October 6, 2025 • 38 mins
Money has been flowing into the AI space, with billions pouring into GPU clusters, data centers, gas turbines, and the infrastructure needed to train and deploy bigger and bigger models. So what comes next? On this episode, we speak with James van Geelen, founder of Citrini Research. We talk to him about his latest field trip to one of OpenAI's new massive Stargate data centers in Abilene, Texas. We also discuss the increasingly complicated financing arrangements that are funding some of these mega projects and the energy needed to power them. Finally, we ask the question which everyone seems to be asking right now: is AI in a bubble?
Read more:
OpenAI Valuation Reaches $500 Billion, Topping Musk’s SpaceX
FOMO Builds as Alibaba Extends $250 Billion AI-Fueled Comeback
Only http://Bloomberg.com subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at bloomberg.com/subscriptions/oddlots
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:02):
Bloomberg Audio Studios, Podcasts, radio News.
Speaker 2 (00:17):
Hello man, Welcome to another episode of the All Thoughts Podcast.
I'm Tracy Alloway.
Speaker 3 (00:21):
And I'm Joe Wisenthal. Joe, Yes, have you.
Speaker 2 (00:25):
Ever wanted to tour a data center?
Speaker 3 (00:28):
I'd love to.
Speaker 2 (00:29):
Part of me thinks it would be really boring, like
if you actually went inside, because it would just be
a bunch of the same cables, like over and over
and over again. But part of me thinks just seeing
the scale of it, yeah, would be amazing.
Speaker 4 (00:42):
You know what I would do if I toured a
data center is I would look at those cables and
every nut and bolt and vec and heating thing and
get every brand name I could on any little piece
of equipment in the entire thing, and then see are
there any of these that are AI play? Is that
the market doesn't really you know, any random screwdriver that's
(01:03):
being used in this operation?
Speaker 3 (01:06):
What is the name? Is it publicly traded?
Speaker 2 (01:08):
So literally picks and shovel?
Speaker 3 (01:09):
Yeah, just exactly.
Speaker 4 (01:11):
Oh here is here is a company that makes the
you know, the doorstop. Are they an AI play? Because
they're going to need a lot of doorstops for all
these doors. That's what I would do.
Speaker 2 (01:20):
Okay, So I guess the Joe hyper Scale data Center
ETF coming to a market near you eventually, that would
be a good idea.
Speaker 5 (01:29):
It's a great idea.
Speaker 2 (01:30):
Well, today we're going to talk about some of these
data centers, yeah, with someone who has actually gone on
a field trip and looked at them, and then we're
going to talk about broader AI because of course with
markets still, you know, basically at records, there's a lot
of concern that valuations are getting out of hand. We've
had these financing deals which just do my head in
(01:52):
in terms of like trying to track who's lending to who,
and who's buying from who and all of that. So
we should talk about.
Speaker 4 (01:58):
It totally because I mean two things, which is one,
we know how important AI is to the stock market. Two,
you know, there's all these sort of complicated financing deals
and the number is changed by the day. So just
today we got news that in the latest open AI
is allowing some of its employees to sell shares at
a half a trillion dollar valuation. It just seems like
it's constant fundraised, constant numbers that go higher and hire
(02:20):
and hire every day. Various estimates for how much capital
expenditure is going to happen in this cycle seem to
go up. I feel like what we really need to
do is have one of these conversations every month and
just get for real, like we need to keep like
it changes so fast and the stake seems so high.
Speaker 2 (02:36):
Yeah, you definitely need frequent updates. Yes. The other thing,
of course, hovering in the background is the importance of
the AI build out to the overall economy, right and
this is the question, like, is the I think Deutsche
Bank actually put out a note on this is the
entire US economy being propped up by AI right now?
Maybe that's not a problem if you know, if the
(02:57):
AI stuff actually gets built and materializes in real money.
But you can imagine if it is a bubble, then
that would be bad.
Speaker 4 (03:04):
The way I see it is, there's essentially two ways
this is going to end up. One is it turns
out it was a bubble, we have a massive recession
and we all lose our jobs. And the other possibility
is that AI is not a bubble, we have AGI
and we all lose our jobs. And so I'm curious
which of the two paths is going to happen.
Speaker 2 (03:21):
Well, I gotta say there are some alternate possibilities. Did
you see the Dallas fed chart.
Speaker 3 (03:27):
I haven't seen that one.
Speaker 2 (03:28):
So it showed GDP per capita and one line was
like for agi and it just goes straightly. Yeah. But
then the other line is what if the robots take
over the world and kill everyone? And that line went
straight down.
Speaker 3 (03:42):
So yeah, okay, okay, lots of good outcomes here.
Speaker 2 (03:47):
Yes, all right, So we have the perfect guest, someone
we've spoken to before, who writes absolutely phenomenal research about
AI and actually digs into literally the nuts and bolts
of everything we're going to be even with. James Van Geln.
He is the author of the Satrini newsletter. James, thanks
for coming back on.
Speaker 5 (04:06):
Thanks for having me. I feel like the last time
we were here we were also discussing whether a is
a bubble or not. And yeah, we get to do
it again.
Speaker 3 (04:15):
Well, it'll be right right.
Speaker 2 (04:17):
Analyst going on field trips is one of my favorite
genres of research. Explain to us how you managed to
do this.
Speaker 5 (04:27):
So drones have become very commoditized, the likelihood of anything
in the world being more than a ten mile radius
from someone with a drone is extremely low and we
kind of use that to our advantage because we've been
sitting in front of a screen watching green numbers go
up and up and up. And it's not my fault
(04:47):
that they named it stargate. It kind of lends itself
to not taking it as serious as one otherwise might.
And someone I've read a tweet that said, well, where's
the proof. There probably should be proof, right, So we
looked at some satellite images and we went, Okay, a
year ago this was dirt, and now there's this thing,
and it's the size of Lower Manhattan.
Speaker 2 (05:08):
Wow. So in Texas it's.
Speaker 5 (05:10):
An apple in Texas, and the I mean this is
actually one of the smaller ones. I mean, if you
look at Louisiana metas Hyperion, it would start the footprint
would start at the top of Central Park and it
would go down to Soho wow.
Speaker 2 (05:25):
So so that's why they call it hyper Scare.
Speaker 4 (05:29):
So just to be clear with Stargate, this is like
the thing that, like Trump and Oracle and probably Soft Bank,
is in the middle, like, what is this complex? Who's
behind it? This thing that you went and toured and
got some drone footage.
Speaker 5 (05:44):
So it's essentially a mix of you know, Oracle Open AI.
There's some involvement from Nvideo, there's Saudi interest with mg X,
there is some financing involved, and it is just one
of currently data centers that are planned with a half
a trillion dollars of investment. That's why we went and
(06:05):
looked at it because when you really think about it,
this represents probably the largest infrastructure build out and effort
since World War Two.
Speaker 2 (06:15):
Where are they getting the energy from?
Speaker 5 (06:17):
So that's that's the best part. That was the first
thing we saw on the drone is you fly over
it and they just built their own natural gas plant
and going like Joe, great idea, just going and looking
at all the but you got to look at the
big stuff too, right and without i mean without power.
(06:37):
These are just kind of you know, hunks of metal.
So you have outside of Stargate aveline ten natural gas turbines.
And the interesting thing is these aren't like the really
good natural gas turbines because if you wanted so, natural
gas turbines fall along simple cycle combined cycle. These are
(06:58):
simple cycle. They're each thirty five megawatts, which is very
much on the lower end. Half of them are from
geev right now, about half of them are from Caterpillar,
a company the Caterpillar on it's called solar turbines. And
the reason why they're not you would think, oh, you're
spending half a trillion dollars on these things. You could
probably get the best thing ever, but that'll take you
seven years.
Speaker 4 (07:16):
Oh right, because there's a there's a natural guest turbine.
Speaker 3 (07:18):
Yeah, shortage.
Speaker 4 (07:19):
By the way, there's a headline.
Speaker 3 (07:22):
So there's a headline.
Speaker 4 (07:24):
An article on the Bloomberg Today by Matthew Griffin, the
hunt for winners in the artificial intelligence gold rush has
landed on an unlikely target, old line industrial equipment maker
Caterpillar in So you got to look at every one
of these brands that's supplying it, and Caterpillar is up
a lot. It's up today and the stock is done
(07:44):
very well. So okay, Caterpillar, who else? What are some
of the what else do you see besides these big
turbines when you go out and tour this facility. So
the stock is a rocket.
Speaker 2 (07:54):
So that's any people like securities are.
Speaker 5 (07:57):
So many people and I mean it's there. I don't
know you've seen like Silicon Valley and there's that scene
where there's like one guy in the data center and
he's like, yeah, you know, night and day don't really
matter much down here, and it's just one guy running.
But in the buildout, I mean build up, there are
every single HVAC technician or guy that knows how to
(08:20):
lay fiber or guy that knows electrical or plumbing within
one hundred mile radius of valley and those are they're
all there, right, There's seven thousand people working on building
this and this is just one site.
Speaker 2 (08:34):
And yeah, so people have been talking about the picks
and shovels phase of AI investment for a while now,
but it seems like stuff is still getting built and
it has further to run.
