Latest / In the Loop AI / The AI Energy Stack: The Actual Industry Winners

Transcript

The AI Energy Stack: The Actual Industry Winners · In the Loop AI

0:00
Ejaaz: For three years, the AI trade has been one simple idea.
0:03
Ejaaz: Buy the chips. NVIDIA became the most valuable company on Earth by making these
0:07
Ejaaz: chips, and they've made countless of investors rich because of this.
0:11
Ejaaz: But this investment thesis now has become overcrowded. Not because AI is slowing
0:15
Ejaaz: down, but because there's another problem that has surfaced.
0:18
Ejaaz: No one can turn these GPUs on. Trillions of dollars being spent in AI,
0:22
Ejaaz: and all these GPUs are collecting dust in data centers.
0:26
Ejaaz: The problem, energy, electrical grids, wiring, networking, the transformers,
0:30
Ejaaz: the web of infrastructure that sits around a GPU that allows it to keep alive,
0:34
Ejaaz: to talk to each other, to transmit petabytes of data between each other.
0:38
Ejaaz: This is the next constraint that hasn't been solved yet and where the majority
0:42
Ejaaz: of the AI capital will eventually flow.
0:44
Ejaaz: Now, there's four physical things that keep a chip alive, essentially.
0:48
Ejaaz: The power to run it, the light to transmit all the data between them,
0:51
Ejaaz: the silicon wiring between them all, and then somewhere to rent it all from.
0:55
Ejaaz: And almost nobody is really talking about this. And on this episode,
0:58
Ejaaz: we're going to unpack that specific layer of the infrastructure stack,
1:00
Ejaaz: the power layer, and why it's so important going forwards in terms of capital.
1:05
Josh: Yeah, and this may feel a little bit different because we're always talking
1:07
Josh: about whose model is the smartest. That's the common conversation.
1:09
Josh: But once you strip away all that abstraction, a token is really just what,
1:14
Josh: is output and that is just a series of electrons flowing through chips like
1:18
Josh: flowing through fiber and then heat being pulled out of the system and all of
1:22
Josh: that requires a lot of energy and electricity uh
1:25
Josh: the money cycle that we're going to talk about has kind of flowed towards this
1:29
Josh: direction as well i mean everyone started with the crowded gpu trade then it
1:33
Josh: flowed to the semis that exist around them then it flowed to the memory trade
1:36
Josh: and now it's kind of moving over to the
1:40
Josh: seemingly the end game bottleneck, which is energy.
1:43
Josh: I mean, US data power center demand, I think is projected to roughly double
1:47
Josh: 80 gigawatts in 2026 to 150 by 2028. So the grid's just not really built for that.
1:54
Josh: And there's no shortage of demand for that. And when we think about energy as
1:58
Josh: an idea, even if you believe that the AI bubble is towards the latter end of
2:02
Josh: it, energy is still something that's not going away.
2:05
Josh: We had energy problems prior to the LLM becoming a big deal from the chat GPT moment.
2:10
Josh: This is just an extension and an exaggeration on top of that.
2:14
Josh: And it creates a lot of interesting opportunities in the marketplace to actually
2:18
Josh: participate in this bottleneck that is now known as, I guess we call it the energy bottleneck.
2:23
Ejaaz: Yeah. And I think a lot of the reason why people don't talk about this is because
2:26
Ejaaz: it's sort of unsexy, you know, not everyone is an electrical grid expert.
2:31
Ejaaz: Not everyone knows how to network wires between these complex GPUs.
2:35
Ejaaz: I'm still trying to wrap my head around the GPU itself, So the fact that we're
2:39
Ejaaz: bringing in all these other things, it's quite complex.
2:41
Ejaaz: Now, if you were to look at a budget of spending for an AI lab or someone that's
2:47
Ejaaz: setting up a data center, only 10% of that budget gets allocated to the power
2:53
Ejaaz: and infrastructure side of things.
2:55
Ejaaz: The irony of it now is it's going close to 100% of the bottleneck of the entire
3:00
Ejaaz: thing. So even though it's not the majority of the cost, still 90% of the cost
3:03
Ejaaz: goes into the actual GPUs itself. That's the most expensive stuff,
3:06
Ejaaz: the memory and the silicon required to build those GPUs.
3:10
Ejaaz: Only 10% is allocated to power, but that is like the main bottleneck that we're facing today.
3:15
Ejaaz: Now, the lead time, if you are an AI lab, you could be the richest,
3:20
Ejaaz: most important person on earth right now.
3:22
Ejaaz: If you want to build a data center in the US, the lead time to get the necessary
3:25
Ejaaz: power to your GPUs is five years.
3:28
Ejaaz: And there are different sectors within like the power thing.
3:31
Ejaaz: It's not just the energy infrastructure grid you need to get access to.
3:34
Ejaaz: You need to get lead times for wiring and to get transformers designed,
3:38
Ejaaz: custom made and delivered to you. All of that takes a lot of time.
3:41
Ejaaz: And we're looking at basically like half a decade.
3:43
Ejaaz: So what we've seen a lot of these AI labs start to do now is figure out what
3:48
Ejaaz: alternatives they could potentially procure to help them get GPUs online.
