January 22, 2026 • 31 mins
In part three of this week’s four-part Better Offline Enshittifinancial Crisis special, Ed Zitron walks you through how the imaginary demand for AI compute is setting up every single data center developer for disaster - and how many major banks will be caught in the contagion.
This series was a ton of work, so please support me by subscribing to my premium newsletter - here’s $10 off your first year of annual https://edzitronswheresyouredatghostio.outpost.pub/public/promo-subscription/84rt762qen
Read along with the newsletter version for links! https://www.wheresyoured.at/the-enshittifinancial-crisis/
YOU CAN NOW BUY BETTER OFFLINE MERCH! Go to https://cottonbureau.com/people/better-offline and use code FREE99 for free shipping on orders of $99 or more.
---
LINKS: https://www.tinyurl.com/betterofflinelinks
Newsletter: https://www.wheresyoured.at/
Reddit: https://www.reddit.com/r/BetterOffline/
Discord: chat.wheresyoured.at
Ed's Socials:
https://www.instagram.com/edzitron
https://bsky.app/profile/edzitron.com
https://www.threads.net/@edzitron
Email Me: ez@betteroffline.com
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 (00:03):
Media. Hi, I'm ed Zetran and welcome to Better Offline.
Speaker 2 (00:19):
That's right, folks, were back for the third part of
our four part series about the terminal stage of the
incitification process, the in shitification of the financial markets. So far,
we've talked about the incertification of the stock market, the
complicity of the analysts and the media, and how venture
capital will be the Canarian the coal mine for any
eventual collapse. Venture capital is both the perpetrator and the
(00:41):
victim of the air bubble, and it's worth exploring how
it came to this point. Now, hang on the gentle voice, folks.
We haven't seen values this big and alone time. These
the biggest numbers we've ever seen. They're simply tremendous open
airs maybe with eight hundred and thirty billion. Can you
believe that they lose so much money? But folks, we
don't mind. Clamisam Morbourn. They call him Klamiel because everybody
he's going to give them one billion dollars and where
(01:02):
we love him dead. Centers are going to have the
biggest deals we've ever seen, even even if we have
to work with Dario. Sorry to those of you who
don't like that voice. It was only about thirty seconds,
you see right Now, AI startups are big exciting news
for the limited partners funding large language model companies. Things
feel exciting because the value of the assets under management
(01:22):
aum are going up, which is nothing dodgy, per se,
but it's just how vcs value things. And if they're
valuing AI stocks, well that's how their fees are getting paid,
because the thing is they've invested in are worth more money,
and now limited partners are paying them a percentage so
that they can sit in an apartment that costs too
much not see friends because they just sit online posting
for eleven hours a day. Yeah, and then they can
(01:44):
lose all the money in the future. I guess investing
early in open Ai allows a bench capitalist or even
an asset manager like Blackstone, which invested in twenty twenty
four to say it as a big holding and a
big increase in its aum AI stocks, make venture capitalists
who bet on them two years ago, Like genius is
on paper, you got in early on open Ai, Athropic, Cursor, Cognition, Perplexity,
(02:05):
or any other company that loves to burn several dollars
per dollar of revenue. You have a big, beautiful number,
the biggest you've ever seen, and your limited partners to
pay you a feed just to manage it. Venture capital
has not seen valuations like this in a long time,
and on paper it feels a lot like a lot
of vcs got in on companies worth billions of dollars
right on paper, that is, on paper, Cognition is worth
(02:28):
ten point two billion, Perplexit eighteen billion, Cursor twenty nine
point three billion, Lovable six point six billion, Coher six
point eight billion, Repic three billion, Glean seven point two billion.
Massive valuations to companies that all basically do products that
open AI or Anthropic or Amazon or Google or any
number of Chinese companies are already working to clone. They're
(02:48):
all losing lots of money too, and have no path
the profitability of any kind. But right now the numbers
are simply tremendous. Folks. I've heard venture capitalists tell me
that there are times when they have to agree to
invest with little to no information or know that they'll
lose the opportunity to another sucker, I mean, investor. I've
heard venture capitalists say they don't have any insight into finances,
(03:11):
but we love the numbers. Folks, they're simply tremendous. We
love it. Let me tell you how not tremendous they are.
