Latest / In the Loop AI / Making Sense of the AI Cycle: Where are We Now?

Transcript

Making Sense of the AI Cycle: Where are We Now? · In the Loop AI

0:00
Josh: If you bought SanDisk stock just a year ago, you're up 40 times your money.
0:05
Josh: Micron, up eight times. Intel and AMD are both up almost four times what you
0:09
Josh: would have put in just a year ago. And we thought this was an AI bubble.
0:12
Josh: We were actually wrong about how it's going to play out. And we have the data
0:15
Josh: to back this up. Now this year, five of the biggest companies in the world are
0:18
Josh: going to spend close to a trillion dollars.
0:20
Josh: As that trillion dollars is spent, it's going to trickle down the stack to a
0:24
Josh: series of layers that are embedded within this AI ecosystem.
0:27
Josh: So we're going to be detectives this episode and track down where all of that
0:31
Josh: trillion dollars in CapEx is flowing.
0:33
Josh: It's clearly leaving the large
0:34
Josh: cap companies. Google is spending 90% of all the money that comes in.
0:38
Josh: Where is it going? Well, we are going to walk you through everything.
0:40
Josh: So hopefully by the end of this, you'll have a pretty good idea of the structure
0:43
Josh: of what the AI investment universe looks like.
0:46
Josh: And you can make up your own mind on where you think the best place is to allocate
0:50
Josh: your dollars to collect the money that's flowing from these large cap companies.
0:54
Ejaaz: Now, I've really had to ground myself over the last couple of weeks because
0:58
Ejaaz: it felt like I'm in the land of make-believe.
1:00
Ejaaz: Look at some of these stocks and how much they've increased over the past year.
1:05
Josh: This is crazy.
1:06
Ejaaz: This is absolutely insane. AMD is up three and a half X and 25% of that move
1:12
Ejaaz: was literally over the last 12 hours, right?
1:15
Ejaaz: Because they reported earnings, they crushed it. SanDisk is up 40%, as you mentioned.
1:19
Ejaaz: ARM, Intel, all of these are in the infrastructure layer of AI,
1:23
Ejaaz: which has become a very consensus trade.
1:25
Ejaaz: And that's for many different reasons. And we're going to address that later on in the stack.
1:28
Ejaaz: But there's one clear thing that changed over the last week.
1:32
Ejaaz: And that came in the form of the major earnings reports for Q1 of 2026 from
1:37
Ejaaz: four of the largest hyperscalers.
1:39
Ejaaz: You've got Amazon, Meta, Microsoft, and Google. Now, there was a problem. There was an issue.
1:44
Ejaaz: People thought that they were going to spend hundreds of billions of dollars.
1:47
Ejaaz: And they originally committed to that being around $630 billion in 2026,
1:51
Ejaaz: but no one knew how they were going to make money.
1:53
Ejaaz: Q1 proved that them spending all that money not only resulted in more profit
1:58
Ejaaz: and more revenue that they had generated, but also that they were going to revise this.
2:02
Ejaaz: They were going to increase their spend in 2026 to a total of $800 billion,
2:07
Ejaaz: with next year projecting $1.1 trillion.
2:10
Ejaaz: This is just from four to five companies. I have to emphasize how nutty and crazy that is.
2:16
Ejaaz: But again, this translates into actual revenue or money being earned.
2:21
Ejaaz: I'm going to show you a few different stocks that kind of demonstrate or portray
2:24
Ejaaz: that. So firstly, Sandisk.
2:27
Ejaaz: 40x over the last year. This is just like an astounding thing where like,
2:32
Ejaaz: you know, a 40x return on any kind of major cap stock is just completely unheard of.
2:36
Ejaaz: Then you have Intel that operates in the CPU kind of infrastructure packaging
2:40
Ejaaz: layer of AI GPUs is up four and a half X. What else have we got? We've got AMD.
2:45
Ejaaz: I mentioned that it's up 320% over the last year, but literally today it is
2:50
Ejaaz: up 17%, absolutely crushing it.
2:52
Ejaaz: And then we have Micron technology, which plays in the AI memory trade is up
2:56
Ejaaz: 7.25x over the last year.
3:00
Ejaaz: And then there's this really cool investment vehicle where there's a bunch of
3:04
Ejaaz: different AI memory trades.
3:05
Ejaaz: You can get access to just a basket via this ETF, Roundhill Memory ETF,
3:09
Ejaaz: something I participated in.
3:11
Ejaaz: That only launched just over a month ago, Josh, and it is up 72%.
3:15
Ejaaz: Just insane gains on these stocks.
3:17
Josh: Do you have FOMO? Because I do. I'm feeling the FOMO. And I think the question
3:22
Josh: and part of the concern that a lot of people are going to be feeling as they're
3:25
Josh: seeing these numbers is like, oh, my God, did I miss this?
