# Data Centers: Financing the AI Buildout with Stijn Van Nieuwerburgh | Markus Academy | Ep. 161

https://www.youtube.com/watch?v=e5YuA53OFQ0

[00:00] So, let's take a quick look at the hyperscaler capex needs.
[00:03] This is the number that Mark has cited at the beginning of the talk.
[00:07] Where we see that about 655 billion on balance sheet capex spending is is is projected for 2026 by these five large hyperscalers.
[00:13] Right?
[00:16] And that's up 58% from last year.
[00:19] So, these are massive numbers.
[00:21] I think it's useful to put those into context and compare them to the operating cash flow of these companies.
[00:25] Right?
[00:27] And so, that's the shaded the light shaded uh bars next to the capex bars from the previous slide.
[00:31] And what you can see is that the ratio of these two, which is the red line plotted against the right axis, has been trending up over time.
[00:40] So, it used to be that capex consumed about 40% of operating cash flow.
[00:45] Today, that number is approaching 100%.
[00:48] So, in other words, the hyperscalers are running out of cash.
[00:52] They're running out of free cash flow, which is the difference between those two numbers.
[00:55] And in fact, this year, for example, Amazon is going to have for the first time in a very long
[01:00] time negative free cash flow.
[01:03] And so, we need to follow the money.
[01:04] We need to understand where the risks reside, who ultimately has claims to the underlying revenues and costs and risks associated uh with that buildout.
[01:22] So, welcome back everybody for another webinar organized by Princeton for everybody worldwide.
[01:27] Today, we talk about data centers and financing the AI AI buildout.
[01:32] But before we go there, I would like to use this opportunity to have a moment of silence for Chris Sims, who was with us um on this channel as well.
[01:43] He was an extraordinary economist and academic intellectual, not only winning the Nobel Prize, many other insights we gained from him.
[01:52] But he was also an extraordinary human being, and I would like just use to have a moment of silence in honor of him.
[02:12] Thank you.
[02:14] So today, as I said, we talk about data centers and financing the AI buildout with Steen van Nieuwenhuijs from Columbia University.
[02:20] Hi Steen, it's great to have you.
[02:21] You're the best person to talk about these things.
[02:25] And uh we're learning from you.
[02:26] Hi Steen.
[02:28] Thank you, Marcus.
[02:31] So of course, we have huge AI investments.
[02:33] The capital expenditures, the CapEx in data centers is projected for this year to be between 660 and 690 billion dollars between Microsoft, Alphabet, Amazon, Meta, and Oracle, and there probably will be even more.
[02:49] And on that's on one extreme.
[02:50] On the other extreme, we have some revenue which supports this uh CapEx and this and expenditures, which typically consists of 20 or 300 a month uh for some subscribers and enterprise versions.
[03:02] So the revenues in 2025 for OpenAI was about 13 billion, for Anthropic about 9 billion.
[03:08] And of course, in order to justify this huge CapEx uh expenditures,
[03:13] uh it requires a lot of growth.
[03:15] But the question is, how is this, you know, this investment, how is it financed?
[03:19] And where's the risk?
[03:21] Where's the allocation of the AI risks?
[03:24] And what we see is that traditionally, until last year, the big hyperscaler, they essentially financed this out of revenue,
[03:32] but it shifted from out of revenue financing to a debt financing, and also involving SPVs, special purpose vehicles, ABS, CMBS, and so forth, and private credit.
[03:44] Things we are familiar with from before the global financial crisis.
[03:48] So we're interested to find out, you You where is the risk of this financing?
[03:52] Of course, if AI revolutionizes everything, it might look like this way,
[03:57] then everything is fine because revenue growth will just hold everything up.
[04:01] But, there might also be some residual risk, and it would be interesting to find out where the risk in the financing is.
[04:07] And there's nobody better answering this than Steen because an expert not only in finance, macro, and also real estate.
[04:15] And we're talking about real estate here with data centers.
[04:18] Steen, the floor is yours.
[04:20] We're looking forward to your presentation.
[04:23] Thank you, Markus.
[04:25] It's a great to be on the podcast and on the on the Markus Academy.
[04:29] I've greatly enjoyed seeing your episodes in the past, so thanks for having me.
[04:34] Um, so, let's get into it.
[04:37] And as you mentioned, I mean, this is about this is fundamentally a story about AI.
[04:41] But, a lot of our conversation about AI is really about the software.
[04:45] It's a software story.
[04:47] It's about how the models are improving, how the algorithms are getting better.
[04:50] It's a story about potential labor displacement risk from AI.
[04:55] But, what I want to argue today is that this build-out of AI is strikingly physical.
[05:00] There's a lot of physical capital where the data sits.
[05:02] After all, data is a sequence of ones and zeros that has to be stored somewhere, and these AI models are running on large and ever larger computers and large data centers.
[05:14] And and so, we want to think about what has
[05:16] changed here and what what is happening on the physical capital side of this of this infrastructural.
[05:23] And so, what I want to do today is make three main points.
[05:28] The first one is that this new wave of AI development is really driving a new wave of physical capital formation that's both unusual in scale and in composition.
[05:39] The second point I want to make is that this is changing this this large-scale investment boom is changing the ownership and the financing of that AI build out.
[05:48] And and so are the economic risks.
[05:50] And so we need to follow the money.
[05:52] We need to understand where the risks reside, who ultimately has claims to the underlying revenues and costs and risks associated uh with that build out.
[06:02] And then the third thing I want to show is that the the new financial architecture that we seem to be converging towards is very different from the traditional funding model.
[06:09] As you just mentioned, uh a lot of the hyperscalers uh used to basically self-fund these data centers and and the
[06:18] and the IT CAPEX associated with it.
[06:20] But more and more this risk is migrating to private credit, to special purpose vehicles, to places where, you know, the end investor is not only more distant from the from the use of those facilities, but potentially also uh you know, in corners are much more opaque, much less easy to measure and to keep track of.
[06:40] And so I think the the the lesson is that it's not just about how large this AI CAPEX boom is, but also how these risks are being distributed throughout this financial system and ultimately uh where where those risks reside.
[06:54] And then I think that is interesting both from an academic perspective, but also from a from a policy perspective.
[07:01] So, let's start and and begin and start with the first question about what is new here about AI and the economics of data centers.
[07:08] Now for that, I want to tell you a little bit about what a data center is without going into too much detail.
[07:14] But basically, you know, a data center is really more than just a box,
[07:19] an industrial warehouse with some servers in it.
[07:21] Right?
[07:23] That sort of used to be the case in in the good old days.
[07:25] But today, these are complex facilities with, you know, sophisticated systems.
[07:30] Um you know, we distinguish between what's called the gray space,
[07:34] which is um the the the physical box, the the the structure, but also all the power, all fiber optic cables that connect into it,
[07:43] um and all the electrical equipment, meaning, you know, the the the the backup power generators, the power lines that come in, also all the cooling equipment, the chillers that, you know, as you can see in this picture in the top left are sitting on the roof of this data center facility.
[07:57] Right, so all of that is called the gray space, and typically it's the landlord of the data center, the owner of the data center that owns that that that physical real estate and and power infrastructure.
[08:11] The tenant of the data center, the renter, if you like, owns everything that's inside what's called the white space.
[08:16] And the white space is all the IT equipment.
[08:18] Okay, so typically the
[08:20] Landlord does not own the IT equipment.
[08:22] It's the tenant who owns the IT equipment.
[08:24] And we'll come back to why this matters.
[08:28] So, um it's useful maybe to just talk a little talk a little bit about sort of the evolution of data centers in order to understand just how different the current moment is.
[08:37] Right, so data centers go back 25 years.
[08:39] They came about when the internet was invented in the late 1990s.
[08:42] And initially, you know, all those internet cables had to connect somewhere in what was called network-dense carrier hotels.
[08:50] The carriers and the telecom companies had to have places where their information could get exchanged.
[08:54] Now, even today these network-dense um data centers still exist.
[08:59] There's a couple of them in every metropolitan area, and they're very valuable critical infrastructure, but they're a tiny part of the data center space.
[09:06] So, the next iteration of this was what's called colocation facilities, where you had essentially companies that needed a place to host their website and and then over time their servers.
[09:17] And often times they would rent one rack of CPUs in a
[09:22] larger facility.
[09:24] So, you would have buildings with many different tenants,
[09:26] multi-tenant, hence the word colocation.
[09:28] And uh these would typically be owned
[09:30] by, for example, publicly listed real
[09:32] estate investment trusts, REITs.
[09:35] Now, the next and and that was sort of
[09:37] the state of the world uh about five
[09:39] years ago.
[09:39] And yes, these facilities
[09:42] were getting bigger and more complicated
[09:44] as more and more corporate use migrated
[09:46] to the cloud, but they were in in they
[09:49] were still sort of similar facilities.
[09:51] What really changed is in the last five
[09:53] years is the rise of these hyperscalers
[09:56] and in particular
[09:57] the growth of AI foundation models and
[09:59] and data centers built for the specific
[10:02] purpose of AI model training and AI
[10:05] model inference.
[10:05] These are typically
[10:07] much larger facilities
[10:09] as I'll tell you in a second with much
[10:10] more complex hardware needs.
[10:14] And are
[10:14] typically occupied by just a single
[10:16] tenant, usually one of the big
[10:18] hyperscalers, Meta or Google or so
[10:21] forth.
[10:21] Right?
[10:21] And so they're they're
[10:23] they're large, complex facilities to accommodate the special need associated with AI.
[10:28] Now, usually as I mentioned the landlord, you know, builds the entire facility including all the electrical and and cooling equipment.
[10:35] You have versions of this where the landlord only provides the land and the building and brings the power to the building but there's nothing inside the building.
[10:43] That's called a powered shell.
[10:45] So, some investors are just investing in that investing in in data centers that way through this powered shell.
[10:51] But, you know, for the most part we should be thinking about these these hyperscale facilities.
[10:56] Now, just to underscore how different the modern infrastructure is, let's just think about sort of your traditional enterprise cloud computing.
[11:03] You know, the key unit of analysis here is energy.
[11:06] It's megawatts and kilowatts.
[11:08] Right?
[11:10] So, basically how much power do these do these do the chips use?
[11:14] And so in a standard rack of 72, let's say, standard CPUs that would use about 5 to 10 kW.