Speaker 5 (08:46):
My kind of favorite we can you know, and I'm
sure we will discuss whether this is a bubble or not,
but at the end of the day, the money is
getting spent. And my favorite kind of setup in the
market is you have something that has a multiple that
reflects kind of a wrong reality or doesn't really fully
integrate secular theme and then you have a cycle that's
(09:10):
really bad. So whether it's kind of the robotics supply
chain goes throughout the automotive cycle and if you guys
are aware, that's been a nightmare. And then you have
some of these A lot of these companies took really
significant drawdowns when Trump got elected. Because this isn't really
tech Capex, it's much more similar to like LNG terminal cabets. Ye,
(09:34):
that kind of entails the same companies that were benefiting
from the IRA oh, and so everyone sold them. And
now it turns out it's kind of like the horseshoe theory,
right you.
Speaker 4 (09:44):
Can now that's super interesting. Like it's just these sort
of like oldline industrials, and they did very well under
the Obama administration because of all the factories and all
that stuff that we talked about. And then it's like, okay,
we're pulling the plug on this, but now there's just
this title wave money. By the way, Caterpillar was a
three hundred and thirty four dollars stock on April second,
(10:05):
right before libration day, which is sort of where we
should benchmark things now, four hundred and eighty six at
an auld time high.
Speaker 3 (10:10):
What are some of the other companies see for real?
Speaker 5 (10:13):
Like preferably the ones that haven't gone up? Yeah?
Speaker 4 (10:15):
Yeah, seriously, Like who else is who else is getting
a piece of this action.
Speaker 5 (10:20):
So something that I want to make sure I mentioned
just to frame this whole thing. If you look at
the building side plans, they have the names of all
the data halls and they are literally each called ludicrous building.
So it's ludicrous building, one ludicrous building.
Speaker 4 (10:35):
Amazing.
Speaker 2 (10:35):
Wait is it Ludicris with a K?
Speaker 5 (10:38):
No?
Speaker 2 (10:39):
That would be great. I mean Ludacris spelled correctly is
also fantastic.
Speaker 5 (10:45):
But the company, I mean really I would encourage anyone
to go and look at this video because the it's
it's just a there is such a disconnect between looking
at green numbers going up like like caterpillar going okay, cool, Yeah,
and then you see it and it is mind blowing.
(11:06):
I mean they have rows of dry and atabiatic coolers
that are you know, hundreds and hundreds for the liquid
cooling loops, and then you have the kind of all
the HVAC, the power generation and transmission. We compile the
list of basically two hundred companies, fifty of which were
(11:26):
high confidence in kind of seeing as part of this
build out. And the interesting thing is you have these
multinational OEMs. This scale of this project is so large
that you can call one of them and say, okay,
we want to go with you know, Mitsubishi heav or
we want to use Eton for this thing, and they
(11:47):
will say, yeah, okay, well we can do sixty percent
of it. I mean, the biggest companies in the world
are Yeah, they're just fighting. They don't have to fight
with each other because the pie is so big.
Speaker 3 (11:58):
How scarce you know, there's all this around the country.
Speaker 4 (12:02):
There's been all this tensions like is there gonna be
enough energy for data centers?
Speaker 3 (12:07):
What about the water? What about the politics? People?
Speaker 4 (12:10):
We see these town halls where the locals do want
a data center. How scarce are locations like Abilene? I
mean there's a lot of empty space in Texas. Could
you see Are there more stargates or more Abilenees out
there waiting where the companies can just say, you know what,
we're just going to build our own infrastructure. We're going
to build our own natural gas plant, and we're just
going to do it in the desert in Texas where
(12:31):
we don't have to worry about nimbi's or local politics
or anything.
Speaker 3 (12:34):
Is that the future for a lot of this stuff?
Speaker 5 (12:35):
Yeah, there's already another facility that's planned in Shackleford, Texas,
which is not too far from Abilene. Abilene is already
planning on expanding by another six hundred megawats. I mean,
that's how we're measuring data centers now, is just in
megawatts and gigawatts. In Louisiana, that's where Meta is building Hyperion.
You have some unidentified or unannounced Midwest facility that'll probably
(12:59):
be in Ohi, Io. Basically, yeah, anywhere that you can
get the gas, anywhere you can get the power, anywhere
that there's the land. There was one planned in Indianapolis.
And it's interesting you bring up water because we wrote
about water a little while ago as well, and we
were basing our projections off of what the water consumption
of like existing data centers were. The advancements in liquid
(13:21):
cooling have changed that a lot, so there was a
huge kind of uproar about, well, this data center is
going to use so much water. And if you look
at it, this was an Indianapolis and if you look
at it, it's about the same amount of water as
a thousand households. Because it's all closed loop liquid cooling. Now,
the water just gets recycled and filtered back, so it's
(13:44):
really just something where the constraint is still power and
getting these kind of turbines, and it's going to increasingly
become a thing where it's behind the meter power generation.
They're just going to build their own power plants. They're
not going to rely on the grid. You guys remember
what happened in Texas and twenty one with a Yeah,
they you cannot underwrite that as a data center, like
(14:08):
you just can't be offline. So that's kind of where
we're headed, is that this power and energy kind of
bottleneck is the bottleneck. One of the more interesting things
we ran across is you look at Xai's data center,
and we went through a lot of the filings of
all these different data centers. They're permitted for fifteen thirty
(14:28):
five megawatt turbines, and there's thermal imaging of the site
that shows they're using thirty five ge Vanova.
Speaker 3 (14:36):
Another one that was a three.
Speaker 2 (14:39):
You're just typing tickers.
Speaker 4 (14:40):
Yeah, there was a three hundred and thirty dollars stock
on April second. That's a six hundred and nine dollars stocks.
Speaker 5 (14:44):
Okay, check out Stemens Energy.
Speaker 2 (14:46):
Thank you, Joe. Oh he's doing it.
Speaker 3 (14:49):
You could keep going.
Speaker 2 (14:50):
All right, I have kind of this isn't an essential question,
but okay, wait, I got it.
Speaker 5 (14:56):
I knew this is going to happen.
Speaker 4 (14:58):
Semens Energy was a fifty six dollars stock on April second.
Speaker 5 (15:01):
That's one hundred and eight.
Speaker 4 (15:02):
That's a string lineup.
Speaker 2 (15:03):
Okay, okay, clearly someone's making money. Like how big are
the walls of these things? And what does security actually
(15:26):
look like? Because you know, we talk about people being
upset about higher energy prices. I understand that maybe some
of the some of them are building out their own
energy capacity, but you know, in a populist mindset, people
might still be upset at AI, still upset that AI
is going to take jobs and things like that. They
could become targets.
Speaker 5 (15:45):
Right, yeah, absolutely, I mean apparently security isn't tight enough
to shoot down at drone yet, oh of course. Yeah.
But when you talk about the walls, like, the interesting
thing is everything in this data center has You can't
just build it like a house. Everything needs to have
the right anti corrosion insulation. You can't just be losing heat.
You're already spending so much money on cooling. So everything
(16:08):
in the building envelope is also kind of the most
expensive thing it could be.
Speaker 2 (16:12):
And I assume it's like a sealed envelope.
Speaker 5 (16:14):
I guess, yeah, exactly, And the opportunity of just how
big is this going to be if you actually believe that.
The people that are building this, and this is kind
of an interesting dynamic of the hyperscalars are in somewhat
of a prisoner's dilemma. We can't speak to whether they
actually believe or not, but the party line is we're
(16:38):
building machine God. And if we're building machine God, then
that's kind of a religious zelotry. So three years ago,
most of these hyperscalers were making net zero. By twenty thirty,
that is not happening. Yeah, there's no way that's happening.
Maybe they buy carbon off, but that's not happening because
if you need to do it now, you're using natural
(16:58):
gas and that then you say, well, once we create
machine God, machine God will fix it, right, and you
can justify a lot of stuff.
Speaker 2 (17:11):
Wait, okay, so let's talk about is AI in a bubble,
because when we're conversing about buildings that are literally called
ludacris and we're talking about machine God will solve all
our problems.
Speaker 5 (17:23):
That sounds a little bubbly.
Speaker 2 (17:24):
Yeah, that's the kind of stuff you look back on
and say, oh, yeah, that was the top I know.
Speaker 5 (17:29):
That we we have to talk about this and that
requires some numbers, so I have say notes. But the
interesting I mean, you can make a case for either side,
and it kind of hinges on not so much what
AI looks like in the future, but more where we
are in the cycle. If the question is is AI bubble? Yeah, probably,
it's a new technology. It's exciting every single time we've
(17:51):
the railroads, is the you know, dot com? Probably if
it's not a bubble, it will be, and it probably
is already. But where are are we in the bubble?
That's the interesting thing for investors. So if you look
at kind of where are we in the financing side.
When we had the dot com bubble, the big thing.
(18:14):
Are you guys familiar with dark fiber? No? So essentially
they were building out this kind of infrastructure for the Internet,
and the conviction of the day was the Internet's going
to be the biggest thing in the world. Correct, And
ninety seven percent of the fiber that was laid was
called dark fiber because it wasn't lit up. It was
just laying there because eventually we're going to need it.
