Latest / In the Loop AI / The NVIDIA Bank: Jensen's $500B Wall Street Deal and GPUs as an Asset Class

Transcript

The NVIDIA Bank: Jensen's $500B Wall Street Deal and GPUs as an Asset Class · In the Loop AI

0:00
Ejaaz: Earlier this week, the most powerful people in finance stood around a table
0:03
Ejaaz: next to Jensen Huang and announced they'd raised $500 billion to buy NVIDIA GPUs.
0:09
Ejaaz: Now, if you're listening to this and you're thinking this is just an AI bubble
0:13
Ejaaz: circular economy type thing, you might not actually be wrong.
0:16
Ejaaz: Larry Fink, the head of BlackRock, actually likened this deal to mortgage-backed
0:21
Ejaaz: debt securities of the early 2000s.
0:23
Ejaaz: And if that sounds familiar, to which he created. Yes, to which he created.
0:27
Ejaaz: And if you're likening that to a kind of like a PTSD flashback,
0:31
Ejaaz: that's because that's exactly what happened in the 2008 financial crisis itself.
0:37
Ejaaz: But if you look at the news in general, if you look at the way that this deal
0:40
Ejaaz: is structured, it might actually hint at something completely different. In fact, the opposite.
0:46
Ejaaz: GPU prices for renting has gone sky high. It's up 40% on the year.
0:50
Ejaaz: And there's not enough GPUs to back a lot of the deals that Microsoft,
0:54
Ejaaz: Google, Anthropic and OpenAI are signing with NVIDIA. So the question that we're
0:58
Ejaaz: going to unpack on the show is, is this very much a bubble back deal?
1:02
Ejaaz: Or is this something completely different that we're missing?
1:04
Josh: A new paradigm of investing, a new paradigm of financial manufacturing and construction.
1:11
Josh: This is a new investable asset class. Yeah, this is an entirely new thing.
1:14
Josh: GPUs. Who would have thought?
1:15
Josh: Michael Burry, the guy who's responsible for the big short, he was like,
1:18
Josh: no, the price of these things are going down only. It turns out he could not have been more wrong.
1:22
Josh: And now Jensen has assembled Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.
1:28
Ejaaz: It's the Avengers of finance. It's the Avengers of finance.
1:32
Josh: It's like Apollo alone has a trillion dollars.
1:35
Josh: Of assets managed. Blackstone has over 1.3 trillion.
1:38
Josh: Brookfield is over a trillion. And combined, that's 3.4 trillion.
1:42
Josh: BlackRock is bigger than all three of those combined.
1:45
Josh: And they're all doing this together. And together, they've signed this thing
1:48
Josh: called a memorandum of understanding. Now, I had to actually look up what this
1:51
Josh: means because I had no idea.
1:53
Josh: A memorandum of understanding, or an MOU, is a formal, usually non-binding document
1:57
Josh: signed by two or more groups.
1:59
Josh: It shows that the groups share a common goal and plan to work together.
2:01
Josh: So this is not a contractual obligation. We have to start with that.
2:04
Josh: This is not a guarantee that $500 billion is going to flow into this new economy.
2:09
Josh: But it is an intention that all of these people are going to be aligned to work
2:12
Josh: together towards funding this next build out of AI. And what I found most interesting
2:17
Josh: is that this was actually orchestrated entirely by Jensen.
2:20
Josh: Jensen reached out to all of these banks himself personally.
2:24
Josh: And he said, hey, I'd like to work together on this thing. And not a single
2:27
Josh: bank that he reached out to said no.
2:29
Josh: So here we are now with a moment on CNBC in which they're all sitting around
2:32
Josh: a table talking about how they are committing $500 billion to this new asset class.
2:37
Josh: And this is unbelievable. This feels like a, for better or worse,
2:41
Josh: a brand new paradigm for the AI era in which now the collective force of the
2:47
Josh: United States banking system is like starting to get behind this.
2:52
Josh: And I should say this is not for AI as a whole. This is purely for NVIDIA as a company.
2:57
Ejaaz: Yeah, and I want to take a moment to actually explain what's happening here,
3:01
Ejaaz: because I think there's a lot of confusion.
3:03
Ejaaz: There's a lot of headlines that people are getting worried over.
3:05
Ejaaz: NVIDIA stock tanks 4% on the news, but I think that's a little too early to
3:10
Ejaaz: judge. So firstly, what does this structure sort of look like?
3:14
Ejaaz: Well, it's what you're seeing on the screen right now, which is essentially.
3:18
Ejaaz: There's a problem in AI right now, which is all these hyperscalers,
3:22
Ejaaz: all these AI labs, Anthropic, OpenAI, Google, Meta, you name it,
3:26
Ejaaz: have spent a lot of money to buy GPUs.
3:30
Ejaaz: The reason why they're doing this is to train and inference brand new AI models,
3:33
Ejaaz: which they have a lot of paying customers for.