[11:23] Contrast that with the modern AI training GPUs, right?
[11:28] Graphic processing units.
[11:29] A rack of those would basically consume 10 times as much power.
[11:33] Think of Nvidia's Blackwell GPUs would consume about 100 kW for a rack.
[11:39] So, it turns out that that's just a very different architecture.
[11:42] You need very different types of power equipment and networking equipment, as well as very different memory in those chips and then electricity to power all of that.
[11:50] Right?
[11:52] And the next generation, which might be coming online at the end of this year, the Vera Rubin Nvidia chips, they would consume another two two x that amount of power.
[12:02] In addition, it becomes more and more complicated to cool these GPUs.
[12:06] If you have one of these simple old, you know, sort of old-school CPUs, you can cool that with air.
[12:11] Uh but these newer GPUs, uh they need to be cooled with liquid initially with water and increasingly with refrigerant-style closed-circuit style refrigerant cooling techniques.
[12:22] So,
[12:23] these are all very bespoke infrastructure investments.
[12:27] Uh what that means is that the old uh data centers cannot simply be repurposed for these new uses because they'll simply lack the power and or the cooling infrastructure to make it work.
[12:38] So, basically, we're going to we're talking about newly built uh newly built facilities.
[12:43] Do you think in the future there will be innovations which make it less energy intensive?
[12:46] So, the innovation will be more on the energy-saving side or not?
[12:50] It could be.
[12:50] So, more efficient chips is definitely one of those one of those directions of innovation that, you know, would potentially change this equation meaningfully.
[12:58] But so far, the direction has been every next iteration of chips has become more and more and more power hungry.
[13:05] I see.
[13:06] And you always have to build a new a new data center.
[13:08] The old ones are not useful for the next generation.
[13:11] Exactly.
[13:11] They're basically technolo- you know, there's technolo- technological obsolescence associated with this sort of recent even fairly recently built data centers.
[13:19] So, if we think about sort of the drivers of demand for compute, for data, and for data centers, there's a whole
[13:25] host of them and I'm plotting them on the left-hand side here and a lot of these are traditional uses like cloud computing or e-commerce.
[13:32] But really the growth in demand for compute is coming from AI.
[13:34] You can see this in the right chart from McKinsey showing that McKinsey expects, you know, about a 22% annual growth for each of the next 5-6 years in compute and 80% of that growth is coming from AI.
[13:50] Be it AI model training or AI model inference.
[13:53] Now, it's important to spend a second on the difference between the two, right?
[13:56] AI model training is about, you know, the large language models, the Claude model and the the ChatGPT model that has like billions or even trillions of parameters.
[14:04] So, those facilities need sort of very specific layouts.
[14:07] They need large dense centers of of of GPUs, lots of them.
[14:14] That's that's a strong strongly growing source of demand about 22% per year.
[14:20] But what's growing even more strongly is AI inference.
[14:22] AI inference is basically, you know, calling these models for
[14:27] applications. So, agenda AI, but even just simply checking the weather app or, you know, asking for directions in in Google Maps, that's all inference.
[14:37] Now, inference needs to be physically closer to population centers because it's more important that it's low latency, that you get a quick response, right?
[14:45] If you're a surgeon using AI, you need an immediate response.
[14:50] And so, what the the type of use also drives the location of these data centers.
[14:56] So, let's Elon Musk talks about moving some data centers or some in the in the sky or above the sky, the atmosphere, uh he talks about the inference or he talks about the training of the models.
[15:11] I mean, I think the the the data centers in space idea, every industry practitioner this is a question that always comes up and every practitioner I've ever seen answer to has always said that this is a ridiculous idea just sort of given how large and heavy it is and and how crazy expensive it
[15:29] would be to move that amount of mass into space.
[15:33] And you know, also the inference right now, we haven't solved that inference problem, right?
[15:37] You would need sort of Starlink and it would have to become much much faster for inference.
[15:39] So presumably that would be for for AI model training.
[15:47] But again, you need a lot and lot of GPUs for that model training and it's not clear how on how on earth would we ever sort of can get those into space in a cost-effective way.
[15:58] So now that we understand sort of the size of these facilities and the need for for new facilities, I want to put some numbers on the cost of these facilities and I think that's sort of interesting.
[16:11] So turns out building one of these modern AI data centers on just the the building side costs about $11 million per megawatt.
[16:20] And the unit of analysis is no longer square feet like it is in every other real estate asset, but it's megawatt.
[16:23] It's all about the power and the power density of these GPUs.
[16:29] So how does that $11 million break down?
[16:32] Very little of it is land, a little bit of it is permits and taxes and stuff like that, but 85% of it is what we call hard costs in real estate.
[16:39] So these hard costs are the building, which is actually only a small part of the hard cost, but then bringing the fiber into that building.
[16:47] And most importantly, the cooling, the sophisticated cooling systems, as well as the sophisticated electrical equipment.
[16:54] In particular, you know, you need full backup cuz these these these data centers are supposed to be online 99.982% of the time.
[17:05] That is the standard agreement between an occupant of these data centers and the landlord.
[17:09] So you need essentially uninterruptible power supply and full backup.
[17:12] So you have massive uh generating uh systems in the uh in the back of the building.
[17:20] So, it turns out that that's actually only part of the cost of building one of these data centers.
[17:24] So, let's do the math for a a a a new state-of-the-art uh AI data center facility.
[17:29] Let's say that's 200 MW, right?
[17:34] that facility will cost $2.2 billion for the for the data center itself.
[17:38] That's that $11 million per MW * 200.
[17:41] You would have to bring additional power infrastructure to the table as well.
[17:45] Some new substations, high-voltage transmission lines, grid interconnection.
[17:48] That would cost you another $400 million.
[17:50] And then, you would have to bring the IT into the building.
[17:53] Now, the IT, remember, is owned not by the owner of the data center, but by uh the tenant, by the user of the data center.
[18:00] But, nevertheless, it's part of the overall capex picture.
[18:01] And that would cost about $5.6 billion.
[18:05] So, the assumption here is that 70% of the overall power of the building would go towards the running the chips.
[18:07] The rest would be for chilling and for uh to the power, the electricity of the building, and so forth.
[18:11] Um and it could cost about $40 million per MW on on IT, right?
[18:13] 70% is of that is the chips, the GPU servers, uh which includes uh not only the logic, but also the memory, high-bandwidth memory uh for GPUs.
[18:26] Some of it is networking fabric,
[18:35] some of it is storage, some of it is racks and integration.
[18:40] So, the total cost of this one facility, this one modern facility would be $8.2 billion.
[18:45] And you could think of it as about a third of that would be the the the the real estate and the power infrastructure.
[18:51] The rest would be the IT.
[18:52] Now, we are planning to build 1,000 of these things.
[18:57] Like, if you look at the actual planned data centers in the United States,
[19:01] uh the planning is calls for about 200 GW.
[19:05] So, that's a thousand times this this 200 MW facility.
[19:09] So, that would cost $8.2 trillion.
[19:12] What what's the time horizon over this for this?
[19:16] So, that's it's a little unclear because this is planned facility, so they're announced and planned is not clear over what a horizon they're planned, but let's call it 7-8 years.
[19:22] Right?
[19:24] And that lines up with predictions from, for example, Moody's, which just looks at the next five years.
[19:29] These predictions, I think, in my view are a little conservative, but they show about three trillion dollars.
[19:34] And again, here you can nicely see the breakdown between how
[19:36] much of that is data center real estate, power, and then the IT.
[19:41] The dark blue is the IT.
[19:43] Okay?
[19:43] So, I think we could debate whether it will be three billion and or or eight sorry, three three trillion or eight trillion and over what horizon, but that is roughly the order of magnitude that we're talking about.
[19:54] Now, if we put that in context, yes, Marcus.
[19:58] But, it could be the chips are overpriced at the moment and Nvidia is making a killing and it might come down and chips price might come down or something.
[20:05] Very true.
[20:05] I think everything in the supply chain is in is is scarce right now and potentially everybody's using their pricing power to increase the cost of all of these things.
[20:14] So, that is definitely the IT component of this uh could potentially come down.
[20:18] Also, as there was more competition coming online, for example, from new chip providers.
[20:24] So, let's take that eight trillion dollars is Go ahead.
[20:29] Perhaps I'll ask you after this slide.
[20:31] Do you have an idea this very US focused?
[20:31] Do you have an idea what the equivalent in Europe would be?
[20:34] Tiny
[20:36] compared to this?
[20:37] >> Yeah, so so there's interesting things
[20:38] about um so, about 70% of all AI compute
[20:43] comes in the United States. The other
[20:44] 30% is the rest of the world. It's
[20:47] growing strongly everywhere. Okay? So,
[20:49] it's growing strongly in Asia. Uh it's
[20:52] growing strongly in Europe. There are
[20:53] interesting data sovereignty laws that
[20:56] essentially make it very hard to
[20:57] exchange data across borders. So, within
[21:00] Europe, for example, we have GDPR
[21:02] um and and that sort of limits
[21:05] uh you know, where the data has to
[21:07] reside. And so, there's large data
[21:08] centers in Frankfurt and in London and
[21:10] in Paris and in in Amsterdam.
[21:13] Uh, and there's also sort of new new
[21:14] sites that are popping up in
[21:16] Scandinavia, for example.
[21:18] So, let's let's take this $8 trillion at
[21:21] face value just to put it in context.
[21:24] Uh, let's think about how large a capex
[21:26] boom that would be relative to GDP. So,
[21:28] what I'm assuming here is 5% nominal GDP
[21:31] growth over the next 8 years. I'm
[21:32] assuming that $8 trillion is an 8-year
[21:34] investment. That gets you to about 2.8%
[21:38] of GDP in terms of an overall investment
[21:40] boom.
[21:41] That's massive. That is larger than the
[21:44] real road boom and the and the sort of
[21:46] the height of the real road boom. It's
[21:47] larger than electrification. It's larger
[21:49] than interstate highways. And it's
[21:51] larger than the telecom fiber boom,
[21:53] which is plotted in this chart here from
[21:54] Apollo. So, even if, you know, that's
[21:56] obviously a speculative number, but it
[21:58] just puts into context, you know, the
[22:00] the magnitudes of what we're talking
[22:01] about. Even today, even in the fourth
[22:03] quarter of 2025, AI accounted for
[22:07] essentially 100% of US GDP growth. There
[22:09] was no contribution, net contribution
[22:12] from any other component of US GDP. So,
[22:15] even today, this is already the key
[22:17] force that is driving the US economy.