(18:37):
That's not the case right now. Every single time that
one of these data centers comes online, you're at one
hundred percent utilization. There's no dark fiber, and we're continuing
to need more. I mean, I don't know if you
guys saw Sora too. More than seventy percent of the
content that we consume is video, so a video model
makes a lot of sense. It's also way more compute intensive,
(18:58):
and the the final form of chat of AI is
not chatbots. It's gonna be video. It's gonna be robotics,
which use video models. So that's kind of like the
bullish side. But I totally see. I have some numbers here,
So we've had some of these circular deals and so far,
(19:22):
I mean, there's hundreds of billions of dollars for this.
Most of the financing is structured much more like toll
roads or LNG terminals. The Metas hyperion was twenty nine
billion dollars. Twenty six billion of that was Pimco on
debt and then blue outl with three billion of equity,
an off balance sheet arrangement, which also was pretty popular
during the dot com bubble. Then you have stuff like
(19:44):
GPU collateralized SPVs core. We've had a seven and a
half billion dollar Blackstone SPV. There's like an eleven billion
dollar GPU backed loan market now, which compared to the
infrastructure thing, I can get that. It's like those GPUs
are gonna be worth like they will depreciate. So yeah, that's.
Speaker 2 (20:05):
That's abs right, that's like asset backsxety Okay.
Speaker 5 (20:08):
Yeah, but when you look at the where we are
just saying where are we in the cycle, Well, there's
been there's been a lot of these balance sheet arrangements.
But in terms of that prisoner's dilemma that we spoke about,
where nobody's going to back off because then maybe they
risk not creating machine GUD. The debt to equity ratio
on most of the hyperscalars is not crazy. Sure Oracle
(20:31):
has the highest, and but they still have a lot
of room to lever up in pursuit of this goal.
The interesting kind of dynamic of in Vidia finances its
customers and then its customers buy Nvidia GPUs, and then
you know that, yeah, that's a bubble dynamic. That's that's
visible circular financing, and but the demand is real.
Speaker 4 (20:52):
Well, this is this is I guess the question. If
the demand is real, why are the videos of the
world financing your customers? Right, intuitively you think, okay, in
videos the demand for chips is maxed out, right, Then
intuitively you think, well, then there's.
Speaker 3 (21:08):
Plenty of money elsewhere.
Speaker 4 (21:10):
So do you have a guess for why in VideA
is compelled to you know, invest in open Ai, invest
in core weave. It's been a long time investor in
core weave and so forth.
Speaker 5 (21:19):
Yeah, I could. I could take the very skeptical view
or the kind of non skeptical view. The non skeptical
view is the best thing for in Vidia is that
we accomplish a GI. So anything that it can do
to get us closer to that massive, massive infrastructure that's
required for that is great for them as quickly able,
as quickly as possible, because, uh, every year that you
don't achieve AGI becomes less likely. So uh, that is
(21:44):
maybe the core factor. And then there's it's good for
them and the and playing into that the because this
this isn't necessarily like the dot com bubble, because the
dot com bubble had fiber and then it had you know, pets,
dot Com and Amazon and all that stuff. This is
all pretty much capex right to tech is capital intensive again,
(22:07):
which means that it's the bust won't necessarily be like
the dot com bubble, where it gives the real players
time to shine. If the CAPEX spending grinds to a
halt because the market goes down, that's the worst thing
in the world. So in a way, it's also it's
in their interest obviously to make sure that that doesn't happen.
Speaker 4 (22:41):
Another question, actually, but I want to ask it now
because I might forget otherwise. A headline I think earlier
this week again headline is almost every day. Can you
explain why Meta, which is building out which obviously hyperscaler,
they have tons of data center capacity, why do they
need to do a deal with Corewave? So I saw
that Meta is going to pay a few billion, fourteen billion,
(23:04):
it's some number for capacity on core Weave. What can
the what do the neo cloudes bring to the table
such that the legacy is like, you know what, we
launched some of your capacity.
Speaker 5 (23:13):
It's basically I would say, anything you can do to
shift that CAPEX away from yourself is great. Right if
you're if you're right now, a lot of this stuff
is funded from cash flow, and yeah, we're starting to
see that financing take its place. But if you the
the you're running a company that's existed for a while
(23:33):
and is one of the biggest companies in the world,
and yes, you could justify taking an ordinate amounts of
risk to do this, But just like anyone else, if
you have a goal and you can kind of shift
away some of that risk to because if you know,
if core Weeve goes bankrupt, Meta is not it's not
It's not the worst thing in the world for Meta.
Speaker 2 (23:51):
What would you be looking out for for signs of
the bubble bursting or I guess the capex FAWCET starting
to turn off.
Speaker 5 (23:59):
So in terms of like the warning signs first, monitoring
used GPU prices I think is a great way to
see what's going on, seeing like contract renegotiations, waiver headlines
from private credit. And also a big one would be
any delays in delivering power make it so that you
(24:20):
finance this thing and then it's not going to be
able to work, right. I do think that another you
really do have to kind of index to the revenue
that's being generated by the AI companies as well, and
you saw open Ai recently announced that they're basically allowing
you to buy things with one click from the inside
the app. And I think that that was also a
(24:42):
big driving factor in the router where they kind of
route your model to the it's setting it up for advertising,
because the other day I asked Google a pretty simple
question and it just gave me the answer with no
advertisements delivered to me. So we need to I think
that there is the killer apps kind of are here
or coming, but that has to happen fast.
Speaker 2 (25:05):
Yeah, the fact that the Internet kind of runs on
ads is an underappreciated reality of our world, and we
did an episode on it way way back in the day.
Is there anyway if I wanted to get exposure to
like the good stuff of AI, the actual money being spent,
but then I don't like all the circular financing arrangements.
(25:26):
Is there a way to hedge that particular risk?
Speaker 5 (25:29):
Yeah, you could buy all the real companies that are
building this kind of data center, and then you could
short Blue Owl and the private credit companies that are
doing the kind of financing aspect of it, because if
it goes bad, they have limited upside and kind of
unlimited downside, and if it continues happening, the companies that
are actually building it have unlimited upside and they're still,
(25:52):
like you said, these oldline kind of industrial that like,
they still have a core business to go back to.
Speaker 4 (25:57):
What's your take on so open ai announced its new
Sora thing that Tracy has a code for, Meta announced
it's Meta AI, which a lot of people is like,
this is slop and I don't we don't need to
get into a debate about whether this is garbage. I
kind of think it is, but that's not actually the
question I'm interested in. The question I'm interested in it
is one way to view these things is you know,
(26:19):
what these are going to be, These video, these AI
videos are going to be really important sort of business
revenue drivers in our pursuit to AGI. And the other
view is, look, if they were close to AGI, they
would not be devoting time to the slop factories. And
this is the sort of question and I'm curious from
your perspective, what do you think is the real logic
(26:40):
behind all of this effort to create these sort of
you know, dom like, oh, a video of a astronaut
riding a horse through space with a you know, cowboy hat.
Speaker 5 (26:48):
On or whatever. Video models are going to become the
most important part of this. You you need, for example,
lms don't really interact with the world, but we're starting
to see in robotics, AI models interact with the world.
So you have these video language action models, and you
can structure it so that there was an interesting paper
(27:08):
on Archive recently that and you know this is video
models like VO three are emergent zero shot learners and understanders.
And what that just means is you can give it
basically like a second of context, and then it can
extrapolate about what's going on beyond that. That becomes extremely
important for robotics. And you can structure these vision language
(27:31):
action models where you have like a slow reasoning model
that is kind of like your conscious mind and determines
what it needs to do, and then you have like
a fast model that actually does the movements and stuff.
AI kind of extends into the realm of you need
to generate as much data as possible, and having you know,
video data and knowing what's good and what's bad is
also very important. But yeah, to your question, open AI
(27:54):
probably does want to show that they're making more money,
and I don't doubt that they're people there that truly
believe that they're building the machine GUD. And I also
don't doubt that they realize if they have a shot
at doing that, they need the market to play long
and they need to keep growing their revenues. So I
think there's a lot of really good counter arguments about
(28:15):
like not necessarily are we in a bubble?
Speaker 2 (28:17):
Yes?
Speaker 5 (28:18):
Is the bubble over? Like? There are good arguments for that.
I don't think one of the I don't think that, oh, well,
they're leaning into like kind of this slops that you know,
it's a company. They're going to do whatever they can
to make money. That's capitalism, you know.
Speaker 2 (28:31):
One of the things I think about another reality of
our current time. This is kind of related, but if
we're talking about video data, YouTube staying power has been
like phenomenal, right, It's been around for decades now, and
we haven't really seen any successful competitors anyway.
Speaker 5 (28:49):
Well, since we wanted to talk about stocks that are
going up, and I know that's another one people kind
of forgot that video takes up a lot more storage
than textas and if we're going to start just like
throwing out the slop on video the part. Basically, one
of the only areas of semiconductors that has gotten ignored
(29:09):
has been storage, not like not like memory like like
uh RAM and stuff, but like hard drives and solid
state discs like Western Digital and uh okay.
Speaker 4 (29:20):
So Seagate, Yes, that's what we're talking about. So this
was a this was an eighty four dollars stock in
late March, and now it's a two hundred and fifty four.
