저자: Kevin · 외국인
요약 보기
샌디스크는 1년간 3825% 급등했으나 PER 8.9로 저평가 주장. 그러나 자산관리 측면 위험(오프밸런스 설비 투자, 선수금 기반 현금흐름, 플래시벤처스 대여금)을 지적하며 고성장 지속을 의심. 회사는 이익 대부분이 가격 인상(출하 +33%, 가격 +67%)에서 발생. 컨센서스는 ASP 15-20% 하락과 연간 성장률 -20% 전망으로, PEG 관점에서 비싸다고 봄. 본인은 포지션 없음을 밝히고 신중한 관망을 권고.
다듬은 글
So, if you invested $100 in SanDisk a year ago, it would be worth almost $4,000 today. This is just madness. Look at the growth chart. SanDisk grew by 3825% in just 1 year. Micron grew by 718%, which is also phenomenal. And the NASDAQ 100 ...well, 26%—that's not bad either, but damn . If you invested that $100 in the NASDAQ 100, you would have $126.
If we invested in SanDisk, we would have almost four thousand. So a lot of people ask me, "Kevin, tell us about SanDisk. What's your target price for SanDisk? What are the risk factors for this company?" Okay, let's get straight to the point. So, what's the point? You see a lot of hype around this company, but there are also real problems.
These are the problems we will now analyze. First, they seem to be taking their chip production off balance. This is a bit strange. This is a hidden risk, but it can also be a brilliant move. They also have very large deposits coming in under the guise of refund obligations related to hyperscalers. This is a bit strange because they increase cash flow now, but increase risks later.
And they also had an investor conference where they talked about high- bandwidth memory versus flash memory. And they've resorted to tricks to make it look more like memory stocks, even though they're actually storage stocks. And I don't mean public warehouses. I mean hard drives, right? Types of solid-state drives.
In short, it works. You know, stocks are feeling really good. Frankly, since Leopold got a margin call, Leo had about $5 billion in SanDisk at the end of June. The end of June, to be fair , was about here, probably this green candle, which is about $2,200 a share. And so, the liquidation that occurred at Leopold's somewhere here, on July 28-29, is half the score for Leopold.
Since then, we have seen a fairly significant recovery in stocks. We are now knocking on the door of the 1836 correction line . We broke through 1676, and frankly, the stock has risen significantly since investor day a few days ago . They have increased by almost 50% since the investor's day. So, there must be some serious fundamental news in this company , right?
Oh, that's what we're going to figure out. We are going to analyze the fundamental indicators. And there is a lot here. You know, we'll talk about those three weird and pessimistic things, but there's a lot that's really, really exciting about this company. So, first of all, remember, this is a data storage company.
This is not a game in the memory chip market. So think about solid-state drives instead, okay? I'll be a little quick about the balance sheet. The balance sheet is pretty good, you know? Some $9.4 billion in cash and accounts receivable . And we have 6.5 billion in long-term and current debts. Almost enough cash to cover all our bills.
Honestly, I'm not too worried about the balance . The income statement is just great, you know? It's like a perfect snapshot of the income statement . And we'll go through the evaluation of everything soon, but that's just a dream. Look at this. Their revenue literally increased 4.7 times. Their expenses have not increased.
Their expenses actually decreased . What? Kevin, are you saying they got five times the revenue with less spending? Yes, that's exactly what happened . And they themselves admitted, "Hey, look, we know that pricing is a big part of this ." I mean, they said that in their fourth quarter growth, two-thirds of the revenue increase was driven solely by price increases.
That's all. Bit shipments, which is how you measure storage growth , are up about 33%, and that's fair, because you're making bigger and bigger chips, so you can sell fewer chips that you can hold in your hands, right? But they may be denser in terms of bit count or due to newer innovations or whatever, it doesn't matter.
The bottom line is that what they're actually selling is bits, storage, 33% growth, prices up 67%. So, most of this is like they have enormous market power now. They said it themselves during the earnings report. They literally used the term " pricing power." I'll open it now. And I'm like: let's go.
