0:04
Hello everyone.
Search this video
0:04
Welcome back to Semi analysis weekly.
0:06
We're here with episode number 21.
0:08
I've got the crew from Singapore to talk about everything Nvidia backs stops including the holy trinity capital offtake and data centers.
0:16
Starting up with Dan chewing away on his lunch. How's it going, man? >> That was dessert. That was apples.
0:21
I'm eating healthy dessert. >> Good, man. Okay.
0:24
With uh with Dan, we've got Kungwen. How you doing? I'm doing great. Doing great. >> Okay.
0:32
And Zayn, how are you, man? >> Yep. I'm doing well. Thank you for asking. >> Okay. Awesome.
0:40
So, in this uh episode, we're going to talk about Nvidia back stops.
0:44
We'll start by discussing the problem statement, explain what a backs stop is, how it's structured, how people are pricing GPU loans, as well as other things to back stop these data center buildouts.
0:54
go through a couple of examples in the Asia-Pacific region, some of the tools that lenders are going to need in the future, and finally the impact on Nvidia's financials.
1:02
So, no better team to dig into this than our our team that works on Cloud TCO every day, including Mr. Dan Nishbal.
1:09
Um, Dan, can you start us off with the problem statement?
1:14
What is that Trinity capital offtake data centers actually mean?
1:16
And why are these back stops so important right now?
1:23
Yeah, let me um let's start off with um you know how we look at capex, right?
1:27
So if you look at AI and data center capex um it's soon going to reach um you know very large levels, right?
1:36
So commumulatively about 11 trillion from 2024 to 2029 um and we're pretty soon going to cross uh 1 trillion of annual capex, right?
1:45
So the debt's already been building up um you know hundreds of billions per year um soon going to reach um over a trillion per year.
1:55
trillion per year. Um so the question is okay I think we well understood we well understand the amount of capex coming through and I think markets are starting to understand given how much debt hyperscalers are issuing um there's a
2:08
bit of angst over whether they continue um but you know the point is um the demand is there right and the demand and the production is there and um everyone's gearing up for 11 trillion of capex and the question is how do we fund it um so how much funding will be needed is about 7.1 trillion. And the way we 1 trillion.
2:26
And the way we look at that is we look at about 75% debt financing.
2:32
And you can think of that as advertising.
2:34
So it's going to roll off.
2:35
So you have a bunch of stacks of debt that that then roll off.
2:37
And typically it's about 5 to 6 years.
2:39
It'll match the contract duration or the expected lifetime um you know of of of the GPU.
2:47
And of course, you know, all these are are are very debt financed.
2:48
Um you know, makes sense.
2:51
It's very capital intensive.
2:52
It should be debt financed.
2:54
Um and um we'll we'll sort of go through some examples later of what the pricing could look like.
2:59
But you know this really is a problem statement and you know the the the other important problem statement to understand is how big this market is going to be in relation to all other markets.
3:10
So here what we're showing is AI debt financing compared to all other markets all other US asset back markets.
3:17
So AI debt financing including data center and GPU IT capex again what we said is going to reach 7 trillion uh by 2029.
3:24
Auto loans, student loans, these are all in like the single digit you know low like 1 or two trillion.
3:31
Um the only market that's bigger is the US mortgage market at 13 trillion.
3:34
So the point is um there's going to have to be a lot more capacity quickly and there's going to have to be a lot more understanding.
3:42
a lot more understanding. Um so the one of the biggest problems so far and we'll we'll sort of talk through the objective of like why this exists right is um you know what can get financed right and what can get financed today is
3:58
mostly if you have a neocloud that has a 5-year offtake with an IG hyperscaler that can be financed um anything else it's very very hard to finance um so you know I'll take a step back and just introduce what we call the the AI project trendy Um that's capital offtake and data center. So any neocloud that has to
4:15
So any neocloud that has to build a project has to figure out how to solve these three things.
4:19
So capital um in order like what I just said in order to get capital and to get lending you need to have a 5-year offtake from an IG hyperscaler.
4:28
But to win that from a hyperscaler a 5-year IG, you first have to demonstrate you know you're going concern um you can find equity to place those deposits and you can definitely secure data center but on the other hand to get data center you got to convince a data center operator to want to rent from you or you build yourself.
4:48
from you or you build yourself. So you know these are these are three legs and um the the easiest solution to this has been um let me find a hypers scale offtaker 5 years get the capital you know I have a small pilot project to sort of convince everyone um so Jordan
5:06
yeah I think that's the that's a quick summary any questions so far >> so Dan the thing that jumped out to me is that um basically this has changed like there's been a significant change where um off the bat at the beginning uh when we saw a bunch of these NEOClouds getting started, they could raise equity. And that was just because
5:24
And that was just because to me the scale allowed them to raise 10 million to hund00 million maybe to get their first data center off the ground.
5:35
But this is like enterprise investors, totally different return profile that they expect and just like a totally different amount of capital you need to raise when these projects are measured in the billions.
5:47
So that's like an obvious fundamental thing that I think has changed and why you now need investment grade hyperscalers signing five-year contracts in order to even be able to raise the money because people don't want to take that amount of risk on, you know, things measured in the billions.
6:02
But maybe are are there other things other than the sheer scale that you're seeing change as people start trying to fund all of these projects?
6:13
>> Well, I think it's it's it's not necessarily the scale, it's it's the volume, right?
6:16
So it's not just um can I do a 10 billion, a 5 billion, um can I do a 100 megawatt, 200 megawatt, it's more can I do 20 of these.
