23 uh you know quad level cell SSDs QLC
20:28 SSDs the SanDisk ones for example the
20:30 highest capability the highest capacity
20:34 one is five uh 256 terabytes and you get
20:37 it in an SSD form factor already. So I'm
20:40 like what is this HMR and he's going to
20:42 give you what 100 terra it's not not
20:44 that great honestly today's day and age
20:47 of course the price point will be lower
20:50 it should be uh but you know HMR yeah
20:53 it's it's been around but it's I don't
20:55 know how relevant it is or how where we
20:57 are on that right now but I wanted to go
20:59 back in the history a little bit more I
21:02 think you'll find this fun so did do
21:04 does the name like Sanjay Meotra ring a
21:06 bell to you
21:09 Uh yes, Micron.
21:10 >> Yes. Micron. Micron. >> Yeah.
21:11 >> Yeah.
21:15 >> Do you know who founded SanDisk?
21:17 >> That person. >> Yes.
21:17 >> Yes.
21:19 >> Serious. What? >> Yes.
21:19 >> Yes. >> What?
21:20 >> What? >> Yeah.
21:20 >> Yeah.
21:22 >> Yeah. It's true. He was he was one of
21:26 the you know founded you know he was one
21:30 of the founding members of SanDisk. Uh
21:34 and you know this was like way before
21:37 his he became the CEO of sand of Micron
21:40 only in 2017 but he was the original >> wow
21:40 >> wow
21:42 >> uh founder of sand. So he's like real
21:45 real memory guy. And another fun bit of
21:48 trivia on the name, right? Uh one of the
21:53 other founders, um whose name I Eli
21:56 Harari, yeah, the one of the other
21:59 co-founders and so they've, you know,
22:01 they were trying to find a name for the
22:02 company or whatever. And then his
22:05 daughter comes in and it's like looks at
22:06 some of the discs lying there, you know,
22:08 you know, these these splatters or
22:09 something and like, oh yeah, that looks
22:12 like the sun. So they decided to call it Sundisk.
22:14 Sundisk.
22:15 >> No way.
22:17 >> You can look up early logos of Sundisk
22:20 and you'll see like a plate like thing
22:22 with like the sun's rays like coming out
22:23 of it. That's their logo. >> Serious.
22:24 >> Serious.
22:27 >> Yeah, we should put a picture if we find one.
22:28 one.
22:30 >> Yes, we should. Totally. Oh, great trivia.
22:31 trivia.
22:33 >> Yeah. Yeah. Yeah. And what happens is
22:36 that later they like sun micros
22:38 systemystems came after them for some
22:40 trademark stuff like oh why you can't
22:42 use sun in it or something so they
22:45 change sun to SanDisk
22:46 came about
22:47 >> fascinating yeah I was wondering I was
22:50 like sun how did it make it sand uh I
22:50 don't know
22:52 >> it changed only like seven years after
22:55 the company was even found like running
22:56 okay seven years after the company was
22:58 like fully functional they changed the
23:01 name to SanDisk thanks to Sun Micros
23:03 Systemystems Nice. Well, we should talk
23:05 to Sa San Sanjay sometime and ask him if
23:08 the sand came from his name, too.
23:09 >> Why not? Yeah.
23:11 >> Yeah. Well, yeah, we should seriously
23:12 see if he would talk to us about the
23:14 history. That's so fascinating that he
23:16 was a co-founder of SanDisk and now he's
23:18 a CEO of Micron. Super interesting.
23:20 >> Yes. Yes. I thought they mentioned that
23:22 as part of a history lesson here.
23:24 >> Oh man, good good history lesson. Okay,
23:27 so SanDisk, they reported earnings.
23:30 Their CEO is not Sanjay. it is someone
23:33 named David Gekckler I believe Sandis
23:36 CEO um so
23:38 >> this quarter they had revenue of almost
23:41 $6 billion they were up 97 their revenue
23:44 was up 97% sequentially so that's just
23:47 quarter over quarter um and of course
23:49 how do you do that you raise prices um
23:52 they're up 251% year-over-year
23:55 um here's a couple things that stood out
23:59 which are pretty crazy gross margin 78 8.4%.