I love it when companies talk about pricing. Especially about pricing power. Look at this. We raised prices significantly in the previous quarter. It's essentially them saying, "Yo, we have a lot of power." And that's true. So, their product is about 2/3 sold out by the end of fiscal year 2028. Although their fiscal year is a bit strange.
It ends in the middle of the year. June 30. Don't even start a conversation about it. I hate it when companies do that. Don't worry about it. I'll help you figure this out, okay? Essentially, in a year, in a year's perspective, you understand? This period is 67% sold out. Like: “Damn it. Well, what's the problem then?
That sounds amazing. "It's like a [ __ ] rocket." And so, if you look at this company just as a snapshot of what it is now, and you're not interested in high-speed flash memory, where there's actually opportunity, you don't care about the liabilities, you don't care about off-balance sheet manufacturing—fine.
Yes. I don't blame anyone for looking at these stocks and saying, "Wow, that looks really good." I mean , frankly, the expected earnings per share is $211. And so, just look at this for a moment. If we take the stock price around 1780 , let's call it this. Let's take, you know, our handy board, okay?
We take the stock price of 1780 and just divide it by, what did we say? About $200 profit. That is, this company trades at a price/ earnings ratio of only 8.9. This is a very, very low figure. Now, many of you who watch my channel know, "Hey, we need to divide this by G." We'll talk about G in a moment, because there are problems there too.
But honestly, the income statement, the balance sheet? Yes, dude. They drove. This is really good. I have to get excited about this. I'm like, "Dude, I'm reviewing this myself." And I'm like : "Let's go." "That's cool." So, then I look at their net profit. This is just madness. They actually went from a loss last year to almost $7 billion in net profit.
Can you imagine that? Like, imagine you're running a company and last year you say, "Yeah, we lost $23 million ." And that was last year. And everyone is like, “Haha, losers, you lost 23 million dollars. Haha. And the next year you're like, "I just made 6.9 billion." Okay, this is really impressive.
Uh, stock payouts here are pretty stable. Okay, let's move on to some of the risks, which we'll talk about a little later, because there are a lot of them, and you need to understand a little bit of the background of this company to get the full picture. So, I'll take you back in time a little bit. Let's start with what the hell is the difference between what they offer and what DRAM is, okay?
Because right now, everything in AI revolves around the so-called " memory barrier." Okay, let's talk about the memory barrier. When you ask an AI about something, let's say a very long contract, for example, you're uploading an 800- page lawsuit or a financial report or whatever. AI has to remember a lot of information.
And now much of this information is stored in so- called high- bandwidth memory, which is really just DRAM stacked together. DRAM is essentially the same crap that's in your computer. Only instead of plugging them into the motherboard slots , those little PCI slots, they actually sandwich them together and then drill a hole so that you can literally get this through-channel that's copper on the inside and silicon on the outside, connecting it all together very quickly and with low latency.
This is DRAM. This is your SK Hynix-style high-bandwidth memory . That's what GB300 stands for, right? Blackwell chips. This is what your advanced LLMs are working on. And DRAM prices have become very high, you know? Pricing in the memory space has become just crazy, you know? So, this is DRAM . What SanDisk does is different.
It's also not SRAM. Quick note: SRAM is just fast, very fast memory, good for voice output, that resides right on the chip. Cerebrus and Groq use these, Nvidia acquired Groq to compete with Cerebrus. This is a completely different topic. She doesn't really play any role in all of this . Okay, so we won't talk about them anymore .
NAND flash is what SanDisk does. This is your solid-state drive for long-term data storage. You turn off your computer, and it will still remember what you talked about. And the premise here, and this is where high-speed flash memory comes in, is that we can take the strategies that are used for DRAM and apply them to flash memory.
I'm oversimplifying, but what if we layered flash memory on top of each other? I'm greatly simplifying, you know? And we can also use these channels. Wow, we can get a much faster flash product. Okay, why should you care? Through the memory barrier. Remember the example with the 800-page document? Of course?