3:36
Ejaaz: But the issue they're facing is the money they've invested, which is now to
3:40
Ejaaz: the tune of $2.6 trillion converted over the next couple of years,
3:44
Ejaaz: I believe, is not enough for them.
3:47
Ejaaz: So much so that they're going into negative cash flow. So what happens when
3:50
Ejaaz: you've spent all the money that you have in your company, in your balance sheet?
3:53
Ejaaz: You need to go to Wall Street. That's exactly what NVIDIA, specifically Jensen, has brokered.
3:59
Ejaaz: He's gone to Wall Street and he said, listen, we need more money to build more
4:03
Ejaaz: GPUs to sell to these different customers so that they can produce their products
4:08
Ejaaz: and services, their new models.
4:09
Ejaaz: And Wall Street has gone back and said, I have an issue with this,
4:12
Ejaaz: Jensen, which is GPUs aren't a versatile asset. like they can only be used for
4:17
Ejaaz: one thing specifically, which is either training a model or inferencing a model,
4:20
Ejaaz: and it's only customer-specific.
4:22
Ejaaz: And Jensen responded to them and said, that's not true at all.
4:25
Ejaaz: GPUs, specifically NVIDIA GPUs, are the most versatile asset out there.
4:30
Ejaaz: You can use it for anything. You can use it for training, you can use it for
4:32
Ejaaz: inference, and you can use it for any model, whether it's Claude,
4:35
Ejaaz: whether it's GPT, whether it's Gemini, whatever. You can use it for it,
4:39
Ejaaz: which means that it's a versatile customer base.
4:42
Ejaaz: Earns a lot of money. And then Wall Street shot back at him and said,
4:46
Ejaaz: well, hang on a second, these GPUs die after a couple of years.
4:49
Ejaaz: And Jensen goes, that's not actually true. In fact, we have 10-year-old GPUs
4:54
Ejaaz: that are being re-signed for another 10 years right now today at a higher price
4:58
Ejaaz: than they sold earlier on.
4:59
Ejaaz: So basically what he's pitched them is this is a new asset class and it can earn a ton of money.
5:05
Ejaaz: And so Wall Street looked at this, some of the biggest financial powerhouses
5:08
Ejaaz: in the world and thought, you know what?
5:10
Ejaaz: He might be right. This is an asset class that can be likened to property or
5:14
Ejaaz: railroads back in the day.
5:16
Ejaaz: And that's why Larry Fink is comparing it to the 2008 mortgage-backed securities.
5:20
Ejaaz: Now, if you're wondering, okay, well, this is like a financial crisis type thing,
5:25
Ejaaz: you might be right, except there was like a few different things going on there,
5:28
Ejaaz: which we'll unpack later in the episode.
5:30
Josh: Yeah, it feels very much like AI compute is equivalent to revenue.
5:34
Josh: And these are very now like durable, appreciating assets that can yield value over time.
5:40
Josh: So when you think of like a bond, per se, the value of a bond is implied to
5:45
Josh: go down over time, the underlying asset like the US dollar due to inflation,
5:49
Josh: but the yield it will come up with is going to outpace that and then some hopefully
5:54
Josh: the construction of a GPU is that not only do you get a yield in terms of the
5:59
Josh: value creation off the back of token generation.
6:01
Josh: But you now also have an asset that is likely going to appreciate.
6:05
Josh: And even in the face that it doesn't, Jensen is giving plunge protection.
6:10
Josh: So 25%. Yes. Depreciation insurance of up to 25% to help the banks get these
6:17
Josh: marginal deals over time. So banks initially were concerned.
6:19
Josh: They don't want to fund this because they don't want
6:21
Josh: nvidia to come out with the new gpu that's a thousand times better than this
6:24
Josh: one and it's going to knock all the margin out from underneath them
6:27
Josh: who knows the roadmap about nvidia gpus better than anyone else it's jensen
6:31
Josh: the guy who's building it so baked into this contract is the idea that
6:34
Josh: jensen will ensure you he will make sure that hey your gpus are not going to
6:38
Josh: fall drastically over time everything is going to be smooth and predictable
6:41
Josh: and we'll work together to fund these companies that don't have
6:44
Josh: the ability to do so so this comes in the form of like this long-term
6:48
Josh: Debt and asset backed structures.
6:51
Josh: And you think of it like if you're not a hyperscaler, if you're not Google who
6:54
Josh: has a couple hundred billion dollars to spend off your balance sheet this year,
6:58
Josh: but you still want to compete in the world of AI, you still need GPUs, these are who you go to.
7:02
Josh: And they will offer you GPUs in exchange for interest on these GPUs.
7:07
Josh: And in the worst case that it doesn't work out, they can just claw back the GPUs.