3:28
Josh: It seems like very obviously we're in some sort of a bubble.
3:30
Josh: But what stage of that bubble are we in?
3:32
Josh: How should we deal with this? How should we navigate this? Well,
3:35
Josh: Brad Gerstner, actually, Altimeter Capital investor, he's like a very prominent
3:39
Josh: thought leader in the space.
3:40
Josh: He went on CNBC yesterday to talk just about this, about where we are in the cycle.
3:45
Josh: And his idea is that this time is different.
3:47
Josh: And these are very famous words. We've heard this time is different a lot,
3:50
Josh: but he explained why. And a lot of it comes from this capital expenditure.
3:54
Josh: I mean, he just mentioned $800 billion is being spent this year.
3:57
Josh: Over a trillion dollars is being spent next year. And a lot of that money is
4:00
Josh: flowing from these large mega cap companies into infrastructure required to
4:05
Josh: build out more AI use cases.
4:07
Josh: As of now, these AI use cases are actually printing a good bit of money.
4:11
Josh: I mean, it comes from that top where we see Anthropic and OpenAI printing billions
4:15
Josh: of dollars of ARR per week. Google just became the most valuable company in the world.
4:20
Josh: And it seems to be like we have this flywheel that is sustainable so long as AI is useful.
4:25
Josh: And in the case that AI continues to be useful and we unlock new use cases that
4:29
Josh: are more valuable than what they are today, then the amount of money people
4:32
Josh: will pay for it will continue to go up.
4:33
Josh: And those capital expenditures that are being priced into these earnings reports
4:36
Josh: can then be priced into these further concentric circles out from the core.
4:40
Josh: So if you'll notice, we didn't mention NVIDIA on one of these most valuable,
4:44
Josh: like the largest gainer stocks, because they're kind of sitting at the top in
4:48
Josh: which a lot of the capex is coming from.
4:51
Josh: So what we're going to do now is kind of walk through the layers and the stack
4:55
Josh: of where this money flows.
4:57
Josh: So if you think about it, all of the money comes in through companies like Google,
5:00
Josh: through companies like Amazon, who are collecting income from retail.
5:03
Josh: Retail is spending on their goods and services.
5:06
Josh: They're delivering a lot of value to them they take that money they redistribute
5:09
Josh: it where is that going is the question we're going to answer so over the last
5:12
Josh: few months we've seen this happen in semis a lot of those um
5:16
Josh: A lot of those companies that you showed us, EJES, are semiconductor stocks. It's moving into CPUs.
5:20
Josh: It's moving into GPUs. But let's start with the layer zero that we have here
5:23
Josh: on the artifact, EJES. What's going on with the layer zero?
5:26
Josh: What is at the base layer of this stack?
5:29
Ejaaz: Okay, so you mentioned that four of the hyperscalers, Google,
5:32
Ejaaz: Microsoft, Meta, and Amazon, are kind of taking inflows from retail.
5:37
Ejaaz: There are currently two ways, or rather two startups and companies,
5:41
Ejaaz: that funnel 99% of that retail. And they're called OpenAI and Anthropic.
5:46
Ejaaz: They're both private companies. So you can't even access them publicly via equities
5:50
Ejaaz: right now, but they contain the bulk of ChatGPT and Claude users, right?
5:55
Ejaaz: So that in itself is kind of scary to kind of read, but the fact is they have
5:59
Ejaaz: paying customers, they have paying enterprises, and that money is where the
6:03
Ejaaz: fountain starts, the waterfall starts.
6:05
Ejaaz: And from OpenAI and Anthropic, it flows down into what we're calling the layer
6:10
Ejaaz: one, which is kind of like the platforms, hyperscalers, and also Model Labs,
6:14
Ejaaz: Google counts itself as one as well.
6:16
Ejaaz: Okay, so you've got Google, which doesn't actually just act as a Model Lab.
6:20
Ejaaz: They have TPUs, they have the infrastructure, they have the distribution layer,
6:23
Ejaaz: they have the cloud infrastructure, which is the main funneler of all this revenue
6:28
Ejaaz: that they're making from their recent earnings.
6:30
Ejaaz: You've got Amazon that's mentioned here as well, which is doing the same with AWS.
6:34
Ejaaz: And you have Microsoft, which is doing the same through Azure.
6:37
Ejaaz: Meta, who's doing the same from their social media platform.
6:39
Ejaaz: And we've included opening an Anthropic here because some of them sometimes
6:42
Ejaaz: play in the infrastructure layer.
6:43
Ejaaz: So the first stack is, okay, we have OpenAI and Anthropic dealing with the retail.
6:48
Ejaaz: That money and revenue flows into the hyperscalers that can provide compute
6:53
Ejaaz: and distribute their models.
