[00:04] SanDisk one year ago, that would be worth nearly $4,000 worth nearly $4,000 today. That's insane. Look at the growth today. That's insane. Look at the growth chart. SanDisk up about 3825%
[00:18] 3,825% in just one year. Micron up 718% which is also phenomenal. and the NASDAQ 100. I mean, 26% is still good, but dang, man. If you put that $100 into the NASDAQ 100, you'd have $126. You put it
[00:35] NASDAQ 100, you'd have $126. You put it into SanDisk, you'd have almost 4 grand. So, this has led a lot of people to ask me, Kevin, please explain SanDisk to us. What's your price target on SanDisk? What are the risk factors for this
[00:49] company? Okay, let's just get right into it. So, here's the thing. You got a lot you've got some real problems. Here are the problems that we're going to break down. First, they kind of offbalance sheet the manufacturing for their chips.
[01:05] of a hidden risk, but it could be brilliant. They also have these really large deposits coming in under the guise of refund liabilities, which have to do with the hyperscalers. Those are kind of weird because they increase cash flow
[01:20] weird because they increase cash flow now, but they increase risk later. And where they talked about high bandwidth memory versus high bandwidth flash. And
[01:33] memory versus high bandwidth flash. And they pulled some misleading tricks to hopefully kind of make them seem a little bit more like a memory stock play, even though they're a storage stock.
[01:46] mean like hard drives, right? Like uh solid state hard drives. Anyway, uh and it's working. You know, the stock has done really well. I mean, frankly, ever since Leopold got Leo margin called on it, uh Leo had somewhere around $5
[02:02] billion uh reported at the end of June in SanDisk. The end of June in fairness was right about here, probably this green candle here. So around 2,200 bucks a share. uh and and so the liquidation that happened for leopold somewhere
[02:18] that happened for leopold somewhere around here July 28th 29th you know have the valuation of that for leopold since then we've had quite a bit of recovery in the stock we're now knocking on the door of the 1836 retracement line here
[02:32] we blew through 1676 and frankly since their investor day a few days ago the stock is up substantially it's up almost 50% since investor day so there has to be like some seriously good fundamental news at this company, right? Well,
[02:48] We're going to break down the fundamentals and there's a lot there's, those three things that are weird and bearish, but there are a lot of things the company. So, remember, first of all, this is a storage company. This is not a
[03:02] memory chip play. So, instead, think solid states. Okay, I'm going to be a little quick on the balance sheet here. The balance sheet pretty dang good. Okay, $9.4 $4 billion cash and cash receivables uh or accounts receivable.
[03:16] We've got 6.5 billion in long and current debt. Pretty much almost have enough cash to pay all of our bills. Really balance sheet not really worried about it. The income statement is freaking glorious. Okay, this is like a
[03:29] dream snapshot of an income statement. And we're going to go through valuation a dream right here. Look at this. Their revenue literally went up 4.7x.
[03:41] revenue literally went up 4.7x. Their costs didn't go up, their costs Their costs didn't go up, their costs actually went down. What, Kevin? So, you mean to tell me they sold they made five times the revenue with fewer costs?
[03:55] [laughter] Yeah. Exactly what happened. And they know, we know pricing is a big part of this. I mean they said that of their Q4 this. I mean they said that of their Q4 growth 23 of the increase of the revenue
[04:10] they saw in the fourth quarter was solely because they raised prices. That's it. Bit shipment so that's how you measure kind of volume growth in storage went up about 33%. And that's fair because you're making bigger and
[04:25] bigger chips. So you could be selling actually fewer, you know, chips that you could be more bit dense, uh, or, you know, with with a newer innovation doesn't so much matter. The point is what they're actually selling, which is
[04:40] what they're actually selling, which is bits storage up 33%. Pricing up 67%. So most of this is like they're just they've got massive PB right now. They call. They literally use the word pricing power. I'll pull it up right
[04:54] pricing power. I'll pull it up right now. Uh, and I'm like, let's go. I love now. Uh, and I'm like, let's go. I love it when companies talk about pricing, especially pricing power. Take a look at this. We took significant pricing in the
[05:07] prior quarter. That's basically them saying, "Yo, we got big pee pe." And it's true. Okay, there's their their product is like twothirds sold out through fiscal 2028, which their fiscal year is a little weird. It's it ends in
[05:22] the middle of the year. June 30th. Don't even get me started on Don't worry about it. I'll help you walk through that. Okay. Basically, a year from now, a year out, okay, that window is sold out 67%. It's like, damn, man.