7:11
Josh: NVIDIA can claw back those GPUs and turn it into their own NeoCloud,
7:14
Josh: give it to another neocloud but the idea is that these assets are valuable they're increasing in money
7:19
Josh: they are transferable in an easy way that you can just take the gpu and plug
7:24
Josh: it in somewhere else or give the actual data center control over to someone else
7:28
Josh: and it's a really lucrative kind of bizarre thing it's like okay if you're a
7:32
Josh: bank 500 billion dollars
7:35
Josh: you have insurance you are now able to allocate this
7:39
Josh: incredibly valuable capital resource to anybody who you want and collect a pretty
7:43
Josh: high rate of return on top of that
7:45
Josh: and nvidia wants in on this too so initially it wasn't for nvidia nvidia now
7:49
Josh: is given the option to backstop up to 25 of each opportunity so that is 125
7:55
Josh: billion dollars at the ceiling
7:56
Josh: and basically now nvidia and co the avengers get to roam around choose who they
8:01
Josh: would like to give these gpus to and wrap it up in a really interesting financial
8:05
Josh: product that they can go off and, I guess, monetize.
8:08
Ejaaz: And I want to stress that this is only for NVIDIA GPUs specifically.
8:12
Ejaaz: Jensen brokered this deal for his company only.
8:16
Ejaaz: And he has a reason to do that because GPUs for the longest time has been very
8:21
Ejaaz: broad based. If you look at some of the GPUs that Google makes or that Meta
8:25
Ejaaz: is making or that even OpenAIR and Anthropic are reportedly working on their own specialized chips.
8:30
Ejaaz: These are exactly what I just said. They're specialized. They can't be used
8:33
Ejaaz: for many other models. It's only used specifically for their things.
8:35
Ejaaz: So it's a much more niche case to create a type of loan or credit-backed security for.
8:41
Ejaaz: Jensen has the opposite issue, which is like, it's too broad.
8:45
Ejaaz: But that makes it an amazing financial asset. So in effect, NVIDIA is sort of becoming a bank.
8:51
Ejaaz: And I wouldn't be surprised if Jensen starts to make a lot of money from this.
8:55
Ejaaz: Over the last couple of weeks, something that he's also started doing is backstopping
8:59
Ejaaz: specific Frontier AI labs and saying, hey, don't worry, I got you.
9:04
Ejaaz: I'll front up the money that you need to purchase my GPUs.
9:07
Ejaaz: And in return, whatever money you make on the products that you're building,
9:12
Ejaaz: you can give me a revenue split from that.
9:14
Ejaaz: I think he signed like a reportedly 10% revenue split from Safe Super Intelligence,
9:19
Ejaaz: which is Ilya Sutskiver's new lab for their breakthrough that they're launching pretty soon.
9:23
Ejaaz: And I think he's gonna do the same for a lot of neoclouts like CoreWeave,
9:26
Ejaaz: Nebius and such like that, that are reporting crazy earnings.
9:29
Ejaaz: I think this morning, CoreWeave reported 464% increase in revenue year upon
9:34
Ejaaz: year, which is just a precursor to like the insane demand that they're seeing right there.
9:38
Ejaaz: But then a question that comes into mind is, where on earth is this money coming from?
9:43
Ejaaz: And on the screen here, it's like the main claimants are pension funds, right?
9:49
Ejaaz: So pension funds who have amassed a large amount of wealth and typically don't
9:53
Ejaaz: invest in high volatile type assets.
9:56
Ejaaz: They kind of stick to real estate, very low interest types of things,
10:00
Ejaaz: are the ones that are going to be backing a lot of this new GPU asset class.
10:04
Ejaaz: And you have the biggest, most
10:05
Ejaaz: powerful financial people in the world that are kind of pushing this on.
10:09
Ejaaz: And so I'm thinking, is this reckless behavior?
10:12
Ejaaz: Well, if you take the word of Goldman Sachs CEO David Solomon,
10:16
Ejaaz: he goes, $500 billion sounds like a lot, but there are $9 trillion in US money
10:21
Ejaaz: market funds and more than $100 trillion in US equities.
10:26
Ejaaz: He has a deep belief in this opportunity and Goldman brings its extraordinary
10:30
Ejaaz: distribution network. Larry Fink, CEO of BlackRock, also says,
10:34
Ejaaz: he said it's a very attractive opportunity with long-dated, long-term returns.
10:38
Ejaaz: They will be talking to pension funds. So it seems like the two most powerful
10:41
Ejaaz: financial connoisseurs in the world are convinced that this new asset class
10:44
Ejaaz: is a very real thing, which means that they've probably looked at the balance sheets.
10:48
Ejaaz: They've probably looked at the revenue demand that a lot of these frontier labs
10:50
Ejaaz: that are meant to be purchasing these things are going to do.
10:53
Ejaaz: And they're looking at it and they're saying, this is an obvious no-brainer.
10:56
Ejaaz: Now, if you're listening to me and you're thinking, dude, this happened with
10:59
Ejaaz: railroads and it didn't work out.
11:01
Ejaaz: This happened with the housing environment, mortgage-backed debt securities
11:05
Ejaaz: in 2008, that didn't work out.
11:07
Ejaaz: I have to say, it's a very different story on our end.
11:10
Josh: Well, that's what I was going to ask you. I was going to say like,
11:13
Josh: hey, obviously they're going to come out and say these things.