That's where you're going, Yeah, okay, yeah, yeah, there's a crazy.
Speaker 2 (29:31):
There's a crazy, but it has gotten attention. But you
think it could go up more?
Speaker 5 (29:36):
I think so, and I think that uh to illustrate
that dynamic that's going to keep happening. Basically, you just
keep looking at stuff where there's people say, oh, well,
you know there's a there's a glut of X and
that's why this isn't going to win from AI. And
then eventually the glut of X doesn't matter. It's a
That's what happened to Nvidia in the beginning of all this.
There was a glut of crypto GPUs and now nobody care, like,
(29:58):
wh when's the last time you heard some want to
be barrassed on a video because there's too many crypto minors.
Speaker 4 (30:02):
Sand Disc was forty eight dollars in April second, that's
one hundred and twenty six dollars stock.
Speaker 5 (30:07):
Pretty crazy.
Speaker 4 (30:08):
These are just these are like just the tools are
just store just basic.
Speaker 3 (30:11):
Yeah, yeah, crazy.
Speaker 5 (30:12):
It's it is and it's it. I guess you could
be worried by that, or you could say this isn't uh,
this isn't like in Nvidia doubling in a month necessarily right,
that it's just people discovering. I think it's a good barometer,
which is important in a bubble, that people are still
paying very close attention. And you start to get worried
(30:34):
when everybody just goes into the cisco of it all.
But these are real companies that have real products. Of course,
there's some silly stuff that's happening, but it's not only
silly stuff. It's not twenty twenty one with the you know, yeah,
I don't need to.
Speaker 2 (30:51):
So speaking of on the ground research, or maybe I
should say in the sky research. When it comes to robotics,
you had a robot dog, right, yeah, and of those
like really advanced ones, we got a unitary robot dog.
Speaker 5 (31:03):
It is super impressive.
Speaker 2 (31:05):
It's also incredibly creepy when it stands up on its
hind legs.
Speaker 5 (31:09):
Yeah, have you seen the video?
Speaker 3 (31:11):
Yeah, I saw your video of it.
Speaker 5 (31:12):
Yeah, it's uh. And Unitry we went and visited Unitry too,
and it's I've seen Boston Dynamics. You know spot robot,
which is actually more impressive. Also could probably murder you
like and and the interesting thing about Unitry is, yeah,
this the robot kind of feels like a weapon. It's
(31:32):
very heavy and stuff, but it's even they're humanoid, like
it can't lift more than eight pounds. And also Unitree
is not focused on like making the brain. They just
want to make the hardware. And you go there and
they were like, where's your light are from? Oh, well
we needed this specific thing, so we just made it ourselves. Okay,
what about your actuators, Well we made those ourselves. Do
(31:55):
anything that they don't make themselves. They get within a
ninety kilometer radius of their headquarters. But where's their head
It's in hangzhoune So it's something where you're really that's
the competition that we have stacked up against us in
the US versus China, and I think the interesting route
that they've taken is it's super advanced. You can code
(32:17):
on it. You can already see videos of people you know,
buying them and making them do specific things, but they
made them not super like, it's not going to kill you.
And that's a great way to get a bunch of
training data because we don't have enough training data of
what it's like to move around the world. We have
the sum history of all human knowledge that's been recorded
for text models, but we need more data for robotics.
(32:41):
And I think that that's like an interesting kind of
thing where they're just kind of sending these out into
the world so that they can get enough data to
actually make it safe enough to have a robot in
your house.
Speaker 3 (32:53):
Can anyone buy a unitary dog?
Speaker 2 (32:54):
Like?
Speaker 3 (32:54):
Are they for sale online?
Speaker 5 (32:56):
Yeah, there's a lead time, but I can get around
it if you wait.
Speaker 2 (33:00):
How much are they?
Speaker 5 (33:02):
So they start at three thousand, and then we got
the super advanced one that you can attach a robotic
arm to so it can manipulate its environment, and that
was eight thousand dollars, and then they're humanoid is sixteen,
which is like less than the cost of the used
time to sither.
Speaker 2 (33:17):
Well, three thousand. Also the price for like a really
good pure bread dogs, I don't know what that.
Speaker 3 (33:23):
People pay a lot of money for actual dogs.
Speaker 2 (33:25):
So yeah, it seems why are you looking at me?
Speaker 3 (33:28):
Why do you think I'm looking at you?
Speaker 2 (33:30):
Okay, James, that was so much fun. Thank you so
much for coming in and sharing with us your your
field trip observations.
Speaker 5 (33:37):
Thanks Levan, Joe.
Speaker 2 (33:52):
I think maybe it's the cynical journalist in me, but
I get really nervous when we start saying it's different
this time.
Speaker 4 (34:00):
Yeah, I mean it's always different. It's always the same
as sort of there are.
Speaker 3 (34:04):
Certainly you know.
Speaker 4 (34:06):
Yeah, I mean, look, it's crazy and it's a lot
of money, and it's still you know, we've been asking
people on the podcast what kind of use value are
you getting out of this technology? And it's still very ambiguous.
Speaker 5 (34:19):
I don't know.
Speaker 4 (34:20):
I don't have It's not my job to time the market,
so I'm glad I don't have to have a view
on all of this stuff.
Speaker 2 (34:26):
But I do think, cop out, Joe, No, I do think.
Speaker 5 (34:29):
I do think.
Speaker 4 (34:29):
For Okay, what's your okay, give us your give us
your price target.
Speaker 3 (34:33):
That's okay, Okay, give us your price time.
Speaker 4 (34:35):
Okay, I do think this is moving so fast and
every day the numbers are so big, and it's the
stakes are so high for all of this that it
does feel like we just need regular updates on the
state of how much money is being spent in this
area and what the eventual payoff will be.
Speaker 2 (34:54):
I want someone to draw a massive, like infographic of
all the financing kills, because I think you could make
a really interesting one.
Speaker 3 (35:02):
Yeah, you definitely could.
Speaker 2 (35:03):
That would probably be illuminating. On that note, I did
think James's suggestion of, you know, shorting the private credit
players and then going along the actual physical hardware ones
that was interesting.
Speaker 5 (35:15):
Yeah.
Speaker 4 (35:15):
No, it's really interesting. Also, I had not realized. I mean,
it is very interesting the way booms or bubbles, whatever
you want to call it. I don't like using the
word bubble because it applies whatever booms or bubbles, the
way they spread out.
Speaker 3 (35:29):
I just think that's extremely interesting.
Speaker 5 (35:31):
Right.
Speaker 4 (35:31):
So, in video takes off in late twenty twenty two,
when people are blown away by chat CHPT and they
say there's incredible and it's being powered by all these
in vidio chips, and then people go on the hunt
right for other things. It's like what else, And even
still today in October twenty twenty five, that spreading out
process continues. And so okay, in the last few months,
(35:53):
maybe it's like a caterpillar, or maybe suddenly people are
realizing that just sort of old fashioned computer storage is
to be very hot. But we're still in this phase
whereas people open their eyes to these things, the spreading
out continues to this day right.
Speaker 2 (36:08):
Well, Also there's an expectation that eventually, you know, normal
companies are going to benefit from it as well and
become more efficient and all of that.
Speaker 4 (36:17):
I just remembered the one question I was going to
ask James, but I could just throw it out as
sort of a spectative question. So he mentioned all of
the workers. Could you ask all of the workers that
are out this or that are building these things? And
it does make me wonder about the sort of crowding
out possibility, Like imagined if you imagined if you had
some other business enable Leen, and you're trying to hire
an HVAC worker right now because HVAC you know, you're like, oh, restaurant.
Speaker 2 (36:40):
And you're competing with like open AI.
Speaker 4 (36:42):
Yeah, you're like, you have a restaurant enabling, and you're like,
you know what, I need to update my air conditioner
or whatever or whatever, and what are you going to
do because all the HVAC workers are at that big
stargate operation. And I do think this is a very
interesting macro question on all kinds of things, which is
the degree to which all of this spending with uncertain
(37:03):
payoffs will have a crowding out and potentially inflationary effect
on the rest of the economy.
Speaker 2 (37:07):
Well, presumably machine God will eventually design HVAC robots, so
problem will be solved.
Speaker 3 (37:13):
Problem will be solved.
Speaker 2 (37:14):
All right, shall we leave it there?
Speaker 3 (37:15):
Let's leave it there.
Speaker 2 (37:16):
This has been another episode of the oud Loots podcast.
I'm Tracy Alloway. You can follow me at Tracy Alloway.
Speaker 4 (37:21):
And I'm Jill Wisenthal. You can follow me at the Stalwart.
Follow our guest James Van Gillen, He's at Sutrini seven.
Follow our producers Carmen Rodriguez at Kerman armand dash O
Bennett at Dashbod and Kale Brooks and Kale Brooks. For
more odd Laws content, go to bloomberg dot com Flash
odd Lots. We have a daily newsletter and all of
our episodes, and you could chet about all of these
topics twenty four to seven in our discord discord dot
(37:44):
gg slash.