High-speed memory is very expensive. You don't want to waste all your HBM just to remember all this junk. If this data doesn't change, just put it on your hard drive and access it when you need it. This is the main idea of what SanDisk is trying to do. They are trying to transform themselves from a regular solid-state drive vendor into a company of the future that helps you make decisions faster and better.
So, the optimistic scenario is that even if we see a slowdown in growth in the storage segment. Fear not, we have the next big thing. Okay, okay. This, this is cool. This can sometimes work for companies. In my opinion, when they are really cheap. For example, okay? And it could turn out to be a complete failure, okay?
Take a look at a company like Enphase . Her shares were unprofitable. The only hope people have for this company right now is twofold. Either interest rates will drop and stocks will go up, or their server power converters will take off in 2027 and 2028. So, you have a cheap company that has a potential chance in AI.
This is Enphase, right ? SanDisk says, " Hey, we have high-speed flash memory." And people think, "Oh my God, if this takes off, it's going to be another S-curve." But you 're already at a very expensive point on the S-curve, right? Interestingly, the high-speed flash memory is expected to be created in partnership with SK Hynix.
So, you know, part of the margin will go to SK Hynix, not SanDisk. I, if I have n't already emphasized this, want to be exceptionally clear. I don't own SanDisk, Micron, Samsung, SK Hynix, neither long nor short, nothing at all , okay? My only motivation here is that you enjoy this content and maybe you'll subscribe to the channel, maybe you'll eventually download the Meet Kevin app and say to yourself: you know what?
Yes, I will download the Meet Kevin app. I will use the Alpha Wire product to see the latest news as it comes out. Or, you know what? I want to see what Nancy Pelosi is buying. So, oh, look at this. Nancy Pelosi, we finally added a stock tracker to this for business people. You know, she's buying Uber and Intel.
Let's see how it goes. That's all. And I only say this here because sometimes people ask, "Kevin, are you getting paid by your competitors?" I understand that people may be disappointed. I'll be completely honest here. No. I look at this company because I think: " What is going on here?" What are we, what are we missing?
And what don't people understand about this company? And people can decide for themselves whether to buy or sell it. We'll talk about target prices in a moment. But, that's a lot, okay? That's a lot of information. If you've made it this far, you're amazing, okay? Now let's move on to some of the problems.
Good. The first thing, which is maybe not so much an ownership issue but is actually very ingenious, is Flash Ventures, okay? So Flash Ventures is a company they co- own with Kioxia. Uh, Kioxia, I'm probably pronouncing it wrong, but whatever. Uh, SanDisk owns 49% of it. And this company, Flash Ventures, you can see here in orange, the notes issued and the distributions from them.
It's actually a company they own 49%. And the idea is that they will go to this company called Flash Ventures. And Flash Ventures will own the factory. Flash Ventures will sell the wafers that SanDisk uses on a " cost plus" model, which is essentially cost plus a small margin so that Flash Ventures can stay in business, presumably based on what we're reading here.
And SK Hynix gets all the benefit of the margin without complicating its reporting. So when they give a loan to Flash Ventures, let's say they're lending $500 million to a company that's estimated to have about $ 15 billion in assets and equipment. You know , so when they send $500 million to this company as a loan, it's actually a disguised capital expenditure.
But look, over the last year, if you look at the cash flow statements of this company, SanDisk, okay? So, here's a year, right? The year is over. If we look here, look at this. It just seems strange to me. Uh, the purchase of fixed assets, real estate, and equipment—we have 177 million. Okay, people look and say, "Oh, well, this is their capital expenditure, almost, you know, $200 million ." But wait a minute, what is this ?
Receivables issuance for Flash Ventures, almost another half a billion dollars. So, you know, two to three times more money than they say they're putting into capital expenditures is actually going into capital expenditures. So it's a little misleading because you're taking the production capacity itself , the equipment, the salaries and all that, you're transferring it to the private side of the business, and you're giving the public the good stuff.