3:53
Ejaaz: Now, you've all heard the news of Meta, Amazon, Microsoft, and Google spending
3:59
Ejaaz: trillions of dollars this year alone.
4:01
Ejaaz: I think it was something along the lines of, actually, it might have hit a trillion
4:04
Ejaaz: dollars for this year, right? Because I think after the last quarterly earnings,
4:08
Ejaaz: they like upped their budget.
4:09
Ejaaz: They haven't set those data centers up. In fact, there have been delays.
4:13
Ejaaz: Stargate from OpenAI, which said that was their biggest project to kind of get
4:17
Ejaaz: compute online, has been delayed multiple times.
4:20
Ejaaz: And the point around this is because of compute, and not many people are talking about this yet.
4:24
Ejaaz: So I think on today's episode, as we walk down this infrastructure stack,
4:28
Ejaaz: and we did a previous episode before, where we covered all the layers of the
4:31
Ejaaz: AI infrastructure stack, we
4:33
Ejaaz: Talked about going from layer zero, which is the model labs down to layer one,
4:37
Ejaaz: which was the hyperscalers, GPU semiconductors, we didn't dig as much into layer
4:41
Ejaaz: five, which is the power and infrastructure side of things.
4:44
Ejaaz: And that's what we're going to do today. And we're going to do it through the
4:46
Ejaaz: lens of four specific companies, starting with Bloom Energy.
4:50
Josh: Yeah. And just before we get into Bloom, there's one thing. Well,
4:53
Josh: one of the things is, I think there's like these bottlenecks within bottlenecks.
4:56
Josh: You mentioned the lead time, but also the actual infrastructure is becoming much more demanding.
5:01
Josh: Vera Rubin, the new chips, I believe, those server racks take...
5:05
Josh: Over 100 kilowatts, which is a 10x increase from what these companies and these
5:10
Josh: data centers were preparing for.
5:11
Josh: And then also before that, I want to kind of loosely define the power for people
5:15
Josh: who are not familiar, because there's watts, there's watts hours,
5:18
Josh: there's gigawatts, it's just like a brief one on one quickly on what everything is.
5:21
Josh: A watt is a rate, it's how fast these data centers draw electricity,
5:25
Josh: a watt hour is an amount, which is the total you drew. So a gigawatt is a billion of those watts drawn.
5:31
Josh: One gigawatt is equal to roughly one large nuclear reactor or enough to power
5:36
Josh: 750 to a million U.S. homes. So this is a tremendous amount of data that we're
5:41
Josh: talking about or a tremendous amount of energy that we're talking about.
5:43
Josh: And these data centers are demanding tens to hopefully soon hundreds of gigawatts.
5:47
Josh: Now back to Bloom Energy.
5:49
Josh: Bloom Energy has a solution for this. It is solving the most painful solution
5:53
Josh: in AI, which is this energy data source.
5:55
Josh: They use these kind of modular energy reactors that allow them to allow these
5:59
Josh: data centers to remove themselves off of the grid and generate a little bit
6:03
Josh: more energy closer to the data center.
6:05
Ejaaz: Yeah, actually, remember, the first time I heard about Bloom Energy at all was
6:09
Ejaaz: in Leopold Ashenbrenner's 13 filings at the end of last year.
6:13
Ejaaz: It was one of his biggest positions. I think in his recent one,
6:16
Ejaaz: I think it makes up about 12.7% of his entire portfolio.
6:20
Ejaaz: So he's very bullish, this particular company. And the reason why is,
6:24
Ejaaz: As I mentioned earlier, the lead times to getting power generation on site for
6:29
Ejaaz: your data center, if you go through the old traditional way, is around half a decade.
6:33
Ejaaz: Bloom energy will reduce that down to 90 days they have a product it's a gas
6:39
Ejaaz: turbine which they can custom build and deliver on site they kind of like fly
6:42
Ejaaz: it in and they reassemble it
6:44
Ejaaz: and it can generate enough power for you to run your gpus now obviously that's
6:49
Ejaaz: A very attractive thing a ton of companies have signed contracts with them now
6:52
Ejaaz: the one caveat that i will say is they haven't really delivered this at scale
6:55
Ejaaz: bloom energy has yet to prove themselves from a manufacturing capacity and delivery capacity, but it is
7:02
Ejaaz: a company that a lot of people like to talk about and kind of speak highly of.
7:06
Ejaaz: Now, the revenue kind of speaks for itself as well, because obviously there's
7:09
Ejaaz: a lot of lead time here, there's a lot of demand here. And as we mentioned,
7:11
Ejaaz: that money is going to flow down.
7:13
Ejaaz: As of Q1, 2026, their revenue hit $750 million.
7:17
Ejaaz: That is up 130% of revenue year over year, which is just like a crazy kind of swing.
7:25
Ejaaz: And of that, they're making around 70 million profit, which isn't large in the
7:29
Ejaaz: context of some of these big hyperscaling infrastructure companies.
7:33
Ejaaz: But for a relatively small company, this is pretty impressive.
7:36
Ejaaz: And Oracle is powering its 2.4 gigawatt data center, primarily using Bloom Energy.
7:43
Ejaaz: So they've signed some pretty major contracts as well.