6:24
Um so when this first started um there were a couple private equity pioneers, right?
6:31
Very famously um Blackstone and Cororeweave.
6:33
Um and at the time it was actually like a 5-year backs stop.
6:36
Um but at the time, you know, it was really only these private equity guys um that were willing to to look at that and um eventually their conclusion was was effectively that okay well it's um you know Microsoft risk or whichever risk um and so um yeah the the the loan was a pretty good rate I think um but it was effectively if the execution went well then it effectively turned into hyperscalic risk. Right.
7:07
Um now I think the the the market has matured and many more lenders are willing to do this which has driven the rates down.
7:14
Um but then to answer your question, what is the limiting factor?
7:18
It is the the sheer quantity cuz even if some lenders are willing to do this and it's more than just private equity, it's actually um you know it's actually getting a lot of lenders to do this even if it's the 5-year backs stop. >> Makes sense. Yeah.
7:35
So we've described back stops at a high level.
7:38
Zayn, do you mind taking us through how these back stops are actually structured?
7:48
>> Well, before that actually before we go to that um I think um let me talk about um you know the the the like other problems, right?
7:55
So like you know I talked about like why um you know I talked about where the market is today, right?
8:01
market is today, right? Today you can actually um only in large quantities finance a 5year backs stop IG right but what what are the problems to that right you know and um what we said a moment ago is to get to 11 trillion of cumulative capex right we can't just
8:17
that that's not going to happen just through a bunch of hypers scale back stops right so that that's the first problem with this market structure is that hypers scale back stops are not infinite so they're not going to be able to absorb 2 3 4 5 trillion and um you know, like not everything can be backed up by hyperscalers. Um, secondly, it's
8:33
Um, secondly, it's what I mentioned you a moment ago, Jordan, that um, you know, private equity and um, private credit have led the charge, but as spreads compress over time, the returns going to be too low for them.
8:46
The capital needs will increase, it's going to need a broader set of capital providers.
8:49
Um, the other problem is is it's it's really like a one-sided market.
8:56
It's it's really only supplying compute to 5-year octakers and no one else.
9:00
So inference providers can't get the GPUs they want.
9:02
Um you know Jordan you could probably talk much better about this than me but there's this total lack of variety like one year two year people can't like you know just just get what they need. Yeah.
9:14
>> It's all the startups.
9:14
I mean, everybody um who's trying to build who's who's left one of the frontier labs and they're trying to build a new model in a new modality like let's say they're working on material science or they're working on drug discovery or they're working on video generation or robotics or world models or anything that the Frontier Labs are not like exclusively focused on.
9:31
They raise money from VCs and the promise that they're going to go rent a ton of GPUs, probably spend 80% or more of the money that they've raised on GPUs.
9:41
and they really want the biggest cluster they can get for about 6 months to a year.
9:46
But the answer that they're getting from Neoclouds right now is that your $80 million isn't good if you can only commit to a year.
9:54
Instead, give me 15 to 20 per year over the next five and the size of the cluster that you expected you were going to get has been divided by four or five and that's how we can actually get a deal done.
10:09
That wasn't really the promise of Neoclouds up front because at that point, you know, you're waiting 6 months and you're spending all this extra money and you're not getting a big enough cluster.
10:19
It's not really a venture style approach where you just like take a big shot, try to train a great model, and then raise an extra round of funding and go even bigger.
10:28
You're basically ending up in a world where you have to run a, you know, long-term sustaining business, which is again not what you raise money for.
10:37
raise money for. or so it looks very different than the way it did 3 4 years ago when people were starting these companies and looking for compute um because really these neoclouds cannot I think it's due to the scale that they're trying to operate at now there's just so many of these companies like you said
10:55
Dan um but nobody's willing to take the risk of speculatively buy buying a billion dollars of GPUs and then hoping that they can that everybody gets another round of funding in a year or two Um, however, obviously Nvidia has recognized that this is a need in the industry and they're going to come in and do something about it. So, maybe you guys can take us through
11:18
So, maybe you guys can take us through what that actually what >> the back stop, right?
11:22
Enter the back stop, right?
11:24
And this is >> uh, you know, Zayn will take us through what exactly it is and how it's going to solve that problem.
11:29
Zay, you want to explain? >> Yeah, for sure.
11:32
Um so this this um the way that this back stop is structured um typically we see it being structured at like a five six year duration.
11:40
Here we have it at 6 years for a GP300.
11:43
And what what Nvidia does is basically they are ready to purchase at these pre-agreed levels um from the NeoCloud and then anything above the Neocloud.
11:55
So, so let's say let's say on a year one basis, right?
12:00
Uh 368 if the NeoCloud's able to go ahead and rent it out for something like $6 instead, then um everything like 368 that goes straight to the NeoCloud.
12:12
Um and anything above the Neo cloud like that is split between the NeoCloud and Nvidia.
12:17
And um the reason why we have this at a declining like RAM profile is because of the naturally expected DK and GPU rental prices, right?
12:25
GPU rental prices, right? Um but you'll notice the average back stop price here we have is only 236 which is um very very low for a GB300 and we'll delve a bit into the rationale for that later on but uh for now suffice it to say that um
12:42
Nvidia is not doing this to to guarantee this new cloud to really high rate of return right um they're more of just doing this to make the new cloud um to make all these clusterable clusters like financable in the first So that's the overall structure of backstar. Uh and with with device a few
12:59
Uh and with with device a few scenarios in which this back could be used.