This is great. Sounds fantastic. Many will say, "Great. I don't want that junk anyway." But the problem is, if Flash Ventures starts losing money because of a downturn or needing to replace equipment or something else, who do they call? They will call their 49 percent partner. And so it's less obvious to investors that SanDisk has this phone call risk.
I mean, I think maybe they could just not respond to their own company and say, "Bye." But then obviously you won't get any more production at all. This is probably a bigger problem. But, you know, many don't realize that they have this off-balance sheet risk because of the way they've set it up. These are just loans to this other company that they own.
Amazingly. That's all. Samsung doesn't do this. Samsung owns its own manufacturing stack. Micron owns its stack. SK Hynix owns its own stack. OK. It's just a different business model. It's a kind of brilliant engineering, but there's something else going on here that's a little strange. So, then I look at the cash flow statement.
And we have this line right here called the refund obligation . This is a very interesting point because the refund obligations do not appear anywhere in their previous 10-K or 10-Q reports. So, my understanding, if I'm reading this correctly, I could be wrong, but I believe they're recognizing this as cash inflows , when in fact it's just upfront deposits that they're receiving from hyperscalers.
Which is cool, but it slightly overstates the amount of cash flow you have today. And if hyperscalers cancel orders and you get a cancellation fee , then people think there's going to be a big cash inflow, but they could have already received that money and spent it on share buybacks. Now, some of this is also not entirely clear, because look at this.
Each of our new business model financing agreements —the hyperscaler contracts— comes through a combination of cash deposits and financial instruments totaling $16.5 billion, which is intended to protect SanDisk if a customer defaults on its purchase obligations. So, basically, if these other companies cancel the order, we'll keep some of the money.
But some of this money has already appeared. Here she is. Look at this. On page four of the earnings report, they actually talk about how they received almost $2 billion in prepayments and deposits for NVM, which are included in cash flow from operating activities. So if someone hadn't listened to that part of the report, they would look at the 7 billion and say, " Wow, they got 7 billion in cash flow on a net profit of 6.9." " These people are just incredible." But wait.
2 billion of that is actually just deposits. So they received deposits, which, yes, remain with them, but it slightly inflates the cash flow figure and their ability to continue to buy back shares. So if these hyperscalers cancel orders and you've already spent that money on share buybacks, then people in the future will say, "Okay, never mind, there are no more hyperscalers, but you're still generating cash flow, right?" Yeah, you guys can still do buybacks, right?
We...we may have had a lot of cash flow because we got a lot of prepayments. And we may have already used them to buy back shares. So, in other words, if the cycle turns against them, they have to some extent already used some of their capital to buy back shares. And it wasn't necessarily the most transparent structure from the beginning.
I'm not saying that's bad. The reason they do this is because they need to build the hardware, right? And they need to send more money to Flash Ventures so they can build it. And they're trying to reduce the cyclicality of business by entering into these contracts under, you know, new business models.
This is important, and it has its advantages. So let's talk about this, because historically there is a year like 2023, and these "puppies" are falling like crazy because you're going through a very difficult recessionary cycle. Nobody wants to go through a recession, but these stocks, memory and storage stocks, are doing it.
And that's why they're trying to stop it. How are they trying to stop it? Look at this. We now have new business model agreements with eight different data center and edge computing customers. We expect our new business models to account for over 50% of our beats sold over the next 12 months and around 67% next year as their calendar starts on July 1st.
Again , strange. I'm trying to help you figure this out. So, in other words, we go from 50% on these contracts over the next 12 months to about 67% in the next 12 months, which is 12-24 months, on these hyperscaler data center contracts that are cancelable. If you're excited that hyperscalers continue to spend money, then cool.
Maybe there's less risk here, right? But this is a big focus on these hyperscalers, especially because they say the following. Look at this in terms of their growth. So, Joe, essentially, we continue to look to grow bits in the mid to high decimal range, right? And we are spending at this level. It was near the end.