7:45
Ejaaz: So Bloom Energy is one of these companies, and there's various different versions
7:49
Ejaaz: of this, where it's not just gas turbines, maybe it's solar energy or nuclear
7:52
Ejaaz: energy. But Bloom Energy is basically the first company that is having an impact
7:56
Ejaaz: almost immediately after this constraint has been identified.
8:00
Josh: Yeah, well, in terms of gas turbines, I mean, that's one way of doing it.
8:03
Josh: And the line to get a gas turbine extends back to 2029. So if you actually want gas turbines.
8:09
Josh: A gas turbine at your data center and you place an order today.
8:13
Josh: And for those who aren't familiar, a gas turbine is basically a jet engine,
8:16
Josh: which is spinning a generator.
8:17
Josh: And it's so funny how this works. It's like the way we generate energy today
8:21
Josh: is the same way we did with like a steam engine way back in the day.
8:23
Josh: You're just basically creating energy to propel a spinning thing.
8:27
Josh: In this case, it's a jet engine.
8:28
Josh: If you want one of those, you're waiting at least three years,
8:30
Josh: probably the next decade. So if you want something in the 2020s, that's not for you.
8:34
Josh: The next option is nuclear or small nuclear reactors.
8:37
Josh: Those are amazing. They're incredible. But if you want them in this decade,
8:41
Josh: again, not possible. Those aren't probably coming online at scale until around 2030.
8:46
Josh: And the licensing and legislation around that is pretty slow.
8:49
Josh: So that leaves us with these fuel cells. And that's what Bloom has.
8:52
Josh: And they're these modular boxes you stand up on site in 90 days instead of three to five years.
8:57
Josh: And so far, people have really been liking them. We've seen that reflected in the stock price.
9:01
Josh: With people desperate for energy, anyone who's able to actually deliver it.
9:05
Josh: A accelerated timeline is going to win. I think about Elon all the time and
9:09
Josh: how they were able to build the Colossus data center and how much premium it
9:13
Josh: had to getting GPUs online quickly.
9:15
Josh: I mean, think about how much Anthropic and Google are paying for these GPUs.
9:19
Josh: The same thing is true with electricity.
9:21
Josh: If a data center is bottlenecked by this power, anyone who's able to provide
9:25
Josh: it for them is going to get paid a tremendous amount of money and a large margin for it.
9:29
Josh: And so far, Bloom has been one of those companies through their fuel cell program
9:32
Josh: that can be deployed in 90 days on site, they're collecting a lot of that value.
9:36
Josh: So I think that's kind of the bull case for Bloom. That's why people are mostly excited about it.
9:39
Josh: It is just a really difficult place for companies because it's hard to make
9:44
Josh: energy. And a company that has figured out a way, I'm not sure this is the end
9:48
Josh: state, but this is certainly a intermediary step.
9:51
Josh: And right now, that's all people need. It's just power today.
9:54
Ejaaz: Now, moving on, one of the other bottlenecks that is appearing,
9:59
Ejaaz: particularly with GPUs talking to each other, is data transfer specifically.
10:04
Ejaaz: Now, if we think of an AI model, typically it has grown exponentially in size
10:10
Ejaaz: and in quantity of data that is required to train a frontier AI model.
10:14
Ejaaz: We've gone from, if you remember, Josh, billions of parameters to trillions
10:19
Ejaaz: of parameters and now tens of trillions of parameters.
10:22
Ejaaz: So the model weights of these actual AI models are huge. People are carrying
10:27
Ejaaz: them around in briefcases.
10:29
Ejaaz: They're like incredibly dense.
10:31
Ejaaz: And when you're training a model, and when you're even inferencing a model,
10:34
Ejaaz: when you're trying to speak to it, when you're trying to send it a prompt and
10:36
Ejaaz: get a response, there is petabytes of data, which is an order of magnitude larger
10:42
Ejaaz: than anything we've ever really spoken about at scale in industry, at least, before.
10:46
Ejaaz: That are transferred between GPUs. So it's not just good enough to have the
10:50
Ejaaz: GPUs and wiring them up and having the power flowing to them.
10:53
Ejaaz: You need to transfer data between them. They need to talk to each other in rapid
10:57
Ejaaz: pace. If they talk to each other slowly, then they won't be able to work as
11:01
Ejaaz: quickly and as powerfully as you'd expect. Think of it like this.
11:04
Ejaaz: Every single way that a GPU requires to function, the lifeblood of it,
11:08
Ejaaz: the energy, the ability to communicate is a potential constraint.
11:12
Ejaaz: And this communication thing is another one.
11:15
Ejaaz: Now, typically, the way you achieve this is through copper wiring.
11:18
Ejaaz: Copper has a really interesting chemical property that allows you to transmit
11:23
Ejaaz: data at lightning speed without heating up, except we've run into a problem.
11:28
Ejaaz: We now have too much data. And so the copper wires are heating up,
11:32
Ejaaz: which causes energy to dissipate and energy efficiency and cost efficiency,
11:36
Ejaaz: therefore, to not really work in the favor of the AI capex that is being spent on this.
11:41
Ejaaz: So people have started scrambling around. They've started looking around for
11:44
Ejaaz: another type of material.