[05:37] Well, what's the problem then? This sounds awesome. This is like a freaking rocket ship. And yeah, if you only look at this company as a snapshot of what it is now and you don't even care about high bandwidth flash, which there's an
[05:52] opportunity there. You don't care about the liabilities, you don't care about the offbalance sheet manufacturing, fine. Yeah, I don't blame anybody for this looks really good." I mean, frankly, their earnings per share are
[06:08] frankly, their earnings per share are expected to be $211. Uh so I mean just look at that for a moment. If you take the stock price of about$,780, let's call it. Let's uh let's go get you
[06:20] know the handy dandy whiteboard here. Okay. And we take the stock price and we go 1780 and just divide it by what do we say $200ish dollars in earnings. I mean,
[06:32] this company is trading for a price toearnings multiple of literally just 8.9 as a PE ratio. That's really, really low. Now, many of you who watch my channel know, hey, we got to divide that
[06:48] by the G. We'll talk about the G in just a moment because there are problems there, too. But I mean, honestly, income statement, balance sheet. Yeah, man. Let's freaking go. This is really good. Like, I got to
[07:03] go through this myself and I'm like, let's go. This is awesome. So, then I is insane. They basically went from a loss of net income last year to almost imagine that? Like, imagine you're running a company and last year you're
[07:18] like, "Yeah, we lost $23 million." And uh that was last year. And everybody's million. Haha." And then next year you're like, "I just made $6.9 billion." [laughter] Okay, that's really impressive. Uh stock
[07:33] compensation's pretty stable over here. All right, let's get into the some of the the we'll get into some of the risks in just a moment because there's there's a lot and you kind of have to understand a little bit of the background of this
[07:45] company to really get the full picture here. So, I'm going to walk this back talking about what the heck the difference is between what they provide difference is between what they provide and what DRAM is. Okay? Because
[07:58] everything in AI right now is circulating around what's called the memory wall. All right? So, let's talk about the memory wall. When you ask artificial intelligence something about let's say a really long contract like
[08:13] you know you go upload an 800page lawsuit or financial filing or whatever lawsuit or financial filing or whatever the AI has to remember a lot of information and right now a lot of that information gets held in what's called
[08:29] high bandwidth memory which is really just DRAM stacked together. DRAM is the same kind of crap that you get in your computer basically, except rather than going into your motherboard slots, those little PCI slots like that, they're
[08:44] basically sandwiching them together like this and then they drill a hole through them. So you could literally have this conduit channel, uh, which is copper in the middle and silicon in the outside. Connect all of these together in a
[08:59] really fast, low latency manner. That's DRAM. That's uh your SKH Highix style high bandwidth memory. That's the stuff that's in the GB300's,
[09:11] right? The Blackwell chips. This is your front tier LLM producing crap. And DRAM prices have gotten really expensive. Okay, pricing in in the memory area has Okay, pricing in in the memory area has gone wild. Okay, so that's DRAM. What
[09:26] SanDisk does is different. It's not SRAMM either. Quick note, SRAMM is just the quick really fast like good for voice inference memory that's right on chip. Cerebrus and Gro use those which Nvidia acquired Gro to compete with
[09:41] Cerebrris totally different topic here doesn't really come into play into any of this. Okay, so we're going to not talk about them again. Nan flash that is what SanDisk is. They're your solid state solid state long-term storage. you
[09:57] turn that computer off, it's still going to remember what you talked about. And bandwidth flash will come in, the premise is, hey, we could take some of these similar strategies that are being used for DRAM, and we can apply some of
[10:11] used for DRAM, and we can apply some of these strategies to flash. Like, I'm way oversimplifying here, but what if we kind of like layer flash on top of another thing of flash? Way oversimplifying. Okay. And we can also
[10:24] use these channels. Wow, we could get a way faster flash product. Okay, why do you care about that? Because of the memory wall. Remember the example of like the 800page document? Okay, high bandwidth memory is really expensive.