6:55
Ejaaz: So very important. Google Cloud, AWS, and Microsoft Azure, all their cloud computing
6:59
Ejaaz: services, distribute the model to all their enterprise customers and governments
7:04
Ejaaz: and customers that actually want to, retail customers that actually want to
7:08
Ejaaz: end up using these things.
7:09
Ejaaz: And then, as you mentioned, Josh, we move into layer two, which is what we're
7:13
Ejaaz: calling kind of the GPUs and semiconductor area, right? Now,
7:17
Ejaaz: this is where the narrative breaks.
7:20
Ejaaz: And classically, we have been told AI is going to do so well.
7:23
Ejaaz: Demand is going to increase. Where's the best way to buy the picks and shovels?
7:26
Ejaaz: Well, it's NVIDIA, of course, right? They're the most valuable company in the world, right, Josh?
7:30
Josh: And that was correct. Yes, they were a $4 trillion company.
7:34
Josh: As of today, that has changed. Google has taken the crown again.
7:37
Josh: And I think it's because we're kind of reaching further out of this risk curve.
7:41
Josh: The money is starting to propagate further out. So as it started,
7:43
Josh: we needed to go from zero to one.
7:45
Josh: We needed to go from no AI to AI. And what was required of that?
7:49
Josh: Lots and lots of GPUs, lots and lots of intelligence. And that's where NVIDIA
7:53
Josh: was most valuable because they were able to provide the GPUs to spin up the
7:56
Josh: AI, to take the industry from zero to one.
7:58
Josh: Now that we have AI, now that we have established business and revenue streams,
8:02
Josh: that money is flowing to optimize the stack.
8:04
Josh: So Jensen and NVIDIA are going to continue to make a tremendous amount of money,
8:08
Josh: but that's kind of baked into the valuation and we've seen that run up.
8:11
Josh: I mean, they're the second most valuable company in the world right now.
8:13
Josh: It's a tremendous amount of market cap that they've absorbed over the last, call it 36 months.
8:17
Josh: Now the time is come to cycle further
8:21
Josh: out this risk curve to move beyond the picks and shovels into this
8:24
Josh: next layer of the stack that even sits below gpu consumption and my understanding
8:28
Josh: he does is it's not just gpus now there's also cpus and that has to do a lot
8:33
Josh: with the the new agentic trend that we're seeing where agents are kind of being
8:37
Josh: proactive and they need orchestrators and while gpus are really good at solving
8:40
Josh: complex uh mathematical problems to produce inference
8:44
Josh: You need smarts and you need brains in order to kind of orchestrate these models.
8:47
Josh: And that's what's key about this next infrastructure layer that's going to really
8:51
Josh: be a pretty big deal in the space.
8:53
Ejaaz: That's exactly it. So I've got this article pulled up where Jensen went on a
8:58
Ejaaz: stage right now and he talked about GPUs and the demand increasing ever so highly.
9:04
Ejaaz: And he goes, consumption is going through the roof for GPUs with the rise of
9:08
Ejaaz: agentic AI in the last several months. Now, I want to unpack this phrase very
9:13
Ejaaz: specifically, because it explains why Intel and AMD are absolutely skyrocketing right now.
9:18
Ejaaz: Agentic AI refers to AI agents. AI agents is kind of like, think of like an
9:23
Ejaaz: instance of a ChatGPT or a Claude code that can just go off and do a bunch of
9:27
Ejaaz: things autonomously for you.
9:29
Ejaaz: So you don't have to type to it, you don't have to speak to it,
9:31
Ejaaz: you don't have to prompt it.
9:31
Ejaaz: It just goes off and works autonomously. Now, you can imagine the market for an AI autonomous worker
9:37
Ejaaz: It's pretty large. The TAM is pretty huge. It's pretty much any sector that
9:41
Ejaaz: involves a computer for now until robots actually become a thing.
9:44
Ejaaz: But there's one kicker. This is the narrative violation, Josh.
9:48
Ejaaz: Guess what you need to allow the AI agents to use the tools to go off and do that work?
9:53
Josh: They need the brain. They need the CPU.
9:56
Ejaaz: The CPU.
9:58
Josh: Computer processing unit versus the graphical processing unit.
10:01
Ejaaz: Right. So let me take you through a little historical context here.
10:04
Ejaaz: Now, if we rewind to, let's take GPT 4.0, right? It was a breakthrough model. Everyone loved it.
10:11
Ejaaz: Can you guess how many CPUs were used to train or inference that entire model?
10:16
Ejaaz: Like, what's the ratio roughly? Do you think how many- CPUs or GPUs?
10:19
Ejaaz: CPUs. So how many CPUs did you need for like the average GPU?
10:23
Ejaaz: Like, what do you think that ratio was?
10:24
Josh: I would guess close to zero.
10:25
Ejaaz: Yeah, you would be correct. We barely even used it.
10:29
Ejaaz: Fast forward to today, that ratio is almost one-to-one. Let me rephrase that. Let me reframe that.