11:46
Ejaaz: And you wouldn't believe it. They've settled on light, optical fibers to transmit data using light.
11:51
Ejaaz: And one of the companies that is moving data as light is known as Lumenta.
11:56
Ejaaz: They're using optical fibers and laser transmitters to move data between chips
12:00
Ejaaz: as light because copper melts at a certain level where data becomes too burdensome, essentially.
12:07
Josh: Yeah, the bottleneck was very much copper and it might still be to some extent
12:10
Josh: because copper is the fastest way of transferring data just because it has so
12:13
Josh: much bandwidth. It's so dense and you could transfer.
12:15
Josh: I mean, now we have clusters of hopefully soon half a million GPUs.
12:19
Josh: I know that's what people are working on.
12:21
Josh: All those gpus need to communicate they need to do so as fast as possible copper
12:25
Josh: has enough bandwidth unfortunately there's this direct correlation between the amount of data
12:30
Josh: and the length it has to travel it's physically impossible to put that many
12:34
Josh: gpus that close to each other to make it feasible there's just not enough space
12:39
Josh: inside of a room to make it happen so as these data centers get larger the
12:43
Josh: distance in which jane needs to travel gets further copper becomes a worse and
12:46
Josh: worse option because i mean every single bit that has to flow through that it causes heat.
12:52
Josh: It causes energy and then like you mentioned the copper it's going to start
12:55
Josh: melting and it's really expensive to cool and it's really difficult so what's
12:58
Josh: the next best thing is photonics you just use light and you use fiber optics
13:02
Josh: and nothing travels faster than the speed of light so
13:05
Josh: it's just a matter of how much bandwidth we could squeeze into those rays of
13:09
Josh: light and by stacking this fiber and by using this glass fiber we're able to actually transfer,
13:14
Josh: information from one ship to another at the speed of light and lumentum is a
13:18
Josh: company that's working on that so i think that's why people are really excited
13:21
Josh: about this company in particular as these data centers grow larger in footprint
13:26
Josh: they are going to require
13:28
Josh: a longer distance in which data has to travel and there's going to be a lot
13:31
Josh: more data which means copper probably won't be ideal and unless we solve this
13:36
Josh: material problem the next best thing
13:38
Josh: is photonics it's just sending it right over a fiber optic line at the speed of light or i should say,
13:44
Josh: many fiber optic lines to store as much data as possible in this transfer.
13:48
Josh: And that's why Lumentum is, I think, one of the more exciting companies.
13:51
Josh: This is part of Jensen's portfolio, if I am not mistaken, right?
13:55
Ejaaz: Yeah, yeah. He's put in about $2 billion into Lumentum Holdings.
13:59
Ejaaz: But to your point, it's important to mention that this is a relatively new type
14:03
Ejaaz: of technology. It hasn't quite hit that exponential wave that NVIDIA GPUs did two years ago.
14:09
Ejaaz: And so we're waiting on that. The bet is that that is going to happen for these
14:12
Ejaaz: particular types of transmitters and receivers. Now, their unit of sales is
14:18
Ejaaz: going up pretty aggressively.
14:20
Ejaaz: Last year, they sold, I think, around like 20 million units.
14:23
Ejaaz: This year, it's now 60 million units. So it's kind of like a two and a half
14:26
Ejaaz: to three X from what they were last year.
14:29
Ejaaz: But they have competitors in the form of Corning, GLW, who NVIDIA also invested $3.2 billion.
14:35
Ejaaz: Now, they're not direct competitors, but Corning creates the actual optical
14:39
Ejaaz: glass fibers that will be used to kind of like transmit light and data for this.
14:44
Ejaaz: And there are many other kind of like companies that are like highly focused on this.
14:47
Ejaaz: It's worth mentioning though that scaling any of these material goods,
14:52
Ejaaz: whether it's the glass fibers, whether it's transmitters, whether it's the power
14:55
Ejaaz: switches or whatever that might be, takes an incredible amount of lead time.
15:00
Ejaaz: Like one of the major critics of the memory bottleneck that everyone is quite
15:05
Ejaaz: familiar with right now is Apple just recently announced that they're raising
15:09
Ejaaz: the prices of all their devices because they can't get their hands on enough memory.
15:13
Ejaaz: NVIDIA is facing the same thing. And when they go to SK Hynix and the memory
15:16
Ejaaz: makers and they say, hey, can you just build more, please? They say, it takes time.
15:21
Ejaaz: The machinery is incredibly detailed. Like it takes a while and we're gonna
15:25
Ejaaz: be in this for a few years. And the optical fiber side of things,
15:29
Ejaaz: the material side of power infrastructure in itself is where this is also playing
15:34
Ejaaz: out. nothing's really changed in that sense.
15:36
Josh: Also, I just sent you a custom link of a post that I'd like to share because
15:40
Josh: this reminded me of an idea that I heard that I think is probably relevant because
15:44
Josh: the metric underneath all of this is joules per bit. It's the energy cost to
15:47
Josh: move one bit. And at scale, light wins decisively.
15:51
Josh: Like there is nothing that flows smoother than light. And it reminded me of
15:55
Josh: this post from John Carmack.