[22:18] You could say that without this AI
[22:19] build-out, we would be in a recession in
[22:21] the United States today.
[22:24] So, that's the first point. It's about,
[22:27] you know, what is a data center? What's
[22:29] new about AI data centers? And how large
[22:32] are the investments associated with
[22:33] these data centers? My second point is
[22:35] about how this is fundamentally changing
[22:37] the ownership and the financing of these
[22:40] data centers. Right? So, it's again
[22:43] important to think about where we come
[22:44] from. We came from a world where these
[22:47] large users of compute would basically
[22:50] own their own data centers. Right? They
[22:51] would build them and own them. That has
[22:53] lots of strategic advantages. If you
[22:55] expect to use your computers a lot, you
[22:57] control your own destiny.
[22:59] You you have full control over it. You
[23:00] can integrate these chips with your with
[23:02] your architecture and your software and
[23:04] so forth in ways that are very specific
[23:06] to your use.
[23:08] Now, this situation is changing.
[23:10] Gradually, these large users, and
[23:12] particularly these hyperscalers, are
[23:13] shifting from owning to renting their
[23:15] data centers. Why is that? Well, as I'm
[23:18] showing you on the next slide, some of
[23:20] it is just too expensive to keep doing
[23:22] all of this on their own balance sheet.
[23:24] The capex needs are huge, and they are
[23:26] running out of operational cash flow.
[23:28] The second reason is they're in a
[23:29] competitive race. They need to get their
[23:31] hands on as much compute as fast as
[23:33] possible. And sometimes if there's
[23:35] available available capacity, it's just
[23:37] easier to lease an existing data center
[23:39] than to start building your own.
[23:41] The third reason is that they may want
[23:42] to preserve financial flexibility for
[23:44] future IT capex. Remember, they're going
[23:47] to be building new data centers in 2028
[23:49] and 2029, and typically they will own
[23:51] that IT on their balance sheet. And so,
[23:53] they need to preserve some fiscal room,
[23:55] some financial space for that.
[23:57] And fourth, and related to that,
[24:00] it's possible that in the future their
[24:01] compute needs change, the architecture
[24:03] of the compute changes. And so,
[24:05] you know, having somebody else do the
[24:06] investment hedges that uncertainty.
[24:10] Enough. And and so, these you could
[24:12] think of these third-party landlords
[24:14] that are now the the landlords that are
[24:16] renting that space to the hyperscalers,
[24:18] they are the residual claimants on the
[24:20] hyperscalers' demand for compute.
[24:21] Because imagine the compute contract at
[24:23] some point, these hyperscalers still own
[24:25] some of their own facilities as well,
[24:27] and they're going to prioritize traffic
[24:29] towards their owned facilities and shift
[24:31] it away from the rented facility. Right?
[24:33] So, these landlords are sort of the
[24:34] residual claimants on that compute.
[24:37] Can you say a few words how quickly this
[24:39] the value of these chips depreciates
[24:41] over time?
[24:42] Yeah, we can we can get into that, but I
[24:44] think nobody knows. I think
[24:48] typically, they are depreciated at least
[24:50] for accounting purposes over a period of
[24:52] like 4 to 6 years. This actually has
[24:54] been lengthening a little bit over time
[24:55] from 4 to 6.
[24:57] Uh now the question is what is economic
[24:59] obsolescence? If every year Nvidia comes
[25:01] out with a great new chip that's better
[25:03] and people always want the latest and
[25:04] greatest chip. It's questionable whether
[25:06] that you can whether that economical
[25:07] life is really four to six years even if
[25:09] these chips are physically able to keep
[25:11] computing for that period. Um that's
[25:14] sort of the first point. The second
[25:15] point is we are now in a period of
[25:17] extreme extreme chip scarcity. So right
[25:20] now it is true that all these old chips
[25:22] are still churning. And so people in the
[25:24] industry will tell you, "Oh, these chips
[25:26] are going to be working for a very long
[25:28] time." cuz that is their current
[25:29] experience. It's not clear whether that
[25:31] will still be true once there's more
[25:33] abundance of these chips in the future.
[25:36] Mhm. Now this shift from owning to
[25:38] renting is happening on the real estate
[25:40] side.
[25:41] It's potentially also happening on the
[25:42] IT side, right? So traditionally the
[25:44] hyperscalers use all of the own all of
[25:46] the IT.
[25:48] But there are new financial structures
[25:49] which we'll get into in a second that
[25:51] make it easier to also start renting
[25:54] that IT. And have somebody else own it,
[25:56] right? So this IT capex may go a similar
[25:59] route and we're in an earlier phase of
[26:00] that but it's also going that route of
[26:02] basically increasingly becoming rented
[26:05] rather than owned.
[26:06] So let's take a quick look at the
[26:08] hyperscaler capex needs. This is the
[26:10] number that Mark cited at the beginning
[26:12] of the talk where we see that about 655
[26:15] billion on-balance-sheet capex spending
[26:17] is is is projected for 2026 by these
[26:20] five large hyperscalers. Right? And
[26:23] that's up 58% from last year. So these
[26:25] are massive numbers. I think it's useful
[26:28] to put those into context and compare
[26:30] them to the operating cash flow of these
[26:32] companies. Right? And so that's the
[26:34] shaded the light shaded bars next to the
[26:37] capex bars from the previous slide.
[26:39] And what you can see is that the ratio
[26:41] of these two, which is the red line
[26:43] plotted against the right axis, has been
[26:45] trending up over time. So it used to be
[26:48] that capex consumed about 40% of
[26:51] operating cash flow. today. That number
[26:52] is approaching 100%. So, in other words,
[26:56] the hyperscalers are running out of
[26:58] cash. They're running out of free cash
[27:00] flow, which is the difference between
[27:01] those two numbers. And in fact, this
[27:03] year, for example, Amazon is going to
[27:05] have for the first time in a very long
[27:07] time negative free cash flow. So, and
[27:09] that is already true for for Oracle, for
[27:11] example. And so, these these companies
[27:14] are quickly running out of cash the
[27:15] cash. They are going to have a lot of
[27:17] future uh IT investments to do. If we
[27:20] assume they want to keep doing this on
[27:21] balance sheet, they're going to need to
[27:22] preserve financial flexibility. That's
[27:24] why they're outsourcing the building
[27:27] part of the data center uh
[27:29] infrastructure and and and and and
[27:31] building to third-party landlords.
[27:34] Now, the other reason why this works is
[27:36] because these hyperscalers are actually
[27:37] very good tenants, right? So, if you
[27:39] sign a lease in real estate, a lease is
[27:41] a hard obligation. It's like It's like
[27:43] issuing a corporate bond. You cannot
[27:45] just walk away from your lease. It's a
[27:46] hard obligation. So, that's why
[27:49] landlords like tenants with good credit
[27:51] that are not going to go bankrupt. And
[27:53] so, that's usually reflected in their
[27:54] strong credit rating. So, we call a
[27:57] tenant a credit tenant if they have a
[27:59] good credit rating. That means they're
[28:00] not going to go bankrupt. They're going
[28:01] to be uh paying their lease. Now, guess
[28:03] what? These hyperscalers have very
[28:05] strong credit ratings. They're all
[28:06] investment grade rated. Um and uh that
[28:10] makes for great tenants.
[28:12] Now, what's interesting is
[28:13] these these hyperscalers are the ones
[28:15] signing the leases, but the actual
[28:17] compute often times is not is actually
[28:20] coming from a different source. It's
[28:21] coming from the uh AI model developers,
[28:24] like the OpenAI's and the Anthropic's of
[28:26] this world. Now, those are startups.
[28:28] They are not credit tenants. They do not
[28:30] have a credit rating. They're in fact
[28:32] cheeky credits. And so, they cannot by
[28:35] in and of themselves essentially lease
[28:37] that data center facility because they
[28:39] don't have the stable cash flows that
[28:40] would make a landlord comfortable uh
[28:42] signing the lease. And so, what they do
[28:44] is is sign contracts with the
[28:46] hyperscalers and access essentially
[28:48] renting the credit rating of the
[28:50] hyperscalers in order to access this
[28:53] this this compute these data centers.
[28:55] Okay.
[28:57] But also of course the hyperscalers
[28:59] could also do some bond issuance and buy
[29:01] the whole thing but they're rather than
[29:03] buy issuing bonds and buying it they
[29:05] prefer to rent it and pass it on to some
[29:08] landlord.
[29:08] >> No, exactly. They could and we'll come
[29:09] back to this. I mean there's an
[29:10] interesting question of why they don't
[29:12] do this and I'll I'll have sort of a
[29:13] very specific and concrete answer for
[29:15] you for you on that and that's one of
[29:17] the key points
[29:19] in this in this in this discussion.
[29:21] So let's have a look at the overall
[29:23] capital
[29:25] that is needed to to finance the the
[29:28] compute needs of these hyperscalers. So
[29:29] these are some numbers that come from
[29:31] Morgan Stanley. They just focus on the
[29:33] four years between 2025 and 2028 and
[29:36] they ask how much would it cost to build
[29:37] all the compute capacity that these five
[29:40] hyperscalers will need over those
[29:41] period. And the answer that they come up
[29:43] with is about $3 trillion similar to the
[29:45] numbers we discussed before. About $1.3
[29:47] trillion is for data centers, $1.6
[29:49] trillion for the IT. And they ask who's
[29:52] going to put up that capital? And the
[29:53] answer they come up with is it's going
[29:54] to be about
[29:56] 50% internally financed by the
[29:58] hyperscalers and 50% externally
[30:00] financed. In the good old days this
[30:02] would have been 100% internally financed
[30:04] but these days are over. And so it's
[30:08] going to be 60% equity in total, 40%
[30:11] debt.
[30:12] That 60% equity again is mostly
[30:15] internal equity. That's the big black
[30:17] bar here but it's a little bit of
[30:18] private equity as well, maybe also some
[30:20] money from sovereign wealth funds for
[30:21] example. And then the debt part is going
[30:24] to be mostly private credit. They think
[30:26] about $800 billion will come from
[30:27] private credit, $200 billion will come
[30:30] from corporate bonds and $150 billion
[30:33] will come will come from securitized
[30:34] debt like ABS and CMBS.