[10:40] You don't want to use up all your HBM just to remember all that crap. If the crap ain't changing, just put it onto a hard drive somewhere and reference it when you need it. That's the premise of what SanDisk is trying to do. They're
[10:54] trying to go from a typical solid state provider to a future where we can actually help you do inference faster and better. And so the bull case here is
[11:06] and better. And so the bull case here is even if we get a growth slowdown in the even if we get a growth slowdown in the storage side, fear not, we have the next storage side, fear not, we have the next big thing. Okay. All right. That that's
[11:18] cool. Now, that can sometimes work for companies in my opinion when they're companies in my opinion when they're really cheap, for example. Okay? And and this could be a total failure. Okay? Look at a company like Nphase. It's been
[11:32] a loser of a stock. The only hope that people have for that company right now is twofold. either interest rates come down and the puppy goes up or their down and the puppy goes up or their [snorts] server uh power converters take
[11:47] off in 2027 and 2028. So here you have a cheap company that has a potential yolo in AI. That's NFACE, right? SanDisk has, hey, we got high bandwidth flash and that's another S-curve. But you're scurving at a really pricey point
[12:03] already, right? Interestingly, high bandwidth flash will be built is expected to be built in partnership with SKHEX. So, you know, some of the margin is going to be going to SKHEX, not uh
[12:18] SanDisk. I now in case I haven't already reiterated this, I want to be exceptionally clear. My I don't own SanDisk, Micron, Samsung, SKH Highix, long or short, no nothing. Okay? My only motivation here is to for you to like
[12:35] this content and maybe you'll subscribe to the channel. Maybe you'll end up like, you know what? Yeah, I'm going to download the Meek Kevin app. I'm going to end up uh using the Alphawire product so I could see breaking news the second
[12:47] want to see what Nancy Pelosi is buying. So, oh look at that. Nancy Pelosi, we for the suits. You know, she's buying That's it. And I just say that here because sometimes people like Kevin, are
[13:01] because sometimes people like Kevin, are you paid by by a competitor? I get people being jaded. [laughter] Just going to be totally honest here. No. Like I'm looking at this company because I'm like what's going on here?
[13:14] Like what do we what what's being missed and what do people not understand about sell it. We'll talk about price targets in just a moment. But that's a lot already. Okay, that's a lot of information. If you have made it this
[13:28] far, you are impressive. Okay, now let's get into some of the problems. All not so much of a property problem, but it's actually really brilliant is Flash
[13:40] it's actually really brilliant is Flash Ventures. Okay, so Flash Ventures is a company that they co-own with Kyokia uh Kyokia. I'm probably mispronouncing that, but whatever. Uh Sandis owns 49% of this. And this company, Flash
[13:54] Ventures, you could see here in orange, notes issued and distributions from, it's basically a company that they own 49% of. And the idea is they're going to
[14:06] go to this company called Flash Ventures. And Flash Ventures is going to Ventures. And Flash Ventures is going to own the FAB. Flash Ventures is going to own the FAB. Flash Ventures is going to sell the wafers that SanDisk uses at a
[14:18] sell the wafers that SanDisk uses at a cost plus model, which is basically cost plus a tiny margin, so Flash Ventures can stay in business presumably based on what we're reading here. And SKH Highix gets all the benefit of margin without
[14:31] gets all the benefit of margin without complicating their sheets. So when they complicating their sheets. So when they give a loan to Flash Ventures and let's give a loan to Flash Ventures and let's say they lend $500 million which is a
[14:43] company that's estimated to have about 15 billion in machinery and assets you know so when they send $500 million to this company as a loan it's really this company as a loan it's really disguised capex but look in the last
[14:57] year if you look at this company's cash flow statement SanDisk. Okay, so here's the year, right? Year ended. If we go down here, look at this. I find this just odd. Uh, purchases of plant, property, and equipment. We've got 177
[15:11] "Oh, okay. Yeah, that's their capital expenditures, almost, you know, $200 million." But wait a minute, what's this over here? Note receivable issuances to Flash Ventures, almost another half billion dollar. So, you know, 2 to 3x
[15:25] putting into capex are really going into capex. So, this is a little bit misleading because you're taking the actual fabricators, the machinery, the payroll, and all of that. You're putting that into the private side of the
[15:39] business, and you're giving the public the good stuff. That's great. That sounds fantastic. A lot of people will be like, "Great. I don't want any of that trash anyway." But the problem is, if Flash Ventures starts losing money
[15:53] because there's a downturn or they need to replace their machinery or whatever, who they going to call? they're going to call their 49% partner. And so it's less obvious to investors that SanDisk has that phone call risk. I mean, I guess
[16:07] own company. Peace, but then obviously you're not going to get manufacturing at problem. But, you know, a lot of people aren't aware that they have that off off balance sheet risk because of how they set this up as just loans to this other