13:04
So the first scenario here is we're running a short-term rental book um for GB300s and obviously that's going to get you pretty high like rental rates here.
13:15
Um so we've we've plotted that out in a six-year average price of 427 starting from 675 in first year.
13:21
Uh so once again the customer will pay the new cloud 675.
13:26
Um the Nvidia back stop price here for the first year will be 368.
13:29
Um and then so at 368 the new cloud takes and then anything above 368 we would assume a certain revenue share between Nvidia cloud.
13:42
So we've synthesized that um in the various scenarios here.
13:45
Um but essentially it's it's pretty intuitive as you would expect.
13:50
Um if we just jump over to the right here on the IR section of course um the more of that percentage revenue share goes to Nvidia then uh the lower the IRR for the NeoCloud.
14:02
Uh but essentially that's um this is what we see the uh the back stoping to do.
14:08
It's it enables the NeoCloud to take a bit more like uh speculative like risk renting to a broader bunch of customers like we just mentioned.
14:20
like we just mentioned. uh renting on a shorter year rental book [snorts] and uh in this case the new cloud doesn't necessarily have to go out and find like a five six year investment grade hypers scale offtaker and is able to rent to you um a broader bunch of consumer bases right >> um so go ahead Jordan
14:40
>> yeah I'm curious obviously Nvidia is kind of double dipping here right because and in one sense it's the back stop but they're going to make money off of that loan and on the other side there's the potential to make upside for them when they can actually go and connect buyers and sell above the back stop price. So maybe can you talk about
14:58
So maybe can you talk about the other side in terms of how the loans are getting priced I guess.
15:08
Well, I think one thing might be interesting like a small detour, right?
15:12
Is if you go through the DSCR analysis because I think the the important thing is like um you know why does the structure uh solve for banks, right?
15:21
Like like why does it why is it solved for them being able to lend um without an offtaker, right?
15:26
So this is how they analyze it.
15:29
Zay, if you want to explain like how banks are able to take this back stop and lend against it. >> Yeah. Yeah, sure. No problem.
15:35
Um so we've run a DSCR analysis of the same of actually just the back stop price itself.
15:44
Um and what we've been hearing from banks is that in order to finance like uh this kind of an arrangement they would need the DSCR ratio to be above 1.
15:54
3 at least for the first couple of years.
15:56
U so here with um we've put in our back stop price and and just on a cluster of 100 megawatt for GB300.
16:03
Um, of course this is illustrative.
16:06
Then you get uh the following DSCR ratios which are quite healthy on the first few years.
16:13
So in other words that it's not sufficient like this Nvidia back stop price is not sufficient for the Neo cloud to earn a very high rate of return on the GPU rentals but it is sufficient to give lenders the kind of comfort that they need uh to be able to finance projects like these.
16:32
Um, so that's where Nvidia aims to land with this kind of a backs stop pricing.
16:40
>> So effectively they're landing landing against the back stop and and what they're thinking about is hey worst case um Nvidia will take the compute.
16:46
um Nvidia will take the compute. So they're thinking in the worst case what is the cash flow I'm going to get and then um if that's my worst case I can lend against that and then if the Neocloud wants to go and um you know not avail the back stop u because again
17:01
right they don't have to right the whole idea by the way is that they don't ever use the back stop like no one actually wants to use the backs stop everyone wants to be above the back stop um neoclouds want to do one year two-year business um but for the banks they're
17:14
like okay what is my worst case scenario and that's how they that's how they um uh lend >> yeah let's say Nvidia you know needed to use a backs stop what kind of you know what what would they even use the GPUs for you know are they just paying money just for the sake of backs stopping this
17:34
GPUs >> plenty of research I mean you've seen the Neotron models come out from Nvidia they do research on all sorts of different smaller things and they've got a massive CI fleet to make sure that anybody's code in PyTorch, VLM, SG lang is actually going to run well out of the box. I mean,
17:54
I mean, in part, any GPUs that Nvidia produces that they rent for themselves and use is going to improve the user experience for other people using those GPUs.
18:10
Um and in in some ways it contributes to the CUDA moat let's say when they have GPUs for themselves to do research and engineering against now there's some limit right um they don't need hundreds of megawws to do research [laughter]
18:27
or um you know they maybe they do they have plenty of ideas but what we've seen from them is that they have then taken those GPUs possibly and started to form consortiums namely the Neotron training collective where you bring together a whole bunch of
18:43
different startups and organizations like thinking machines and mistrol and things like that and say hey you guys can use our GPUs you can contribute if you're willing to work on open source and so I think they have plenty of uses for these GPUs it's very rare that
19:00
you're going to see any GPU cluster that's just sitting around idle collecting dust but clearly the point of Nvidia's business is to rent this to other people who are interested in paying something beyond what the backs stop would be at the base level. And
19:11
And anybody uh at Nvidia would prefer to rent it out to a hot startup that's interested in training their model on their specific domain with all their expertise for five bucks an hour in year 1 instead of having Nvidia research engineers using it for $3.
19:29
68 or something like that.
19:31
Um, and anytime that they can support a Neo cloud, actually renting these GPUs to these startups will guarantee that the Neo Cloud is going to be able to grow their business and then buy more Nvidia GPUs in the future.
19:45
So, it's a I mean, I'm making the the best case for what they're doing here, I think.
19:50
But, um, the generally speaking, Dan, I think you called them the central bank in the article of AI, right?
20:00
And uh certainly for startups and certainly for Nvidia GPUs that's true.