[30:37] Now it's important to remember that the
[30:39] hyperscalers are usually owning the IT
[30:41] on balance sheet, right? So, they want
[30:43] to they own the IT. And so, most of the
[30:45] internal equity will actually go towards
[30:47] financing IT. And much of the external
[30:50] debt will go towards financing the data
[30:51] centers. So, by by shifting some of the
[30:54] debt some of the financing externally
[30:56] through debt, what that's allowing you
[30:58] to do is actually get much more leverage
[31:00] on the on the real estate part on the
[31:02] data center building and power
[31:04] infrastructure part. Where while the
[31:06] overall leverage in the system here
[31:08] looks low, it looks it's 40%. The the
[31:10] data center leverage will be much closer
[31:12] to 70% because most of the IT will be
[31:15] equity financed and most of the real
[31:17] estate will be debt financed.
[31:20] So, let's have a look at the different
[31:22] parts of this of this capital stack
[31:24] especially on the debt side. So, one
[31:26] important source will be corporate debt.
[31:28] And sure enough, the hyperscalers issued
[31:30] a record amount of corporate debt in
[31:33] just the fourth quarter of 2025. You can
[31:35] see the light blue bar just jumping up
[31:38] massively about 150 billion dollars in
[31:42] in corporate debt issuance. Right? So,
[31:43] some of it is coming through traditional
[31:46] unsecured corporate debt on the balance
[31:48] sheet of these hyperscalers.
[31:51] But, increasingly
[31:52] >> What's the maturity of this debt? Is it
[31:53] other sources? Yes.
[31:56] Do you know what the maturity of this
[31:57] debt is? Is this long-term debt?
[31:59] So, typically long-term debt and I'll
[32:01] give you a very concrete example of it
[32:03] in a second where it will be essentially
[32:05] 25-year debt.
[32:08] So, could be long-term debt. Typically,
[32:10] when we're going to build these new
[32:11] facilities and we're going to build them
[32:13] bespoke for a particular hyperscaler, we
[32:16] we are going to want to
[32:18] you know, match the maturity of that
[32:20] building roughly with the maturity of
[32:22] the debt.
[32:23] And so, these are the most important
[32:24] >> best use of this building if you don't
[32:27] use it for
[32:28] IT? Can you use it for anything else or
[32:30] that's
[32:31] >> Well, look, it's a storage facility. You
[32:33] could use it as a logistical warehouse.
[32:35] That obviously would have much lower
[32:36] value. It A lot of it depends on where
[32:39] it's located. If it's an inference
[32:40] facility that's located close to a
[32:42] population center, uh, it might be worth
[32:44] a lot more than if it's somewhere in the
[32:46] middle of nowhere in Louisiana. Okay?
[32:48] But, let's agree that's sort of the
[32:50] residual value of that building if it's
[32:53] not going to be used for computing is is
[32:55] very low relative to its current value.
[32:59] So, one of the new sources of finance is
[33:02] is securitized finance. And of course,
[33:04] we're very familiar with mortgage-backed
[33:06] securities from the good old
[33:08] uh, GFC days. Uh, but here we're talking
[33:10] about uh, data center backed
[33:13] facility. So, CM There's two flavors.
[33:15] There's CMBS and there's ABS. So, CMBS
[33:18] is collateralized by the mortgages on
[33:20] these data centers, right? So, this this
[33:22] is usually when the data center is fully
[33:24] built, it's fully leased up, it's it's
[33:26] it's basically now generating a stable
[33:28] flow of cash flows. That's when we
[33:30] finance it with permanent debt, with a
[33:32] long-term mortgage, and then we can put
[33:34] a bunch of these long-term mortgages
[33:35] into a special purpose vehicle and sell
[33:37] claims on that.
[33:39] Typically, however, there is very little
[33:41] diversification in these CMBSs. There's
[33:43] usually only a single tenant, and these
[33:45] facilities are so large that often there
[33:47] is a single borrower in these
[33:49] facilities. So, these are called SASB
[33:51] deals, single asset single borrower. It
[33:53] The collateral is usually either a
[33:55] single data center or a single borrower.
[33:58] Could be a REIT that owns a couple of
[33:59] different data centers, for example, and
[34:01] puts that in one CMBS trust, but usually
[34:04] the tenants are high quality, but
[34:05] there's very little diversification
[34:07] among the tenants.
[34:10] A second and and and that and that
[34:12] size of that market just jumped from $3
[34:14] billion to $11 billion last year, and
[34:17] it's expected to grow again very
[34:19] strongly in 2026.
[34:21] The second flavor of this is
[34:23] asset-backed
[34:24] structures, where the key difference is
[34:25] that the underlying collateral is not a
[34:27] mortgage. It's basically a claim to to
[34:29] cash flow, not to collateral value. And
[34:32] so, this could be for example, you sign
[34:34] a compute contract and that's going to
[34:36] generate cash flow you can then
[34:37] collateralize that cash flow. Or it
[34:39] could be you have GPUs and you get a
[34:42] loan to finance these GPUs from a bank
[34:45] and then the bank puts that loan that is
[34:47] collateralized by the GPUs into an asset
[34:50] back security. Okay? That sort of
[34:52] activity has also grown strongly to
[34:54] about 15 billion dollars last year and
[34:55] again as we've already done 5 billion
[34:58] dollars of this in just the first few
[35:00] months of 2026.
[35:02] Related to this
[35:06] Excuse me? Numbers from the global
[35:08] financial crisis. Do you have a
[35:10] reference point for that?
[35:11] So
[35:12] >> small?
[35:12] >> This is still very small, right? Because
[35:14] remember in terms of residential
[35:16] mortgage back securities, subprime
[35:18] mortgages at the peak in 2007,
[35:21] we had about 2 trillion dollars of
[35:25] issuance of these subprime mortgage back
[35:27] securities. So this is
[35:28] still several orders of magnitude
[35:30] smaller. Nevertheless, it's growing
[35:32] quickly and it's just one part of the of
[35:34] a complex capital stack.
[35:36] So one new development is these GPU
[35:38] backed securities where now there's a
[35:42] basically a special purpose vehicle that
[35:43] owns the GPUs itself, right? So it's it
[35:46] owns the GPUs and that's that's what it
[35:48] has and then it basically leases out
[35:50] these GPUs. You could imagine a
[35:51] hyperscaler buying GPUs and then doing a
[35:53] sale and leaseback.
[35:55] Uh
[35:56] you know, say and and basically sell
[35:58] them to a to another company and then
[36:00] basically the the the cash flows from
[36:02] this from this lease would be servicing
[36:04] the debt, the GPU based debt. And this
[36:07] is one of the ways in which I think this
[36:09] in the future, it's already beginning to
[36:11] happen. But this is an increasing tool a
[36:13] growing tool for these hyperscalers to
[36:15] move also some of the IT capex off their
[36:18] balance sheet. They've already been
[36:19] massively moving the real estate
[36:22] exposures, ownership off their balance
[36:24] sheet and leasing the the buildings. Now
[36:26] they're going to increasingly be leasing
[36:27] the GPUs as well.
[36:29] Right? Now, when we think about the risk
[36:31] in those GPU backed securitizations,
[36:33] well, you know, there's obviously severe
[36:36] technological obsolescence risk. Right?
[36:38] We just talked earlier about the
[36:40] lifespan of these GPUs, maybe it's only
[36:42] 4 to 6 years. Often times the the pool
[36:45] of these is very correlated with each
[36:47] other, there's very limited
[36:48] diversification. And you know, we have
[36:51] sort of the risk of of quick of quick
[36:53] obsolescence potential.
[36:55] But nevertheless, this is a tool that is
[36:57] increasingly being used.
[36:59] And then the largest source of credit of
[37:01] debt in this space is private credit,
[37:03] right? Private credit is sitting on a
[37:05] vast amounts of money. They just raised
[37:07] $1.3 trillion of private credit in the
[37:10] last few years. They're trying to deploy
[37:12] 300 billion, let's say, every year.
[37:15] So far in the data center space, they've
[37:17] deployed about $200 billion
[37:19] of private credit. And the reason this
[37:22] is so good is that this is a match made
[37:24] in heaven. Data centers and private
[37:26] credit are a match made in heaven. The
[37:28] reason is they're so very capital
[37:30] intensive, and so they need a lot of
[37:32] money, and private credit is sitting on
[37:33] a lot of money.
[37:35] Second,
[37:36] um private credit is not averse to a lot
[37:38] of structural complexity. These are
[37:40] sophisticated funds that have a lot of
[37:41] freedom in what they do. But they like
[37:43] to lend to investment grade credit,
[37:45] which is exactly what these hyperscalers
[37:47] are.
[37:48] So, we're going to see a lot more of
[37:50] this. Morgan Stanley predicts $800
[37:51] billion of private credit will flow into
[37:54] into data centers, and it's and it's
[37:56] already happening sort of at full speed.
[37:59] Um one thing to keep an eye on is that
[38:01] this private credit space, as it comes
[38:03] to data centers, is heavily
[38:04] concentrated. There's essentially five
[38:06] companies that uh have a 70% market
[38:09] share. Blue Owl is the largest,
[38:10] Blackstone is the second largest. We
[38:12] have the usual suspects, Apollo, KKR,
[38:14] and Brookfield, uh that are very active
[38:16] in this space.
[38:19] So, now that we understand not only the
[38:21] large sums of money, but also where the
[38:23] capital is coming from. I think it's a
[38:25] good time to ask, you know, is this new
[38:27] arrangement, where the risks is being
[38:29] spread through private credit, through
[38:31] asset back securities, are there are
[38:33] there potential issues with that? Are
[38:35] there potential vulnerabilities that
[38:37] come from this as it comes to potential
[38:40] increases in leverage and potential
[38:42] increases in opacity?
[38:44] And I think it would be very useful to
[38:45] do a little case study here, because I
[38:47] think it's very illustrative. And this
[38:49] is not just a random case study. This is
[38:51] the largest data center financing deal
[38:53] we have seen
[38:55] to date. And this could very well become
[38:57] a landmark deal for how future data
[39:00] centers will be funded. So, let me give
[39:01] you a little bit of background here. So,
[39:04] this is a large data center facility
[39:06] that Meta was building in Richland
[39:08] Parish, Louisiana, in the middle of
[39:10] nowhere. It's a gigantic facility. It's
[39:12] 2 gigawatts. Remember, earlier my
[39:15] example was 200 megawatts is a very
[39:17] large state-of-the-art facility. This is
[39:19] essentially 10 of those on one campus.