[16:21] set this up as just loans to this other company that they have. Weird. That's it. Samsung doesn't do this. Samsung owns their stack. Micron owns their stack. SKH Highix owns their stack. Okay, it's just a different business
[16:34] model. It's kind of brilliant engineering, but there's more that's going on here that's a little weird. So, then I look at the cash flow statement. Uh, and we've got this line right here called refund liability. Now, this is a
[16:47] really interesting one because refund liability does not show up anywhere in liability does not show up anywhere in their prior 10K or their prior 10 Q's. this correctly, and I could be wrong, but I believe they're recognizing this
[17:01] as cash flow coming in, and those are really just upfront deposits that they're getting from the hyperscalers, they're getting from the hyperscalers, which is cool, but it kind of inflates
[17:15] which is cool, but it kind of inflates how much cash flow you have today. And if the hyperscalers cancel and you get your cancellation fee, then people of cash, but they may have already gotten that cash and used it on stock
[17:30] gotten that cash and used it on stock buybacks. Now, some of that is not entirely clear either because look at this. Each of our new business model financial agreements, those are the contracts with the hyperscalers,
[17:43] come through a combination of cash deposits and financial instruments totaling 16.5 billion, which is intended to protect SanDisk if a customer fails to satisfy its purchase obligations. So basically, if these other companies
[17:57] basically, if these other companies cancel, we keep some of the money, but some of that money has already shown up. Here it is. Take a look at this. On page four of their earnings call, they actually talk about how they received
[18:10] almost $2 billion of NBM prepayments and deposits, which are included in cash flow from operations. So, if somebody didn't listen to that part of the earnings call, they would look at this $7 billion
[18:25] and go, "Oh, wow. They made $7 billion of cash flow on net income of 6.9. These people are amazing." But wait, $2 billion of that were really just billion of that were really just deposits. So they got deposits
[18:39] that, yeah, they got to keep, but it sort of pumps up what that uh dare I say um cash flow number looks like and their capacity to keep doing buybacks. So if those hyperscalers cancel and you already spent that money on doing
[18:54] "Okay, whatever. The hyperscalers are gone, but you're still generating cash flow, right? uh you guys could still do buybacks, right? Uh you uh we we may have had a lot of cash flow because we got a lot of prepayments and we may have
[19:08] already used that to buy the stock. So, in other words, if the cycle turns on them, they've kind of already used up some of their capital to do buybacks and that wasn't necessarily the clearest structure up
[19:21] front. I'm not saying that's bad. The reason they do this is because you have to build out the machinery, right? and and they got to send more money to Flash out. And they're trying to reduce cyclicality at the business by doing
[19:35] these contracts with, you know, new business models. This is important and there's an upside here. So, let's talk about this because historically, you get about this because historically, you get like a 2023 and these puppies crash like
[19:49] crazy because you go through a really dirty down cycle. Nobody wants to go dirty down cycle. Nobody wants to go through a down cycle, but these stocks, memory and uh these storage stocks, they do. So, they're trying to kill that. How
[20:03] are they trying to kill that? Take a look at this. We now have new business model agreements with eight diverse data centers and edge customers. We expect our new business models to represent more than 50% of our bits that we sell
[20:17] for the next 12 months and about 67% of our bits in the year after the next 12 months because their calendar starts July 1st. It's again weird trying to help uh make make sense of that. So in other words, we're going from you know
[20:31] other words, we're going from you know the next 12 months here 50% going into these contracts to the next 12 months thereafter. So 12 to 24 months 12 months thereafter. So 12 to 24 months from now, uh we're going to about 67%
[20:44] into these cancellable data center hyperscaler uh contracts. As long as you're hyped up about hyperscalers continuing to spend, cool, maybe less risk here, right? But this is a lot of concentration on those hyperscalers,
[21:00] especially since they say the following. Look at this about their growth. So Joe, basically we continue to be committed to build bits in the mid to high teens, right? And we're spending at that level. This was near the bottom.
[21:15] call, a little more than halfway into their earnings call. This is actually a really important line from their CFO because what they're saying is even with these contracts, we're really only committing to increase volumes by call
[21:31] it mid to high. So call it 16 to maybe 17%. Right? Otherwise they would just say high. Mid to high is usually around 16 17%. So call it 16%. So even with these really big contracts that they're signing, they're actually telling you
[21:46] they're only expecting volume, bit volume to go up 16%. in sort of this longer term like okay 16% volume growth means either pricing
[21:59] power has to keep skyrocketing in order to get this or we're not going to be growing at 4.7% anymore. Remember at the beginning how I'm like, "Oh yeah, yeah, yeah, guys, revenue almost 5x. This is freaking awesome. This is lit." Yeah.