20:04
They are acting as a central bank who in some ways is a market participant um but in other ways is like a external party that is just watching everybody else buy and sell from each other and and trying to grow the market, grow the pie in total um rather than pick winners, right?
20:26
>> And it's it's not just that, right?
20:27
They're also taking hyper they're also taking data center leases.
20:30
There was a report by which they may be buying fiber in order to then reallocate to neoclouds.
20:37
There's a lot of areas that they realize, you know, this trinity like AI is is just growing way faster than the Trinity can organically develop.
20:46
So the central bank's purpose is to step in and support the economy um when private credit can't, which is the case here, right?
20:54
It's just growing way too fast. >> Yeah.
20:58
And let me maybe refine something that I said a second ago where I said that they're not picking winners.
21:02
Um they're absolutely picking winners here.
21:05
[laughter] Um and much more so than the central bank does, even though the central banks can sometimes pick winners of who they're going to bail out.
21:12
Um is not waiting for bailouts.
21:15
They're actively going and picking the winners.
21:18
And those winners are the people that they like the best.
21:19
And who do they like the best?
21:21
The people that buy the most of their stuff.
21:22
So, this is a tool that they're using um as an active market participant to support people, not necessarily the same way that a central bank just kind of sits back and sets rates. >> Yeah.
21:35
Why don't we uh Zane, do you want to jump down to the bullseye diagram?
21:38
I think that'll illustrate Jordan's point really, >> really well because Nvidia can decide who are they going to sell to.
21:45
So, Zane, do you want to explain our our bullseye concept?
21:49
>> Yeah, that plays into what Jordan said.
21:51
[laughter] Actually, now that I realize it, it's like I didn't think about that, but it is a bullseye, right?
21:56
It's their target, right? >> Yeah. In a way. In a way. Yeah.
21:58
This is um this is shout out to probably the the least technical chart in the whole article, but but I I hope at least we understand this.
22:09
[laughter] So, I think like the way that we kind of think of it is is um is a few concentric pools of demand, right?
22:17
like um these are basically um people that can buy Nvidia uh Nvidia's hardware but um in in the first and the widest pool you have just all buyers generally um and here the pool is just for Nvidia to sell hardware to them and get the margin on top of that right so um this is the first layer the first circle um of course the inner circle is is Neocloud says that's the blue one and They take those trips and they turn them into a rental business.
22:50
Um, and they are repeat buyers of Nvidia hardware.
22:54
They continue to buy them, continue to rent them out, make them available to the ecosystem.
23:00
Um, the red circle is new clouds that are SCPs.
23:03
So here is your Nvidia certified cloud partner tier and then Nvidia gives gives them reference designs and priority allocation and so on.
23:13
But also in return, Nvidia gets um standardization and and a sort of stickiness with the SCPs, right?
23:20
Um and lastly, you also have um the innermost circle in green, which is the Neoclouds, their SCPs with a back stop.
23:29
So here um it's it's probably the stickiest, right?
23:36
Because Nvidia, they help to credit support that cluster via the back stop that we just mentioned, but they also take a cut of the revenue above the back stop.
23:46
And so like this one-time hardware sale um ends up being a a recurring margin business because of the revenue share as well.
23:57
So that's uh this is our bullseye diagram.
24:00
And uh of course in city you have the NCPs with the back stock. >> Makes perfect sense.
24:04
Who are they going to sell to?
24:05
People where they can double dip, right? To your point, Jordan.
24:10
>> Yeah, they're not exactly um throwing darts at this or if they are, they're hitting the 180 every time. So, yeah, bullseye.
24:17
Um anyway, there's lots of uh people in the [laughter] in this industry right now that's pursuing a backs stop, let's say, and uh trying to, you know, maneuver so that they can get this.
24:34
It's it's definitely sought after.
24:36
There's only so much to go around.
24:39
Um maybe the one comment I can make is that neoclouds that are NCPs in the green circle in the middle like you said they they get standardization stuff.
24:48
People can't just choose to be an NCP.
24:51
They are granted the privilege by Nvidia.
24:53
And that privilege represents itself as a skew like a part number on a bomb that represents an increased percentage that you are paying Nvidia for your GPU servers and networking.
25:09
And you're also committing to purchasing all of the Nvidia switching and all the Nvidia software.
25:18
And you know, like you said, standardization.
25:22
um also means no competition in these uh designs.
25:27
You're you're not going to be able to just put put Arista switches in your backend network here, right?
25:32
You're going to be using the Spectrum X switches if you want to back stop.
25:35
You can't use somebody else's transceivers.
25:39
You got to use the Nvidia branded LinkX transceivers that are the exact same as the cheaper ones that are four times less expensive, right?
25:45
But um I mean it's a way to keep the keep the ball moving from for them. So, uh, good segue.
25:55
Let's keep the ball moving on our our side as well.
25:56
Kong, we can talk about how these GPU loans are actually being priced by the banks and then maybe we can give some examples of back stops that you guys are most familiar with.
26:08
So, we can talk about like the actual projects instead of just this at a at a high level saying just NeoCloud over and over. >> Yeah.
26:17
And we also have a view on where where the backs stop lending might price.
26:20
So yeah, it' be interesting. >> Yeah.
26:23
So I guess some context into thinking about credit spreads is that the way that you would price a credit bond is you have the underlying you know whatever your government business is right so US treasuries maybe at 10 years now they are like priced at 4. 3%.
26:38
And then you have to think if I'm going to lend money to a company how much riskier are they than the US government for example.