10:35
Ejaaz: You need one CPU core per GPU, and that trend is going to flip over the next
10:42
Ejaaz: six months where the number of CPUs will outweigh GPUs.
10:46
Ejaaz: So basically, Intel and AMD, who kind of built their bread and butter of profit
10:51
Ejaaz: margins off of CPUs that were used for gaming and a bunch of other stuff,
10:55
Ejaaz: now has found themselves in an absolute goldmine of an industry which requires
11:00
Ejaaz: the things that they've been building for over decades.
11:03
Ejaaz: So Intel and AMD are like, well, okay, I'll spin up as many CPUs as you want.
11:07
Ejaaz: And Jensen's like, I need all these CPUs so I can spin up all these data racks.
11:10
Ejaaz: Now, to help you understand why it is needed specifically, Josh,
11:14
Ejaaz: you mentioned AI agent orchestration, right?
11:16
Ejaaz: So let's say you spin up a bunch of AI agents. They need to,
11:19
Ejaaz: guess what, interact with other AI agents. They also need to use tools.
11:24
Ejaaz: And most importantly, they need to use these tools faster than humans can themselves.
11:29
Ejaaz: Josh, on yesterday's episode, you mentioned what you loved about Codex specifically
11:34
Ejaaz: was the browser use and the fact that it's so quick.
11:38
Ejaaz: You love that. The reason why it's able to do that is because it has access to more CPUs.
11:44
Ejaaz: It allows you to kind of run C compiler and a bunch of other things.
11:47
Ejaaz: So the long story short is CPUs are in huge demand. And that's why Hymdi stock is up 15% today.
11:53
Ejaaz: And it's probably going to be up what's up three and a half X over the last couple of years.
11:57
Ejaaz: So they just reported their earnings. Intel did the same about two weeks ago.
12:00
Ejaaz: And there's a huge demand for CPUs in general.
12:03
Josh: And this is kind of the natural extension of the way that AI is moving.
12:07
Josh: This hasn't always been the case, but we went from basically LLMs,
12:10
Josh: right? Where we just used GPUs. It was a chatbot. You type into an interface.
12:14
Josh: Then we went into reasoning and chain of thought where there was a lot more
12:17
Josh: tokens needed in order to answer the same question.
12:20
Josh: Then we moved into the agentic era where things like OpenClaw and the claw agentic
12:25
Josh: swarm operating system came into play. And that's when this trend changed again.
12:30
Josh: So assuming these trends are going to continue changing, the paradigms are going
12:33
Josh: to continue to shift. But the one singular core truth throughout this is that
12:37
Josh: we need more tokens. And no matter how we get them is up for debate,
12:42
Josh: like how those are going to be generated.
12:43
Josh: But the fact is that no matter what we do, every paradigm shift has resulted
12:47
Josh: in a gigantic increase in more tokens, which seems like is a trend that you continue to bet on.
12:53
Josh: And one person who bet on this pretty bigly was Donald Trump and the US government,
12:59
Josh: like us, for the first time, like we're making money.
13:02
Josh: Trump bought Intel and he's up what, 500% so far?
13:05
Ejaaz: 500 percent um so uh the government
13:08
Ejaaz: uh about a year and a half ago
13:12
Ejaaz: took a 10 stake in intel and
13:15
Ejaaz: their primary reasoning for that was we have
13:18
Ejaaz: too much reliance currently in the u.s on this one company called tsmc which
13:22
Ejaaz: is based in taiwan which china presumably wants to take over at some point it's
13:25
Ejaaz: too much of a geographical and national risk and so we wanted to kind of bring
13:30
Ejaaz: a bunch of tsmc's capabilities on shore and that was expressed in the form of
13:34
Ejaaz: intel who doesn't just build CPUs, by the way,
13:37
Ejaaz: they're working on building a bunch of frontier GPUs. And that's a story for another day.
13:41
Ejaaz: But the point is, Trump bought 10% via the government, and they are up 5x.
13:45
Josh: You know what price you bought at $20.47? It's trading at $111 today is what
13:53
Josh: the ticker is trading at.
13:54
Josh: So shout out to the US government pumping our bags and winning on behalf of
13:58
Josh: the people. That's very exciting.
14:00
Ejaaz: Yeah. Okay, so let's continue down this waterfall. So what we've started off
14:03
Ejaaz: with is Anthropic and OpenAI as the retail that flows down into the hyperscalers,
14:08
Ejaaz: the Google, the Amazons, the cloud distributors, their revenue margins are expanding.
14:13
Ejaaz: Okay, this is amazing, but they rely on semiconductors.
14:16
Ejaaz: They need Jensen Huang's GPUs, but Jensen Huang needs all these CPUs because
14:20
Ejaaz: all these AI agents are using all these different tools.