[39:22] Um potentially in the future it'll be
[39:24] future further expanded to 5 gigawatts.
[39:27] And let's just focus on this first
[39:29] stage, this 2 gigawatts. That's supposed
[39:31] to come online in 2029. They're building
[39:33] it as we speak.
[39:35] Now, Meta initially developed the began
[39:39] this this project as the 100% owner of
[39:41] this facility.
[39:42] Uh
[39:43] and but it's changed gears. In
[39:46] the fall of last year, it basically sold
[39:48] an 80% stake in this in this project to
[39:51] Blue Owl, the private credit firm we
[39:53] just talked about.
[39:55] Is this a training center or
[39:57] inference
[39:59] center? This will be mostly a
[40:02] a training facility. I see. It's in the
[40:04] middle of nowhere, so it's not going to
[40:06] be very good as an inference facility.
[40:08] It will be mostly a training facility.
[40:10] I see. So, Blue Owl paid $2.5 billion
[40:13] for a 20% equity stake.
[40:15] I I For sorry, for an 80% equity stake
[40:18] and Meta retained 20% equity.
[40:20] And then in October on October 16th,
[40:23] this facility So, they formed a joint
[40:25] venture, Blue Owl and Meta, which was
[40:26] called Binyan.
[40:28] And and it's a JV, it's a special
[40:30] purpose vehicle. There's nothing else in
[40:31] it except for this data center.
[40:33] This special purpose vehicle, Binyan,
[40:35] issued 27.3 billion dollars in debt.
[40:39] This was the single largest investment
[40:41] grade bond ever issued in the United
[40:43] States history.
[40:45] It was rated A+, so single A,
[40:48] by Standard & Poor's. It only had one
[40:51] credit rating, which is very unusual for
[40:52] such a large bond. And that single A
[40:54] rating is only one notch below Meta's
[40:56] own corporate credit rating.
[40:59] This bond has lots of other special
[41:01] features. To your earlier question,
[41:02] Mark, this is a long-term bond. It's
[41:04] maturity date is 2049, which is
[41:07] basically the same end date as the
[41:09] lease. We'll talk about the lease in a
[41:11] second. This bond is also amortizing,
[41:13] which is very unusual for large
[41:15] corporate bonds. And its coupon was
[41:17] 6.58%. Now, I did some calculations and
[41:19] I convinced myself that this 6.58%
[41:22] is about 120 basis points higher
[41:25] than the yield to maturity that Meta
[41:28] would have paid had it financed that
[41:29] security with unsecured corporate bonds.
[41:32] I mean, cuz I looked at some recent
[41:33] unsecured corporate bonds.
[41:35] And that 100 So, this it's expensive
[41:37] capital.
[41:41] That 120 basis points, if you do the
[41:43] math, amounts to 5.6 billion dollars in
[41:46] extra interest payments over the life of
[41:48] the loan.
[41:49] Mhm.
[41:50] The other very important thing to
[41:52] realize here is the enormous leverage in
[41:54] this structure, right? So, there's 27
[41:56] billion dollars of debt in 30 billion
[41:59] dollars of asset value. So, that's a 90%
[42:02] leverage ratio. It's fair to say that if
[42:04] Meta has done had done this on balance
[42:06] sheet, they would have never leveraged
[42:07] it 90%.
[42:09] In addition,
[42:10] >> Who's buying these bonds? We'll talk
[42:12] about this in a second. These bonds.
[42:13] We'll talk about this in a second. So,
[42:16] the the On addition, there's enormous
[42:18] structural complexity here. I'm not
[42:19] going to go through this diagram. This
[42:21] diagram is just meant to show you and it
[42:23] should remind you of some of the
[42:25] diagrams we saw after the GFC
[42:27] of the very complicated
[42:29] interactions between all the different
[42:31] LLCs and the layers upon layers of
[42:33] of of LLCs.
[42:35] So,
[42:36] suffice to say this structure is
[42:38] bankruptcy remote from Meta's balance
[42:41] sheet, which means that the investors
[42:42] and bondholders in this bond have no
[42:44] recourse to
[42:46] to Meta. They only have recourse to the
[42:48] balance sheet of this Bene SPV.
[42:51] Okay. So, let's talk a little bit about
[42:53] who who bought these bonds to Marcus's
[42:54] question. So, first of all, when these
[42:57] bonds were issued, they were par bonds.
[42:58] The price immediately jumped from $100
[43:01] to $110,
[43:03] which meant that the yield came in,
[43:05] you know, quite a bit, about 70 basis
[43:07] points or 80 basis points.
[43:09] PIMCO
[43:10] was actually had bought 18 billion of
[43:13] that 27 billion dollars of debt. And
[43:15] PIMCO immediately sold these bonds and
[43:17] basically realized that 10% capital
[43:19] gain. So, they made it they real not
[43:22] only on paper, but in actuality made $2
[43:24] billion in profit from this transaction.
[43:28] Even today, this bond this bond price
[43:30] has just fell a little bit and just a
[43:31] few days ago this bond was trading at
[43:32] $106. So, investors are still believing
[43:36] that this is a valuable bond. It's not
[43:38] It's still trading above par. Um and
[43:42] you know,
[43:43] you know, and that means that
[43:44] >> bought it right away and sold it, but
[43:46] who bought it from PIMCO? Do you know
[43:47] who bought it?
[43:48] >> know. We do not know who bought it from
[43:49] PIMCO. That's a good question.
[43:52] And this is one of my points. We would
[43:53] we would To understand who holds the
[43:55] risk, we would need to be able to answer
[43:57] those type of questions. Who holds this
[43:58] bond now that PIMCO sold it to? Maybe
[44:01] through 13F filings we can sort of
[44:02] unravel this depending on who it is, but
[44:05] this would be a sort of a lot of
[44:07] detective work because none of this
[44:08] stuff is is reported.
[44:10] There's other interesting
[44:11] um
[44:12] I don't want to call them obfuscations,
[44:14] but other interesting sort of uh
[44:15] disclosure limitations here because this
[44:17] bond is a 144A private placement. Um
[44:20] it's never going to be part of
[44:21] Bloomberg's corporate or aggregate bond
[44:23] indices even though it's the largest
[44:24] bond ever issued. And because it's a
[44:26] 144A for life issue
[44:29] uh that was limited distributed, which
[44:31] means only a few investors got to buy
[44:33] it. It does not need to re report
[44:35] annually to the SEC. So, the structure,
[44:38] normally there would be annual
[44:39] reporting, but here there's not because
[44:41] of this limited distribution feature.
[44:43] So, there's a lot of structural
[44:44] complexity here that sort of limits the
[44:46] transparency uh in this bond.
[44:50] Now, the next point is is I think
[44:52] interesting and very important. So,
[44:54] let's think about Meta. So, Meta
[44:56] was still the developer of this of this
[44:58] data center even though they only own
[45:00] 20% of it. They bear all the
[45:02] construction risk contractually. All the
[45:04] cost overruns, all the time delays. If
[45:06] the GPUs don't come in, it's Meta's
[45:07] problem.
[45:09] Meta is also the only tenant. And it
[45:11] signed not one 20-year lease, but five
[45:14] four-year leases. Now, you may say,
[45:16] "What's the difference? Five four-year
[45:17] leases or 20 one 20-year lease?" Well,
[45:19] there is a big difference.
[45:21] Right? These four-year leases were
[45:22] structured to coincide with sort of
[45:24] roughly the life of an IT cycle. This is
[45:26] sort of, you know, what we talked about
[45:27] before.
[45:28] And every four years Meta essentially
[45:30] has the option to scale down. There's
[45:31] nine of these buildings. Each one has
[45:33] their own lease. They can say, you know,
[45:35] four years from now or eight years from
[45:36] now they can say, "We only want seven of
[45:38] the nine buildings anymore." Mhm.
[45:40] Now, if they do Obviously, nobody would
[45:42] build this data center with such a loose
[45:44] commitment from a tenant.
[45:46] Built data center builders need
[45:48] long-term tenant commitments. So, how
[45:50] did they get around that? They promised
[45:52] what's called a residual value
[45:53] guarantee. So, the idea is that if Meta
[45:55] decides they don't want one of these
[45:57] buildings anymore in four years,
[45:59] then that will automatically trigger a
[46:00] sale process for that building and then
[46:02] will be some realized market value for
[46:04] that building from selling it.
[46:06] And Meta will owe the difference between
[46:09] a minimum guaranteed value number that
[46:11] is contract that is written in the
[46:12] contract and the sale value. So, if the
[46:15] mark if the building is not worth so
[46:16] much anymore, then Meta will owe the
[46:18] difference. And these minimum guaranteed
[46:20] value amounts are set such that the bond
[46:23] can be paid off at those points in time.
[46:25] So, in theory, there's enough money
[46:27] coming in even if Meta exits the lease
[46:29] that the bond can still be paid back.
[46:31] But, of course, there's counterparty
[46:32] risk. What if Meta is not around 8 years
[46:34] from now or 12 years from now?
[46:37] So,
[46:38] so
[46:39] this structure of signing these
[46:42] successive 4-year leases and making this
[46:45] residual value guarantee has some very
[46:47] interesting accounting implications. It
[46:50] turns out that under GAAP accounting,
[46:53] because these lease renewals are not at
[46:55] least 70% likely,
[46:58] Meta need not put these future lease
[46:59] obligations on its balance sheet. Only
[47:02] the first lease, not the future ones.
[47:04] But, because the residual value
[47:06] guarantee is also not very probable,
[47:08] Meta also does not need to account for
[47:10] the residual value guarantee on its
[47:12] balance sheet.
[47:13] So, that uh that sort of logic that GAAP
[47:15] And these are the rules under GAAP. Meta
[47:17] is not cheating here. These are the
[47:18] rules under GAAP.
[47:20] Because of that,
[47:21] the credit credit rating agency that
[47:23] rated this deal, S&P, decided not to
[47:25] include either obligation in their
[47:27] normal calculation of adjusted debt. So,
[47:29] the Why does the rating agency follow
[47:31] this GAAP rule? That that makes no
[47:33] sense.