26:47
So maybe if let's say I think that Microsoft is only you know just a tiny bit riskier than the US government I might charge them extra 50 basis points or extra 0.
26:58
5% to lend them money at the rate they'll lend the US government.
27:04
It's sort of this interesting idea where you can decompose the level of risk when you're lending to a company where you can sort of understand why whatever credit is priced the way it is.
27:14
So for example, if you take a look at this core reef uh chart here, the context to this is that you know core reef has priced multiple kind of depth that's available in the market today.
27:28
The line that you see at the fair top, the core reef in the blue or purple line, that's corre unsecured bonds.
27:36
So that's trading the market today at just probably over a 9% yield.
27:42
And you can see the brown dotted line below what Cor would receive what Cory will pay on interest if they manage to project finance that.
27:55
So in this situation the the DEP is secured by the GPUs and they have a investment grade customer and you can see below all the different multicolored lines representing the different investment grade hyperscalers you know showing the kind of customers that corre.
28:17
So you can see the three you know big brackets there that's how you would decompose the credit risk of a core reef debt for example.
28:24
So you can see illustratively if we think about core reef meta DDTL 4. 0.
28:29
So that was priced at about 225 bits over so far or what that means is it was priced as 2.
28:37
25% in interest more risky than what the US government that illustratively would be.
28:48
And if you try to decompose at 225 basis points obviously some of it will be the risk of meta.
28:56
So meta obviously has to pay corre if not cor is not going to get money to pay the bank.
29:02
So that 97 basis points is what meta bonds are trading in the market over the US government bond and that is the risk that is implicitly taken by KI when they have meta as a customer.
29:15
Then that backs the question right.
29:17
So if we know meta is good for this 97 basis points of risk, why is that additional call it 105 basis points on top of that?
29:26
Well, that 105 basis points would then all be reflected as core reef execution risk.
29:32
So what happen if core reef doesn't stand up the cluster on time?
29:37
Now what happens if core reef you know misses that srs constantly?
29:41
What happen if core reef has issues with reoccurring or installing the GPUs?
29:46
That's the kind of risk that banks are taking off when they price this additional 105 basis points.
29:54
And then you know that comes into all rate of 225 and then we see the unsecured bonds an additional 400 basis points over that uh execution risk that we talked about.
30:04
So these bonds are not secured by the GPUs, but they're just merely, you know, that that subordinate to the specific GPU debt that's being financed.
30:17
that's being financed. So what this means is that okay let's say if cor goes into bankruptcy you know the first dep they pay off is the GPS that are secured with the loans that already backing them and then whatever money that's left over you know then you go towards the unsecured bonds u held by dep investors
30:35
in core fun secured bond so this is how we think about decomposing the various spreads within uh different new cloud deps and I think that's quite interesting because that let lets us think okay in the future if there are more and more of this kind of debts being raised, what would be the appropriate rate they should land at? >> Yeah, makes perfect sense. I mean, in my
30:55
>> Yeah, makes perfect sense.
30:55
I mean, in my view, like it it's quite simple to see that hyperscalers also have an entire other business that's a free cash flow monster and Cory's entire business is this GPU rental thing.
31:10
So, they're obviously going to have different credit risk, but the spread is uh is fascinating when you uh kind of break it down like that in the chart. >> Yeah, exactly.
31:23
You know, the hyperscalers are good for the money.
31:26
They you know, they're getting unlimited monies from different all different kind of businesses.
31:30
But for core reef, you know, the second method doesn't pay, you know, you know that something's going to go wrong. >> Yeah. >> Yeah.
31:37
Let's And then I think Con you had a pretty good chart where we did the 5-year pricing by deal structure where we we kind of um yeah this one um so I think in this one we a little more explicitly um try and decompose these different things and then uh suggest where we think it would be.
31:53
Um so on the right one you can think this is a bit more comparable to unsecured lending to core reef where you're taking platform risk and execution risk.
32:02
Um and then you've got the base base rate which is just risk-f free on the left which is the other extreme.
32:09
You're not taking platform risk as much.
32:10
You're just taking execution risk um which is the back stop or or I should say the not the backs stop the hyperscaler offtake um 5-year octate contract.
32:20
And then I think Nvidia is somewhere in the middle where you got to believe in their execution.
32:25
You're taking Nvidia credit risk, but then there's also an element of do they manage to create a good pool of customers.
32:30
And I think the whole idea is we want to get the market um comfortable with something in the middle.
32:36
And you you can think of this as um like a trial period um an initial period for them to really understand how the NeoCloud runs its business.
32:45
And and and banks are not just going to look at the Nvidia backs stop and call it a day.
32:49
They're going to start asking questions about, okay, well, how are you going to build a book of customers?
32:53
How are you going to build this business?
32:55
Because the the Nvidia backs stop is not going to be around forever.
32:58
The point is, it's only meant to be there for the time being to allow banks and everyone the time to really understand the NeoCloud business model, really understand the risks, and then one day they'll have to lend on on on a standalone basis, just like they lend to every other business, right?
33:13
just like they lend to semiconductor fabs, just like they lend to watch factories.
33:18
Like every business has duration.
33:20
You know, it's like a bit ridiculous when people say, "Oh, but there's duration risk and depreciation."
33:26
It's like that's every business on Earth.
33:28
You know, every business on Earth doesn't know if they're going to have a sale the next day. Like most businesses. >> Yeah.
33:35
Most people aren't building bridges, I guess. Yeah. >> Yeah. Yeah.
33:37
You know, even like Hermes doesn't know next month they're going to sell bags.