It's about halfway through their income statement, a little more than halfway through. This is actually a very important phrase from their CFO. Because they're saying that even with these contracts, we're really only committing to increasing volumes in the mid to high range, say 16 to maybe 17%, right?
Otherwise they would just say "tall". The mid or high range is usually around 16-17%. So, let's say 16%. So even with these very large contracts that they're signing, they're actually telling you that they expect bit volume growth of only 16% in the long term. Like, okay. A 16% volume increase means that either pricing power has to continue to grow rapidly to get that, or we won't be growing at 4.7% anymore.
Remember at the beginning , how I said, "Oh yeah, yeah, yeah." Guys, the income is almost 5 times higher. This is incredibly cool. This is fire. "So, well, why are you predicting that you're only going to grow 16% in bits? So, in other words, any growth over 16% has to be driven by pricing power. So 16% bit growth plus pricing power equals revenue growth.
Got it? So 16 % plus PP equals revenue growth. Let's call that the RR. Okay. Okay. That gives you some insight into some of the issues and why high-speed flash is so important to this company. So what we have in terms of risk is off-balance sheet assets . Okay, a lot of companies have that. I mean, Meta has off-balance sheet debt.
That's fine, as long as people know about it, you know? Okay, we have some off-balance sheet assets, it's a little weird with manufacturers. Maybe that's not a problem. Then, we have these upfront payments that support the cash flow of the company, the upfront cash flow. Okay, cool. I mean, they need it for investment, right?
Like, okay, whatever we call it. It's just two little notes here. Then we're talking about 16% growth. Okay. I mean, that's kind of low. Why are we talking about 16% growth? It should be growing much more. It should be growing 4.7 times. That's how people get to, "Oh my God, it's grown 4.7 times, and it's only selling for eight times the earnings." What's going on?
It should be selling for much more." And then we get to something, uh, strange, I would say, in the way they branded high-speed flash memory. I think they did something a little fishy here with the argument about high-speed flash memory. This is going to get a little technical now, but I'll try to simplify it as much as possible and maybe demonstrate it graphically in the simplest way possible, because there's just too much information.
Okay. So, high-speed flash memory. Again, we've already outlined the idea a little bit, but the gist of it is, "Hey, let's do some research and see how we can, you know, actually make more money." In fact, that's exactly what they talk about during the earnings call when asked about Nvidia. On the earnings call, they're asked about the CMX data storage architecture, right?
So Nvidia has this next-generation storage concept with CMX. Basically, Nvidia's idea is that instead of using our server racks, where there are all these Blackwell 200-type chips, and they all have high-speed memory on them, why do n't we just go to another part of the data center that just has a bunch of solid-state hardware, which is basically your NVMe, you know?
Solid-state equipment. So, let's call it storage. Think of it almost like a wardrobe. So you have employees working in this closet. And when high-speed memory needs to address high-speed flash memory, or in this case, just storage, you can do that. You can do this over an Ethernet connection, as Jensen envisions, or through a high-speed flash architecture.
These are completely different things, you know? Jensen's strategy here, what I just drew. This is Jensen's strategy. This is related to CMX. High-speed flash memory is another product that will help us implement a similar strategy. It's Jensen saying, "Oh my god, we're hitting a ' memory wall.'" How are we going to solve this?
Well, we'll solve this with the help of the Nvidia repository. We will use the CMX infrastructure and overcome the " memory wall." You know, Jensen is smart, right? He sees a problem, something that could slow down the business, and he takes on its solution to the fullest. Very smart guy, right? It is worth giving him credit.
SanDisk is trying to do this too. So, SanDisk held this presentation for investors. Let's open its screenshot. In a presentation to investors, they showed a kind of simulation they ran between high-speed flash memory and high-speed memory. And the idea here is , wait. They seem to be comparing themselves to high-speed memory, which has skyrocketed in popularity because it is vertically multi-layered DRAM.
Its prices have risen rapidly . Nvidia can't get enough of it . Nvidia has probably booked 33% of the high-speed memory market supply for the next few years. They need it in their GPUs . SK Hynix, Samsung and Micron produce it. I think Samsung and SK Hynix are probably the biggest manufacturers in terms of percentage of revenue.