[47:34] It doesn't, but that's the rule. So, in
[47:36] other words, either of these two
[47:37] obligations will materialize with
[47:39] probability one,
[47:40] but neither scenario is reflected on
[47:43] Meta's balance sheet. So, I call this a
[47:44] probability vacuum, right? Now, you may
[47:47] say, "Is that a big deal?"
[47:49] Well, let's have a look at the numbers.
[47:51] This is from a very interesting Moody's
[47:52] report that came out in a couple of
[47:54] weeks ago.
[47:56] Mhm. The five hyperscalers,
[47:57] collectively, have about a trillion
[48:00] dollars in lease obligations.
[48:02] Only 1/3 of which is actually reported
[48:05] on their balance sheet. So, 2/3 of it,
[48:08] $660 billion,
[48:10] are lease obligations that they have
[48:12] already signed but not yet begun, which
[48:15] are not reflected on their balance
[48:16] sheet. They will be reflected on the
[48:18] balance sheet if and when they begin,
[48:20] but right now they're nowhere to be
[48:22] found.
[48:23] So, let me summarize.
[48:25] Data centers have been around for a long
[48:26] time, but they've gotten a lot bigger
[48:28] and a lot more expensive.
[48:31] The data the the hyperscalers used to
[48:33] finance these things on balance sheet.
[48:34] They can no longer do this. It's too
[48:35] much, and moreover, if they did, it
[48:38] would potentially increase their
[48:40] leverage,
[48:41] uh threaten their credit rating, and
[48:43] more importantly, threaten their high
[48:45] equity valuation multiples. These
[48:47] companies trade in the stock market
[48:49] today like growth companies, like
[48:50] software companies,
[48:52] not like infrastructure companies, which
[48:54] would have much lower PE multiples.
[48:56] So, if they were to move some of that
[48:59] CapEx on balance sheet, their multiple
[49:02] would compress, and that would have a
[49:04] massive impact on the dollar value of
[49:06] equity,
[49:07] much more than the extra few billion
[49:09] dollars of extra financing cost off
[49:12] balance sheet.
[49:13] And so, that is fundamentally the reason
[49:15] why they make this this this this this
[49:18] this decision to essentially move a lot
[49:20] of the CapEx off balance sheet.
[49:22] Now, the the the the repercussions of
[49:24] all of these are severe. First of all,
[49:27] it allows for much higher leverage off
[49:28] balance sheet. Think about the 90% LTV
[49:31] on Hyperion. It uh it leads to this no
[49:34] accounting recognition of future leases.
[49:36] It's a form of a regulatory arbitrage.
[49:39] Right now, there's no impact on their
[49:40] credit rating, and they get the
[49:42] optionality of potentially not renewing
[49:44] their future leases. So, you can think
[49:46] of this residual value guarantee as a
[49:48] bundle of put options, uh but and then
[49:51] they're expensive, but at least there's
[49:52] some optionality there that's valuable
[49:54] to the hyperscalers. And then finally,
[49:56] this complexity of this Hyperion deal
[49:59] sort of illustrates that, you know,
[50:00] we've added a lot of opacity to the
[50:02] system and it becomes unclear who's the
[50:04] actual owner of the of of the underlying
[50:07] risk.
[50:08] So, in my
[50:09] >> requirement that the rating agencies has
[50:10] to follow the GAAP rules. No, that's was
[50:13] their decision, in a sense.
[50:14] >> And you know, I think the rating
[50:15] agencies are internally debating this
[50:16] whether they want to change this method
[50:18] and and sort of maybe uh adjust that for
[50:20] at least probabilistically for those for
[50:23] one of those scenarios. I think it would
[50:24] be a good a good thing to do, in my
[50:25] opinion.
[50:27] So, the final thing I want to do in my
[50:29] last few minutes is to think about now
[50:30] that we know what the who sort of or or
[50:33] at least we sort of understand that the
[50:34] risks have become more dispersed, it
[50:36] seems important to understand sort of
[50:37] what are the risks that we're talking
[50:39] about. And I want to be clear, there's a
[50:42] a very easy bullish case to be made for
[50:45] the risks are maybe not so large, right?
[50:47] And this is what the practitioners would
[50:48] like you to believe. They would tell you
[50:50] the modern economy is data-centric, DCs
[50:52] are critical infrastructure, AI's going
[50:55] to be great, it's going to be ample
[50:56] revenue. Uh, every AI company has case
[51:00] use cases that are 12-mi long and
[51:02] growing. They would also tell you that
[51:04] we have because of that we have years of
[51:06] growth ahead both in compute and in new
[51:09] data center development. Especially how
[51:12] how hard it's been to build new data
[51:13] centers with all the shortages in power
[51:15] and in chips and in server racks and so
[51:17] forth.
[51:18] They would also tell you that right now
[51:19] the data center market is very strong,
[51:22] vacancy rates are at an all-time low,
[51:24] rent growth is at an all-time high that
[51:27] will offer cushions margins for the
[51:29] current owners of these data centers.
[51:31] And they will tell you that even though
[51:32] there's a lot of new building, there's
[51:34] no speculative development. In real
[51:35] estate, we always worry about
[51:37] speculative development, by which we
[51:38] mean building something before you know
[51:41] who's going to occupy it. Right? That's
[51:44] not happening here. Nobody's putting a
[51:45] spade in the ground before they have a
[51:47] long-term lease in hand.
[51:49] And so, you know, not only do they have
[51:51] a lease, they have a lease from a credit
[51:52] tenant, from some of the strongest
[51:54] balance sheets on Earth.
[51:56] And so, that, and you know, arguably
[51:58] creates predictable cash flow. As long
[52:00] as these hyperscalers don't default, um
[52:03] you know, that's money good.
[52:05] Now, the data center people will go
[52:06] further and they will tell you, even if
[52:08] one of those major tenants were to
[52:09] default, compute is compute. There's so
[52:12] much potential other potential users for
[52:14] these data centers and for these GPUs
[52:16] that are currently, uh you know, crowded
[52:18] out that will gladly take their place.
[52:21] I think you can also make the academic
[52:23] case that um equity and debt investments
[52:26] in AI and data center companies are a
[52:28] hedge against the displacement risk of
[52:30] AI. Right? We all hear a lot about the
[52:32] labor market disruptions. We also know a
[52:34] lot about the other business sector
[52:36] disruptions. Think about the sell-off in
[52:38] software as a service a few weeks ago.
[52:41] AI and data centers hedge those risks.
[52:44] And then, we also know from research, a
[52:46] lot of long line of research coming out
[52:48] after the financial crisis, that there's
[52:50] certain logic to structured finance and
[52:52] to um structural complexity and this new
[52:54] financial architecture that I sketched,
[52:57] uh you could defend it on the basis
[52:59] that, you know, maybe we're shifting the
[53:01] risks to those entities who can best
[53:02] withstand those risks. Those could be
[53:04] unlevered pension funds, unlevered bond
[53:06] mutual funds, private credit funds. Many
[53:08] of these invest their money on behalf of
[53:10] wealthy investors. If they blow up, it's
[53:12] a bunch of rich people who lose their
[53:13] money. It's better than if those risks
[53:15] reside in, for example, a highly
[53:17] leveraged
[53:17] >> don't know who is holding this stuff.
[53:19] But that's I'm making the bullish case
[53:21] here. Okay? It requires knowing who
[53:23] holds that risk and sort of it requires
[53:25] that the holders, the ultimate holders,
[53:26] are these unlevered institutions. Yes.
[53:29] Now, I think there's
[53:31] a lot of risks, and even though the
[53:33] bullish scenario might be a likely
[53:34] scenario, it's incumbent upon us to
[53:36] think about those risks and think those
[53:38] through and think about the
[53:39] repercussions of it. So, there's a few I
[53:41] have in mind. One is that the
[53:43] hyperscalers face rising costs of
[53:45] capital. And this is sort of the
[53:47] scenario where CAPEX keeps growing
[53:48] relative to operational cash flow. The
[53:50] hyperscalers decide to keep funding IT
[53:52] at least on their balance sheet. And
[53:54] potentially these future lease
[53:55] obligations that we talked about
[53:57] eventually come on balance sheet. So,
[53:59] all of a sudden these hyperscalers are
[54:01] much more levered. That will increase
[54:02] their cost of debt. Maybe it will
[54:04] trigger a credit rating downgrade. It
[54:06] will also increase their cost of equity
[54:07] cuz their multiples will compress. These
[54:09] hyperscalers will become more like
[54:10] infrastructure firms, less software
[54:12] firms. And that in turn will increase
[54:14] their cost of leasing these facilities
[54:16] cuz they will not be as good a credit as
[54:17] they were before. Now, all of this
[54:19] ultimately depends on the productivity
[54:22] of that investment, right? The return on
[54:24] that investment. How fast can the
[54:25] revenue from AI actually grow?
[54:28] But I don't think we know that because
[54:30] right now we're subsidizing lots of AI
[54:33] use with free accounts for the most
[54:34] part. And we don't actually know what
[54:36] the demand is at a price that, you know,
[54:39] exceeds the marginal cost of compute.
[54:41] Uh the other point about revenue is that
[54:43] if the economic life of these
[54:44] investments is of the IT at least is
[54:46] maybe 5 years, then the revenue growth
[54:48] actually has to come in in the near
[54:50] term.
[54:52] The second risk is tenant credit risk.
[54:54] Many data centers have a single tenant.
[54:55] There's a lot of concentration risk.
[54:57] Could one of these hyperscalers implode
[54:59] potentially in the next 5 years? That
[55:01] could happen. The credit market, the CDS
[55:03] market seems to think that that
[55:04] probability has gone up in the last few
[55:07] months.
[55:08] Now, um one of the large AI model
[55:11] companies like OpenAI or Anthropic,
[55:13] which is a heavy user of of compute, uh
[55:16] could also uh become financially
[55:17] fragile. You know, think about Anthropic
[55:20] being banned by the US government
[55:21] recently. So, you have shocks like that
[55:23] that could potentially undermine the
[55:25] case.