15:56
Josh: He is the guy who created Dune. He's like this internet legend.
16:00
Josh: And basically, John was saying that you can store data within this light if
16:05
Josh: it is transferred over a certain.
16:08
Ejaaz: Length of space. Oh, Elon responded to this, right?
16:09
Josh: Yes. It is amazing. So when we think about memory and we think about data transfer
16:14
Josh: in general, I think fiber optics is a really interesting industry to consider
16:19
Josh: that can go much further than I think what we imagine it can do today.
16:23
Josh: Because during the time in which this light or these bits are being transferred at the speed of light
16:28
Josh: you can transfer a lot of data and if that's spun over a long enough distance
16:32
Josh: it can actually store data in a way that is
16:35
Josh: very fast and accessible way faster than we get in memory and RAM because it
16:40
Josh: is actually moving at the speed of light and there's a lot of interesting science
16:43
Josh: experiments and I think general experimentation as it relates to,
16:47
Josh: fiber in general not only as a way of transferring data but as a way of storing
16:51
Josh: data should you stack it and extend it over a long period of time and i just
16:54
Josh: want to highlight this is
16:55
Josh: like lumentum is hot right now but i think the industry of fiber and fiber optics
16:59
Josh: and moving things around at the speed of light is ultimately where we're going
17:03
Josh: to land on we talk about data centers in space
17:05
Josh: They're going to be traveling at the speed of light. It has the lowest amount
17:07
Josh: of friction. It has the highest velocity of information.
17:10
Josh: If we can experiment and kind of iterate on the way that we do that,
17:13
Josh: I think that's something worth noting.
17:14
Josh: And Lumentum being the fiber optic company may stand to benefit over a much
17:18
Josh: longer period of time than I think people think based on this current bottleneck.
17:21
Josh: I just want to highlight that because I was like, huh, that's pretty cool.
17:24
Josh: That reminded me of something I saw back in the day.
17:26
Ejaaz: Oh, I love it. Now, you would think that the entire problem with the power infrastructure
17:31
Ejaaz: stack is just generating the power, right, Josh?
17:34
Ejaaz: But there's also delivering the power, making sure the right amounts of power
17:39
Ejaaz: go between the chips at this time. And then when they are really in demand,
17:42
Ejaaz: delivering even more power, right?
17:44
Ejaaz: That takes incredibly complex design patterns. Now, most people probably aren't
17:49
Ejaaz: aware of this, but NVIDIA isn't a chip manufacturing company.
17:53
Ejaaz: They're a chip design company.
17:55
Ejaaz: They design what the silicon wafer is going to look like. And then they send
17:58
Ejaaz: that over to their friends in Taiwan and TSMC.
18:00
Ejaaz: And they actually manufacture the thing for them. Now, one key component of
18:04
Ejaaz: manufacturing this design is figuring out how the power is transmitted between
18:09
Ejaaz: these chips and within the chip itself.
18:11
Ejaaz: And this is a huge role that is played within the power infrastructure because
18:16
Ejaaz: it determines where the power gets dispersed. It's an incredibly hard thing to figure out. Now...
18:22
Ejaaz: About three weeks ago, there was this very popular compute conference called
18:27
Ejaaz: Computex, which was basically NVIDIA GTC 2.0. Jensen was like the bell of the bell there.
18:34
Ejaaz: So he goes on stage and he talks about this company, Marvel,
18:37
Ejaaz: and he says, this is the next trillion dollar company. And this isn't investment
18:41
Ejaaz: advice, by the way. This is literally verbatim what he said in Taiwan.
18:45
Ejaaz: And the stock subsequently went up 76% after he mentioned this.
18:51
Ejaaz: And I was looking up on this just to see, you know, have they like,
18:53
Ejaaz: you know, rescinded since then?
18:55
Ejaaz: Josh, they broke into the S&P 500 based off of this single use.
19:00
Ejaaz: And the reason the TLDR of why Marvel is attracting so much attention and capital right now
19:05
Ejaaz: is because they are experts in designing that part of the chip,
19:10
Ejaaz: which determines where power gets dispersed, how it gets dispersed and where
19:14
Ejaaz: it gets used within the chip and between chips.
19:17
Ejaaz: And if you're wondering, oh, okay, well, they're just solely reliant on NVIDIA, the answer is no.
19:21
Ejaaz: Broadcom and every other single AI company that is trying to design their own
19:26
Ejaaz: chip and diversify away from NVIDIA is using Marvel.
19:31
Josh: Wow, it's up 80% on the month I'm seeing in one month since Jensen mentioned this company.
19:38
Josh: I actually don't know much about Marvel and I feel bad about it because clearly
19:41
Josh: I missed out. But it's one of those things where it's been very lucrative to
19:44
Josh: copy trade the people in authority.
19:46
Josh: It's like if someone who is running the largest company in the world says this
19:51
Josh: company is going to a trillion dollars, people are going to react.
19:54
Josh: And you have to just trust the fact that he has a good reason to do this.
19:57
Josh: I mean, Jensen seems like he holds pretty strong generally on his morals.
20:02
Josh: He's not trying to pump and dump. He genuinely believes this.