Speaker 2 (37:44):
Odd lots And if you enjoy odd Lots, if you
like it when we talk about the physical realities of AI,
then please leave us a positive review on your favorite
podcast platform. And remember, if you are a Bloomberg subscriber,
you can listen to all of our episodes absolutely add free.
All you need to do is find the Bloomberg channel
on Apple Podcasts and follow the instructions there. Thanks for listening.
Bloomberg Audio Studios, Podcasts, radio News.
Speaker 2 (00:17):
Hello man, Welcome to another episode of the All Thoughts Podcast.
I'm Tracy Alloway.
Speaker 3 (00:21):
And I'm Joe Wisenthal. Joe, Yes, have you.
Speaker 2 (00:25):
Ever wanted to tour a data center?
Speaker 3 (00:28):
I'd love to.
Speaker 2 (00:29):
Part of me thinks it would be really boring, like
if you actually went inside, because it would just be
a bunch of the same cables, like over and over
and over again. But part of me thinks just seeing
the scale of it, yeah, would be amazing.
Speaker 4 (00:42):
You know what I would do if I toured a
data center is I would look at those cables and
every nut and bolt and vec and heating thing and
get every brand name I could on any little piece
of equipment in the entire thing, and then see are
there any of these that are AI play? Is that
the market doesn't really you know, any random screwdriver that's
(01:03):
being used in this operation?
Speaker 3 (01:06):
What is the name? Is it publicly traded?
Speaker 2 (01:08):
So literally picks and shovel?
Speaker 3 (01:09):
Yeah, just exactly.
Speaker 4 (01:11):
Oh here is here is a company that makes the
you know, the doorstop. Are they an AI play? Because
they're going to need a lot of doorstops for all
these doors. That's what I would do.
Speaker 2 (01:20):
Okay, So I guess the Joe hyper Scale data Center
ETF coming to a market near you eventually, that would
be a good idea.
Speaker 5 (01:29):
It's a great idea.
Speaker 2 (01:30):
Well, today we're going to talk about some of these
data centers, yeah, with someone who has actually gone on
a field trip and looked at them, and then we're
going to talk about broader AI because of course with
markets still, you know, basically at records, there's a lot
of concern that valuations are getting out of hand. We've
had these financing deals which just do my head in
(01:52):
in terms of like trying to track who's lending to who,
and who's buying from who and all of that. So
we should talk about.
Speaker 4 (01:58):
It totally because I mean two things, which is one,
we know how important AI is to the stock market. Two,
you know, there's all these sort of complicated financing deals
and the number is changed by the day. So just
today we got news that in the latest open AI
is allowing some of its employees to sell shares at
a half a trillion dollar valuation. It just seems like
it's constant fundraised, constant numbers that go higher and hire
(02:20):
and hire every day. Various estimates for how much capital
expenditure is going to happen in this cycle seem to
go up. I feel like what we really need to
do is have one of these conversations every month and
just get for real, like we need to keep like
it changes so fast and the stake seems so high.
Speaker 2 (02:36):
Yeah, you definitely need frequent updates. Yes. The other thing,
of course, hovering in the background is the importance of
the AI build out to the overall economy, right and
this is the question, like, is the I think Deutsche
Bank actually put out a note on this is the
entire US economy being propped up by AI right now?
Maybe that's not a problem if you know, if the
(02:57):
AI stuff actually gets built and materializes in real money.
But you can imagine if it is a bubble, then
that would be bad.
Speaker 4 (03:04):
The way I see it is, there's essentially two ways
this is going to end up. One is it turns
out it was a bubble, we have a massive recession
and we all lose our jobs. And the other possibility
is that AI is not a bubble, we have AGI
and we all lose our jobs. And so I'm curious
which of the two paths is going to happen.
Speaker 2 (03:21):
Well, I gotta say there are some alternate possibilities. Did
you see the Dallas fed chart.
Speaker 3 (03:27):
I haven't seen that one.
Speaker 2 (03:28):
So it showed GDP per capita and one line was
like for agi and it just goes straightly. Yeah. But
then the other line is what if the robots take
over the world and kill everyone? And that line went
straight down.
Speaker 3 (03:42):
So yeah, okay, okay, lots of good outcomes here.
Speaker 2 (03:47):
Yes, all right, So we have the perfect guest, someone
we've spoken to before, who writes absolutely phenomenal research about
AI and actually digs into literally the nuts and bolts
of everything we're going to be even with. James Van Geln.
He is the author of the Satrini newsletter. James, thanks
for coming back on.
Speaker 5 (04:06):
Thanks for having me. I feel like the last time
we were here we were also discussing whether a is
a bubble or not. And yeah, we get to do
it again.
Speaker 3 (04:15):
Well, it'll be right right.
Speaker 2 (04:17):
Analyst going on field trips is one of my favorite
genres of research. Explain to us how you managed to
do this.
Speaker 5 (04:27):
So drones have become very commoditized, the likelihood of anything
in the world being more than a ten mile radius
from someone with a drone is extremely low and we
kind of use that to our advantage because we've been
sitting in front of a screen watching green numbers go
up and up and up. And it's not my fault
(04:47):
that they named it stargate. It kind of lends itself
to not taking it as serious as one otherwise might.
And someone I've read a tweet that said, well, where's
the proof. There probably should be proof, right, So we
looked at some satellite images and we went, Okay, a
year ago this was dirt, and now there's this thing,
and it's the size of Lower Manhattan.
Speaker 2 (05:08):
Wow. So in Texas it's.
Speaker 5 (05:10):
An apple in Texas, and the I mean this is
actually one of the smaller ones. I mean, if you
look at Louisiana metas Hyperion, it would start the footprint
would start at the top of Central Park and it
would go down to Soho wow.
Speaker 2 (05:25):
So so that's why they call it hyper Scare.
Speaker 4 (05:29):
So just to be clear with Stargate, this is like
the thing that, like Trump and Oracle and probably Soft Bank,
is in the middle, like, what is this complex? Who's
behind it? This thing that you went and toured and
got some drone footage.
Speaker 5 (05:44):
So it's essentially a mix of you know, Oracle Open AI.
There's some involvement from Nvideo, there's Saudi interest with mg X,
there is some financing involved, and it is just one
of currently data centers that are planned with a half
a trillion dollars of investment. That's why we went and
(06:05):
looked at it because when you really think about it,
this represents probably the largest infrastructure build out and effort
since World War Two.
Speaker 2 (06:15):
Where are they getting the energy from?
Speaker 5 (06:17):
So that's that's the best part. That was the first
thing we saw on the drone is you fly over
it and they just built their own natural gas plant
and going like Joe, great idea, just going and looking
at all the but you got to look at the
big stuff too, right and without i mean without power.
(06:37):
These are just kind of you know, hunks of metal.
So you have outside of Stargate aveline ten natural gas turbines.
And the interesting thing is these aren't like the really
good natural gas turbines because if you wanted so, natural
gas turbines fall along simple cycle combined cycle. These are
(06:58):
simple cycle. They're each thirty five megawatts, which is very
much on the lower end. Half of them are from
geev right now, about half of them are from Caterpillar,
a company the Caterpillar on it's called solar turbines. And
the reason why they're not you would think, oh, you're
spending half a trillion dollars on these things. You could
probably get the best thing ever, but that'll take you
seven years.
Speaker 4 (07:16):
Oh right, because there's a there's a natural guest turbine.
Speaker 3 (07:18):
Yeah, shortage.
Speaker 4 (07:19):
By the way, there's a headline.
Speaker 3 (07:22):
So there's a headline.
Speaker 4 (07:24):
An article on the Bloomberg Today by Matthew Griffin, the
hunt for winners in the artificial intelligence gold rush has
landed on an unlikely target, old line industrial equipment maker
Caterpillar in So you got to look at every one
of these brands that's supplying it, and Caterpillar is up
a lot. It's up today and the stock is done
(07:44):
very well. So okay, Caterpillar, who else? What are some
of the what else do you see besides these big
turbines when you go out and tour this facility. So
the stock is a rocket.
Speaker 2 (07:54):
So that's any people like securities are.
Speaker 5 (07:57):
So many people and I mean it's there. I don't
know you've seen like Silicon Valley and there's that scene
where there's like one guy in the data center and
he's like, yeah, you know, night and day don't really
matter much down here, and it's just one guy running.
But in the buildout, I mean build up, there are
every single HVAC technician or guy that knows how to
(08:20):
lay fiber or guy that knows electrical or plumbing within
one hundred mile radius of valley and those are they're
all there, right, There's seven thousand people working on building
this and this is just one site.
Speaker 2 (08:34):
And yeah, so people have been talking about the picks
and shovels phase of AI investment for a while now,
but it seems like stuff is still getting built and
it has further to run.
Speaker 5 (08:46):
My kind of favorite we can you know, and I'm
sure we will discuss whether this is a bubble or not,
but at the end of the day, the money is
getting spent. And my favorite kind of setup in the
market is you have something that has a multiple that
reflects kind of a wrong reality or doesn't really fully
integrate secular theme and then you have a cycle that's
(09:10):
really bad. So whether it's kind of the robotics supply
chain goes throughout the automotive cycle and if you guys
are aware, that's been a nightmare. And then you have
some of these A lot of these companies took really
significant drawdowns when Trump got elected. Because this isn't really
tech Capex, it's much more similar to like LNG terminal cabets. Ye,
(09:34):
that kind of entails the same companies that were benefiting
from the IRA oh, and so everyone sold them. And
now it turns out it's kind of like the horseshoe theory,
right you.