33:43
>> So, what's the difference?
33:43
But actually I was thinking about that point you raised right and an interesting counterpoint is that okay each Hermes back is small relative to the overall cost of the factory.
33:51
So for example on the contrary if you think of we work you know they can't feel like in a single office lease that's a huge proportion of the cost that they have paid for the building.
34:02
>> So it would be similar in this situation right if let's say I stand up a 200 megawatt cluster and I can't find customers that's obviously a way bigger issue than you know if I can't sell one of M bag. Yeah, totally.
34:11
Um, and and that that's like there's a whole spectrum.
34:17
You got businesses which are fortunate enough to have like a foundry maul, so then they're not the ones who wear the appreciation.
34:23
Um, you know, we work a great example. You got property.
34:27
I think the like to your point, there is there is a spectrum.
34:30
Um, but it doesn't mean no lending at all.
34:32
Um, you know, there there are kind of like worst things which which are super variable like airlines, but there's a ton of lending for airlines.
34:38
lending for airlines. um you know there could be a pandemic any any time there could be travel disruptions that that's one of the hardest things to get right and there's still plenty of lending airlines um and it's even worse because it's like 15 year and then you don't even know like you're selling tickets
34:54
like two two you know weeks in advance um but but I think one one one thing you do have a point on is is it's a very brisk depreciation so if you do get if you do get utilization holes then there's just not a lot of time to fix it right if you're not executing then Then time time ticks very quickly, right? Do
35:09
Do you think that's fair con? >> Yeah, exactly.
35:13
I think that was what I was thinking of like, you know, you have little time to make it right, right? If something goes wrong. >> Yeah. Yeah.
35:19
Like if you're if you're commercial property, you have 99 years to figure it out.
35:24
>> I think the other point is that like banks also, they're a bit scared to underwrite, you know, what happens if something goes wrong, right?
35:30
something goes wrong, right? So if let's say like a building lease you know the the the building developer defaults on that debt you know they do work they're thinking like okay how much can I sell this building for in two years 3 years time but for banks you know when they
35:44
approach GPUs they don't have that same mindset right it's more like you know I have no clue what these GPUs are going to work in two years three years so you know I'm just going to take it as as zero >> yeah I think Um, but you're you're right like residual value of offices is also a lot easier. Residual value of airplanes a
36:04
Residual value of airplanes a lot easier because it's not like every every year Boeing has like something that flies twice as fast, right?
36:12
[laughter] Or as fast as you know except in like in tokens [laughter] >> in 9 years.
36:19
Although the seats get twice as twice, you know, half as much every year.
36:24
I wish we had MOS laws for the internet.
36:29
>> In 9 years, we just start teleporting.
36:31
Yeah, we just uh [laughter] double every year how fast this is. Okay.
36:35
Um can you guys give a real example here?
36:38
I mean, we've talked about it conceptually across these these uh these other examples and comparisons, but you walk through a couple of real examples in the article specifically in Australia and Indonesia with a couple of neocids.
36:52
Do you mind talking through uh those examples?
36:58
>> Yeah, I can talk about uh PHMAS is the largest one in the world.
37:00
Um so it's a 360 megawatt project in in Batam.
37:03
Um and um Yosharon actually announced theirs first and then um Fermis followed uh shortly thereafter just uh I think a week before we wrote this this article.
37:19
Uh gosh, I don't remember.
37:19
But but anyway um you know like like I said um you know pharmacist is kind of interesting because they started off with um actually um a cluster in Singapore um that that was you know very much like NeoCloud v1.
37:35
0 know, you know, raised a bunch of capital and and they they sold for the trinity originally uh because they had SCT GDC um invested in them.
37:46
So they had the data center sorted out um and they had good investors and they could take a bit of risk um and then they parlayed that into um the large cluster in Melbourne.
37:55
Um and then also they have one in Tasmania and there what they did to solve the Trinity is they actually self-built the data centers.
38:06
Um and it sidesteps the whole conversation around um uh uh data sets.
38:13
We can talk a little bit more about that.
38:14
Um and I believe the next one um is actually not going to sidestep that.
38:19
They're going to work with Day One on that facility.
38:20
Um and again, yeah, that's that's kind of the um um you know, we think um we think if that's the objective as we outline, if we're right about the objectives, then it's going to be pretty interesting because it's going to be one of the largest sources of compute in the region and it's going to be broadly available to all sorts of folks.
38:42
folks. um you know assuming that again you know our assessment of the spirit of this thing is that the spirit is not to take the backs stop and do a 5-year offtake right you probably don't need to pay Nvidia half the revenue right um so it almost ser it almost stands the
38:58
reason anyone doing this is probably going to uh you know want to do something a bit a bit broad-based and Sharon too so I think we'll see I think we'll see more in the region I really feel that this region is somewhat underserved by capacity in general. Um,
39:12
Um, and there's certainly a lot of capacity in data center in general that's coming up.
39:18
So, um, yeah, that's what it's Yeah, it's like high level here.
39:25
It's unbelievable to see Australia, not the biggest region in the world and Indonesia.
39:31
Australia getting 72 megawws with a planned 102 to scale to that's 55,000 GPUs by the middle of next year.
39:40
[sighs] and um firm in Indonesia.