So, if you are looking for such a pure investment option in HBM, many are investing in SK Hynix. That's why we saw the Korean market explode because of this, right? But high-speed flash memory, which as we explained also uses those small drilling channels, is being touted by SanDisk as having 8 to 16 times the capacity of high-speed memory.
And they assure that it can be done for a much lower price. So, that's the proposal. The tradeoff is: "Hey, you know that really expensive thing called HBM that Nvidia can't get enough of?" Yes. We'll do something 8-16 times better and cheaper ." It was their investor day. And people are like, "Oh my God, the stock is going to skyrocket!" You need to buy more SanDisk!
"But there were certain details that were, shall we say, skewed in SanDisk's favor. SanDisk got a lot of credit for that. Fair enough , but only from those who had the technical expertise to pull it apart. The first thing they did was run Qwen on a model with 480 billion parameters in BFloat16. That's a lot of numbers.
Sorry. I know, that's a lot. I'll try to explain it as simply as possible. Most LLMs today don't run on BFloat16. Most of them are quantized and divided by four, okay? So most LLMs actually run on a quarter of the size that was used in this experiment. So a lot of people criticized the use of B16. The next thing they criticized was that SanDisk kept the throughput stable at 12.8 terabytes per second.
Well, why would they do that? Why did they make these two settings: keep the throughput stable and use Bfloat16? Well, the reason is, like many people point out is to hide the main drawback of high-speed flash memory: speed. When you run a model with 480 billion parameters on BF16, you need a terabyte of VRAM, which is about 31 5090 cards, 11 Blackwell 6000s, or 4 GB300s.
So you have to connect a bunch of GPUs, creating a bottleneck that doesn't really exist in AI. If you compress a model, as is usually done, it will fit into a regular GB300, actually only one. You only fit that huge model into one GB300. So SanDisk has simply ignored where their technology, high-speed flash memory, fails.
And that's probably why one of their partners, Kioxia, believes that high-speed flash memory has no place in data centers. It belongs in mobile networks and towers. communications, not in data centers. So there's some disagreement here about where this technology starts and ends and where it's applied.
And some people look at this and say, "Listen, SanDisk, why did you have to set everything up in such a weird way?" Because this shouldn't be a confrontation between HBM and HBF . We could use both." It sounds like you're being a little bit sly. So some people say, "Oh, this is just a tool to pre- sell the next S- curve, the next S- curve of growth." Why would they want to do that?
Well, because the predictions for G are not very encouraging. We'll come back to that in a moment. But let's look a little bit more into high-speed flash memory so you can have a better idea of what it is. We'll go into it for maybe two or three minutes. I don't want to go too deep into it, but I want you to understand high-speed flash memory, because it's not going to go away.
I actually think it has a purpose. It has a lot of purposes, and there are certain uses. But bandwidth is very important, as is latency. And there's a great graph that will help you understand that, especially when we think about LLM. So let me quickly open this graph. This graph shows how latency accumulates.
There it is. Okay. So look at this. The high-speed flash architecture has to wait for a full read and write cycle every time it collects data or information. That's very slow. It doesn't do it in parallel, like you would do with prefetching information. When do you do that? You see, in the Sandisk example at Investor Day, they prefetched or essentially loaded all the information in advance.
They knew exactly what weights they were going to use. They knew exactly what they were going to do with the model. And they essentially allowed the high-speed flash settings to run in parallel, indistinguishable from each other. That's how they got to about 98% efficiency of the high-speed memory. That's how they did it.
You already knew what you were going to ask. But the problem is that the high-speed memory in the MOE models slows down, or the high-speed flash memory is very slow in those models. Why? So that in MOE you ask: do we access this part of the data set, or this one, or this one, or the other one? Well, it depends on the situation.