Speaker 4 (09:44):
Can now that's super interesting. Like it's just these sort
of like oldline industrials, and they did very well under
the Obama administration because of all the factories and all
that stuff that we talked about. And then it's like, okay,
we're pulling the plug on this, but now there's just
this title wave money. By the way, Caterpillar was a
three hundred and thirty four dollars stock on April second,
(10:05):
right before libration day, which is sort of where we
should benchmark things now, four hundred and eighty six at
an auld time high.
Speaker 3 (10:10):
What are some of the other companies see for real?
Speaker 5 (10:13):
Like preferably the ones that haven't gone up? Yeah?
Speaker 4 (10:15):
Yeah, seriously, Like who else is who else is getting
a piece of this action.
Speaker 5 (10:20):
So something that I want to make sure I mentioned
just to frame this whole thing. If you look at
the building side plans, they have the names of all
the data halls and they are literally each called ludicrous building.
So it's ludicrous building, one ludicrous building.
Speaker 4 (10:35):
Amazing.
Speaker 2 (10:35):
Wait is it Ludicris with a K?
Speaker 5 (10:38):
No?
Speaker 2 (10:39):
That would be great. I mean Ludacris spelled correctly is
also fantastic.
Speaker 5 (10:45):
But the company, I mean really I would encourage anyone
to go and look at this video because the it's
it's just a there is such a disconnect between looking
at green numbers going up like like caterpillar going okay, cool, Yeah,
and then you see it and it is mind blowing.
(11:06):
I mean they have rows of dry and atabiatic coolers
that are you know, hundreds and hundreds for the liquid
cooling loops, and then you have the kind of all
the HVAC, the power generation and transmission. We compile the
list of basically two hundred companies, fifty of which were
(11:26):
high confidence in kind of seeing as part of this
build out. And the interesting thing is you have these
multinational OEMs. This scale of this project is so large
that you can call one of them and say, okay,
we want to go with you know, Mitsubishi heav or
we want to use Eton for this thing, and they
(11:47):
will say, yeah, okay, well we can do sixty percent
of it. I mean, the biggest companies in the world
are Yeah, they're just fighting. They don't have to fight
with each other because the pie is so big.
Speaker 3 (11:58):
How scarce you know, there's all this around the country.
Speaker 4 (12:02):
There's been all this tensions like is there gonna be
enough energy for data centers?
Speaker 3 (12:07):
What about the water? What about the politics? People?
Speaker 4 (12:10):
We see these town halls where the locals do want
a data center. How scarce are locations like Abilene? I
mean there's a lot of empty space in Texas. Could
you see Are there more stargates or more Abilenees out
there waiting where the companies can just say, you know what,
we're just going to build our own infrastructure. We're going
to build our own natural gas plant, and we're just
going to do it in the desert in Texas where
(12:31):
we don't have to worry about nimbi's or local politics
or anything.
Speaker 3 (12:34):
Is that the future for a lot of this stuff?
Speaker 5 (12:35):
Yeah, there's already another facility that's planned in Shackleford, Texas,
which is not too far from Abilene. Abilene is already
planning on expanding by another six hundred megawats. I mean,
that's how we're measuring data centers now, is just in
megawatts and gigawatts. In Louisiana, that's where Meta is building Hyperion.
You have some unidentified or unannounced Midwest facility that'll probably
(12:59):
be in Ohi, Io. Basically, yeah, anywhere that you can
get the gas, anywhere you can get the power, anywhere
that there's the land. There was one planned in Indianapolis.
And it's interesting you bring up water because we wrote
about water a little while ago as well, and we
were basing our projections off of what the water consumption
of like existing data centers were. The advancements in liquid
(13:21):
cooling have changed that a lot, so there was a
huge kind of uproar about, well, this data center is
going to use so much water. And if you look
at it, this was an Indianapolis and if you look
at it, it's about the same amount of water as
a thousand households. Because it's all closed loop liquid cooling. Now,
the water just gets recycled and filtered back, so it's
(13:44):
really just something where the constraint is still power and
getting these kind of turbines, and it's going to increasingly
become a thing where it's behind the meter power generation.
They're just going to build their own power plants. They're
not going to rely on the grid. You guys remember
what happened in Texas and twenty one with a Yeah,
they you cannot underwrite that as a data center, like
(14:08):
you just can't be offline. So that's kind of where
we're headed, is that this power and energy kind of
bottleneck is the bottleneck. One of the more interesting things
we ran across is you look at Xai's data center,
and we went through a lot of the filings of
all these different data centers. They're permitted for fifteen thirty
(14:28):
five megawatt turbines, and there's thermal imaging of the site
that shows they're using thirty five ge Vanova.
Speaker 3 (14:36):
Another one that was a three.
Speaker 2 (14:39):
You're just typing tickers.
Speaker 4 (14:40):
Yeah, there was a three hundred and thirty dollars stock
on April second. That's a six hundred and nine dollars stocks.
Speaker 5 (14:44):
Okay, check out Stemens Energy.
Speaker 2 (14:46):
Thank you, Joe. Oh he's doing it.
Speaker 3 (14:49):
You could keep going.
Speaker 2 (14:50):
All right, I have kind of this isn't an essential question,
but okay, wait, I got it.
Speaker 5 (14:56):
I knew this is going to happen.
Speaker 4 (14:58):
Semens Energy was a fifty six dollars stock on April second.
Speaker 5 (15:01):
That's one hundred and eight.
Speaker 4 (15:02):
That's a string lineup.
Speaker 2 (15:03):
Okay, okay, clearly someone's making money. Like how big are
the walls of these things? And what does security actually
(15:26):
look like? Because you know, we talk about people being
upset about higher energy prices. I understand that maybe some
of the some of them are building out their own
energy capacity, but you know, in a populist mindset, people
might still be upset at AI, still upset that AI
is going to take jobs and things like that. They
could become targets.
Speaker 5 (15:45):
Right, yeah, absolutely, I mean apparently security isn't tight enough
to shoot down at drone yet, oh of course. Yeah.
But when you talk about the walls, like, the interesting
thing is everything in this data center has You can't
just build it like a house. Everything needs to have
the right anti corrosion insulation. You can't just be losing heat.
You're already spending so much money on cooling. So everything
(16:08):
in the building envelope is also kind of the most
expensive thing it could be.
Speaker 2 (16:12):
And I assume it's like a sealed envelope.
Speaker 5 (16:14):
I guess, yeah, exactly, And the opportunity of just how
big is this going to be if you actually believe that.
The people that are building this, and this is kind
of an interesting dynamic of the hyperscalars are in somewhat
of a prisoner's dilemma. We can't speak to whether they
actually believe or not, but the party line is we're
(16:38):
building machine God. And if we're building machine God, then
that's kind of a religious zelotry. So three years ago,
most of these hyperscalers were making net zero. By twenty thirty,
that is not happening. Yeah, there's no way that's happening.
Maybe they buy carbon off, but that's not happening because
if you need to do it now, you're using natural
(16:58):
gas and that then you say, well, once we create
machine God, machine God will fix it, right, and you
can justify a lot of stuff.
Speaker 2 (17:11):
Wait, okay, so let's talk about is AI in a bubble,
because when we're conversing about buildings that are literally called
ludacris and we're talking about machine God will solve all
our problems.
Speaker 5 (17:23):
That sounds a little bubbly.
Speaker 2 (17:24):
Yeah, that's the kind of stuff you look back on
and say, oh, yeah, that was the top I know.
Speaker 5 (17:29):
That we we have to talk about this and that
requires some numbers, so I have say notes. But the
interesting I mean, you can make a case for either side,
and it kind of hinges on not so much what
AI looks like in the future, but more where we
are in the cycle. If the question is is AI bubble? Yeah, probably,
it's a new technology. It's exciting every single time we've
(17:51):
the railroads, is the you know, dot com? Probably if
it's not a bubble, it will be, and it probably
is already. But where are are we in the bubble?
That's the interesting thing for investors. So if you look
at kind of where are we in the financing side.
When we had the dot com bubble, the big thing.
(18:14):
Are you guys familiar with dark fiber? No? So essentially
they were building out this kind of infrastructure for the Internet,
and the conviction of the day was the Internet's going
to be the biggest thing in the world. Correct, And
ninety seven percent of the fiber that was laid was
called dark fiber because it wasn't lit up. It was
just laying there because eventually we're going to need it.
(18:37):
That's not the case right now. Every single time that
one of these data centers comes online, you're at one
hundred percent utilization. There's no dark fiber, and we're continuing
to need more. I mean, I don't know if you
guys saw Sora too. More than seventy percent of the
content that we consume is video, so a video model
makes a lot of sense. It's also way more compute intensive,
(18:58):
and the the final form of chat of AI is
not chatbots. It's gonna be video. It's gonna be robotics,
which use video models. So that's kind of like the
bullish side. But I totally see. I have some numbers here,
So we've had some of these circular deals and so far,
(19:22):
I mean, there's hundreds of billions of dollars for this.