I had a source recently suggests that nobody who invested
pasted open AI's one hundred billion dollar valuation have any
information rates. And what that means is they literally do
not know anything about the company. They don't know anything
other than what the company will tell them. That the
(03:32):
company doesn't have to tell them shit, which means they
don't know the revenues, they don't know the costs. They
only know what new start new ridiculous stat Sarah Friar,
CFO of open ais, farted out. They posted recently that
twenty billion dollars annualized revenue. Yeah. I think that's complete nonsense,
But they want to do that so that you go
open ai made twenty billion dollars in twenty twenty five.
(03:54):
Fuck that, I don't even think they made ten. Now,
venture capitalists would argue, of course that they'd say I'm insane.
They would say they would think me mad, and they
would say that the growth is obviously there, while pointing
to whatever startup has made one hundred million dollars in
annualized recurring revenue, which is eight point three million dollars
a month, by the way, all while not discussing the
(04:16):
underlying operating expenses. The idea I believe is that the
current state of AI spending is only set to increase
the next year, and that will somehow lead to fixing
the margins. I think venture capitalists staunchly refuse to learn
anything other than investing growth and then profit from growth,
even if profiting from growth doesn't appear to be happening anymore.
(04:39):
In reality, venture capital shouldn't have touched LMS with a
shitty stick because the margins were obviously blatantly bad from
the very beginning. We knew open ai would lose five
million dollars in the middle of twenty twenty four. The
same venture capital climate would have fucking panicked, but instead
chose to double dribble and kadruple down. I believe that
massive valuation drawdowns are a certainty for AI companies, and
(05:01):
by that I mean these companies will end up being
revalued and the number will be smaller than the big
number everybody fell in love with. Look, venture capitalists, I
gotta ask you a question, what happens if open AI dies?
Do you think this will make investors interested in funding
or acquiring other AI startups. How much longer are we
(05:21):
going to do this? When will Venture Capital realize it's
setting itself up for a disaster And what exactly is
the plan? Open AI and anthropic will cuck the lake's
dry like an Nvidia GPU named after Nancy Reagan? How
is this meant to continue and what will be left
when it does? The answer is simple, there won't be
any money for venture capital for a while. Those AI
(05:42):
holdings are going to be worth at best fifty percent
if they retain any value at all. Once one of
these startups die, Once there's a major casualty, a panic
will ensue, sending venture capitalists scrambling to get their holdings
acquired by which I mean sold to someone else, until
there's little or no investor interest left. Why would limited
partners ever trust venture capital after this? Why would anybody?
(06:04):
Because based on the past four years, it doesn't appear
that venture capital is actually good at investing money. It
just got lucky year after year until there were a
few ideas that they could sell for one hundreds of
billions or billions of dollars, And now we're out of them,
we're out. I think we're out. I think there's probably
still some, but I think the era of the unicorn
might be dead. For a while, venture capital believed it
(06:27):
knew better as it turned its back on basic business fundamentals,
starting with Clubhouse, Crypto, the Metaverse, and now Generative AI.
Yet they're far from the only fuck wits on the
Dickhead Express. Because yes, I am going to talk about
data centers. Per Bloomberg, there were at least one hundred
and seventy eight point five billion dollars in data center
credit deals and in the US in twenty twenty five,
(06:49):
rivaling the two hundred and fifteen point four billion dollars
invested in US venture capital in twenty twenty four and
the one hundred and ninety seven point two billion dollars
invested in US venture capital through August seventh to twenty
twenty five, and over one hundred billion dollars more than
the sixty point sixty nine billion dollars of data center
deals done in twenty twenty four. That's that's really worrying.
(07:10):
It's like more than more than more than one hundred
percent more. I'm very worried, and I'm going to tell
you why, using a company called core Weave, core Weave
that I've been actively warning about, warning people about since March.
And I get a little crazy around corwaf because I
have had the core Weave conversation a lot, and I've
seen people straight up just steal the title of my
(07:33):
articles as they come to this realization six to seven
months later. But cor Weave is a mess, and corwy
was obviously always a mess. And everyone has been saying
Corwave's fine because they went public. Core Weave is not fine.