24:01 8.4%.
24:03 That's like gross margins of a software company.
24:05 company. >> Yeah.
24:05 >> Yeah.
24:07 >> Yeah. Totally. Um
24:10 >> I have my notes here that it was 51.1%
24:12 the prior quarter. >> Okay.
24:12 >> Okay.
24:14 >> And you know the revenue estimate was
24:16 supposed to come in like I don't know
24:19 4.8 billion or something.
24:21 >> They come in a billion above that. I'm like
24:22 like
24:26 >> what like they are up 250% yearover-year
24:28 in like revenue
24:29 >> and you know the funny thing is that
24:31 they haven't shipped like that many
24:34 extra bits this is all pricing
24:36 >> yeah this is not as much you you think
24:40 like oh they sold maybe that much more
24:42 you know to account for like 97% quarter
24:44 over quarterart increase no not really
24:46 you can't bring on that much capacity
24:47 that quickly
24:49 >> I mean you're talking about making
24:52 wafers and stuff spend in a fab and that
24:53 stuff doesn't move in the time frame of
24:55 a quarter. So this is all price increases.
24:56 increases.
24:58 >> Totally. Wow. So you said their margins
25:00 were 51% last quarter and now they're up
25:01 to in the 70s.
25:04 >> 70s and I think they projected above 80
25:05 next quarter. >> Nuts.
25:05 >> Nuts.
25:06 >> The guidance is above 80.
25:09 >> They're going to make Yeah. Nvidia look
25:11 like they have work to do.
25:13 >> Nvidia is only like 75, right?
25:16 >> Yeah. Right. Totally. Totally. Of
25:17 course, Nvidia is a very very large
25:20 company and has had these margins for
25:22 quarter after quarter after quarter. So,
25:25 the question is um why why is Sandis
25:28 crushing it? Um and the what they the
25:30 story that I heard on the earnings call
25:33 was that they're they really believe
25:35 that the NAND market is transitioning
25:38 from commodity spot business to
25:41 something less cyclical. Um, and so the
25:43 the points that they made on the call,
25:47 and I'll read some quotes here, too. Um,
25:49 they have five multi-year supply
25:53 partnerships signed. Um, these three new
25:55 ones in quarter two, three. They'd
25:57 announced two previously and they they
25:59 had announced three more. Um, and those
26:02 three new ones account for $42 billion
26:04 of RPO remaining performance
26:06 obligations. You can kind of think of it
26:09 like a backlog. And this was the first
26:12 time they've actually disclosed that.
26:15 Um, so clearly there's long-term
26:18 agreements for many years and people are
26:19 like signing up and these customer
26:23 commitments um already cover onethird of
26:26 the fiscal year 27 bits. So it's not
26:28 just people signing up for 2026, but
26:29 they're already signing up and paying
26:32 for and committing to 2027 bits, which
26:34 actually have financial guarantees
26:36 backing them. So it's not, you know, it
26:38 people are putting their money where
26:39 their mouth is. Oh yeah, the CEO
26:41 Gekckler said he wanted to drive that
26:43 those long-term commitments to above 50%
26:45 of bits and already that one of the
26:48 longest contracts was for five years.