20:05
Josh: He has made a sizable investment. What is the total amount? It's a couple of
20:08
Josh: billion dollars, right?
20:09
Ejaaz: I did $2 billion. Yeah. So he's up, he's up pretty big now.
20:14
Josh: Man, my God, exactly. That's the point. And it's like, yeah,
20:16
Josh: he, he actually has the ability to influence the success of this company,
20:20
Josh: not just by talking about it, but by actually entering and the technology into his company.
20:24
Josh: Same thing happens when you talk about the person in charge of the United States
20:27
Josh: of America. Like when Trump says something, it generally, it goes up generally
20:31
Josh: as a good reason. People are looking for projects to invest in,
20:34
Josh: for companies to invest in.
20:35
Josh: And Marvel has been one of the most recent winners. Now, there is another section
20:40
Josh: of this stack in which you have to rent out the compute. Say that you are not,
20:45
Josh: you don't have hundreds of billions of dollars of CapEx at your disposal this year.
20:49
Josh: Sorry to hear that. Chances are you're not winning the AI race,
20:51
Josh: but you can at least make an impact. You could do something novel,
20:54
Josh: something interesting in AI. And if that's the case, you're going to need to
20:58
Josh: rent that compute from someone.
20:59
Josh: There's a bunch of companies that we've talked about on the show in which you can rent compute from.
21:03
Josh: We talked about Iren, a lot of the NeoClouds, but we have a new one to talk
21:07
Josh: about today, which is called Nebius.
21:09
Josh: Ejaz, can you explain to everyone what Nebius is, what makes them so different and unique?
21:13
Ejaaz: It's actually not what makes them different or unique. It's just,
21:16
Ejaaz: it's a scarce service that they supply.
21:18
Ejaaz: So if you've heard of CoreWeave or if you've heard of Iron that Josh just mentioned,
21:25
Ejaaz: they're known as a neocloud.
21:28
Ejaaz: A neocloud is basically a company that will come in, they will construct the
21:33
Ejaaz: data center for you, they will handle all the wiring, they'll make sure that
21:36
Ejaaz: the GPUs talk to each other, they'll make sure that the networking and transformers
21:40
Ejaaz: and receivers, they'll make sure that the copper doesn't melt.
21:43
Ejaaz: And they'll make sure that these beautiful, precious GPUs that you just spend
21:47
Ejaaz: billions of dollars on function optimally.
21:50
Ejaaz: The reason why someone will go to a Nebius or a CoreWeave is simple.
21:55
Ejaaz: Google, Amazon, and all of those companies, they don't want to focus and spend
21:58
Ejaaz: time operating and maintaining and designing the flow of a data center.
22:03
Ejaaz: They want someone else to handle them. That's what Nebius does.
22:06
Ejaaz: And they do it pretty damn well. I believe Nebius's background is they operated
22:10
Ejaaz: a bunch of data centers that focus primarily on, maybe it was mining
22:15
Ejaaz: in the Web3 side of things, maybe it was something else, but they pivoted pretty
22:18
Ejaaz: aggressively to focus on AI specifically and their expertise of managing hardware.
22:23
Ejaaz: And most importantly, the software that talks to each of these GPUs and the
22:27
Ejaaz: different hardware components is their expertise. So the answer,
22:30
Ejaaz: truthfully, is they do exactly what CoreWeave does. They do exactly what Iron
22:34
Ejaaz: does, but they do it at a massive scale.
22:36
Ejaaz: And their Q1 revenue is up almost 700% on the year. It's just insane.
22:41
Josh: Yeah, it's crazy. I saw it at $400 million dollars and this is the thing with the demand
22:45
Josh: issues is that like if you build it they will come it doesn't matter who you
22:48
Josh: are one fun fact about nebius is they were previously known as yandex or this
22:53
Josh: is what was built from the remains of yandex
22:55
Josh: if you are from russia you probably know this because that was very much the
22:58
Josh: google of russia up until 2024 when they severed and they started to build a new cloud
23:03
Josh: so that's a yeah they are familiar with
23:06
Josh: managing data and storing data and uptime with servers wow yeah so that was
23:11
Josh: interesting and then i think one of the things that makes this company special
23:14
Josh: again nvidia is a backer nvidia has placed
23:17
Josh: two billion dollars of direct equity into this company nvidia should just like
23:20
Josh: man if they weren't making gpus jensen could just make it as a vc because this.
23:24
Ejaaz: Dude is making
23:25
Josh: Pretty unbelievable,
23:28
Josh: yeah two billion yeah maybe we just screenshot this portfolio and i just send
23:32
Josh: it to my broker and say can you buy all of these for me just split it across
23:35
Josh: my whole portfolio because,
23:37
Josh: clearly jensen knows what he's doing but the the company is doing remarkably
23:40
Josh: well and because they have contracts from companies that you actually have heard
23:44
Josh: of they have 20 billion dollar contract from microsoft they have a 27 billion dollar contract from.