All right, Let's talk about core Weave. Corwave is something
called a neocloud. It's a company that builds data centers
(07:56):
by renting out, and they build data centers they film
full of GPUs and they rent them to people. They're
in something called AI Compume, And as I explained a
few months ago, they do so by building data centers
backed by endless debt. And I quote myself saying, up
a neocloud is expensive. Even if the company in question
already has data centers, as core We've did with its
scryptocurrency mining operation, AI requires completely new data center infrastructure
(08:20):
to house and call the GPUs, and those GPUs also
need paying for, and then there's the other stuff like power, water,
and the other bits of the computer, the CPU, the motherboard,
the memory of the storage, and the housing and the
power supplying all that good stuff. As a result, these
neoclouds are forced to raise billions of dollars in debt,
which they collateralize using the GPUs they already have, along
with contracts from customers which they use to buy more GPUs. Corwei,
(08:44):
for example, has twenty five billion dollars in debt on
an estimated five point three five billion dollars in revenue,
losing hundreds of billions of dollars in a quarter. You
know who also invests in these neoclouds in Video, the
people who make the GPUs in VideA is also one
of Corwey's largest customers, accounting for fifty percent of its
revenue in twenty twenty four, and just signed a deal
(09:05):
to buy six point three billion dollars of any capacity
the care we can't otherwise sell to someone else through
twenty thirty two, an extension of a one point three
billion dollar deal from twenty twenty three reported by The
Information oh And in Video was the anchor investor of
Corweav's IPO, putting in two hundred and fifty million dollars.
Corwave is also one of the largest providers AI compute
(09:25):
in the world, and its business model is indicative of
how most data center companies make money. In fact, i'd
argue that they are probably doing better because they have
the backing of everyone. Now you may think that's a
too big to fail thing. Not really, It just means
that they're kind of the village bicycle of AI compute grossness. Aside,
let me explain how this works. First, core we've signed
(09:46):
contracts such as it's fourteen billion dollar deal with Meto
or it's twenty two point five billion dollar deal with
open Ai before it has the physical infrastructure to actually
service them. It then raises debt using this contract as
collateral orders GPUs from VideA. Those take three months to
arrive and then another three months to install, at which
point monthly client payments begin. And I should also add
(10:07):
that that's kind of taken out of deferred revenue that
the customers already put down. To really simplify this, data
center developers are raising money months up to a year
before they ever expect to make a penny of revenue,
not profit. In fact, I can find no consistent answer
to how long a data center takes to build, and
the answer here is pretty important because that's how the
(10:28):
money is going to get made from the data center's
for building fucking everywhere. You may also notice that monthly
payments in the chart that I linked, which I should
have mentioned, begin at six to thirty months, a curious
and broad blob of time. You see, data centers are
extremely difficult to build, and the concept of an AI
data center is barely a few years old, with the
concepts of hundreds of megawats in one data center campus
(10:49):
entirely made up of aigpus barely two years old jpus,
which means basically everybody building one is doing so for
the first time, and even experienced developers are running into problems.
For example, core Scientific, Corweave's weird partner organization it tried
and failed to buy that is truly different despite the
similar name, has been trying to convert it's Denton, Texas
(11:11):
scriptocurrency mining data center into an AI data center since
November twenty twenty four, specifically so that Corewave can rent
it to Microsoft for Open AI. This hasn't gone well,
and this is when things get weird. The Wall Street
Journal reported a few weeks ago that Denton had been
racked with several months of delays thanks to rainstorms preventing
contractors and pooring contry contrete. I'm going with it. It's contrete.
(11:34):
Now we're calling it contry concrete. I originally, within this
script wrote down that this was a possible thing. I
thought that this was normal. However, I should add something,
this is this may actually be bollocks. So I'm a psycho,
and I went and I looked up whether. I looked
up weather in Abilene, Texas, and it turns out that
(11:57):
there were like eleven days with rain total, and there
was only I think on June third and September twenty second,
there were rainstorms or thunderstorms of any kind. I think
that there are only two days where rain was over
point one inches. The people I've talked to about this
deal are saying it wouldn't be weather, it would actually
just be money. Nevertheless, those are things that are stopping
(12:19):
them building a two hundred and sixty megawat data center,
and that's fucking nothing compared to the Giga. What's the
open Ai claims to want to build. And what this means,
by the way, is that call we've can't actually get
paid by open Ai because per its contract, customers don't
have to stop paying until the computer is actually available.
This is a very important deal to know for literally
(12:40):
any data center development you've ever heard of or seen.