26:50 And so the quote that I thought was
26:53 pretty telling on how SanDisk is
26:56 perceiving this um from the CEO, he
26:59 said, "Last quarter, we were engaged in
27:01 discussions with customers on multi-year
27:04 supply partnerships, what we refer to as
27:07 new business models or NBMs. Um I am
27:08 pleased to share that we have
27:09 successfully advanced those
27:11 conversations with five multi-year
27:12 partnerships signed so far. They are
27:14 structured to lock in committed supply
27:16 for our customers and committed
27:18 financials for SanDisk. So kind of the
27:20 customers are prepaying almost. Our
27:23 customers commitments are backed by firm
27:25 financial guarantees. These partnerships
27:27 support durable structurally higher
27:29 earnings in a significantly more
27:31 predictable and less cyclical business
27:33 for SanDisk. We believe this marks a
27:35 fundamental evolution of our business
27:36 which is centered on deeper customer
27:38 alignment, enhanced visibility, and
27:41 long-term value creation. So, so far in
27:43 this quote, he said, "We're signing up
27:44 people for the long run and they're
27:47 paying for it." There hasn't yet, in my
27:49 opinion, been an argument for why it's
27:52 not cyclical. It could just be demand is
27:56 crazy and people are signing up. Um,
27:59 as we talked about earlier with Samsung,
28:02 when you're a commodity, when you're not
28:03 a commodity, you're actually competing
28:05 on um performance, for example. and he
28:07 hasn't talked about that yet. But then
28:08 later there was a quote that started to
28:10 get into this which I thought was
28:13 interesting. Um he said as AI models
28:14 scale from billions to trillions of
28:16 parameters and deployments advance from
28:18 simple inference to deep reasoning and
28:21 increasingly agentic systems NAND has
28:23 become a critical component of the
28:25 underlying infrastructure. inference
28:27 optimizations such as KV cache along
28:30 with workloads like rag require
28:32 substantial high performance low latency
28:34 flash to deliver real-time
28:35 responsiveness and quality of user
28:38 experience and then he goes on to say
28:41 basically NAND flash is emerging as the
28:43 only economically viable solution to
28:45 deliver that capacity performance and
28:48 efficiency and then he said you know
28:49 goes on to make the argument that
28:52 SanDisk is the best on these metrics and
28:54 therefore um that's why you know they're
28:57 capturing this value. So I want to throw
29:00 it over to you Vic. Do you think this is
29:03 SanDisk? Is this is this AI you know
29:05 truly needing lots of NAND and
29:08 performance matters to your economics to
29:11 your token cost? Um or and bits are
29:13 therefore no longer interchangeable.
29:16 Um or or is this just a demand thing
29:18 where it's like, hey, they're locking up
29:19 people for five years because they have
29:23 supply and it's sort of
29:27 preventing true competition on like
29:28 performance latency that that kind of thing.
29:29 thing.
29:31 >> Yeah. Yeah. This is this is good. Did
29:34 you look at uh the Deep Seek V4
29:36 announcement? Did you look into what is
29:37 happening there?
29:38 >> Oh, a little bit. But please tell us
29:40 about it. It's it's related
29:44 >> perfectly to LAN. You can see basically
29:48 how uh DeepSeek V4 has compressed KV
29:50 cache massively compared to the previous
29:54 version 3.2 2 I think and
29:56 you would think like why would you need
29:58 NAND now if your KV cache is compressed
30:02 the problem is that the KV cache still
30:05 when you're doing agentic AI and agentic
30:08 multi-turn AI none of it fits in HBM
30:11 you're running hundreds of agents they
30:13 all have to have long context long
30:15 running context multi-turn context all
30:19 of those key value cache pairs are just
30:23 too much to store in HBM or DAM. So if
30:25 you see the DeepSc V4 inventions
30:28 recently, it is optimized to be an
30:31 entirely SSDcentric
30:34 inference system where uh KV cache is
30:36 stored almost entirely on SSDs.