11:15
Josh: I mean, we've seen them manipulate markets for a long time. We just saw what
11:18
Josh: Citadel did to Leopold. It's like everyone is very clearly out in their own
11:22
Josh: best interest. So if we look at this deal, okay, they're not going to say it's
11:25
Josh: anything less than exceptional.
11:27
Josh: So how do we kind of vet this? How do we fit this into a specific piece of context
11:32
Josh: that I guess we could reference?
11:32
Josh: And there's an interesting example of like aircraft finance versus mortgage
11:37
Josh: finance, because this is something that has happened in the past where
11:41
Josh: when you have an expensive standardized asset that's transferable between operators
11:45
Josh: and has a lot of demand for these secondary markets, it creates this interesting marketplace that
11:51
Josh: I think is much more comparable to aircrafts than mortgages.
11:54
Josh: And I'll explain. So like a Boeing 747 or 737 or whatever, that's been built
11:59
Josh: like 20 years ago. And trust me, you've flown on these. The airlines kind of
12:02
Josh: suck. You're flying in some old planes.
12:04
Josh: That is just as valuable today
12:06
Josh: As it was 20 years ago, because they're able to derive so much value from it.
12:11
Josh: It does the same exact job.
12:13
Josh: The same plane that was built today is doing the same job that was built 20
12:16
Josh: years ago. And sure, perhaps you would prefer to fly on the newer plane.
12:20
Josh: But the reality is, is that tickets are sold out on the 2005 plane and the 2025 plane.
12:25
Josh: And when you think of GPUs, they exhibit a lot of the same traits and characteristics
12:30
Josh: as an airplane, where it's expensive, standardized, it's transferable.
12:33
Josh: It has a lot of liquidity in secondary markets.
12:35
Josh: And I think this is an interesting way of looking at it relative to mortgage
12:39
Josh: finances, which is where we got in trouble. And this isn't the first time this has happened before.
12:43
Josh: There is something that has been done similar to this with Broadcom,
12:46
Josh: where like that Google Anthropic structure actually runs through this thing
12:50
Josh: called an SPV, a special purpose vehicle that buys TPUs and leases them with
12:54
Josh: Broadcom providing the residual value guarantees,
12:57
Josh: and then Apollo and Blackstone supplying the private credit to fund all of this.
13:00
Josh: So people have experimented with these structures, before. It has worked.
13:05
Josh: We haven't seen it at this scale. I mean, the alarm bells are partially ringing.
13:09
Josh: I'm like, just out of instinct, like out of an e-joke reaction,
13:12
Josh: like, oh, wow, this is a lot of money. This is a lot of powerful people who
13:16
Josh: can very much control and sway the way the market moves.
13:19
Josh: So far, it seems like a pretty reasonable thing. It's like, hey, we need GPUs.
13:23
Josh: GPUs are transferable. They're kind of like they're fungible,
13:28
Josh: I guess. I'm like thinking of the word. I'm like, well, this feels kind of crypto adjacent.
13:31
Josh: There's like these fungible assets that can be transferred that are valuable.
13:34
Josh: So I don't know. There's a chance this goes over. OK.
13:37
Ejaaz: And it's important to not extrapolate too far into the future,
13:41
Ejaaz: like what we can feasibly attain from the data, which, by the way,
13:46
Ejaaz: is publicly available. If you're listening to this and you don't believe anything that we're saying.
13:50
Ejaaz: Maybe we should actually link to a bunch of sources. Maybe we'll link this artifact
13:52
Ejaaz: that you're seeing on the screen right now.
13:54
Ejaaz: The data is all available through quarterly earnings of every single company
13:57
Ejaaz: that is leading at every layer of the AI stack. So you can see the data,
14:02
Ejaaz: digest it yourself and figure it out for yourself.
14:04
Ejaaz: But what I will say is when you look at the 2008 financial crisis,
14:08
Ejaaz: when you look at the railroad crisis, when you look at the telecom crisis back
14:12
Ejaaz: then there was a huge amount of oversupply.
14:16
Ejaaz: Which didn't have the back demand that it stated it had so 2008 people assumed
14:22
Ejaaz: that property prices would just keep going up and at some point
14:26
Ejaaz: that got two head over heels if you look at the railroad they built too much
14:29
Ejaaz: if you look at the telecom they built too many cables right in this case
14:34
Ejaaz: we're constrained by a few things number one physically it takes so many different
14:38
Ejaaz: substrate layers to build a gpu.
14:41
Ejaaz: Every single layer right now in the world of physical atoms is incredibly constrained.
14:45
Ejaaz: There's not enough. The GPU demand is overweight, the actual supply that is available.
14:51
Ejaaz: Number two, the AI demand, which is driving GPU demand, is accelerating much
14:56
Ejaaz: faster than the supply itself can.
14:59
Ejaaz: So if you look at memory as a basis for this, they can increase capacity.