[22:14] Well, why then are you forecasting that you're only going to grow bits at 16%. So in other words, any kind of growth above 16% above 16% has to be from pricing power. So 16% bit
[22:29] growth plus pricing power equals revenue growth, right? So 16 plus% plus PP growth, right? So 16 plus% plus PP equals revenue growth. Call it RG. Okay.
[22:42] equals revenue growth. Call it RG. Okay. Fine. What this does for you is it sets Fine. What this does for you is it sets up some of the issues and it sets up why high bandwidth flash is so important to this company. So what we have is we have
[22:56] in the downside risk we have the offbalance sheet assets. Okay, a lot of offbalance sheet debt. That's fine as long as people know about it, you know. assets. Kind of weird with the fabricators. maybe it's not a problem.
[23:12] fabricators. maybe it's not a problem. Then we've got these upfront payments Then we've got these upfront payments that bolster the cash flow of the company. Upfront cash flow. Okay, cool. I mean, they need it to make their
[23:24] whatever we want to call it. Like, those are just two little notes here. Then we are just two little notes here. Then we were talking about 16% growth. Okay, I talking about 16% growth? It should be growing a lot more than that. It just
[23:37] grew 4.7x. This is how people get to like, oh my gosh, it grew 4.7x and it's only pay selling for eight times earnings. What's going on? They should be selling for so much more money. And then we get to something uh weird, I'll
[23:53] call it, with how they branded high bandwidth flash. I think they did something a little sussy here with the high bandwidth flash argument. Now, this is going to get a little bit technical, but I'm going to do my best to simplify
[24:08] this and and maybe sort of like graphically demonstrate it uh in the simplest way possible because it's a lot. Okay, so high bandwidth flash. Again, we already kind of drew out a little bit what the idea is, but the
[24:23] idea is, hey, let's get into inference and see how we can, you know, really pitch some more dollar hollers. In fact, they talk about just that on their earnings call when they're asked about Nvidia. On their earnings call, they're
[24:36] Nvidia. On their earnings call, they're asked about asked about the CMX storage architecture, right? So, Nvidia's got this storage next concept with CMX. this storage next concept with CMX. Basically, Nvidia's idea is instead of
[24:50] taking our server racks where we've got all these like Blackwell 200 chips and they all have high bandwidth memory on them, why don't we just like talk to a
[25:03] different part of the data center that just contains a bunch of solidstate hardware. That's going to really be your NVME. All right, solid state hardware.
[25:15] So, let's call this the uh storage. Think about it almost like a coat closet. So, you got like workers and employees working in a coat closet. And then when the high bandwidth memory wants to call on some high bandwidth
[25:29] flash uh or in this case just just a storage, you could do that. You could do that through the Ethernet connections in the vision of Jensen or you could do it in the high bandwidth flash architecture.
[25:44] Very different things. Okay, Jensen strategy is right here what I just drew. This is the Jensen strategy. This has to do with CMX. High bandwidth flash is a different product that can help us address a similar strategy. This is
[26:00] hitting the memory wall. How are we going to solve it? Well, we're going to going to solve it? Well, we're going to solve it with Nvidia storage. We're going to use the CMX infrastructure and we'll solve the memory wall. You know,
[26:13] Jensen's smart, okay? He sees a problem, something that's going to hold up the business and he's going to go to town on trying to solve it. It's a really smart guy, okay? You got to give it to him.
[26:26] SanDisk is trying to do that as well. So SanDisk ended up doing this investor day presentation. Let's pull up a screenshot of it. In the investor day presentation, they showed a sort of simulation that they did between high bandwidth flash
[26:43] and high bandwidth memory. And the idea here is wait, they seem to be comparing themselves to high bandwidth memory, which has been skyrocketing because it's that vertically stacked DRAM. Prices have been skyrocketing for that. Nvidia
[26:57] can't get their hands on enough of it. Uh Nvidia's probably got 33% of the supply for high bandwidth market on lock for the next few years. They need it in their GPUs. SKH Highix, Samsung and Micron make this. I think Samsung SKH is
[27:12] Micron make this. I think Samsung SKH is probably the largest producer uh as as a looking for like that pure play kind of HBM play, a lot of people have been we saw the Korean market sort of blow up on this, right? But high bandwidth
[27:26] flash, which we've explained also uses those little drilling channels, is being pitched by SanDisk as having the having pitched by SanDisk as having the having somewhere between 8 to 16 times the
[27:41] somewhere between 8 to 16 times the capacity of high bandwidth memory and they're pitching that you could do it at a fraction of the cost. So that's the pitch. The sales pitch is, hey, you know that really expensive thing called
[27:55] HBM that Nvidia can't get enough of? Yeah, we're going to do something that's Yeah, we're going to do something that's 8 to 16 times as good and cost less money. That was their investor day. And so people were like, "Oh my gosh, the
[28:09] stock's going to go to the moon. We got to buy more SanDisk." But there were some details that were a little, let's just say, skewed in favor of SanDisk. Sandis got slammed for this, rightfully
[28:24] so, but only by some people who had the technical sort of background to break this apart. The first thing they did is they used Quen uh on a $480 billion
[28:36] they used Quen uh on a $480 billion billion parameter model in Bflat 16. They're a lot of numbers. I'm sorry. I know it's a lot. I'm gonna try to keep it as simple as possible. Most LLMs today are not running on Bflat 16. Most
[28:55] of them get quantized and divided by four. Okay. So, most LLMs are actually running on a fourth of the size that was used here in this experiment. So, a lot
[29:07] of people really criticize this B16 usage. The next thing they criticize is usage. The next thing they criticize is that Sandis kept the bandwidth right that Sandis kept the bandwidth right here stable at 12.8 terabytes a second.