30:40 >> It is pretty amazing actually. And
30:44 earlier uh after GTC when uh Jensen
30:47 announced the inference context storage
30:50 system which now I think he calls CTX uh
30:52 context storage or whatever. So you
30:56 basically this is a bunch of uh SSDs in
30:59 a rack that is sitting there and
31:02 connected uh via like high bandwidth
31:04 fabrics to the GPUs so that they can
31:07 offload the KB cache matrices into this
31:11 this like SSD storage that is becoming
31:14 uh very important and at that time I
31:16 wrote a substract post pointing out how
31:17 this is going to change the inference
31:21 tokconomics forever because one of The
31:24 ways you can save on inference cost when
31:26 you're using, you know, the API. You can
31:30 look this up for any model is that
31:34 it really depends on a few things. First
31:36 of all, you've got the input tokens. You
31:38 got a certain number of tokens uh cost
31:41 per million tokens. Then you've got the
31:45 output number of u dollars per out
31:47 million output tokens, right? And it's
31:49 usually like a four or five is to one.
31:50 Output tokens are like five times more
31:53 expensive than input tokens because it
31:55 go under goes this like reasoning this
31:56 thinking process, right? The thinking
31:58 process costs a lot more money. So the
32:00 output tokens are more expensive than
32:02 input tokens. In the middle you have
32:05 another pricing point called the cash
32:08 hit or cash miss. So what that means is
32:13 that if you have KV cache and your uh
32:16 inference is able to reuse that KV cache
32:19 as much as you possibly can, right? You
32:21 have a cash hit and when you have a cash
32:23 hit, your cost of inference goes down
32:27 massively. So what people do now is to
32:29 basically have a bunch of system
32:31 instructions right up front and then
32:33 they'll like only append to the bottom
32:35 so that you can maintain your cache
32:37 hierarchy completely. So this is like
32:40 cash aware uh inferencing >> and
32:42 >> and
32:44 deepseek for example like there was some
32:48 metric saying that like in agentic use
32:51 you can make it use 95% of cash hit
32:54 rates. So 95% of the time it's hitting
32:57 cash. It in very few cases it even goes
33:00 to 99% it's hitting cash. So even now if
33:02 you go to the deepseek uh API pricing
33:05 page and see how much the cost has
33:10 dropped for the Deepseek V4 Pro. Okay,
33:14 it has gone down to a fraction of a cent
33:17 per million tokens. Okay, you can
33:18 compare that to Opus. I don't remember
33:19 the number right now but it's like
33:22 significantly high. So what this says is
33:27 basically SSD driven uh inference is
33:30 basically the only way forward. Deepseek
33:32 has kind of portrayed that clearly now
33:36 and you cannot store all of this
33:38 information on DM on or HBM. It's just
33:41 too much. It's too much and you almost
33:44 have an infinite storage of SSDs. You
33:46 know the cost per unit is like an order
33:49 of magnitude higher when it comes to DAM
33:51 versus SSD. So you've got this
33:54 essentially free storage and now you
33:56 have like capacity is a solved problem
33:58 if you go to SSDs. The only problem is
34:01 then uh you have this bandwidth
34:02 bottleneck because you can't access
34:06 stuff as fast uh from SSDs as you are right
34:06 right
34:08 >> you know so I've been meaning to read
34:10 this new paper that's come out from the
34:12 deepseek team called I have it on my
34:14 screen right now it's called dual path
34:16 breaking the storage bandwidth
34:19 bottleneck in LLM especially agentic
34:23 LLM. So I want to see exactly how all of
34:25 this ties in together. So I've basically
34:28 given a surface level overview of
34:31 justifying why SanDisk is saying that
34:34 nan storage is so important in the
34:36 future of agentic you know AI because
34:39 deepseek is already a data point that is
34:40 heading in this direction.
34:42 >> There's a long answer to your rather
34:42 simple question.
34:44 >> Oh this is this is so good. Okay so I'm
34:45 going to summarize it and then I have a
34:47 follow-up question for you. Okay, so
34:49 you're saying, hey, look, Deepseek is
34:52 showing us that uh SSDs are more
34:54 important than ever. NAN flash are more
34:55 important than ever. And if you think
34:59 about the tokconomics, it's way more
35:01 economical when you've got these cash
35:03 hits versus cash misses. So you got to
35:06 think really hard about can we have a a
35:09 very big cache or memory hierarchy and
35:11 can we store as much as possible and and
35:14 people are going to be very incentivized
35:16 even the end customers using APIs
35:17 running agents and stuff are going to be
35:18 incentivized to actually think about
35:20 memory and think about caching. Um
35:23 therefore SSD market is going to just
35:26 continue to grow. Um but my question for
35:30 you then is okay SSD market SSD TAM
35:33 going to explode. All good news. Invest
35:36 in these companies. But why SanDisk?