15:02
Ejaaz: This is the top docs, top memory manufacturers can increase capacity around 20% per year.
15:08
Ejaaz: But demand is compounding at 45 to 60% per year.
15:12
Ejaaz: So if you do the math, if that continues, you're going to be in a constrained
15:16
Ejaaz: supply state for like at least until 2028 or 2029 until some of these other
15:21
Ejaaz: chip fabs can get increased. So we're still kind of in a holding period, right?
15:25
Ejaaz: Now, if you look at the backlog for some of these companies.
15:29
Ejaaz: You might be like, well, customers don't want AI as much as these guys are making
15:33
Ejaaz: it out to be. They're just kind of like pushing their bags.
15:35
Ejaaz: Well, look at Google's backlog. It doubled this year in a matter of months.
15:39
Ejaaz: It is now at $460 billion, and we're at the halfway mark of this year.
15:43
Ejaaz: It's probably going to increase even more. This is the case across Microsoft
15:46
Ejaaz: and a bunch of other hyperscalers as well.
15:48
Ejaaz: Then if you look at the GPU rental prices, my favorite thing,
15:51
Ejaaz: which kind of came out this morning, Josh, or maybe yesterday,
15:55
Ejaaz: CallWeave had their earnings, and they said, we recently signed an A100 contract that extends into 2029.
16:00
Ejaaz: For those of you who don't know, and A100 is an NVIDIA GPU that was created in 2020.
16:05
Ejaaz: And its lifecycle prediction back then was three and a half years.
16:09
Ejaaz: Now they're predicting that it's actually going to be good to go until 2029.
16:13
Ejaaz: That's because it's not just being used for bleeding edge training.
16:15
Ejaaz: It's being used for inference and a bunch of other stuff. So the point is,
16:18
Ejaaz: these GPUs are very versatile, NVIDIA specifically, and that's why they raised a crap ton of money.
16:22
Josh: Yeah, and more valuable over time. It's this really bizarre thing in which the
16:27
Josh: useful lifecycle of a hardware object is increasing instead of depreciating for the first time.
16:33
Josh: And we've never really seen this phenomenon at scale before.
16:36
Josh: But like you mentioned, there's not many reasons in which it's going to slow
16:39
Josh: down in the near term future. When I look at this, I'm kind of looking at it
16:43
Josh: like around the corner. And then you can't really see around the next corner,
16:47
Josh: but we have an idea of the first corner. And that first corner is sold out supply
16:51
Josh: for at least 2027, likely 2028.
16:55
Josh: And then by the end of the decade, 2029, 2020, 2030, we start to run into larger
17:01
Josh: constraints, mostly around energy and power.
17:04
Josh: And we start to like run into resource constraints that we don't quite have now.
17:08
Ejaaz: So are you basically saying there's like multiple corners, Josh?
17:11
Ejaaz: Like I'm curious, like, for the GPU specifically, do you think it's like six
17:15
Ejaaz: months? Do you think it's also like 12 months? Like, what's your guess if you had to...
17:19
Josh: Well, there's a somewhat clear trajectory for the next 24 months,
17:23
Josh: like 18 to 24 months in terms of...
17:26
Josh: It seems fairly predictable where we know how many, like the lithography machines,
17:30
Josh: we know how many chips they can create.
17:32
Josh: Then we know how many of those chips can be packaged into usable chips.
17:36
Josh: And then we know roughly how many data centers can be built that can actually
17:40
Josh: turn those chips on and power them.
17:41
Josh: And you can somewhat project that out up to 24 months loosely,
17:45
Josh: very loosely, because there is only so much throughput for these machines.
17:50
Josh: So if you assume everyone's operating at full capacity, you can kind of work
17:53
Josh: those numbers backwards and understand like, OK, they're sold out and they're
17:56
Josh: still not going to be enough to satisfy the demand, assuming the demand continues,
18:00
Josh: which there is no signs of slowing down.
18:02
Josh: All these use cases for AI, particularly around agentic AI, require a tremendous
18:07
Josh: amount of inference and even including efficiency upgrades to the software,
18:10
Josh: something similar to what we imagine SSI is working on.
18:13
Josh: There's still a huge amount of demand that will fulfill that this is jevin's
18:16
Josh: paradox which we really should name jensen's paradox because that seems to be
18:20
Josh: a little more accurate in terms of how this is working
18:22
Josh: but we can kind of project out till then and we know all right gpus fully sold
18:26
Josh: out fully constrained after that things get a little
18:30
Josh: more hairy right it's because you have to assume by that time we'll have something
18:33
Josh: similar to agi asi self-recursive improvements we should be getting a lot of
18:38
Josh: innovation breakthroughs around efficiency and software and we don't really know what
18:42
Josh: the power market's going to look like. We're not sure if we're able to make enough
18:46
Josh: Energy to satisfy the demand of the gpu centers that are being projected out
18:50
Josh: into 20 to 30 so it seems like this is a very long duration thing that's going
18:56
Josh: to need to play out but in the short term in the intermediary term it seems like i mean
19:01
Josh: i'd like to find the steel man against this because this seems important and
19:04
Josh: we should talk about like what are the possible ways in which this breaks
19:07
Josh: but just looking at demand of inference and our capability of serving inference
19:12
Josh: and there is a huge mismatch in the case of inference demand that's just not
19:16
Josh: going to be met for a really long time.