39:42
and um firm in Indonesia. Their third major project let's say after Melbourne and Tasmania is Batam which is 360 megaww that's you know tens of billions of dollars of capital expense all going to Indonesia to build out GPU capacity which again is is just it's just a massive amount of
40:08
investment in a single area that not a lot of people think about when it comes to data centers and Dan, to go back to your point at the very beginning of this conversation, um that's how we add up to trillions of dollars is projects like this all around the world that need to get financing. >> Yeah. And it's interesting because we're >> Yeah.
40:29
And it's interesting because we're having there's lots of conversations happening right now um between banks uh from banks that that are are are asking you how do we how do we do this for the first time?
40:39
Um, even with a backs stop, this is the first time they're really thinking seriously about understanding NeoCloud.
40:46
Um, and so, you know, there's a bunch of tools that we think they're going to need.
40:51
And I can I can talk quickly about that.
40:52
Um, I guess, um, if you scroll, um, let's see.
40:59
I think scroll up, right? Oh, yes. Um, yeah.
41:03
Tools that GPU lenders are need. Yeah.
41:05
um they're going to need a bunch of tools.
41:08
If they're going to put trillions to work, they're going to need a few things.
41:12
They're going to need to know what is a what is a fair market price for rental.
41:15
Um so we've actually been working on that for um almost 2 years, almost 3 years in fact.
41:23
Um we've been tracking bilateral contract price across the entire term structure.
41:27
Um and these are this is deal information um that is is basically impossible.
41:33
You can't get it by scraping.
41:36
um you know, you're not going to find a posted.
41:38
You really just have to gather it uh by just surveying a lot of NeoClouds and asking like, "Hey, you know, what do you think um a B200 could go for?"
41:46
Um and then Jordan also um you know, helps out a lot in terms of connecting people to compute.
41:50
Um so, you know, him and him and the Sams have a pretty good sense of where things will transact.
41:57
Um and that's this all goes into composing this index.
42:00
Um it's it's a very different approach than others.
42:02
It's not very scraping focused, but you know, we really believe that every data point there's a story and we try and uh recount those stories.
42:11
We try and we try and explain that a lot of the research we do and get a sense for the market.
42:14
Um the second thing is um which NeoClouds are going to execute well um which ones are going to have the lowest execution risk.
42:23
So if we scroll down, this is this is what Jordan does.
42:25
U maybe Jordan, you should explain why this is so important to whether you would want to lend. >> Yeah.
42:32
Yeah, I mean to us like the there's basically a ratings agency that's missing here um when it comes to the quality of these uh the the product that these people provide and then therefore the quality of the debt and um the only way that you can get an understanding of that is to do hands-on testing and to talk to the existing customers in a real you know unbiased way and that's exactly what we do with cluster.
42:58
So for everybody listening stay tuned for cluster 3. 0.
43:00
know uh testing deadline is August 1st so we're almost there and we'll be putting out an article in mid August to update these rankings based on our hands-on testing and interviews with over 150 customers now of these Neoclouds.
43:12
Um you know in some sense it's hard because you've got to go and talk to over 200 of these Neocloud providers now to get a sense of what everybody offers.
43:24
And like I said over 150 customers that we're connected with.
43:29
We're always trying to connect with more.
43:30
So that number is growing, but in another sense, it's easy to get a sense for who's got their stuff figured out and who doesn't.
43:38
Um, we have a whole checklist of criteria that's public on our website.
43:44
We've told many of these Neoclads for months or even years about certain things that they're missing that are big differentiators for why people would want to buy from them and would improve the quality of service that they can provide.
43:54
And many people just don't do it. Really simple stuff.
43:58
So in that sense it's pretty easy to differentiate between people who have health checks and people who don't.
44:06
People who have monitoring dashboards and people who don't.
44:07
People who offer managed slur and managed Kubernetes and people that don't.
44:11
And as they expand into other businesses, many of these people serving tokens as a service or serving RL environments and just going up the stack or many of them going down the stack where they do self-built data centers like you're referring to with Fermas who need to be able to assess their competency across all of those different facets of the business and that's exactly what we do with cluster max.
44:30
So anyway, it's a fun project to work on cuz we just get to learn so much.
44:35
And Dan, to your point, when you're trying to draw a curve to show the pricing of stuff and how it changes over time, some of these individual stories are worth so much more and should be weighted so much more heavily when compared to, you know, one spot price that anybody can get from hitting a public API to see that that vendor has chosen to price that GPU instance type at a certain price on a certain day.
45:03
um talk to somebody else who's rented, you know, who signed a $20 million GPU contract and talk to them about the decisions that they were making at that time and you're going to get a much better sense for, first of all, what that contract is priced at.
45:15
Second of all, what they care about when they're choosing a a provider and how they're going to discount certain providers that can't provide a high level of service versus pay a premium for those who can.
45:27
Um, and then compare that to some on demand price that somebody's paying for a single H100 that they're going to rent for 6 hours to train their anime bot or something on.
45:36
Very, very different business.
45:38
So anyway, yeah, I think in summary, the ratings agency is basically missing here and we're doing our best to help people solve that.
45:45
Yeah, I thought this this section of the note was pretty precient because I think call it two weeks after we released this note, Navius announced they raised their first debt facility, right?
45:56
debt facility, right? And then you know turned out that it was at a similar rate to corre suggesting that banks thought that execution risk you know was similar to corre which you know if you look at the table we have here we have cor and eB just side by side >> makes sense >> and it's pricing of debt but also
46:15
pricing of their products um and it's going to be important that if you're lending um you the question is are you lending to like a Hermes are you lending to uh like uh I don't know what's it like an H&M right [laughter] are they going to be priced high or low and but that's a real thing
46:34
because I if they're doing um you know Kubernetes slurum right versus like just bare metal or like whatever like there's a price difference right Jordan >> price difference between what you can deliver an SLA and so you know it's important for lenders to know what they're getting into >> yeah look I mean you guys have have
46:51
called this out that there's this backwardation curve where people who have GPUs for sale that are available right now are charging a massive premium and those who have, you know, have delivery in 3 months or 6 months provide a discount just to lock in the customers. And the big variable that people play
47:05
And the big variable that people play with right now is the prepay amount.