23:51
Josh: Meta and then they have a backlog that's approaching 50 billion dollars for
23:55
Josh: the years 2027 all the way out to 2031 and this is the thing that we talk about a lot too is that the,
24:02
Josh: the useful life of a gpu has extended so far
24:06
Josh: that i think these companies are getting revalued and re-underwritten based
24:09
Josh: on that because a lot of the time
24:11
Josh: people would say that oh if you have a gb100 today it's going to be worth 10
24:17
Josh: of what it is today five years from now and it's a five-year depreciation schedule
24:21
Josh: the reality is is that gpu from three to four years ago
24:24
Josh: is worth more today per hour than it was back then
24:27
Josh: and so long as this demand curve continues these gpus the useful life of them
24:31
Josh: is going to continue to extend out over longer and longer durations,
24:35
Josh: making the inventory of these companies more and more valuable.
24:38
Josh: And if that trend continues, everyone's going to have to continue to re-underwrite
24:41
Josh: these companies based on this new depreciation scale.
24:44
Josh: And that is going to continue to bring the value of the company up.
24:47
Josh: They have the scarce resource, which is the ability to actually build these
24:50
Josh: data centers, which means they have the land, they have the power, they have the shell.
24:53
Josh: And this is another company that is doing it pretty well. So Nebius is a good
24:58
Josh: example. Neoclads in general are a good.
25:03
Josh: Insane demand that is insatiable it's like no one can actually provide enough
25:07
Josh: energy enough gpus that are powered online and any company that's able to contribute
25:12
Josh: to that is just going to continue to grow so long as this demand stays there
25:16
Josh: and there's no signs of it's slowing down.
25:18
Ejaaz: You know as we talk about this it's so interesting to see this sort of
25:23
Ejaaz: It's like this entire ecosystem that has built around the gpu as gpus are basically
25:30
Ejaaz: the new gold bar right there's this whole economy around this.
25:34
Ejaaz: This is news that I was seeing earlier this week that the Chicago Mercantile Exchange, CMC,
25:40
Ejaaz: I believe, is setting up a compute futures trading index so that you can hedge
25:47
Ejaaz: trades and do trades based off of compute and GPU infrastructure.
25:52
Ejaaz: It is becoming a financial instrument that you can eventually borrow money off
25:57
Ejaaz: of. That's what NVIDIA and a ton of other companies are doing.
26:00
Ejaaz: And so it's this kind of like pseudo or proxy investment vehicle where it's just based around GPUs.
26:06
Ejaaz: And this whole ecosystem is now blossoming around it, whether it's neoclouds,
26:10
Ejaaz: whether it's optical light fibers, whether it's the silicon itself,
26:13
Ejaaz: whether it's the chip design that helps
26:15
Ejaaz: transmit all this kind of power, it is formed around the gravitational pull
26:19
Ejaaz: of these GPUs is pretty insane.
26:22
Ejaaz: One final note that I wanted to mention for Nebius specifically,
26:25
Ejaaz: And they have the same advantage that CoreWeave and Iron has is permitting and
26:29
Ejaaz: licenses, like yawn, right? Like who the hell cares?
26:32
Ejaaz: As we mentioned earlier, a lot of the five-year lead time to get like power
26:37
Ejaaz: and energy to your GPUs is regulatory.
26:41
Ejaaz: It's red tape. It's trying to jump through the government hoops to get all of this figured out.
26:47
Ejaaz: Nebius, Iron, Coreweave all have this permitting in advanced stages.
26:53
Ejaaz: And that's why they're signing these multi-billion dollar deals with Meta,
26:57
Ejaaz: Google, and all the other hyperscalers because they have this advantage.
27:01
Ejaaz: It's all a regulatory game. It's all a hardware game. Whoever can figure out
27:06
Ejaaz: a way to arbitrage the red tape that they're currently facing will end up winning pretty hugely.
27:12
Ejaaz: And if you look at what leopold ashenbrenner is investing and if you look at
27:16
Ejaaz: what uh jen so puang is investing the people who are on the ground that are
27:19
Ejaaz: seeing these bottlenecks in real time that can kind of give you an idea of where
27:23
Ejaaz: that ai capital is eventually going to flow
27:25
Josh: Yeah and then one one final point on the neoclabs which i think is fun and underrated
27:29
Josh: and a lot of people don't recognize what the incentive is to go with them as
27:33
Josh: if you don't need the gpus but if you're a large company like meta for example
27:36
Josh: who has that 27 billion dollar deal with nebius
27:38
Josh: you can actually write off that spend as an expense spread over years instead
27:43
Josh: of just like hammering your free cash flow with this increased capex today.
27:46
Josh: And for a lot of earnings reports, we see the first numbers we go to look at
27:50
Josh: are the capex. How much are they going to be spending on building out compute?
27:54
Josh: And if that number is too high based on what the market believes they can actually
27:57
Josh: build, then they get crushed for it.
27:59
Josh: So a way to kind of obfuscate that is by investing in these neoclouds,
28:03
Josh: signing these deals that can then be extended out over the years and don't immediately
28:07
Josh: crush your balance sheet like we're seeing a lot of these companies doing um
28:10
Josh: google being one nvidia being one where they just raised 25 billion dollars
28:13
Josh: so it's an interesting financial instrument as well that these companies are using but i think
28:17
Josh: that is mostly it in the hardware energy
28:20
Josh: stack that's where we are um how do we.