19:17
Josh: So those H100s from 2020, or the A100s, I should say, from 2020,
19:21
Josh: are still going to be useful in 2027, 2028.
19:25
Josh: And that is particularly valuable when you're investing in GPUs at this scale.
19:29
Ejaaz: Well, actually, now that you say it, a lot of that inference demand,
19:32
Ejaaz: at least in the next six months or so, is going to come from AI agents.
19:37
Ejaaz: I'm in no doubt about that.
19:39
Josh: You said the word agents. It's funny you should mention agents because we have
19:43
Josh: something to say about agents from our sponsor of this episode, Ledger.
19:47
Josh: If you're building with AI agents, you are probably worried about security and
19:51
Josh: rightfully so because as we've seen recently, these agents have been kind of
19:54
Josh: doing some funky things.
19:55
Josh: So Ledger has this three-step approach to solving this. The first is that the
19:58
Josh: agent proposes a change, then the human approves the change,
20:01
Josh: and then the Ledger signer actually enforces this change.
20:04
Josh: There's a three-step process to make sure your agent doesn't do anything you
20:08
Josh: do not want it to do they have this thing called the ledger agent stack which
20:11
Josh: is an open source software stack that gives you a series of tools to help you navigate
20:15
Josh: your journey with your agents it works with cloud code it works with codex cursor anywhere that you use
20:20
Josh: your ai models it's available now it is open source you can find the link in
20:23
Josh: the description down below and thank you so much to ledger for sponsoring this episode the agent
20:28
Josh: barrage? I guess we have to talk about the downside effects,
20:32
Josh: right? What could happen if things don't go as well as planned?
20:37
Josh: And where does that risk actually live? So, Ijaz, it seems like this has been
20:41
Josh: fully prepared, lovely by our clawed artifact right here. So,
20:44
Josh: what is actually the downside risk? What do we need to look out for when we're
20:48
Josh: evaluating how to invest around this?
20:49
Ejaaz: Okay. So, we are the Limitless Show. And as anyone who's listened to us for
20:53
Ejaaz: a while knows, we are grounded or we are trying to ground ourselves a lot more from the bullish case.
20:58
Josh: I mean, if you got to say it. Yeah.
20:59
Ejaaz: So there are a few ways where this can obviously go wrong.
21:04
Ejaaz: And I want to kind of like walk through some of these and get your take on this,
21:07
Ejaaz: Josh. So number one, the thing that's like blaring to me is this is all based
21:13
Ejaaz: on the fact that AI demand not only is sustained.
21:17
Ejaaz: So you have paying customers to buy cloud subscriptions, GPT subscriptions,
21:21
Ejaaz: companies paying tens to hundreds of millions of dollars a year for API access. But...
21:26
Ejaaz: That it increases. Right now, it's increasing at a crazy rate.
21:29
Ejaaz: We see all these quarterly earnings, revenues compounded between 100% to 500% year upon year.
21:34
Ejaaz: It is insane, but that's not sustainable. It's not going to keep doing that.
21:37
Ejaaz: It'll presumably eventually plateau.
21:39
Ejaaz: So if that does plateau, or in worst case, if that plummets,
21:43
Ejaaz: then these companies are going to need fewer GPUs, which means that Jensen's
21:48
Ejaaz: $500 billion, these debt-backed securities are going to be in less demand.
21:52
Ejaaz: And that's where you might see a default.
21:54
Ejaaz: The second major thing here is that the GPUs themselves depreciate a lot faster.
22:01
Ejaaz: And that has been the Michael Burry, the guy that did the famous big short back in 2008.
22:06
Ejaaz: That's been his view this entire time. He says that the upgrade cycle for a
22:10
Ejaaz: lot of these NVIDIA GPUs are actually a lot shorter than what Jensen NVIDIA claims.
22:16
Ejaaz: However, in practicality, this seems to not be the case.
22:19
Ejaaz: However, NVIDIA is now releasing a lot of GPUs at a much more higher frequency
22:25
Ejaaz: rate, which means that they're going to replace more of the GPUs in the prior
22:29
Ejaaz: market, in the prior cycle, and they'll start flooding the market.
22:32
Ejaaz: My countess of that is simply you can't make GPUs that quickly.
22:36
Ejaaz: It takes a lot of technical expertise, and it is limited by the likes of TSMC
22:40
Ejaaz: and wafer capacity and a bunch of other technical stuff, which I don't want to get into on this show.
22:44
Ejaaz: So I'm struggling to actually see how these two factors might actually be triggered.
22:48
Ejaaz: But I don't know if you have a different opinion, Josh.
22:50
Josh: Yeah, the thing that I'm looking out for most is the return on invested capital
22:54
Josh: from the large hyperscalers.