What's our first result? Then what's our next result? Then another result? What's the user input or something else? All of these delays add up and hurt high-bandwidth flash in a way that it doesn't hurt high-bandwidth memory (HBM). Just to give you an example of the speed difference between the two.
What takes 1 second for high-bandwidth memory takes 17 minutes for high-bandwidth flash. This is a hugely exaggerated example, because high-bandwidth memory doesn't actually do anything for a second. All of this happens in the blink of an eye. Imagine: you blink your eyes and 300 milliseconds have passed.
High-bandwidth memory can do 3 million round- trip requests in that time. High-bandwidth flash can do only 3000. That's a 1000x speed difference. So high-speed flash is, frankly, very slow for unpredictable tasks. Expert mixture models break that order. If you open the same contract, the same 800-page legal document, high-speed flash easily wins.
Because high- bandwidth memory is reaching its limit. It's like, we do n't have enough resources . High-speed flash? No problem, buddy. We'll help you with the storage system. We know exactly where this data is. We'll find it and get it to you at flash speed . But high-speed flash loses where HBM is used .
HBM is sold for training, fine-tuning, and rewriting all the models that you have. Or for processing the responses of expert mixture models. So think about it, okay? And that's where people get a little disappointed with Sandisk . So memory with high- throughput is training, it's writing. You're changing the underlying data because you're changing the weights, changing the whole algorithm, or working with an expert mixture model.
High-speed flash is a conclusion, which is not bad, but think of it more as a read-only tool . Then it's the fastest, because the write speed is very slow because of the physics of how flash works. Writing is considered 10 times slower than reading, because you're literally injecting energy into the cell, not just reading its charge.
That's how you get the data. And the more you overwrite the data, the more you wear out the flash memory cells, which is true of any flash memory, and the more you limit the long-term performance. So it breaks down. So high-speed flash memory wears out. It's much slower for training, writing, and working with expert mixture models.
Yes , it's cheaper, but it's only cheaper for certain purposes. Where do you think Where is most of the money being spent on AI today? I'm sorry to say that. We know where AI is going. AI should be going . That's where AI is going. That's where AI is going. That's why companies like Cerebras, I think, are exciting, because they're focused on inference.
High-speed flash can handle some of the inference, but not all of it, because of the mix of experts problem. But what is actually being spent on this cutting-edge field? Who's spending the most money? Well, the biggest spenders on all of this are obviously companies like Anthropic and OpenAI. And what are they buying ?
They're buying tools that help them train their models better. That's why we're seeing such high prices for high-bandwidth memory . And that's why a lot of people are saying, "Hey, SanDisk, you're comparing yourself to high-bandwidth memory ." Well, you have a purpose in AI data centers, but you're not high-bandwidth memory.
. Don't compare yourself to our high- bandwidth memory." While I think it's cool that they're pushing high-speed flash. But there's a reason they're doing it. I think what's happening is that flash prices have been on an S-shaped curve, and the growth rate is starting to level off. We'll see in a moment.
They're trying to move from the SanDisk peak to creating a new vertical of growth. That's your S-curve, made up of parts. That's why there's so much hype around high-speed flash . But the downside is that the money from the advanced LLMs is not being spent there, it's being spent on high-bandwidth memory .
It's not a replacement for high-bandwidth memory for training, which is a huge expense. It's not a replacement for the "mixture of experts" models that are very common now, especially in advanced development. It's a great tool for specific tasks that don't change. Okay. Great. That explains this nice graph in a nutshell.
So where are we at? are we on the growth rate expectations for this company? Okay. This is where the general consensus gets a little bit bleak, but maybe the consensus is wrong. I'll start by opening up the Mizuho report on this. Mizuho actually says that the company could see a decline in average selling prices.
Look at this. We think the consensus estimate for average selling prices for calendar year 2027 , which is 6 months out and the next 12 months, is going to be down 15-20%. So in other words, Mizuho is already predicting that SanDisk's prices will be down. And the average selling prices are actually , well, let me, let me clarify this.