[29:21] Now, why would they do that? Why would they do or make these two adjustments keeping the bandwidth stable and using Bflat 16? Well, the reason is a lot of
[29:33] people point out to cover up the failure of high bandwidth flash, which is speed. When you put a 480 billion parameter When you put a 480 billion parameter model on this BF-16, you require a
[29:47] terabyte of VRAM, which would be like 315090s or 11 Blackwell 6000s or 4GB300s. So you have to link together a bunch of GPUs and you're creating this bottleneck that isn't actually real in artificial intelligence. If you compress
[30:02] it like you usually do, you fit it into a normal GB300 and basically just one of them. That giant model you fit into just one GB300. one GB300. And so what they've done is SanDisk
[30:15] ignored where their technology high bandwidth flash breaks down. And this is potentially why one of their partners, Kyokia, which is a Sandis partner, actually believes high bandwidth flash doesn't
[30:29] belong in the data center. It belongs in mobile networks and cell towers, not data centers. So there's there's some disagreement here about where this technology begins or and ends and where it's used. And some people look at this
[30:43] have to set it up in such a weird way?" Because this doesn't need to be HBM Because this doesn't need to be HBM versus HBF. we could use both of them. It kind of seems like you might be being a little disingenuous.
[30:56] And so some people say, well, it's just a tool for them to pre-ell the next Scurve, the next growth Scurve. Why would they want to do that? Well, it's because the G, the forecast estimates are not that good for the G.
[31:14] let's understand a little bit more about high bandwidth flash just so you have a little bit more color on this. Uh we'll go deep on this maybe for two or three minutes. I don't want to go super deep in it, but I want you to understand high
[31:28] bandwidth flash because it's probably not going to go away. I actually think it has its purposes. It's got a lot of purposes and there's some uses here, but bandwidth is really important and as is latency. And there's a good chart that
[31:43] can help you understand this especially when we think about LLMs. So let me pull that chart really quickly. That chart is going to be how latency compounds. Here going to be how latency compounds. Here it is. Okay. So take a look at this. A
[31:57] uh high band with flash architecture is going to have to wait for a full cycle going to have to wait for a full cycle read and write every single time it uh goes and collects data or information. that is really slow. It doesn't do it in
[32:14] parallel as you might if you could prefetch information. prefetch information. When would you do this? Because see in the SanDisk example they had during investor day, they prefetched or
[32:28] basically preloaded all the information. They knew exactly what weights they were they were going to do with the model. And it basically let you run the high bandwidth flash setup in parallel, which is indistinguishable. That's how they
[32:42] got to about 98% of the efficiency of high bandwidth memory. That's how they pulled it off. You already knew what you were going to ask for. But the problem with this is high bandwidth memory when you get into mixture of expert models
[32:57] slows down or uh high band with flash slows down a lot when you get into high um mixture of expert models. Why? Because ane is really where you're going, all right, are we going to ask this part of the data set or this one or
[33:12] this one or this one? Well, it's going to depend what's our first result, then what's our next result? Then what's our next result? What's the user inputting or whatever? All of this latency stacks up and hurts high bandwidth flash in a
[33:26] way that it doesn't hurt high bandwidth memory. Just for an example of the speed difference between the two. What takes high bandwidth memory one second takes high bandwidth memory one second takes high bandwidth flash 17 minutes. That's
[33:41] an extremely scaled up example because high bandwidth memory ain't really doing happening in like the blink of an eye. Think about it like you blink your eyes. That's 300 milliseconds gone. High bandwidth memory can do 3 million back
[33:56] and forth requests in that time frame. High band with flash can only do 3,000. That's a 1,000x difference in speed. So high band with flash, frankly, is really slow for nonpretermined
[34:12] purposes. Mixture of expert models break this. You go pull up the same contract, that same 800page, you know, legal contract or whatever, high bandwidth flash can easily win because high bandwidth memory caps out. like we don't
[34:28] have enough high bandwidth flash. No problem, man. We got you on the storage system. We know exactly where that data is. We will go pick that data up and get it to you in flash speeds. [laughter] Uh but high bandwidth flash really fails
[34:44] where HBM is sold. HBM is sold for training, fine-tuning, and rewriting all of the different models that you have or working these mixture of expert LLM uh
[34:58] answers. So, think about this, okay? And this is where people get a little frustrated with SanDisk. So, high bandwidth memory, think train, think writing. So you're changing underlying data because you're changing weights,
[35:12] you're changing the whole algorithm uh or you're doing mixture of experts uh or you're doing mixture of experts work. High bandwidth flash is inference which is fine but it's basically think of it more of like a readonly tool.