As of cour we've SQ three twenty twenty five earnings,
they're sitting on one point one billion dollars into thirst
revenue that's income for service is not yet rendered are
(13:03):
from nine hundred and fifty one million in Q two
and four hundred and thirty six million in Q one
twenty twenty five. This means deposits have been made, but
the contract is yet to be serviced. Now, I'm also
a curious little critter, so I went and found the
nine hundred and twenty one page two point six billion
dollar DDTL delayed draw Term Loan three point zero loan
(13:23):
agreement between core Weave and banks including Morgan Stanley, Mitsubishi,
UFG and Gold and Sachs, and in doing so I
learned something well. First of all, open ai appears to
have net three sixty payment terms from Corwave, meaning that
it can literally pay a year from invoice. Second, Corweve
is required to maintain something called a contract realization ratio
(13:44):
of point eighty five times, meaning that core Weave has
to make at least eighty five cents of every expected
dollar or it is in default of its loan. This
is important to note because it means that if say
open Ai decides not to pay up in a year,
corwave will be up shit creek without a paddle. Now,
I apolog that suggests that core Weave is an already
up shit creek, or that it might have a paddle
(14:05):
of some sort buried inside in Vidia's latest earnings on
page seventeen, there's a little clue at home, and I
quote in the third quarter of fiscal year twenty twenty six,
we entered into an agreement to guarantee of partner's facility
lease obligations in the event of their default. The agreement
allows our partner to secure a limited availability facility lease
backed by our credit profile in exchange for issuing US warrants.
(14:28):
The maximum gross exposure is eight hundred and sixty million,
which is reduced as the partner makes payments to the
leaser of a five years. The partner has placed four
hundred and seventy million dollars in escrow and executed an
agreement to sell the data center cloud capacity, mitigating our
default risk credit where credit is due. Eaglide analysts Just
Dario caught this in November, but in core weaves condensed
to consolidated balance sheets, there sits a four hundred and
(14:50):
seventy seven point five million dollars line item under restricted
cash and cash equivalence non current, though this might not
be in Vidia's escrow. This number shifted from six hundred
and seventeen in Q one to three hundred and forty
million in Q two. It lines up all too precisely,
and who else would in video be guaranteeing In any case,
(15:10):
corweav is likely to get the best deals in data
centered debt outside of Oracle. It has top tier financiers
who I'll get to in a little bit, the fallbacking
in Video who is both an investor customer and apparent
financial backstop, and also Core with as a customer of
theirs and the ability to raise debt quickly. Corweb's deals
are likely indicative of how data center financing takes place,
and those top tier financiers it's they may basically been
(15:34):
in every single deal. It's actually it's actually really worrying
when you, I mean, you lay it all out, which
is exactly what I'm going to do. So who's actually
paying for this shit? I went and dug through a
pile of twenty six prominent data center loan deals, including
the proposed thirty eight billion dollar debt package that Oracle
Advantage data Center partners are raising for Stargate Shackelford and Wisconsin,
(15:55):
Stargate Ablene, New Mexico, and of course soft Banks fifty
billion dollar bridge loan, which I included for a reason
that will become obvious shortly, and multiple core wave loans.
I've found a few commonalities, and forgive me, I'm about
to say a lot of names and numbers. You might
want to pull up the newsletter to read along with
it like a sing along. Now, Blue Owl, Blue ol,
We love blue ol folks. They were present in every
(16:17):
single Stargate deal other than the thirty eight billion dollar
deal that's being raised by Advantage that is yet to close.
Blue Hour was also involved in a one point three
billion dollar Australian data set and a debt package by
Virtue of owning Stack infrastructure. Remember that name MUFG MUG
also Mitsubishi UFJ Financial Group is their full name, but
(16:37):
I like saying. MUG was present in seventeen out of
twenty six of the deals, including three separate core weafinancing,
Stargate New Mexico an eighteen billion dollar deal, the alleged
Stargate thirty eight billion dollar deal, SoftBank's bridge loan which
they had to raise by the way, to fund open AI,
and a five billion dollar green loan package for Vantage
Data Centers, who you might have just heard me say. JP.
(16:59):
Morgan Chase was involved in eight deals, and they were
involved in some of the largest. Two cory was October
twenty twenty four financing their third delayed draw Tone loan
and November financing as well the funding behind Stargate, Apple
in the thirty eight billion dollar Oracle deal, and Blue
Owl's acquisition of IPI Partners in twenty twenty four. And yes,
they were also part of soft banks bridge loan. Now,
(17:21):
let's not forget Deutsche Bank, who've never done anything dodgy. Ever,
do not google what Deutsche Bank has done with banks
or particular fellas Deutsche Bank, they were involved in soft
Bank's Bridge loan and three smaller deals data center in Seoul,
Cory's twenty twenty four debt couise, November financing, and a
data center deal in Latin America, as well as a
six hundred and ten million dollar data center project in
(17:42):
Virginia and a billion euro of data center project in
Germany that was invested in within Video BNP Perribus seven deals.