Most of the financing is structured much more like toll
roads or LNG terminals. The Metas hyperion was twenty nine
billion dollars. Twenty six billion of that was Pimco on
debt and then blue outl with three billion of equity,
an off balance sheet arrangement, which also was pretty popular
during the dot com bubble. Then you have stuff like
(19:44):
GPU collateralized SPVs core. We've had a seven and a
half billion dollar Blackstone SPV. There's like an eleven billion
dollar GPU backed loan market now, which compared to the
infrastructure thing, I can get that. It's like those GPUs
are gonna be worth like they will depreciate. So yeah, that's.
Speaker 2 (20:05):
That's abs right, that's like asset backsxety Okay.
Speaker 5 (20:08):
Yeah, but when you look at the where we are
just saying where are we in the cycle, Well, there's
been there's been a lot of these balance sheet arrangements.
But in terms of that prisoner's dilemma that we spoke about,
where nobody's going to back off because then maybe they
risk not creating machine GUD. The debt to equity ratio
on most of the hyperscalars is not crazy. Sure Oracle
(20:31):
has the highest, and but they still have a lot
of room to lever up in pursuit of this goal.
The interesting kind of dynamic of in Vidia finances its
customers and then its customers buy Nvidia GPUs, and then
you know that, yeah, that's a bubble dynamic. That's that's
visible circular financing, and but the demand is real.
Speaker 4 (20:52):
Well, this is this is I guess the question. If
the demand is real, why are the videos of the
world financing your customers? Right, intuitively you think, okay, in
videos the demand for chips is maxed out, right, Then
intuitively you think, well, then there's.
Speaker 3 (21:08):
Plenty of money elsewhere.
Speaker 4 (21:10):
So do you have a guess for why in VideA
is compelled to you know, invest in open Ai, invest
in core weave. It's been a long time investor in
core weave and so forth.
Speaker 5 (21:19):
Yeah, I could. I could take the very skeptical view
or the kind of non skeptical view. The non skeptical
view is the best thing for in Vidia is that
we accomplish a GI. So anything that it can do
to get us closer to that massive, massive infrastructure that's
required for that is great for them as quickly able,
as quickly as possible, because, uh, every year that you
don't achieve AGI becomes less likely. So uh, that is
(21:44):
maybe the core factor. And then there's it's good for
them and the and playing into that the because this
this isn't necessarily like the dot com bubble, because the
dot com bubble had fiber and then it had you know, pets,
dot Com and Amazon and all that stuff. This is
all pretty much capex right to tech is capital intensive again,
(22:07):
which means that it's the bust won't necessarily be like
the dot com bubble, where it gives the real players
time to shine. If the CAPEX spending grinds to a
halt because the market goes down, that's the worst thing
in the world. So in a way, it's also it's
in their interest obviously to make sure that that doesn't happen.
Speaker 4 (22:41):
Another question, actually, but I want to ask it now
because I might forget otherwise. A headline I think earlier
this week again headline is almost every day. Can you
explain why Meta, which is building out which obviously hyperscaler,
they have tons of data center capacity, why do they
need to do a deal with Corewave? So I saw
that Meta is going to pay a few billion, fourteen billion,
(23:04):
it's some number for capacity on core Weave. What can
the what do the neo cloudes bring to the table
such that the legacy is like, you know what, we
launched some of your capacity.
Speaker 5 (23:13):
It's basically I would say, anything you can do to
shift that CAPEX away from yourself is great. Right if
you're if you're right now, a lot of this stuff
is funded from cash flow, and yeah, we're starting to
see that financing take its place. But if you the
the you're running a company that's existed for a while
(23:33):
and is one of the biggest companies in the world,
and yes, you could justify taking an ordinate amounts of
risk to do this, But just like anyone else, if
you have a goal and you can kind of shift
away some of that risk to because if you know,
if core Weeve goes bankrupt, Meta is not it's not
It's not the worst thing in the world for Meta.
Speaker 2 (23:51):
What would you be looking out for for signs of
the bubble bursting or I guess the capex FAWCET starting
to turn off.
Speaker 5 (23:59):
So in terms of like the warning signs first, monitoring
used GPU prices I think is a great way to
see what's going on, seeing like contract renegotiations, waiver headlines
from private credit. And also a big one would be
any delays in delivering power make it so that you
(24:20):
finance this thing and then it's not going to be
able to work, right. I do think that another you
really do have to kind of index to the revenue
that's being generated by the AI companies as well, and
you saw open Ai recently announced that they're basically allowing
you to buy things with one click from the inside
the app. And I think that that was also a
(24:42):
big driving factor in the router where they kind of
route your model to the it's setting it up for advertising,
because the other day I asked Google a pretty simple
question and it just gave me the answer with no
advertisements delivered to me. So we need to I think
that there is the killer apps kind of are here
or coming, but that has to happen fast.
Speaker 2 (25:05):
Yeah, the fact that the Internet kind of runs on
ads is an underappreciated reality of our world, and we
did an episode on it way way back in the day.
Is there anyway if I wanted to get exposure to
like the good stuff of AI, the actual money being spent,
but then I don't like all the circular financing arrangements.
(25:26):
Is there a way to hedge that particular risk?
Speaker 5 (25:29):
Yeah, you could buy all the real companies that are
building this kind of data center, and then you could
short Blue Owl and the private credit companies that are
doing the kind of financing aspect of it, because if
it goes bad, they have limited upside and kind of
unlimited downside, and if it continues happening, the companies that
are actually building it have unlimited upside and they're still,
(25:52):
like you said, these oldline kind of industrial that like,
they still have a core business to go back to.
Speaker 4 (25:57):
What's your take on so open ai announced its new
Sora thing that Tracy has a code for, Meta announced
it's Meta AI, which a lot of people is like,
this is slop and I don't we don't need to
get into a debate about whether this is garbage. I
kind of think it is, but that's not actually the
question I'm interested in. The question I'm interested in it
is one way to view these things is you know,
(26:19):
what these are going to be, These video, these AI
videos are going to be really important sort of business
revenue drivers in our pursuit to AGI. And the other
view is, look, if they were close to AGI, they
would not be devoting time to the slop factories. And
this is the sort of question and I'm curious from
your perspective, what do you think is the real logic
(26:40):
behind all of this effort to create these sort of
you know, dom like, oh, a video of a astronaut
riding a horse through space with a you know, cowboy hat.
Speaker 5 (26:48):
On or whatever. Video models are going to become the
most important part of this. You you need, for example,
lms don't really interact with the world, but we're starting
to see in robotics, AI models interact with the world.
So you have these video language action models, and you
can structure it so that there was an interesting paper
(27:08):
on Archive recently that and you know this is video
models like VO three are emergent zero shot learners and understanders.
And what that just means is you can give it
basically like a second of context, and then it can
extrapolate about what's going on beyond that. That becomes extremely
important for robotics. And you can structure these vision language
(27:31):
action models where you have like a slow reasoning model
that is kind of like your conscious mind and determines
what it needs to do, and then you have like
a fast model that actually does the movements and stuff.
AI kind of extends into the realm of you need
to generate as much data as possible, and having you know,
video data and knowing what's good and what's bad is
also very important. But yeah, to your question, open AI
(27:54):
probably does want to show that they're making more money,
and I don't doubt that they're people there that truly
believe that they're building the machine GUD. And I also
don't doubt that they realize if they have a shot
at doing that, they need the market to play long
and they need to keep growing their revenues. So I
think there's a lot of really good counter arguments about
(28:15):
like not necessarily are we in a bubble?
Speaker 2 (28:17):
Yes?
Speaker 5 (28:18):
Is the bubble over? Like? There are good arguments for that.
I don't think one of the I don't think that, oh, well,
they're leaning into like kind of this slops that you know,
it's a company. They're going to do whatever they can
to make money. That's capitalism, you know.
Speaker 2 (28:31):
One of the things I think about another reality of
our current time. This is kind of related, but if
we're talking about video data, YouTube staying power has been
like phenomenal, right, It's been around for decades now, and
we haven't really seen any successful competitors anyway.
Speaker 5 (28:49):
Well, since we wanted to talk about stocks that are
going up, and I know that's another one people kind
of forgot that video takes up a lot more storage
than textas and if we're going to start just like
throwing out the slop on video the part. Basically, one
of the only areas of semiconductors that has gotten ignored
(29:09):
has been storage, not like not like memory like like
uh RAM and stuff, but like hard drives and solid
state discs like Western Digital and uh okay.
Speaker 4 (29:20):
So Seagate, Yes, that's what we're talking about. So this
was a this was an eighty four dollars stock in
late March, and now it's a two hundred and fifty four.
That's where you're going, Yeah, okay, yeah, yeah, there's a crazy.
Speaker 2 (29:31):
There's a crazy, but it has gotten attention. But you
think it could go up more?
Speaker 5 (29:36):
I think so, and I think that uh to illustrate
that dynamic that's going to keep happening. Basically, you just
keep looking at stuff where there's people say, oh, well,
you know there's a there's a glut of X and
that's why this isn't going to win from AI. And
then eventually the glut of X doesn't matter. It's a
That's what happened to Nvidia in the beginning of all this.