They're indicating that the consensus estimate for average selling prices is going to be down 15-20%. In fact, they think that average selling prices are going to be flat or up a little bit . So let's clarify this. The consensus—it's kind of Wall Street's estimate—of prices falling. Remember how important the prices were for this company?
Mizuho actually thinks we'll be at or slightly above that. The current actual growth projections for this company are not that impressive. The current projections look like they're written down. We're taking EPS of 211.34 and expecting it to grow 22.1% next year. Oh, that sounds great, Kevin. That's half the PEG ratio.
You 're right. That's less than half the PEG. That sounds really cheap , doesn't it? But wait, the forecast for next year is minus 30.9. The forecast for next year was actually 20. Wait a second. Let me, uh, that's next. Uh, 20.9, minus 20.9. The forecast for next year is 30.9. And the forecast for the year after that is minus 49.
Whoa. So now, all of a sudden, when you add up the growth rates for this company, you can look at that and say, "Hey, it's got a PEG ratio of PE nine or eight." Whatever you want, it doesn't matter. We're going to divide that PE by the growth rate, which in this case, if I average these numbers, is about -78.7 %.
Divided by, uh, that's divided by four. Divided by four. That's if I add all of that up and divide by four. That gives me a growth rate of -19.68% per year. So in other words, a PE of 9 looks good, but if your forecast is shrinking by almost 20% per year, then your PEG ratio goes to infinity. I mean, any negative growth rate means an infinite PEG ratio.
So that's where the problem lies. The consensus on Wall Street is that this company is going to start losing its pricing power. But that's the problem, because if all of a sudden their pricing power, their ability to raise prices, if ASPs, average selling prices, go down, the only thing that's helping us grow is the bit volume.
But the bit volume growth is expected to be only mid- to-high- teens . In this case, 16% is what they're forecasting. That's what they wrote in their earnings report. So now you actually have a company that looks really cheap. They seem to be spending a lot of money on share buybacks, and they have a solution for everything.
They have a solution to the memory wall problem, and to be fair, they say there are a lot of ideas for that, right? They're being upfront about it. They're like , "Look." They wrote, "There are a lot of different ideas." This is about solving the memory wall problem. See? They're literally talking about it .
We think that the output is limited by memory. Everyone has different ideas. It's similar to what I was talking about with Jensen. Jensen has this idea: "Hey, let's just use the CMX architecture to sample data from memory ." SanDisk says, " Hey, let's use high-speed flash ." Others They say, "Hey, why do n't we just use smaller models?" Right ?
There are a lot of ideas for solving the storage problem . But here's the main problem. It's the growth of bit volumes in the mid to high tenths of a percent. Because if prices start to come down, we're actually going to start accumulating problems. And I'm not trying to say that this is a bad investment.
But when I look at it, I think, "Wow, I learned a lot about SanDisk that I didn't know before, right?" I have off-balance sheet risk . It's the factories. It's the capital expenditure that's actually hidden right now. It's hidden capital expenditure. I have upfront payments from the hyperscalers that are partly being used to buy back shares, but I don't know if that's going to be sustainable.
Also, the presentation of high-speed flash is a bit misleading. And also, the consensus is that the average selling price is going to be down, and the bit growth is going to be 16%. So to be able to say, "Hey, let's invest in SanDisk," you have to say, "I think that's a mistake." You have to say, " I think the average selling price is going to be up, as Mizuho says." I think the bit growth is going to be higher.
And I believe that high-speed flash, which I intentionally tried to explain as simply as possible. Maybe I did n't do the best job, but I think I explained it simply. I did it because it's important to the story. So if If you believe in this story and understand everything else, SanDisk could be a great opportunity.
I think there's too much information here to digest, and I'd probably advise waiting and seeing what happens with high-speed flash, especially since it's still being co-produced with SK Hynix. So all the benefits if the technology is successful won't go to SanDisk. That's it . If you enjoyed this video and found it useful, please subscribe to the channel.
I do my best to continue to do my research and give you as much context as possible, because I consider it my job and I enjoy it. I hope you enjoyed it. .
Comments 0