14:22
Ejaaz: Okay, so we need CPUs. Now, can you guess? There's another component,
14:26
Ejaaz: Josh, that makes up 50% of the bill of materials cost for a GPU. 50%.
14:33
Ejaaz: Guess what it is?
14:34
Josh: What does every single GPU on the planet need? Jensen, what does he need?
14:38
Josh: What is his bottleneck choke point that doesn't work?
14:41
Josh: Like the company doesn't work in the absence of it's memory.
14:44
Josh: Memory is the biggest thing in the world. It's impossible to make enough memory. No one has it.
14:49
Josh: And therefore any of the memory stocks that you have invested in,
14:52
Josh: in the last year have gone absolutely nuclear.
14:54
Josh: And these are probably among the biggest winners. When you think about SanDisk
14:57
Josh: delivering a 40 times return, Micron is looking like a seven times return.
15:02
Josh: Memory has been the choke point of all of these because memory is the most important thing.
15:07
Josh: It's where we get our context windows from. It's where a lot of the training data is stored.
15:11
Josh: Memory is the next layer of the stack.
15:13
Ejaaz: Yes, exactly. And if you thought NVIDIA had a monopoly on GPUs,
15:18
Ejaaz: let me introduce you to the secret monopoly of memory.
15:21
Ejaaz: So there are basically four companies which dominate in the AI memory landscape.
15:27
Ejaaz: Their names are Micron, which is a U.S.
15:29
Ejaaz: Company, SK Hynix, which is, by the way, the biggest memory provider,
15:33
Ejaaz: Korean company, Samsung, second biggest, and then we have SanDisk.
15:37
Ejaaz: Now, you might assume that they all make the same types of memory. Three of them do.
15:42
Ejaaz: Micron, SK Hynix, and Samsung make something called high-bandwidth memory.
15:47
Ejaaz: This is the premium memory that goes into making the GPUs, the Rubin Ultras,
15:53
Ejaaz: the fashionably new GPUs that NVIDIA releases every single year.
15:58
Ejaaz: That incorporates 50% of it.
16:00
Ejaaz: Our bill of materials is this fancy high bandwidth memory.
16:03
Ejaaz: It is an extremely complex thing to make. The supply is super constrained and
16:08
Ejaaz: only these three companies dominate in making it.
16:11
Ejaaz: And that is for a different reason. I'll explain the memory cycle later on.
16:15
Ejaaz: But then you have SanDisk, which is up 40%. And the reason why they're up 40%
16:19
Ejaaz: is because there's a second type of memory which is required by these AI models.
16:23
Ejaaz: And it's called NAND or N-A-N-D, NAND flash. Now, here's the difference between the two.
16:30
Ejaaz: High bandwidth memory, which is the original GPU premium, basically allows you
16:35
Ejaaz: to move data really quickly in the AI models.
16:37
Ejaaz: Think about it, right? These AI models are like 10 trillion parameter large,
16:41
Ejaaz: at least Mythos and GPT 5.5.
16:43
Ejaaz: They need access to data very quickly, and it's clunky. It's very big.
16:47
Ejaaz: High bandwidth memory basically solves that. It allows you to kind of store
16:50
Ejaaz: the memory and like move that memory really quickly.
16:52
Ejaaz: But there's the second type of memory, which SanDisk specializes in, called NAND.
16:56
Ejaaz: And what that does is when you're having a conversation, Josh,
16:59
Ejaaz: have you ever noticed that the models now are really good at maintaining the
17:03
Ejaaz: context and remembering things
17:05
Ejaaz: that you mentioned like a few sentences before? Have you noticed that?
17:08
Josh: For a longer period of time, yes. The context windows have expanded quite a bit now.
17:11
Ejaaz: Yes, exactly. Now, the main enabler for that is this NAND storage,
17:16
Ejaaz: which is kind of like a temporary memory storage.
17:18
Ejaaz: It's more of a commodity. It's not as sexy as high bandwidth memory,
17:21
Ejaaz: but it's super important.
17:23
Ejaaz: SanDisk dominates that entire sector, which is why it's up 40X.
17:27
Ejaaz: So memory has famously gone under a huge supply constraint.
17:32
Ejaaz: These providers are sold out, and I'm not exaggerating this, until the end of 2028.
17:37
Ejaaz: They have signed customer contracts until the end of 2028.
17:41
Ejaaz: They don't have enough memory, they don't have enough supply,
17:43
Ejaaz: and so ramping that up is going to be the key focus over the next couple of years.
17:46
Josh: And this has been really felt in the retail market as well. If you use memory
17:49
Josh: for traditional use cases, like building PCs or just general consumer products,
17:53
Josh: all of those have either been delayed or the prices increased because it is
17:57
Josh: so difficult to get your hands on this memory.
17:59
Josh: And when you think about this memory, there's a pretty simple analogy that I
18:02
Josh: was using earlier today when it came to describing and understanding how this works.