35:37 >> Not advice.
35:39 >> Oh yes. Not not this is not investment
35:40 advice. This is not investment advice.
35:42 We are just thinking very hypothetically here.
35:43 here.
35:46 >> Um do your own due diligence. You know
35:49 don't hold Vick and I accountable. Okay.
35:53 So uh the SSD TAM is going to explode
35:56 >> or is exploding. But is there is a bit
35:59 is an SSD bit a bit a bit or is is
36:01 SanDisk's bit better than someone else's?
36:04 else's?
36:08 I think a bit is a bit I mean is is NAND
36:10 is a rather established technology.
36:13 There isn't a controller
36:15 actually there is a controller
36:18 difference controller for the QLC NAND
36:21 is called Stargate and they're still
36:24 working on it and usually the more bits
36:26 you add so in going from a triple level
36:30 cell uh to a quadle cell you go from
36:32 having nine states a single cell has
36:34 nine states in a triple level cell in a
36:37 quad level cell it has 16 states a
36:39 single cell and now depending on how you
36:41 program the You should be able to
36:42 distinguish between 9 or 16 different
36:44 states otherwise you don't know what bit
36:46 it's holding. Is it holding like a
36:50 pattern 10 or pattern 15. So the
36:52 complexity of the controller for this
36:54 NAND gets significantly higher even when
36:56 you add a single bit like if you go to
36:59 pentle cells which exists in research
37:02 mode you will go to 32 levels. Now you
37:04 have to distinguish between 32 levels
37:06 you know that makes it very difficult.
37:08 Um I have a whole article on how all of
37:12 this works but yes there is a difference
37:15 in controller and how they how it works
37:17 and all that but it's not really a
37:19 differentiating factor. Okay the
37:22 controller needs to work well and as far
37:24 as I can understand it please correct me
37:26 if you know there are any storage
37:28 experts out there who actually work on
37:31 this stuff on a day-to-day basis they
37:33 always know better but as far as I know
37:34 a bit is a bit.
37:36 Yeah, we should follow up and bring on
37:39 someone Micron, SanDisk, Samsung,
37:41 whoever is listening out there. Of
37:43 course, to talk memory, but also to talk
37:45 storage. Let's I would love to talk
37:47 storage here and hear like a product
37:49 manager's argument for a bit is not a
37:51 bit because it probably comes down to
37:56 things like reliability or pricing per
38:01 bit or or other other factors than just
38:03 straight up read latency or right latency.
38:04 latency.
38:06 >> Yeah. Yeah. Yeah. Yeah. Maybe there's
38:08 some controller magic there because a
38:11 lot of the latency comes from adjusting
38:14 the voltages just right to read you know
38:16 this particular state of the cell. So
38:18 what it does is it it iteratively
38:20 programs the right voltage to reach that
38:22 state that takes time you know so if
38:23 there's like some fancy new algorithm
38:26 that can go quickly and you know that
38:28 you can reduce the latency that way that
38:30 may be a differentiating factor
38:34 >> actually from IEDM last year and if I
38:36 think we spoke about this on a past episode
38:37 episode
38:39 >> we spoke about a totally different kind
38:42 of cell where you can dramatically
38:46 increase uh the the the reliability
38:48 while still having like 36 states to a
38:49 cell, you know.
38:51 >> That's right. Yes. Circle one or something.
38:51 something.
38:52 >> The circle one. Yes. Yes.
38:54 >> Yes. Yes. All right. Listeners, go check
38:56 out our backlog. >> Yeah.