22:56
Josh: It feels like they just run the world. They're spending all the capex.
22:58
Josh: They are basically floating the entire economy right now.
23:01
Josh: And if the returns on that investment start to go down, for example,
23:05
Josh: that seems like a very scary thing. So looking at Google's earnings reports,
23:08
Josh: we see like, okay, They have $514 billion dollars.
23:15
Josh: They're spending. Can they keep returning revenue on that on schedule?
23:19
Josh: If the answer is yes, if there's still revenue to be made on AI spend, that is amazing.
23:24
Josh: In the case that that turns and we start seeing earnings reports from companies
23:27
Josh: who are spending huge amounts of capex saying, our margins are actually shrinking.
23:31
Josh: Our revenue is not coming at the multiple that we expected.
23:35
Josh: That seems to be a red flag because that will slow down spending significantly across the board.
23:38
Josh: Basically, we want to make sure that all this stays profitable.
23:40
Josh: So we want to make sure that Inference demand is actually continuing.
23:43
Josh: Companies are actually able to meaningfully monetize this.
23:46
Josh: The enterprises that are spending billions, hundreds of billions of dollars
23:49
Josh: a year on AI spend, we need to make sure that they're actually getting value.
23:52
Josh: Otherwise, they're going to cut those contracts. That is probably the most important thing.
23:56
Josh: The second is just monitoring the rental rates. Like currently,
23:59
Josh: a lot of companies are terrified to re-sign their long-term GP rental deals
24:02
Josh: because the price that they're going to get them at this time around is going
24:05
Josh: to be double the price that they got originally.
24:07
Josh: That is a phenomenon that like no one was really expecting, but here we are.
24:11
Josh: And if that trend continues as well that's something i'm kind of looking out
24:15
Josh: for so i'm looking at what is the hour double
24:17
Ejaaz: Like three years from now right i mean it.
24:19
Josh: Might it might but i'm saying this this is just something that like you should
24:23
Josh: keep a close eye on you know how long these
24:24
Ejaaz: Contracts are for josh that they're signing i know.
24:26
Josh: They vary quite a bit like some are short they're like a year some are longer
24:31
Josh: out to like three years maybe um but i they're they're variable and i know that
24:36
Josh: when the time has is coming to kind of re-sign this contract
24:40
Josh: the price is higher not lower for the same supply so ensuring that this
24:45
Josh: continues this trend continues and even if it doesn't continue making sure it
24:48
Josh: doesn't flip negative because that might change things and granted nvidia has
24:52
Josh: your 25 plunge protection service you're available but you don't really want to put that 125
24:58
Ejaaz: Billion dollars by the way for those of you trying to do the math.
25:01
Josh: That's a lot of money man a lot of money um and then third is just like
25:05
Josh: what the yields actually is from these gpus like how much the real yield yeah
25:10
Josh: the actual real yield and
25:11
Josh: these are also like those mous this is an assigned deal and we've had something
25:15
Josh: similar to this before remember that crazy project back in the day called project stargate where elon and
25:21
Josh: masa or not sorry not elon sam altman and masayoshi son and even donald trump
25:27
Josh: all stood in an office together and they said
25:29
Josh: we're going to spend x billion dollars on this data center build out
25:32
Josh: it hasn't really happened as planned so this is not a contractual obligation
25:36
Josh: to spend 500 billion dollars this is a hey dude
25:40
Josh: We're all rich. We'll commit to like $500 billion and we'll see how it goes.
25:43
Josh: And that's kind of what they have. It's a handshake deal to build this new financial
25:47
Josh: economic instrument around the GPU, particularly as it relates to NVIDIA.
25:51
Josh: So huge win for NVIDIA, probably a large win for a lot of the companies that
25:56
Josh: are not able to afford this, and probably a huge win for the banks.
25:59
Josh: At the end of the day, they seem to always win. And that's kind of what the deal is here.
26:04
Ejaaz: I am really struggling to think about a world, an alternative scenario,
26:10
Ejaaz: where AI doesn't require GPUs, specifically the monopolistic GPUs from NVIDIA.
26:19
Ejaaz: They just have such a stronghold on the entire market.
26:23
Ejaaz: And even if you have some kind of novel LLM architecture that gets created in
26:28
Ejaaz: the future that completely disrupts the current paradigm, you're still going
26:31
Ejaaz: to need hardware to run these things.
26:32
Ejaaz: And that hardware is very much GPUs that are being designed and created by Jensen Huang.
26:37
Ejaaz: So however way I skin this cat, I still think that you're going to need these GPUs.
26:42
Ejaaz: You still need token generation. Gavin Baker has made this point across so many
26:46
Ejaaz: other podcast episodes in the last two weeks that it's ingrained in my head at this point, right?
26:49
Ejaaz: And then the other thing I think about is, okay, well, if NVIDIA becomes a bank
26:54
Ejaaz: themselves and they start taking revenue splits from all these frontier AI labs,
26:59
Ejaaz: that's a completely new revenue line for NVIDIA.