Cory was delayed draw Tone loan three Stargate New Mexico,
Stargate Wisconsin and Texas IPI partners of Blue Owl, that
data center deal in Seoul, and another data center in Chile.
Morgan Stanley. You remember Morgan Stanley, don't you? While they
(18:04):
were in eight of these fucking things. Core's October twenty
twenty four loan loan, three November loans stargeting New Mexico
Stargate with Advantage. They're thirty eight billion dollar one I
just talked about, and one with EQT Edge Connects where
they I don't I don't need to go into detO.
It's another fucking data center. Oh and they were also
in SoftBank's Bridge loan, which they needed to to find
(18:25):
open AI. Now here's another name that you're going to
hear in the future, SNBC, Sumi Tomo, Mitsui Banking Corporation,
and I must say the Japanese nyler name of fucking company. Anyway,
they were in seven deals, all notable cor Weeve's DDTL
three point zero and November Financing loans, Stargate, New Mexico,
Stargate Texas, Wisconsin. That's that thirty eight billion dollar one
(18:46):
A data center in Rowan, Maryland, also involving mug TD
Securities and HSBC, as well as data center deals in
Chile and Latin America you've already heard about. Oh and
guess what Soft Banks Bridge loan which they used to
invest in open Ai. There are a lot more data
centered deals than these, but I wanted to tell you
exactly how centralized they are. I also really need you
(19:08):
to know how worrying that is, because literally every part
of this pie is being funded by like eight to
nine banks. If open ai is raising money for a date, well,
they wouldn't raise it would be Oracle or core Weave
or have you raising for a data center. One of
these people is in it or companies. They're not people.
It's really worrying. It's worrying how centralized this is, and
(19:30):
the largest deals, such as the thirty eight billion Stargate Texas,
Wisconsin deal or the eighteen billion dollar Stargate New Mexico
deal both involved Gold and Sachs b and Peter Ribas,
SMBC and fuf MUFG, and all four of those companies
have at some point funded core Weave. In fact, everybody
appears to a funded core Weave at some point. City Bank, Credit, Agricole, Society, General,
(19:52):
Wells Fargo, Carlile, Blackstone, black Rock, Barkley's, Magnatar, and Jeffrey
is the name of few, and if you've been watching
the news recently, I have heard that term term word
jeffries to refer to the company that invested in first
brands and their exposure anyway. Of the forty banks and
(20:25):
financial institutions I research twenty four have at some point
loan to or organized dead for Corwave. Of those institutions, Blackstone,
Deutsche Bank, JP, Morgan Chase, Morgan, Stanley, MUFG, and Wells
Fargo have done so multiple times. Core Weave is a
deeply unprofitable company, saddled with incredible debt and deteriorating margins.
With one of the largest clients paying net three sixty
(20:46):
and as I've said, it is arguably the best financed
data center company in the world with the best chances
of survival. What I'm getting at is that most data
centered deals are likely much worse than the terms that
corwav faces and are likely financed in a similar way
whereas where a client is signed for data center capacity
that doesn't exist, such as when nebuists raised four point
(21:07):
three billion dollars through a share sale and convertible notes
read loans to handle its seventeen point four billion dollar
data center contract with Microsoft and yees what Dobin Sachs
acted as the lead underwriter on the deal, with assistance
from Bank of America, City Group, Morgan Stanley, all three
of which have invested in core Weave. AI. Data centers
(21:27):
are expensive, required debt due to the massive cost of
construction in the cost of GPUs, and all take at
least a year, if not two, to start generating a
single dollar of revenue, at which point they also begin
losing money because it seems that renting out GPUs is
really unprofitable. Didn't think, no, no one, no one fucking
think to check that one Every single major bank and
(21:47):
financial institution has piled hundreds of millions, if not billions
of dollars into building data centers that take forever to
even start generating money, at which point they only seem
to lose it. Were still in Video sells GPUs and
a one near upgrade cycle, meaning that all of those
data centers being built right now are being filled with
Blackwell chips, and by the time they turn on and
Video will be selling its next generation via Ruben chips,
(22:10):
baking them obsolete. Now now now, now, no, no, no.