28:23
Ejaaz: How do we do how do we do i'm curious like people have been asking for this
28:27
Ejaaz: episode for a while i i feel like we did a good job kind of shepherding the
28:31
Ejaaz: story between these companies? Josh, do you feel the same?
28:34
Josh: I feel pretty good about it. I think there's this like,
28:38
Josh: the layers go so deep and you could even go deeper than this with like the actual
28:42
Josh: um manufacturing of this and like who is supplying all the raw materials in
28:48
Josh: but i think this is about as low as you need to get
28:50
Josh: because this is the this is the interesting part past this you start getting
28:53
Josh: into all the science and physics that i'm not sure it's super interesting but i feel good about this
28:57
Josh: i actually learned a lot in preparing and recording for this episode
29:00
Josh: as it relates to just like where the bottlenecks are who's responsible for them
29:04
Josh: i didn't even know what Marvel was until Jensen talked about them on stage.
29:07
Josh: So we're learning a lot about them. There's a lot of alpha and coming across
29:10
Josh: these early. I mean, all of the companies so far,
29:12
Josh: they've done pretty well. It's no lie. Like people know about this.
29:15
Josh: This is not this like deep alpha, but there is a really great opportunity.
29:19
Josh: And these companies are building valuable goods and services for companies that
29:24
Josh: need them. So yeah, I think that's probably it. That covers the whole thing.
29:29
Josh: We went through the stack all the way down to the bottom. We previously covered
29:32
Josh: the higher parts of the stack. If you're interested in that episode, you can find it.
29:36
Josh: We'll link it in the description. It's also in our channel. You may have heard
29:39
Josh: it if you've been around.
29:40
Josh: But yeah, I think that's pretty much it. That is covering the base layer,
29:43
Josh: the energy infralayer of this AI stack.
29:47
Ejaaz: There is one more layer, Josh, that I, not today, not next week,
29:52
Ejaaz: but maybe in a few months time that we need to prep hard for,
29:55
Ejaaz: which is the substrate layer.
29:57
Ejaaz: That is the one at the bottom of the bottom. That is what we're talking.
30:01
Ejaaz: We're going back to old school days. We're talking about raw materials here
30:05
Ejaaz: that actually help form the materials, the silicon itself, the wiring that we've
30:10
Ejaaz: been speaking about, where do those materials come from?
30:12
Ejaaz: And like fun little nugget, an Easter egg before you even hear anything about
30:17
Ejaaz: this episode is the majority of those materials, they're in Japan.
30:21
Ejaaz: And Japan has a company that manufactures toilets.
30:25
Ejaaz: It's known as Toto. In fact, if you look at your toilet right now,
30:28
Ejaaz: we filmed an episode about Toto.
30:30
Ejaaz: They happen to have machines which get access to this really funky material
30:35
Ejaaz: that is only accessible in Japan or primarily based in Japan that is used to make GPUs.
30:41
Ejaaz: It is an essential component in NVIDIA's manufacturing stack.
30:44
Ejaaz: Just a little fun fact for you. We're going to make an episode on that eventually.
30:47
Josh: There's a teaser. What do toilet bowls and AI have in common? Ceramic is one answer.
30:52
Ejaaz: Billions of dollars of market cap.
30:54
Josh: Yeah, maybe we could do this again. We'll do it for the furthest down the stack.
30:58
Josh: It's amazing how granular it gets. Like when we talk about these two nanometer
31:02
Josh: chips that are coming down the line, the difference is a matter of atoms.
31:06
Josh: Like they are literally constructing on the atomic level, five atoms.
31:09
Josh: And if one of them is in the wrong spot, then these things don't work.
31:12
Josh: So the precision required to make this happen is is like it's unbelievable there's
31:16
Josh: nothing else in the history of humanity that has ever been so complex as this
31:21
Josh: there's just nothing that humans have made at least and,
31:24
Josh: going down the stack it's just unbelievable to see and the resulting output
31:27
Josh: is you could type into your chat box and you get intelligence out and it seems
31:31
Josh: imperceivable from magic so that is the ai energy stack episode i hope you enjoyed
31:36
Josh: if you did please be sure to
31:39
Josh: leave a comment share with a friend who you think might be interested be like
31:41
Josh: yo bro you should check out this company.
31:44
Josh: These guys on Limitless talked about it and they made a pretty interesting case.
31:47
Josh: Again, none of this is investment advice. I'm not exposed to half these companies
31:50
Josh: for better or worse, but we should start following your own book.
31:56
Ejaaz: We should just copy
31:57
Josh: Trade our own.
31:57
Ejaaz: Episodes because everything that we mentioned has been going nuclear.
32:01
Josh: That's the current state of the union. So if you invest, listen, that's on you.
32:05
Josh: We're too stupid to take our own advice sometimes, But that is the episode.
32:09
Josh: Thank you all so much for watching.
32:11
Josh: Don't forget to share, like, subscribe, leave a comment on something we missed.
32:15
Josh: Is there another one? Is there another layer of the stack that we're missing?
32:17
Josh: Is there another company that is going to go nuclear?
32:21
Josh: That'd be good to know. Just out of curiosity. But yeah, I think that's it.
32:24
Josh: So we'll see you in the next one.
32:25
Ejaaz: See you guys.