27:01
Ejaaz: So when I think about this with my investing hat on, I'm thinking.
27:05
Ejaaz: Okay, not only is NVIDIA supplying the foundational element that is required
27:10
Ejaaz: to run and inference these GPUs, train these GPUs, but they're also
27:14
Ejaaz: being the ones that are driving cost down per token, right? So like they're
27:19
Ejaaz: doing this with their CUDA software mode.
27:20
Ejaaz: And then I think about the financing side of things. So they're being the financiers
27:24
Ejaaz: of this entire thing as well.
27:26
Ejaaz: Now, that does sound like a house of cards if the demand wavers, if the demand plummets.
27:31
Ejaaz: And I can easily see the market being very volatile and reacting to any kind
27:34
Ejaaz: of headline like they did to this initial headline but i don't know it just
27:37
Ejaaz: seems very bullish to me on nvidia at least and yeah i don't really know how
27:41
Ejaaz: to think about it yeah yeah.
27:42
Josh: And i mean like in this case like i do kind of lean on the opinions of people
27:46
Josh: who are more in the know than me yeah someone like elon who is now exclusively
27:50
Josh: committed to purchasing only nvidia gpus for the new data center build out
27:54
Josh: and they are effectively the best data center builders in the world so i you
27:58
Josh: have to like have a little bit of trust in the opinions of the
28:00
Josh: true experts who are in the arena doing things when i look at that and i see
28:04
Josh: like they exclusively want nvidia and they are building the best fastest most
28:07
Josh: efficient data centers i'm like okay that's pretty good signal like micro hard
28:11
Josh: the new data center that the spacex ai team is working on
28:14
Josh: is i think like a third the footprint of macro
28:17
Ejaaz: Hard right no micro hard.
28:18
Josh: No no there is there's micro hard is
28:20
Ejaaz: A micro hot.
28:21
Josh: Yes and micro hard is about a third of the footprint i might be getting this
28:24
Josh: wrong half or a third of the footprint of macro hard easy but it contains the
28:27
Josh: same cluster of 200 000 gpus wow they just figured out how to do it much more
28:31
Josh: efficiently and much more dense so
28:34
Josh: there's a huge amount of innovation clearly they know things that the rest of
28:37
Josh: the industry does not and when they come out the biggest purchaser right
28:40
Josh: yeah and they're committed exclusively to the nvidia gpu wow and when vera rubin
28:46
Josh: comes out at scale man oh my god i keep saying this for like holy smokes
28:49
Ejaaz: Those models are gonna be insane those models are gonna be absolutely insane.
28:52
Josh: So buckle up good time to be good time to be nvidia good time to be a GPU.
28:58
Josh: Yeah, that's the update. So is this a house of cards?
29:02
Josh: Is it all going to come tumbling down? Is this financial innovation in a new
29:06
Josh: era of the United States of GPUs?
29:09
Josh: Let us know in the comments down below if you enjoyed this episode.
29:11
Josh: Don't forget to share it with a friend who might also enjoy this episode.
29:14
Josh: Ejaz and I, cool story. We were walking down the street last night after dinner
29:17
Josh: and two people walked up to us.
29:19
Josh: They were like, hey, you guys are those podcast guys. You host the show.
29:21
Josh: And we were like, yeah, we do. Like, cool. Thanks for watching.
29:24
Josh: So that's like so cool whenever that happens. And that is because of you sharing
29:28
Josh: with your friends, letting everyone else know that the show exists.
29:31
Josh: And if you enjoyed it, you know, don't forget. Thumbs up. You could subscribe.
29:34
Josh: You could give us a five-star review on your favorite podcast platform.
29:37
Josh: Any parting thoughts, EJs?
29:38
Ejaaz: Yeah. Homework for you guys. If you see us on the street, don't be shy.
29:41
Ejaaz: Come and say hi. We want to meet you guys.
29:43
Josh: Don't be shy. Say hi.
29:44
Ejaaz: Yeah. It's lovely to meet you guys. We've now met people. We've now met listeners,
29:49
Ejaaz: in New York and we've met them in San Francisco.
29:53
Ejaaz: I have people calling in from Europe to my family talking about these random
29:56
Ejaaz: guys that yap about AI. Turns out it's us. Like it's really cool to see the
30:02
Ejaaz: momentum that we're getting here.
30:03
Ejaaz: And it's all thanks to you guys. So if you're one of these people that care
30:07
Ejaaz: passionately about what we talk about and tune in every day, thank you so much.
30:12
Ejaaz: And if you're not, turn on notifications, please. Subscribe to us.
30:15
Ejaaz: We bring the best news, hot, fresh out the oven, every single day,
30:19
Ejaaz: or rather four times a week.
30:21
Ejaaz: And we would love to hear from you. Leave us a comment, DM us on X.
30:24
Ejaaz: And yeah, I think that's it. Thank you so much for listening.
30:27
Josh: See you next time.