47:08
So, um, having those milestone payments as you're going there, so these neoclouds don't need to, uh, raise as much bridge financing or take out a construction loan or, um, just change the structure of the debt that they're raising because they can get these prepayments from these customers that trust them.
47:28
I mean, that that changes their financials completely.
47:30
So, that's a big thing on the pricing.
47:32
the pricing. But after you get past that, then there, you know, two people offering the exact same GPUs in the similar location with the exact same, you know, headline price are actually going to deliver very, very different
47:45
things, which is what we've talked about on previous episodes to the tune of people willing to pay in some cases 20 or a 30% premium for the exact same quantity of GPUs just because they stay online and they perform uh compared to some other providers who are less trustworthy. And so um that's
48:00
And so um that's represented in this chart.
48:03
The high level is is represented there.
48:04
But for anybody who's looking to dig into more details, please get in touch with us because we do a lot of this consulting work for uh for people who are making these critical decisions. Right?
48:14
When when you're making a a decision on where to send $50 million of the VC money that you've raised and kind of bet your company on some NeoCloud, you really need to be able to trust them.
48:26
And and we're just providing people with that peace of mind.
48:29
to know what they're getting or to just think through the decision, right?
48:33
How to negotiate, how to get through the SLA discussion, how to def, you know, write a contract, um, what sort of things to get the the NeoCloud provider to commit to.
48:42
I mean, these are these are big questions that we help people answer. >> Yeah.
48:48
The last piece of the puzzle is, if you scroll down a little bit, is inference X.
48:52
Um, a lot of people ask, what is the residual value of a GPU?
48:54
Um, it gets into what's a GPU worth?
48:57
Um, you could argue what a GPU is worth is what it can generate and that's how many tokens.
49:03
So, inference X tells you how many tokens per second you can generate per GPU.
49:06
Folks can make assumptions on how many how much the tokens are going to cost.
49:10
Um, but you know, this is this is really important because it's the fundamental revenue generating capacity.
49:17
Um, you know, just as much as how much uh bags cost, how much flight tickets cost, how much nuts and bolts cost.
49:21
This is really the the unit cost the atomic um unit of revenue for the entire industry.
49:29
>> Yeah, everybody pays for tokens right now.
49:31
Uh some special customers rent by the GPU hour when they're the ones training the models.
49:35
But the whole ROI on the whole industry depends on people buying tokens.
49:40
And that's this chart right here.
49:44
>> I thought I thought inference is taking over from training though, Jordan.
49:45
I thought training is going down. >> Yeah. Okay.
49:50
You want to you want to have a whole another podcast right now where I rant about the uh shift from training.
49:55
>> For the record, we don't actually think that.
49:56
For the record, [laughter] >> no, but did you see the two charts that I found uh from back in 2018? >> Yeah.
50:04
A couple of our our favorite forecasters who are linear extrapolators instead of exponential extrapolators the way we are at semi- analysis.
50:11
In 2018, we're projecting that in 2025, one day inference might be a $10 billion market.
50:20
Yeah, >> one they're right. >> Yeah.
50:23
So we we've and they had inference at double training at that point.
50:27
Um and actually edge inference was going to be 10 I think 5 billion and and data center training was only going to be 4 billion.
50:37
This was uh this was our our our peers at McKenzie let's say um who are who are making these forecasts.
50:44
I mean, I I'm not sure anybody in 2018 could have gotten the order of magnitude right and seen the explosion like this, but the ratios were just never making sense that people would someday decide this is the final model we've ever trained.
50:58
We're not going to invest in any more training.
50:59
We're only going to buy GPUs for inference. This doesn't make sense.
51:04
The Frontier Labs believe they're building the machine god.
51:08
All the evidence to me points to the fact that they are and they're going to try to invest as much as they possibly can in training.
51:17
Inference is a means to buy more GPUs to train the next mod. >> Wow. Feels like a MMO RPG.
51:22
It's like you buy items to kill monsters to buy items to kill monsters, right?
51:29
>> Never ending meat cycle.
51:31
>> That that probably describes life, right?
51:33
you know, it's like [laughter] we're all just killing time.
51:40
[laughter] >> Okay, guys.
51:42
Anything you uh you think that we haven't covered on this episode?
51:46
I think we've uh gone off the rails on quite enough tangents on our tour. >> Yeah.
51:50
>> Tour of the Holy Trinity, right?
51:50
I think we've subjected the listeners who are still hanging on to enough talk about GPU debt and financing rates.
51:56
Um, if you guys are interested in seeing anything fun or clever or witty, we can do it right now.
52:04
I give you a three count. Two, three.
52:08
Okay, that's a no from the Singapore team. We've had enough wit. >> We are not smart. No.
52:12
[laughter] >> For today's episode, >> our wit drained out. Sorry. >> Yeah.
52:19
Well, I appreciate everybody taking the time to listen in here, guys.
52:22
good job walking us through everything to do with the trinity capital offtake in data centers and what what Nvidia's doing with these GPU debt back stops. Hopefully it was fun.