[55:26] And and and and not so often talked
[55:28] about risk, which I think is important,
[55:29] is development and operational risks of
[55:31] these data centers, right? So,
[55:32] development is about I'm building the
[55:33] facility,
[55:35] uh you know, I'm already maybe uh issued
[55:37] a big corporate bond, but all of a
[55:38] sudden I get delayed for I you know,
[55:40] there's something going wrong with my
[55:41] power infrastructure. Maybe I have to
[55:43] build my own power plant, and that is
[55:45] increasingly the case. 50% of all the
[55:48] new data center capacity, the power is
[55:49] generated on site, not on the grid.
[55:52] There could be bottlenecks in IT
[55:53] infrastructure. Maybe I can't get
[55:56] the the chips or the memory and so
[55:58] forth. So, I've already issued bonds.
[56:00] Those bonds need to be repaid with the
[56:02] cash flows from the leases, but I
[56:03] actually cannot build the facility
[56:05] because I
[56:05] there are no GPUs to anywhere to be
[56:07] found. And so, how am I supposed to
[56:09] service my debt if there's no lease
[56:10] revenue because there is no chips?
[56:12] So, I think that's a potentially a risk
[56:14] that uh we need to think about. Um and
[56:17] then similarly on the operational side,
[56:18] the cost of electricity could go up
[56:20] meaningfully. Remember that comes out of
[56:21] the pocket of the tenant.
[56:23] Uh or things like greenhouse gas
[56:25] emissions or water usage rules could
[56:27] tighten in the future. Some of these
[56:29] facilities use millions of gallons of
[56:31] water per day.
[56:34] Fourth risk is collateral values,
[56:35] technological obsolescence risk. We
[56:37] talked a bunch about this already. It's
[56:38] obvious on the IT side. Um it's also
[56:41] important on the data center side. This
[56:43] cooling and power infrastructure could
[56:44] become obsolete if the technology
[56:46] changes, either because there's a lot
[56:48] more power density in the future, or
[56:50] there's just a radically new compute
[56:52] technology that emerges. Quantum
[56:54] computing, for example. There was an
[56:55] interesting article in the Wall Street
[56:56] Journal about it. Edge computing.
[56:59] Uh again, Apple, for example, has this
[57:01] vision that every cell phone can be
[57:03] their own little
[57:04] uh data center. And so, we don't need
[57:06] these big central data centers anymore
[57:08] if most people can do their most of
[57:09] their compute and their inference on
[57:11] their own phone.
[57:12] Uh we could have large in increases in
[57:14] chip efficiency, AI inference chips, for
[57:16] example.
[57:17] And then I think another important part
[57:19] of this is misallocation. We've We're
[57:21] now building a lot of these large
[57:22] massive AI training facilities, but what
[57:24] if we mostly need AI inference
[57:26] facilities? How easy is it to switch
[57:28] from one use to the next?
[57:30] Um you know, these places are these data
[57:32] centers are for example in different
[57:34] locations. So, you know, latency is an
[57:36] issue. Um,
[57:37] so that sort of limits conversion
[57:39] options.
[57:41] The last part is a bit more of a sort of
[57:43] a finance point that I want to make,
[57:44] which is if you look at data centers as
[57:47] an investment, data centers used to be
[57:49] this utility-like
[57:51] investment, low beta in the language of
[57:53] finance. And and what I'm plotting here
[57:56] is the the stock market beta of REITs,
[57:58] of data center REITs. And as you can see
[58:01] indeed the data, you know, this is based
[58:02] on 36-month rolling window regressions
[58:04] and a five-factor model. I'm just
[58:06] showing you the stock market beta.
[58:08] Indeed, data centers were trading at a
[58:10] beta of 0.5, like utility companies.
[58:13] But that beta has been rising to one.
[58:15] What that means is that these data
[58:17] centers have become heavily entangled
[58:19] with the rest of the economy, and that's
[58:21] maybe not surprising cuz a lot of the
[58:23] stock market is being driven by these
[58:25] by these AI companies, but data centers
[58:27] are part and parcel of that. So, they're
[58:29] not really sort of a good hedge
[58:31] against at least against the overall
[58:32] stock market, and they've become
[58:34] systemically much more risky than they
[58:36] used to be.
[58:37] And part of this is sort of the
[58:38] circularity in this in this sector where
[58:40] Nvidia makes
[58:41] strategic investments in OpenAI and in
[58:43] CoreWeave, and then OpenAI makes large
[58:45] revenue commitments to Oracle, which in
[58:47] turn promises to buy Nvidia. All of this
[58:50] is creating a lot of correlation in the
[58:52] system and essentially a very large
[58:55] principal component. If there's ever any
[58:57] shock to AI revenue risk, that whole
[58:59] ecosystem is heavily exposed to it.
[59:03] If you if you also thought that the high
[59:05] oil price does it and the electricity
[59:06] price might go up because of that, too.
[59:09] That's just a question.
[59:09] >> Yeah, so electricity prices are an
[59:11] important risk factor here because
[59:13] again, the tenants are bearing that
[59:14] risk, not so much the landlord. The
[59:16] landlord gets to pass on that risk to
[59:18] the tenants, but like all these
[59:20] hyperscalers would would suffer from
[59:21] that, absolutely.
[59:24] So, I'm almost done. Let me finish with
[59:26] a couple of reflections on on risk
[59:27] shifting. So, we talked about this
[59:28] notion that the financial risks of AI
[59:30] and data centers are shifting from the
[59:32] hyperscalers to these private pools of
[59:34] capital to private credit funds and pen-
[59:37] and pension funds and insurers that are
[59:39] buying these asset-backed securities and
[59:41] mortgage and mortgage-backed securities.
[59:43] And ultimately, households are standing
[59:45] behind those pension funds, right? They
[59:47] are the beneficiaries of these pension
[59:48] funds and the and the insurers. They're
[59:50] the investors in these credit funds.
[59:51] They're the depositors in the banks.
[59:53] And so, ultimately, that risk will
[59:55] potentially filter down to the
[59:56] household. We know that this
[59:58] originate-to-distribute model did not
[01:00:00] work so well in the great financial
[01:00:01] crisis. There's a recent academic paper
[01:00:03] that argues with some early evidence
[01:00:05] that that's the case for data centers as
[01:00:07] well.
[01:00:08] We also know that technological
[01:00:10] innovations like AI in general result in
[01:00:14] rising wealth inequality. The reason is
[01:00:16] that it you know, that basically, you
[01:00:19] know, some of the for contractual and
[01:00:21] incentive reasons, not all the risk can
[01:00:23] be shared, and the innovators are sort
[01:00:25] of disproportionately benefiting from
[01:00:27] those innovations. So, maybe AI stocks
[01:00:29] are a decent hedge against some of the
[01:00:32] risks for the average households, again,
[01:00:34] to the extent that they can access those
[01:00:36] investments, but they're definitely not
[01:00:37] a perfect hedge sort of by by that
[01:00:39] logic.
[01:00:40] So, interesting implications for wealth
[01:00:42] inequality here as well.
[01:00:44] So, let me conclude.
[01:00:46] What do we need to look out for as this
[01:00:48] story unfolds? Because we are early in
[01:00:50] this investment boom. And so, you know,
[01:00:53] as we look for signs of fragility, what
[01:00:55] should we look out for? We should look
[01:00:57] first and foremost towards top-line
[01:00:59] revenue growth from AI. Is it coming in?
[01:01:01] Is it growing quickly? Because it needs
[01:01:02] to grow very quickly to support all of
[01:01:04] this CapEx, and it needs to come in come
[01:01:06] in soon.
[01:01:07] The second point is we need to look for
[01:01:09] signs of growing leverage in the system
[01:01:11] and financial opacity. A lot of this
[01:01:14] will happen off balance sheet of the of
[01:01:16] the hyperscalers. It's no longer enou-
[01:01:18] enough to look at the balance sheet of
[01:01:19] the hyperscalers.
[01:01:21] We need to look at unaccounted lease
[01:01:22] commitments. We need to look at weaker
[01:01:24] loan underwriting over time. And then we
[01:01:26] need to always think about this risk
[01:01:28] circularity, this systematic risk
[01:01:30] because of all these entanglements of
[01:01:31] all the players in this industry. And
[01:01:33] then I think we need to look for signs
[01:01:35] of technological disruption. What if
[01:01:37] quantum computing makes a big jump and
[01:01:38] becomes a viable computational
[01:01:40] technology? What if we have much better
[01:01:42] chips? What if we have abundant power?
[01:01:44] How does that change this ecosystem? So,
[01:01:47] in conclusion, there's lots of
[01:01:49] interesting thing things to think about,
[01:01:50] not just for policy makers, but also for
[01:01:53] academics thinking about where where are
[01:01:55] the risks hiding. We don't have good
[01:01:57] visibility into that. We need much
[01:01:59] better efforts at at systematically
[01:02:02] collecting those data. We need better
[01:02:03] accounting frameworks. We need better
[01:02:05] risk models. We need better models for
[01:02:07] the valuation of that data center and IT
[01:02:09] collateral and depreciation.
[01:02:11] And then I think there's interesting
[01:02:12] questions in corporate finance about,
[01:02:15] you know, how do we think about the
[01:02:16] value of these hyperscalers when they
[01:02:17] become more physically capital
[01:02:19] intensive, but but not actually don't
[01:02:23] sort of potentially own but shift the
[01:02:24] physical capital to others while still
[01:02:26] becoming economic while still remaining
[01:02:28] economically dependent on that physical
[01:02:30] capital.
[01:02:31] Uh we need to think about financial
[01:02:33] stability. I think about sort of
[01:02:35] parallels with earlier infrastructure
[01:02:37] booms and often busts.
[01:02:39] And there's interesting local public
[01:02:40] finance implications about this AI
[01:02:42] bailout thinking about local
[01:02:43] externalities, thinking about the price
[01:02:45] of power, uh thinking about uh subsidies
[01:02:48] to attract these tenants is an
[01:02:49] interesting recent paper by Gargano and
[01:02:51] Gecolete on this as well.
[01:02:53] So, I compiled a few references for you
[01:02:55] if you're interested.
[01:02:56] Um and with that, let me uh thank you
[01:02:58] for uh giving me this opportunity.
[01:03:02] Great
[01:03:03] Great. Thanks a lot, Austin. Let me also
[01:03:04] mention that you also wrote a paper
[01:03:07] which will be available for downloading
[01:03:08] as well. So,
[01:03:10] which summarizes um all the points you
[01:03:12] touched upon as well today.