18:05
Josh: So high bandwidth memory, like you mentioned, it's known as HBM for short.
18:08
Josh: It's basically a series of DRAM stacks stacked on top of each other.
18:11
Josh: And you can think of those as this like this huge desk right next to the AI
18:15
Josh: model that they can very quickly reference.
18:16
Josh: And then the NAND is the file cabinet that's kind of right next to it that is
18:20
Josh: more persistent that holds these ideas for a longer period of time it's larger
18:23
Josh: but it is a bit slower and the dynamic between the two is really interesting
18:27
Josh: now one of the things that was
18:29
Josh: kind of a narrative violation as i was reading about it is the idea that
18:33
Josh: Deep Seek can actually create more efficient models in terms of how much they
18:37
Josh: use memory, and yet the demand for memory goes up.
18:40
Josh: So Deep Seek, they published this famous paper that allowed them to get basically
18:43
Josh: frontier level intelligence using a small fraction of the amount of memory that
18:47
Josh: traditional AI labs have used. This seems like a bad thing.
18:51
Josh: We need less memory because we've become more efficient, but the reality is,
18:54
Josh: is that the inverse actually happened, where now we have greater memory efficiency,
18:59
Josh: but far greater memory demand.
19:01
Josh: So maybe you could explain this dynamic that's happened with this,
19:04
Josh: because there's a strange deep seek paradox going on that I think is a narrative
19:07
Josh: violation in which people kind of look for when seeing things that can pop this bubble.
19:12
Josh: And the reality is, is that this is this is not even a little small tear in the bubble.
19:16
Josh: It's actually improving the quality of the money spent. How is this working
19:19
Josh: with this memory dynamic here?
19:21
Ejaaz: Yeah, it's basically Jevin's paradox. So the problem that you're explaining
19:25
Ejaaz: is if 50 percent of this GPU thing is reliant on this one component supplied
19:31
Ejaaz: by four different companies, they have to work on a workaround, right?
19:35
Ejaaz: They're probably going to create GPUs or models that don't rely high on memory.
19:39
Ejaaz: And Deep Seek version 4, which was released a few weeks ago,
19:41
Ejaaz: was the instantiation of that.
19:43
Ejaaz: It uses, I think, like 5% to 15% of a Claude Opus 4.7 model,
19:48
Ejaaz: which is a drastic reduction, which may lead you to think that memory stocks
19:51
Ejaaz: are going to crumble, except the actual opposite happened.
19:55
Ejaaz: So I've got this block of text here, but I'm going to explain it in very simple terms.
19:58
Ejaaz: What they found was the architecture unlock that DeepSeq v4 created actually
20:05
Ejaaz: ended up using more of that NAND flash memory because their architecture change
20:11
Ejaaz: was, we'll just use more agents.
20:13
Ejaaz: We'll let more agents do the thinking before we give someone an answer to their prompt.
20:17
Ejaaz: And it resulted in a smarter effort. But that didn't decrease the reliance on
20:22
Ejaaz: HBM or memory in general.
20:24
Ejaaz: Increased it overall. You'll see here that each prompt required 157 rounds and
20:31
Ejaaz: a bunch of token contacts, all of that interfacing with these memory components.
20:34
Ejaaz: So the actual opposite happened. And this is Jeven's paradox playing out in
20:38
Ejaaz: reality, where if you assume that the demand for goods, or rather the cost of
20:42
Ejaaz: goods goes down, you'll assume demand goes down because it's like cheaper,
20:45
Ejaaz: you don't make as much money.
20:46
Ejaaz: But the in fact, opposite happens where demand goes way higher,
20:50
Ejaaz: because now you can do more things for cheaper. So Jevon's paradox playing out here.
20:54
Josh: Okay, next layer of the stack, power generation and infrastructure.
20:58
Josh: How are we going to plug in these GPUs? Well, we need infra for it.
21:01
Josh: Otherwise, they're just going to be sitting there dark with no power and no
21:04
Josh: ability to turn them on. This is layer five of the stack.
21:07
Josh: This is where you will recognize names, perhaps like Bloom Energy,
21:11
Josh: which we have famously mentioned.
21:12
Josh: Our boy Leopold is up like what, four or five times return on Bloom Energy.
21:16
Ejaaz: Yeah, like two bills.
21:18
Josh: Yeah, like doing pretty well. And this is the part of the stack that is,
21:21
Josh: I guess, core and sits a little bit downstream of the GPUs. But again,
21:24
Josh: contingent on it working.
21:26
Josh: Tell me a little bit about this company, particularly, I mean,
21:29
Josh: Corning. We had a deal with Corning Opio today and NVIDIA, which is pretty big.
21:32
Josh: I would imagine most people listening to this have never even heard of Corning before.