38:58 >> Yeah.
39:00 >> Okay. This has been a super good memory
39:02 deep dive. Let's let's carry on real
39:04 quick. Okay. So, let's talk hyperscaler
39:07 earnings quick. Um, so, hey, it was a
39:09 good quarter for Samsung. It was a good
39:13 quarter for SanDisk. Um, memory, storage
39:15 price through the roof, demand through
39:17 the roof. What does this mean for the
39:18 hyperscalers? We kind of already hinted
39:21 about it. Um, Microsoft
39:25 disclosed um $190 billion of capex for
39:28 2026. And the CFO, Amy Hood, said
39:32 roughly $25 billion of the calendar 2026
39:34 capex is specifically higher component
39:37 pricing. So, that's pretty crazy. We are
39:40 only a few years removed from someone
39:42 like Microsoft spending a total of $25
39:45 billion per quarter uh just on all of
39:47 their capex. And now we're talking about
39:48 in a year they're going to spend like
39:50 kind of a quarter's worth of capex just
39:53 on higher memory storage component prices.
39:54 prices. >> Um
39:55 >> Um
39:58 >> just give it to the memory guys.
40:01 >> Yeah, right. Pretty much. Uh hey, good
40:03 good time to be a memory guy. Meta Meta
40:07 raised their capex. Um and Zuck on the
40:09 call said most of the raise is higher
40:11 component costs particularly memory pricing.
40:14 pricing.
40:17 Google has a higher capex although they
40:21 didn't talk about that as much. Um
40:24 Sundar did say that their cloud their
40:26 cloud business was up 63%
40:29 um which was insane and Sundar said that
40:31 that would have actually been higher if
40:34 they were able to meet demand. So, still
40:36 talking about demand and capacity being
40:38 the the bottleneck there. And then
40:41 Amazon, they didn't uh talk specifically
40:44 about component prices and impacting
40:46 their capex, but but check this out.
40:49 This was an interesting qu like uh
40:51 commentary that came out. The CEO Andy
40:54 Jasse was asked if memory constraints um
40:57 are negatively impacting them. And he
41:00 answered by saying that for their cloud
41:02 business for AWS memory constraints are
41:05 actually driving cloud growth. And here
41:06 and that feels very counterintuitive.
41:08 And here's what he said. Who knows how
41:11 big this is in aggregate but he said one
41:12 of the interesting things that we see
41:13 right now with the change in price and
41:16 in supply on things like memory is it is
41:18 actually a further impetus pushing
41:20 companies who have been on premises
41:22 infrastructure into the cloud. And it's
41:24 because these suppliers, the memory
41:26 suppliers are prioritizing their very
41:28 largest customers, the hyperscalers,
41:30 cloud providers. And so therefore,
41:33 there's a number of enterprises who
41:36 can't get onrem infrastructure. It's
41:38 actually pushing them to speed up their
41:40 move to the cloud.
41:41 >> So I thought that was pretty interesting.
41:42 interesting.
41:44 >> And I think people are generally
41:47 comfortable in using like AWS cloud.
41:50 It's it's a very easy switch to go and
41:53 do your stuff on there because it has
41:55 been the the bread and butter for like
41:57 pretty much the software industry for
41:59 the last decade, right?
42:01 >> Yeah, totally. It it would be
42:02 interesting to dive in with them because
42:04 at this point you ask, well, who's still
42:06 running running on prem? And it's it's
42:09 got to be people who are still just
42:11 thinking about like, oh, what data do I
42:13 keep on prem because I'm in the
42:15 financial industry or health industry or
42:17 something. But even those folks are
42:20 still moving to the cloud and they're
42:22 getting pulled there faster because the
42:24 only place they can get compute and
42:26 memory is actually from the cloud
42:30 players which is very very fascinating. >> Yes.
42:30 >> Yes. >> Yeah.
42:31 >> Yeah.