[01:03:15] Let me ask you
[01:03:17] if if you think you know there's a lot
[01:03:19] of risk involved, but do you think the
[01:03:21] social benefits outweigh the financial
[01:03:23] risks we are entering here or
[01:03:26] in a sense that you know what we had in
[01:03:28] the internet boom we also had you know
[01:03:30] uh
[01:03:31] some some
[01:03:33] global crossings which pulled you know
[01:03:35] across the globe some cables which was
[01:03:38] very useful at the end, but do you think
[01:03:41] it will be similar in this case there
[01:03:43] will be corporate losses there will be
[01:03:44] bankruptcies, but ultimately from a
[01:03:47] social perspective these investments are
[01:03:49] useful to have. Yeah, part of me is
[01:03:51] tempted to answer part of me is tempted
[01:03:53] to answer yes to that question. I think
[01:03:55] one difference though is people will
[01:03:56] often talk about the fiber optic cables
[01:03:59] that were
[01:04:00] there was a lot of speculative
[01:04:01] development in those there were way too
[01:04:02] many of them laid that before they were
[01:04:04] used, but 10 years later they're all
[01:04:05] used. Now, the problem is right now we
[01:04:08] don't have speculative development in
[01:04:10] data centers, but imagine we did. The
[01:04:12] problem is that IT might depreciate a
[01:04:14] lot more quickly. So, I don't it's not
[01:04:15] clear to me that if we don't use these
[01:04:17] GPUs for 5 years that they'll still be
[01:04:20] very valuable 5 years later. So, the
[01:04:22] depreciation schedule may be a little
[01:04:25] different here, but I think one thing
[01:04:27] that is true is as long as there's not
[01:04:29] too much leverage in the system
[01:04:31] uh like in the internet boom um you
[01:04:34] know, I think that should sort of uh
[01:04:37] be good news versus the GFC where we had
[01:04:40] tons of leverage. Right? So, I think
[01:04:42] what makes the difference between it's
[01:04:43] an economic recession and some rich
[01:04:45] people lose a bunch of money, financial
[01:04:47] sector doesn't collapse.
[01:04:48] Uh that was sort of the case in the
[01:04:49] internet boom that was not the case in
[01:04:51] the GFC. So, a lot will depend on just
[01:04:54] how much leverage we put on these assets
[01:04:56] and these are real estate assets and you
[01:04:57] can already see in these off balance
[01:04:59] sheet vehicles there's 90% leverage in
[01:05:02] the Hyperion deal, right? Maybe that's
[01:05:04] an extreme example, but that's what we
[01:05:05] need to be uh mindful of because if that
[01:05:08] becomes sort of a common occurrence,
[01:05:10] then I think we are sort of edging
[01:05:12] closer to a GFC type scenario.
[01:05:16] So in the global financial crisis before
[01:05:19] that, you saw a lot of investment money
[01:05:21] coming out of Europe to finance these
[01:05:23] you know, subprime mortgages and all
[01:05:25] that.
[01:05:26] Do you see a similar Do you know
[01:05:27] anything about it? Who is buying So we
[01:05:29] talked about this. Who is buying these
[01:05:30] bonds and this risky part of this
[01:05:32] portfolios?
[01:05:34] Is there a lot of international capital
[01:05:35] coming to the United States to finance
[01:05:37] it as well?
[01:05:38] Absolutely, right? So a lot of this gets
[01:05:41] channeled through these private credit
[01:05:43] funds. For example, they have a very
[01:05:45] They have a global investor base. Mostly
[01:05:47] an institutional investor base.
[01:05:50] I would not be surprised if a lot of
[01:05:51] German banks, for example,
[01:05:53] invested in these data centers through
[01:05:56] private credit funds. We also know
[01:05:58] there's sovereign wealth money flowing
[01:05:59] into this space, usually on the equity
[01:06:01] side.
[01:06:02] So yes, I think the answer is yes.
[01:06:04] Again, we need better visibility into
[01:06:07] you know, the ultimate owners of the
[01:06:09] equity and the debt in those in those
[01:06:11] but I am sure there is a a non-trivial
[01:06:13] fraction of international money flowing.
[01:06:18] And do you think the European and the
[01:06:20] Chinese data centers are funded in a
[01:06:23] similar way or that's hard to say at
[01:06:24] this stage?
[01:06:25] Some of the European data centers are
[01:06:27] built by Google and some of the US
[01:06:30] hyperscalers. Yes, absolutely. So a lot
[01:06:32] of the data centers abroad are built by
[01:06:34] the same hyperscalers in the United
[01:06:36] States, absolutely, right? So these
[01:06:39] these companies like Google and
[01:06:40] Microsoft have sort of a global
[01:06:41] footprint. They're building data centers
[01:06:42] in Indonesia and in Malaysia
[01:06:45] and in Frankfurt. So I'm not sure there
[01:06:47] are large builders of data centers
[01:06:51] in in Europe and in Asia outside of
[01:06:53] China. I bet you in China there's sort
[01:06:55] of their whole a whole separate
[01:06:57] ecosystem of of data center providers.
[01:06:59] But this is something that there's
[01:07:00] relatively little
[01:07:02] known about
[01:07:03] and I think it would be again a a
[01:07:06] uh to look into more more deeply.
[01:07:10] I guess there's some investment on on
[01:07:11] resilience that you said if something
[01:07:12] changes that we can adjust
[01:07:14] easily and reorient these data centers
[01:07:18] or is it really let's focus fully on
[01:07:22] training or inference? Well, I mean,
[01:07:23] these things are
[01:07:25] you know, they are convertible, right? I
[01:07:26] mean, these these these these chips
[01:07:28] still work to compute anything. I think
[01:07:29] the question is about latency, for
[01:07:31] example, like if you
[01:07:33] again, if you're that surgeon that is
[01:07:34] using some AI model and you need
[01:07:36] instantaneous feedback, is that data
[01:07:39] center in Louisiana going to be good
[01:07:41] enough to help the surgeon in Boston or
[01:07:42] is that latency too too high for that
[01:07:45] use case? And there there's a range of
[01:07:47] use cases. For some things you don't
[01:07:48] mind waiting uh you know, half a second.
[01:07:51] Other things need to be right away and
[01:07:53] maybe the latency technology, the fiber
[01:07:54] optic cables, for example, might improve
[01:07:57] in the future to make those sort of
[01:07:59] remote data centers also useful for AI
[01:08:02] inference. That's possible. That would
[01:08:03] be a big That would be important. Uh but
[01:08:06] right now there's a sense that these are
[01:08:07] not perfectly substitutable assets and
[01:08:09] at the minimum there is uh expenses,
[01:08:11] there are investments that need to be
[01:08:12] made to turn one into the other.
[01:08:14] Uh also these Blackwell These Blackwell
[01:08:17] chips that are used for model training,
[01:08:18] they're very expensive. Somebody called
[01:08:20] them the Ferrari of uh data of of chips,
[01:08:23] whereas what people need for inference
[01:08:25] is often not a GPU but a CPU. You need a
[01:08:28] Prius, not a Ferrari. So, we've invested
[01:08:30] in a lot of Ferraris. Could you drive a
[01:08:32] Ferrari when you wanted to drive a
[01:08:34] Prius? Yes, it's still a car. You can
[01:08:36] still use it, but you completely
[01:08:38] overinvested for what you actually need.
[01:08:41] Yes.
[01:08:44] Okay, so typically we would like to end
[01:08:45] with a positive note.
[01:08:47] Well, you gave us a bullish case. Is
[01:08:49] this your positive note?
[01:08:50] Yeah, I think
[01:08:51] >> Or should be Let me ask you this way.
[01:08:55] We're going through a revolution right
[01:08:57] now. It's It's This is the next
[01:08:58] industrial revolution. It's massive in
[01:09:00] scale.
[01:09:01] It's potentially going to change the
[01:09:03] world Um and it requires vast amounts of
[01:09:06] capital and we're using all the tools in
[01:09:09] our back, all the structural purpose
[01:09:11] vehicles and all the different sources
[01:09:13] of capital to make this happen. I think
[01:09:15] we've learned that we got to be careful
[01:09:18] when that happens. There's all these
[01:09:19] previous infrastructure booms that ended
[01:09:22] in tears. We want to avoid that outcome
[01:09:24] to the extent possible.
[01:09:25] But fundamentally, this is something
[01:09:28] that is
[01:09:29] sort of the key key asset class in real
[01:09:32] estate today. That's all any real estate
[01:09:34] investor wants to talk about today is
[01:09:36] data centers.
[01:09:37] You know, all of us and our kids are
[01:09:39] using these devices and using this
[01:09:41] compute all day long. It's going to make
[01:09:42] all of our research so much more
[01:09:44] productive.
[01:09:46] But there's a physical place where that
[01:09:47] needs to happen, right? And that that
[01:09:49] requires a lot of investment and we just
[01:09:51] got to be careful how we invest it.
[01:09:55] I'll make you the king of regulation.
[01:09:58] And you have two measures to undertake
[01:10:00] to ensure financial resilience. What
[01:10:03] would you do?
[01:10:05] I would begin by having
[01:10:07] reporting requirement for um
[01:10:10] you know, who's holding these risks so
[01:10:12] we have much more transparency into it.
[01:10:15] That would tell us a lot about are these
[01:10:18] are the claims ultimately ending up in
[01:10:19] the hands of unlevered resilient
[01:10:22] investors or on very levered balance
[01:10:24] sheets. I think that's number one.
[01:10:26] Fix these accounting
[01:10:28] fix these accounting
[01:10:30] gimmicks as well.
[01:10:32] I think that those two things would go
[01:10:33] sort of a long way in the short run to
[01:10:36] to give us more visibility into into the
[01:10:38] risks that may be building.
[01:10:43] Thanks a lot. This was really
[01:10:45] fascinating and you illuminated us about
[01:10:48] the dangers and the opportunities.
[01:10:51] And um
[01:10:52] hopefully the regulators will pick up
[01:10:54] some of your recommendations and we can
[01:10:56] avoid
[01:10:57] some financial risks down the road.
[01:11:00] Thank you, Marcus.
[01:11:00] >> Thanks again. Thanks to everybody for
[01:11:02] joining us.