21:36
Ejaaz: Yes, exactly. So I have to be upfront. This is where I'm weakest in this stack or this.
21:42
Ejaaz: So I have to be upfront and maybe we'll do an episode later down when we've
21:46
Ejaaz: done a bit of research. But basically, if you look at the stack so far,
21:50
Ejaaz: it flows down from retail, it goes through semiconductors, it goes through CPUs,
21:53
Ejaaz: and it goes through memory.
21:54
Ejaaz: Where does the puck flow next? Well, there's an issue.
21:57
Ejaaz: I buy all these GPUs, I buy all these CPUs, I set them all up neatly in a warehouse,
22:03
Ejaaz: in a stack, in a server rack.
22:05
Ejaaz: But I don't have enough power to power the thing. Or if I do have the power,
22:09
Ejaaz: I don't know how to regulate the power to the right GPUs and the right CPUs
22:13
Ejaaz: at the right time to prevent a blackout, but not overheat it.
22:17
Ejaaz: And then I have a cooling system.
22:18
Ejaaz: There's a whole architecture around the GPUs and the CPUs operating at optimal form.
22:23
Ejaaz: There was an article that was released by The Information this week,
22:27
Ejaaz: which showed that XAI, which has the largest cluster of GPUs,
22:31
Ejaaz: over a million high-grade NVIDIA GPUs, only utilize 11% of its power.
22:37
Ejaaz: That's because they don't have enough power generation or the chips or the architecture
22:41
Ejaaz: to allow this power to flow directly through GPUs.
22:45
Ejaaz: I think this is where the puck is flowing next.
22:48
Ejaaz: And today's announcement, where NVIDIA is partnering with Glassmaker Corning,
22:52
Ejaaz: which is focused on optics specifically, is part of that puzzle.
22:55
Ejaaz: But there's also another piece, which
22:57
Ejaaz: is basically the power suppliers or infrastructure providers on its own,
23:01
Ejaaz: mainly referenced by companies like GE, Vinova, CEG, Constellation Energy,
23:06
Ejaaz: which we had on our stack just now,
23:07
Ejaaz: which not only supply the power and make sure the power ends up at the data
23:12
Ejaaz: center, but also regulates and makes sure that the power enters at the right
23:16
Ejaaz: time to the CPUs, to the GPUs, to make sure they're optimally performing.
23:19
Josh: And this is layer five of the six-layer stack, the sixth and final layer being the raw materials.
23:24
Josh: Now, this is kind of the bare-bones foundational layer. When you think about
23:28
Josh: The AI stack from first principles, what is required at that foundation,
23:31
Josh: it is the raw materials in to get the intelligence out.
23:34
Josh: We've turned sand and silicon into thought.
23:37
Josh: And throughout that entire process, there is a tremendous amount of technology empowering it.
23:41
Josh: Now, we are not proficient currently in materials, but if you would like us
23:45
Josh: to be, and if this is an episode of interest, we can go deeper on layer five
23:49
Josh: and layer six of the stack.
23:50
Josh: Because I was looking at a chart this morning from of lithium carbonate,
23:53
Josh: which is just a very critical component to a lot of the AI stack.
23:56
Josh: I'm looking at a chart from November where it was priced at $75,000,
23:59
Josh: and now it is at $187,000.
24:02
Josh: So the gain and return on some of these materials has been just unbelievable.
24:06
Josh: And this is the sixth and final layer of the stack. So perhaps in a future episode,
24:09
Josh: but hopefully this has given you kind of a loose orientation of how you could
24:13
Josh: think about the trickle down economics of AI, how it kind of starts from this
24:17
Josh: layer zero and works its way through the infrastructure.
24:20
Josh: Now, EJs, we have to kind of orient ourselves in reality.
24:24
Josh: Good things don't last forever and there's
24:27
Josh: a high probability that this one doesn't perhaps this time
24:30
Josh: is different but there is a case to be made that it is not because
24:32
Josh: traditionally memory has done well in the past there
24:35
Josh: have been a series of memory booms that have done very well every single one
24:39
Josh: of which has followed up with a bust so where would you say we are currently
24:44
Josh: um in the cycle or at least what are the downsides that people should be looking
24:48
Josh: out for and aware of that would let them know that hey maybe things are looking a little frothy here
24:54
Josh: there's some cause for concern.
24:55
Ejaaz: Yeah, so the first factor is the boom and bust cycles, particularly with memory,
25:01
Ejaaz: which I think is, what was it, layer three, layer four of the sec that we just described.
25:05
Ejaaz: It has traditionally gone through many boom and bust cycles,
25:08
Ejaaz: which I'm showing on our screen here through this chart.
25:12
Ejaaz: Now, what isn't represented here is.
25:15
Ejaaz: At the start, on the left-hand side of this chart, we had 14 key memory providers.
25:22
Ejaaz: Fast forward to today, and we have three for high bandwidth memory and one for NAD.