42:32 >> If that's the only place you can get it,
42:34 that's where you go. I mean it doesn't
42:35 matter what you want, right?
42:37 >> It's like yo, sorry compliance team. I
42:39 know we're we're dotting all the eyes
42:42 and crossing all the tees, but literally
42:44 we have to we have to make this move
42:45 happen right now.
42:47 >> Yes. If you can buy memory from the
42:49 memory guys, then maybe.
42:51 >> Yeah. Yeah. Exactly. Exactly. Yeah. You
42:52 go, you go find it. You go find it.
42:55 Okay. Um, so, okay. AI accelerators were
42:57 talked a lot about across the
42:59 hyperscalers. Google said, you know,
43:01 publicly on their earnings call that the
43:04 TPU will be sold in a merchant capacity
43:06 for the first time and at multi-gawatt
43:08 scale. And and then actually if you look
43:11 in the 10 Q there was some risk language
43:13 now confirming that
43:17 >> the the risk to doing that is getting
43:19 cos and HBM allocation.
43:22 >> So I thought that that was interesting
43:24 >> and just go to EIB. Forget about co-ass.
43:26 >> Hey there you go. Yeah. Go to EIB.
43:26 >> That's right.
43:28 >> Intel. Intel for the win.
43:30 >> Yes. I hope to write something about
43:32 E-IB here shortly soon comparing it to
43:36 like co-was L which also has bridges but
43:38 to to unpack that for people um AWS
43:41 talked about trrenium um at a 20 billion
43:44 run rate growing triple digits and of
43:48 course people are not generally
43:50 asking to buy tranium and they're not
43:53 even necessarily uh renting it directly
43:56 but they are consuming Amazon Bedrock
43:57 and all these services that run on top
43:59 of tranium
44:01 Um, Tranium 2 largely sold out, Tranium
44:05 3 nearly fully subscribed and Jasse said
44:09 on the call that um, because they do
44:12 custom XPUs, they will save tens of
44:14 billions of dollars of capex savings
44:16 each year, which translates into several
44:18 hundred basis points of operating margin
44:20 advantage versus relying on merchant
44:23 chips. And on the one hand, normally I'd
44:24 been like, "Wow, that's so incredible.
44:25 They're going to save tens of billions
44:27 of dollars." But then right away I
44:28 thought, "Oh, so they can pay for their memory."
44:31 memory."
44:32 >> Yeah. Yeah.
44:35 >> We're making our own chips so we can pay
44:37 the memory guys. Yeah. Right now a funny
44:38 cycle. I don't know what what's what
44:40 what what is happening right now, but
44:41 yeah, we'll see.
44:45 >> Totally. And so um then Meta mentioned
44:48 of course they've had tons of uh press
44:50 releases building up to this quarter
44:53 about 1 gawatt worth of MTIA with
44:55 Broadcom and they had showed their road
44:57 map of you know four chips in two years
45:01 uh MTIA XPUs. We had a great
45:03 conversation with Meta recently on that
45:05 with Matt Steiner and hopefully we'll
45:07 have more getting into it further. Um
45:09 they now Meta did mention significant
45:12 deployments with AMD and then also
45:15 running on some new Nvidia. So they were
45:18 talking uh up all their multi- silicon
45:21 vendor uh partners. Um so with that
45:23 we've run long. I think we should call
45:25 it quits here. Thank you everyone for
45:28 listening. Um YouTube YouTube folks
45:30 thank you for your comments. We see that
45:33 you would like Vic to explain again from
45:36 our Google TPU one. Why is it seven hops
45:37 with I think it was with boardfly and
45:40 not 16 hops. So we will follow up with
45:43 you. We'll draw a picture sometime. So
45:45 keep writing your comments, keep writing
45:46 your questions. Thanks for everyone for
45:47 listening, watching us on YouTube,
45:50 watching us on X and reach out. Uh
45:52 thanks again. We'll talk to you guys later.