There was a glut of crypto GPUs and now nobody care, like,
(29:58):
wh when's the last time you heard some want to
be barrassed on a video because there's too many crypto minors.
Speaker 4 (30:02):
Sand Disc was forty eight dollars in April second, that's
one hundred and twenty six dollars stock.
Speaker 5 (30:07):
Pretty crazy.
Speaker 4 (30:08):
These are just these are like just the tools are
just store just basic.
Speaker 3 (30:11):
Yeah, yeah, crazy.
Speaker 5 (30:12):
It's it is and it's it. I guess you could
be worried by that, or you could say this isn't uh,
this isn't like in Nvidia doubling in a month necessarily right,
that it's just people discovering. I think it's a good barometer,
which is important in a bubble, that people are still
paying very close attention. And you start to get worried
(30:34):
when everybody just goes into the cisco of it all.
But these are real companies that have real products. Of course,
there's some silly stuff that's happening, but it's not only
silly stuff. It's not twenty twenty one with the you know, yeah,
I don't need to.
Speaker 2 (30:51):
So speaking of on the ground research, or maybe I
should say in the sky research. When it comes to robotics,
you had a robot dog, right, yeah, and of those
like really advanced ones, we got a unitary robot dog.
Speaker 5 (31:03):
It is super impressive.
Speaker 2 (31:05):
It's also incredibly creepy when it stands up on its
hind legs.
Speaker 5 (31:09):
Yeah, have you seen the video?
Speaker 3 (31:11):
Yeah, I saw your video of it.
Speaker 5 (31:12):
Yeah, it's uh. And Unitry we went and visited Unitry too,
and it's I've seen Boston Dynamics. You know spot robot,
which is actually more impressive. Also could probably murder you
like and and the interesting thing about Unitry is, yeah,
this the robot kind of feels like a weapon. It's
(31:32):
very heavy and stuff, but it's even they're humanoid, like
it can't lift more than eight pounds. And also Unitree
is not focused on like making the brain. They just
want to make the hardware. And you go there and
they were like, where's your light are from? Oh, well
we needed this specific thing, so we just made it ourselves. Okay,
what about your actuators, Well we made those ourselves. Do
(31:55):
anything that they don't make themselves. They get within a
ninety kilometer radius of their headquarters. But where's their head
It's in hangzhoune So it's something where you're really that's
the competition that we have stacked up against us in
the US versus China, and I think the interesting route
that they've taken is it's super advanced. You can code
(32:17):
on it. You can already see videos of people you know,
buying them and making them do specific things, but they
made them not super like, it's not going to kill you.
And that's a great way to get a bunch of
training data because we don't have enough training data of
what it's like to move around the world. We have
the sum history of all human knowledge that's been recorded
for text models, but we need more data for robotics.
(32:41):
And I think that that's like an interesting kind of
thing where they're just kind of sending these out into
the world so that they can get enough data to
actually make it safe enough to have a robot in
your house.
Speaker 3 (32:53):
Can anyone buy a unitary dog?
Speaker 2 (32:54):
Like?
Speaker 3 (32:54):
Are they for sale online?
Speaker 5 (32:56):
Yeah, there's a lead time, but I can get around
it if you wait.
Speaker 2 (33:00):
How much are they?
Speaker 5 (33:02):
So they start at three thousand, and then we got
the super advanced one that you can attach a robotic
arm to so it can manipulate its environment, and that
was eight thousand dollars, and then they're humanoid is sixteen,
which is like less than the cost of the used
time to sither.
Speaker 2 (33:17):
Well, three thousand. Also the price for like a really
good pure bread dogs, I don't know what that.
Speaker 3 (33:23):
People pay a lot of money for actual dogs.
Speaker 2 (33:25):
So yeah, it seems why are you looking at me?
Speaker 3 (33:28):
Why do you think I'm looking at you?
Speaker 2 (33:30):
Okay, James, that was so much fun. Thank you so
much for coming in and sharing with us your your
field trip observations.
Speaker 5 (33:37):
Thanks Levan, Joe.
Speaker 2 (33:52):
I think maybe it's the cynical journalist in me, but
I get really nervous when we start saying it's different
this time.
Speaker 4 (34:00):
Yeah, I mean it's always different. It's always the same
as sort of there are.
Speaker 3 (34:04):
Certainly you know.
Speaker 4 (34:06):
Yeah, I mean, look, it's crazy and it's a lot
of money, and it's still you know, we've been asking
people on the podcast what kind of use value are
you getting out of this technology? And it's still very ambiguous.
Speaker 5 (34:19):
I don't know.
Speaker 4 (34:20):
I don't have It's not my job to time the market,
so I'm glad I don't have to have a view
on all of this stuff.
Speaker 2 (34:26):
But I do think, cop out, Joe, No, I do think.
Speaker 5 (34:29):
I do think.
Speaker 4 (34:29):
For Okay, what's your okay, give us your give us
your price target.
Speaker 3 (34:33):
That's okay, Okay, give us your price time.
Speaker 4 (34:35):
Okay, I do think this is moving so fast and
every day the numbers are so big, and it's the
stakes are so high for all of this that it
does feel like we just need regular updates on the
state of how much money is being spent in this
area and what the eventual payoff will be.
Speaker 2 (34:54):
I want someone to draw a massive, like infographic of
all the financing kills, because I think you could make
a really interesting one.
Speaker 3 (35:02):
Yeah, you definitely could.
Speaker 2 (35:03):
That would probably be illuminating. On that note, I did
think James's suggestion of, you know, shorting the private credit
players and then going along the actual physical hardware ones
that was interesting.
Speaker 5 (35:15):
Yeah.
Speaker 4 (35:15):
No, it's really interesting. Also, I had not realized. I mean,
it is very interesting the way booms or bubbles, whatever
you want to call it. I don't like using the
word bubble because it applies whatever booms or bubbles, the
way they spread out.
Speaker 3 (35:29):
I just think that's extremely interesting.
Speaker 5 (35:31):
Right.
Speaker 4 (35:31):
So, in video takes off in late twenty twenty two,
when people are blown away by chat CHPT and they
say there's incredible and it's being powered by all these
in vidio chips, and then people go on the hunt
right for other things. It's like what else, And even
still today in October twenty twenty five, that spreading out
process continues. And so okay, in the last few months,
(35:53):
maybe it's like a caterpillar, or maybe suddenly people are
realizing that just sort of old fashioned computer storage is
to be very hot. But we're still in this phase
whereas people open their eyes to these things, the spreading
out continues to this day right.
Speaker 2 (36:08):
Well, Also there's an expectation that eventually, you know, normal
companies are going to benefit from it as well and
become more efficient and all of that.
Speaker 4 (36:17):
I just remembered the one question I was going to
ask James, but I could just throw it out as
sort of a spectative question. So he mentioned all of
the workers. Could you ask all of the workers that
are out this or that are building these things? And
it does make me wonder about the sort of crowding
out possibility, Like imagined if you imagined if you had
some other business enable Leen, and you're trying to hire
an HVAC worker right now because HVAC you know, you're like, oh, restaurant.
Speaker 2 (36:40):
And you're competing with like open AI.
Speaker 4 (36:42):
Yeah, you're like, you have a restaurant enabling, and you're like,
you know what, I need to update my air conditioner
or whatever or whatever, and what are you going to
do because all the HVAC workers are at that big
stargate operation. And I do think this is a very
interesting macro question on all kinds of things, which is
the degree to which all of this spending with uncertain
(37:03):
payoffs will have a crowding out and potentially inflationary effect
on the rest of the economy.
Speaker 2 (37:07):
Well, presumably machine God will eventually design HVAC robots, so
problem will be solved.
Speaker 3 (37:13):
Problem will be solved.
Speaker 2 (37:14):
All right, shall we leave it there?
Speaker 3 (37:15):
Let's leave it there.
Speaker 2 (37:16):
This has been another episode of the oud Loots podcast.
I'm Tracy Alloway. You can follow me at Tracy Alloway.
Speaker 4 (37:21):
And I'm Jill Wisenthal. You can follow me at the Stalwart.
Follow our guest James Van Gillen, He's at Sutrini seven.
Follow our producers Carmen Rodriguez at Kerman armand dash O
Bennett at Dashbod and Kale Brooks and Kale Brooks. For
more odd Laws content, go to bloomberg dot com Flash
odd Lots. We have a daily newsletter and all of
our episodes, and you could chet about all of these
topics twenty four to seven in our discord discord dot
(37:44):
gg slash.
Speaker 2 (37:44):
Odd lots And if you enjoy odd Lots, if you
like it when we talk about the physical realities of AI,
then please leave us a positive review on your favorite
podcast platform. And remember, if you are a Bloomberg subscriber,
you can listen to all of our episodes absolutely add free.
All you need to do is find the Bloomberg channel
on Apple Podcasts and follow the instructions there. Thanks for listening.
Hosts And Creators
Joe Weisenthal
Tracy Alloway
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Breakfast Club
The World's Most Dangerous Morning Show, The Breakfast Club, With DJ Envy, Jess Hilarious, And Charlamagne Tha God!