25:26
Ejaaz: So four. And the reason why is every single boom and mainly the bust has crushed
25:32
Ejaaz: out specific companies to leave like the three that we have right now.
25:36
Ejaaz: And so history would tell us that the same thing is going to happen this time.
25:39
Ejaaz: Now, I'm conflicted here, Josh, because I'm sitting here as a podcaster and
25:43
Ejaaz: saying this, but I listened to three podcast episodes from the memory execs
25:48
Ejaaz: from all of these different companies yesterday.
25:50
Ejaaz: Guess what they said? They said, this time is different. They said,
25:53
Ejaaz: this time is a very unique opportunity because not only is AI surging demand
25:58
Ejaaz: for all of these memory chips and they've got payments for all of these things,
26:03
Ejaaz: on the front end, the AI products themselves are being paid for.
26:07
Ejaaz: They're actually making money. Now, the definition of a bubble is it's levered.
26:11
Ejaaz: It's a bunch of hoopla, right? There isn't actually any money coming in.
26:15
Ejaaz: This disproves that narrative, the bubble narrative, because there's money being
26:18
Ejaaz: paid off. None of these companies are levered up.
26:20
Ejaaz: That 800 billion CapEx number that we started this episode with is all coming
26:23
Ejaaz: from cash flows that they have right now or cash reserves that they have right
26:27
Ejaaz: now. No one is levered up. No one's borrowing money for this.
26:29
Josh: Yeah, in 2026 and 2027, I mean, that's $2 trillion of guidance that comes from cash flow.
26:35
Josh: It comes from Google spending 90% of the money they actually make.
26:38
Josh: So there isn't leverage. This is somewhat sustainable.
26:42
Josh: And now that we have the rough CapEx guidance, people can kind of price in what
26:46
Josh: this will look like when it impacts the market.
26:48
Josh: And that is the current layout of the land.
26:52
Josh: This is everything you need to know about the full AI stack,
26:55
Josh: what has been doing well, what hasn't.
26:57
Josh: And I guess the question for you is now that you're oriented or whoever is listening
27:01
Josh: is what part of the stack are you most interested specifically what companies
27:05
Josh: are most interesting because there is this gigantic hot ball of money that is
27:09
Josh: moving its way through the market for the first time in a very long time.
27:12
Josh: This is a paradigm shift for the last decade or two.
27:15
Josh: All of the money has kind of accumulated at the top of these funnels has accumulated
27:18
Josh: to Google to Amazon to these large mega cap companies, but now it's working its way out.
27:23
Josh: And where is that money going to end up? It is TBD. We've seen semiconductors.
27:26
Josh: We've seen memory do really well will they continue tbd where
27:30
Josh: else is the money flowing to we're not sure but the goal
27:33
Josh: of this episode is just orientation right you kind of now have a lay of the
27:36
Josh: land you have an idea of where everything is and can maybe go out and pick some
27:40
Josh: winners so i guess the prompt for this episode is who are the winners who's
27:43
Josh: going to win this um battle of capex accumulation is is what i would love to
27:49
Josh: know because i will plan and allocate my portfolio accordingly
27:53
Ejaaz: Yep, same here. My bags is the answer, Josh.
27:56
Ejaaz: But right now, I think that we are kind of in a bubble, but it's a different
28:02
Ejaaz: type of bubble than what we're used to because no one's levered up,
28:05
Ejaaz: as I mentioned earlier on.
28:06
Ejaaz: And it remains to be seen. Like every quarterly earnings, I'm going to evaluate
28:09
Ejaaz: everyone's spec sheet and see if they're actually making money from this thing.
28:13
Ejaaz: Q1 earnings, which released over the last two weeks, tells us that it is.
28:16
Ejaaz: And listen, there are always levered bets everywhere.
28:19
Ejaaz: Like if you don't buy any of the companies that we You mentioned there are smaller
28:22
Ejaaz: cap stocks that you might go after.
28:24
Ejaaz: They are high risk with all of these different types of investments and the
28:27
Ejaaz: narrative can change, as we know, in AI in a second.
28:29
Ejaaz: So the long story short is, neither of us know, but we will be keeping track
28:33
Ejaaz: of everything and we'll be updating you.
28:35
Ejaaz: Josh's prompt earlier on is genuine for me as well.
28:39
Ejaaz: If you want to hear more about lower layers of the stack, we will do the research
28:42
Ejaaz: and we'll deliver you that episode.
28:43
Ejaaz: Let us know in the comments and let us know what stocks you're investing.
28:46
Ejaaz: But aside from that, Josh, I think we're done.
28:48
Josh: That's it. Thank you guys so much for watching. As always, share it with your
28:51
Josh: friends who might be interested in an episode like this, and we will see you guys in the next one.
28:55
Ejaaz: See you guys.