[35:29] That's when it's the fastest because see writing speeds are uh are really slow because of the physical way flash works. Writing is considered to be 10 times slower than reading data in flash because you're literally pushing energy
[35:44] into a cell rather than just sensing what's the energy on that cell. That's how you're pulling the data. And the more you rewrite that data, the more you degrade high bandwidth flash cells, any kind of flash cells that's true for, and
[35:57] the more you actually limit the long-term use or efficacy of the high bandwidth flash in the first place. So uh this decays so high bandwidth flash
[36:09] decays uh it's much slower for training writing a mixture of experts models. Yes it is cheaper but it's cheaper for certain purposes. Most of the AI spend today where do you think it is? I mean I hate to say it. We
[36:26] think it is? I mean I hate to say it. We know where the AI usage goes. AI usage should be going right here. This is where AI usage goes. That's why companies I think like Cerebrus are exciting because you know you're
[36:39] focusing on the inference side. High bandwidth flash. We could focus on parts because of the mixture of experts problem. But where is the frontier money really being spent? Who's spending the most money? Well, who's spending the
[36:53] most money? Well, who's spending the most money in all of this is obviously like anthropic and open AI. And what are they buying? They're buying tools to help them train better. That's why we're seeing such premiums on high bandwidth
[37:08] memory. And that's why a lot of people are saying, "Hey, SanDisk, you guys are comparing yourself to high bandwidth memory." Well, you have a purpose in the AI data center. You are not high bandwidth memory. Don't compare yourself
[37:21] bandwidth memory. Don't compare yourself to our high bandwidth memory folks. Now, I think it's cool that they're pitching high bandwidth flash. There's a reason why I think they're doing it though. So, I think what's happening is prices for
[37:34] this flash storage have scurved up and the growth rate is starting to plateau. See that in just a moment. What they're trying to do is go from peak SanDisk to
[37:46] trying to do is go from peak SanDisk to let's get a new growth vertical going. That's your compounding scurve. That's why there's so much excitement about high bandwidth flash. Again though, the downside is the money
[37:59] LLMs. It's being spent on the high bandwidth memory. This is not a replacement for high bandwidth memory for training, which is where a lot of replacement for mixture of model experts, which are really common right
[38:13] now, especially on the frontier. This is a fantastic tool for specific things that don't change. Okay, great. That explains this little
[38:25] Okay, great. That explains this little fancy chart in a nutshell. So, where do fancy chart in a nutshell. So, where do we sit with growth rate expectations for this company? All right. Well, this is where things get unfortunate for the
[38:38] consensus, but maybe the consensus will be wrong. I'm going to start by pulling be wrong. I'm going to start by pulling up the Mazou piece on this. Mazou actually argues that this company could potentially see average selling prices
[38:53] come down. Look at this. We believe consensus calendar 2027 average selling starting in 6 months from now, the next 12 months are down 15 to 20%. So in
[39:08] other words, Mazu is already forecasting that prices are going to come down for SanDisk products and average selling prices are actually they think well the prices are actually they think well the cons let me let me clarify this they are
[39:23] pointing out that the consensus estimates are that average selling prices are going to come down 15 to 20%. They actually think average selling They actually think average selling prices will be flat to slightly up. So,
[39:35] prices will be flat to slightly up. So, let's clarify that consensus. This is let's clarify that consensus. This is sort of the Wall Street forecast. Prices down. Remember how important prices were to this company?
[39:47] Mizu actually thinks we're going to be flat to slightly up. The current actual forecasts for growth for this company, they're not that great. The current
[40:00] they're not that great. The current forecasts are as written. We are going forecasts are as written. We are going to take uh earnings per share of 21134 on EPS and we expect that to grow by 22.1% next year. Oh, that sounds great,
[40:14] Kevin. That's a half peg. You're right. It's less than a half peg. Seems really cheap, right? But wait, the next year's estimate is negative 30.9. The next year's estimate was actually
[40:32] 20. Hold on one sec. Let me uh That's the next one. Uh 20.9 - 20.9. The next the next one. Uh 20.9 - 20.9. The next year's estimate 30.9 and the next year's estimate after that 49.
[40:46] Yikes. So now all of a sudden when you put together the growth rate for this company, you might look at this and say, "Hey, it's got a 9 PE ratio or an 8 PE ratio, whatever you want. It doesn't matter." We're going to divide that PE
[40:59] ratio by a growth rate that in this case if I average these numbers is like if I average these numbers is like -78.7%.
[41:12] up divided by four that works out to a growth rate of per year 19.68%. So in other words, a 9p ratio looks good, but if your forecast is shrinking
[41:25] by nearly 20% per year, then your PEG ratio is infinite. I mean, any negative growth rate is going to have an infinite peg uh, you know, peg ratio. So that's where there's a problem. The Wall Street consensus is this puppy's going to start
[41:40] shrinking its pricing power. But that's an issue because if all of a sudden their pricing power, their ability to increase prices, if ASPs, average selling prices go down, then the only thing that's helping us grow is bit
[41:56] volume. But bit volume growth is only expected to grow mid to high teens. In this case, 16% is what they forecast. That's what they wrote in their earnings call. So now what you actually have is a company that looks really cheap.
[42:14] They look like they're spending a ton of money on stock buybacks and they've got money on stock buybacks and they've got a solution for everything. They've got a solution for the memory wall problem, which in fairness, they say there are a
[42:28] they're honest about this. They're like, look, they wrote where they said there's a lot of different ideas. This has to do with the memory wall solution. See, they inference is memory bound. Everybody's got different ideas. That's kind of like
[42:43] Jensen has this idea like, "Hey, let's just have uh, you know, the CMX architecture go pull uh from data storage." We've got SanDisk saying,
[42:55] "Hey, let's use high bandwidth flash." Uh, we've got other people going, "Hey, [laughter] Right? Like there are a lot of ideas for Right? Like there are a lot of ideas for the storage solution, but the big
[43:09] problem is right here. It's that mid to high teens of bit growth because if pricing starts shrinking now, we basically collect problems. And I'm not
[43:21] trying to say this is a bad investment, but when I put this together, I go, man, I've got a lot of things that I didn't know about regarding SanDisk that now I know about. Right? I've got off balance sheet risk. That's the fabs. That's the
[43:38] capex that's basically being hidden right now. It's a hidden capex. I've got right now. It's a hidden capex. I've got upfront payments from the hyperscalers buybacks, but I don't know if those are going to last. Then I've got a little
[43:52] bit of an misleading presentation on high bandwidth flash. Uh, and then I have consensus estimates that say average selling prices are going to come average selling prices are going to come down and bit growth is going to be 16%.
[44:07] So to really argue, hey, let's go invest in SanDisk, what you kind of have to do in SanDisk, what you kind of have to do is say, I think this is wrong, you have to say, I think average selling prices like Mazu are going to go up. I think
[44:23] that uh bit growth is going to be higher and I believe that high bandwidth flash, which I purposefully tried my best to explain in a simple way. I probably think I did it in a simple way. The reason I did that is because it matters
[44:39] for the growth story. So, if you could believe that growth story and your understanding of all the other stuff, hey, Sandis could be a great opportunity to me. There's a little much here to digest and
[44:54] I'd probably be [music] a bigger fan of waiting a little bit to see what happens with high bandwidth flash high bandwidth flash especially since high band with flash is being co-produced with skhix anyway. So all of the benefits if it
[45:06] takes off don't actually go to SanDisk [music] anyway. So that's it. If you consider subscribing to the channel. do my best to keep doing [music] a ton of context as possible because I think that's my job and I kind of like it. So,
[45:22] >> Why not advertise these things that you told us here? I feel like nobody else see how it goes. you. >> Kevin Pra there, financial [music]
[45:34] analyst and YouTuber. Meet Kevin. Always great to get your take.