You've probably heard that Verra Ruben, the next GPU from
in Video, will use the same ranks called Oberon as Blackwell,
which is true to an extent, but won't be true
for long as in Video intends to shift to their
Kuiper racks in twenty twenty seven, hoping to build one
megawat it rax, which involve entire racks full of power supplies,
(22:34):
meaning that all of those data centers you see today,
whenever it is they get built, if they get built,
will be full of racks incompatible with the next generation
of GPUs. This will also decrease the value of the
assets inside the data centers, which will in turn decrease
the value of the assets held by the firms investing
Stargate Aplin, the one invested in by JP Morgan, Blue
Oul and Primary Digital Infrastructure and Society General, the one
(22:55):
that's been heavily delayed and won't be ready until the
end of twenty twenty six of the earliest fall to
the brim to the gills with two year old GB
two hundred Blackwell racks. Hell yeah, baby woo by the
beginning of twenty twenty seven, when this shithole eventually opens,
I should really say potentially opens, because let's be honest,
(23:16):
I'm not confident Stargate Apolene will be obsolete, as will
any and all data centers filled with Blackwell GPUs, as
will any and all data centers being built today. Every
single one takes one to three years and hundreds of
millions or billions of dollars to build, but probably raised
in damp, and every single one faces the same kinds
of construction delays and better yeah, almost all of them
(23:37):
will turn on in roughly the same timeframe. All right, Look, folks,
I gotta admit I don't really get our money works.
I'm not economist, but I do know that supply and
demand has an effect on pricing. What do you believe
happens to the price of renting at Blackwell GPU when
all of these data centers come on? Do you think
(24:00):
it will it will mean that the price goes up
or down. Also, while we're on the subject, what do
you think happens if there isn't sufficient demand? And demand
is a real question, by the way. Right now, open
ai makes up a large chunk of the global sale
of compute, at least eight point six seven billion dollars
of a zero revenue which is Microsoft's cloud platform through
(24:21):
September twenty twenty five, and they're part of about twenty
two point five billion dollars of Corway's backlog, thirty eight
billion dollars of Amazon's backlog, and so on and so forth,
and made based on my reporting from last year, just
over four point five billion dollars through the end of
September twenty twenty five. Open ai can't afford to pay anyone,
and nowhere is that more obvious than when it negotiated
(24:42):
year long payment terms of core with Otherwise, when you
remove the contracts signed by Hyperscale as an open Ai,
which I do not believe has the money to pay anybody.
Based on my analysis, there was less than a billion
dollars of AI compute revenue in twenty twenty five, or
zero point five eighty three percent of the money spent
on day center credit deals in the US. Is that good?
(25:05):
Hyperscalar revenue is also immediately questionable with Microsoft's steel with
Nebeus per their six K filing set to default in
the event that Nebius cannot provide the capacity it's sold
of its unfinished Fineland, New Jersey data center, which is
being built by Data One, a company that's never built
an AI data center, with a CEO that had his
LinkedIn location set to United Arab Emirates as a very
(25:28):
recently and indeed was in Arabic until very recently as well,
and they have funding from a concrete firm that is
also a vendor on the project. I also believe that
Microsoft is setting Nebeus up to fail based on discussions
with sources direct with direct knowledge of plans for the
violent New Jersey data center. Nebius has agreed to timelines
that involve having eighteen thousand video B two hundred and
(25:48):
B three hundred GPUs up and running by the end
of January for a total of fifty megawats, with another
eighteen thousand, b three hundreds due by the end of May.
On speaking with experts in the field about how viable
he is, the plans are two laugh and one to
me to fuck off. If Nebius fails to build the capacity,
Microsoft can walk away, much like open Ai can walk
away from Stargate in the event that Oracle fails to
(26:09):
build it on time, as reported by Anissa Gardisi of
The Information in April twenty twenty five, and I believe
that this is the case for literally any data center
provider that's building a data center for any signed up tenant.
This is another layer of risk to data center developers
that nobody bothers to fucking discuss, because everybody loves seeing
these big, beautiful numbers, Except the numbers might have become
a little too beautiful for sun. On December seventeenth, the
(26:32):
Financial Times report that the Blue Ou Capitol had pulled
out of the ten billion dollar Stargate Michigan data center project,
citing and I quote concerns about its rising debt and
artificial intelligence spending to quote the ft again. Blue Oul
had been discussions with Lender's and Oracle about investing in
the planned one gig what data center being built to
serve open AI in Saline Township, Michigan. What debt, you
(26:55):
may ask? While Blue Owl, formerly the loosest legs in
data center financing, was in always six hundred million dollar deal,
the seven hundred and fifty million dollar deal as well
for its planned Virginia data center with Teresa Technology Parks,
a four billion dollar corps with have data center project
in Pennsylvania, Stargate Abilene, Stargate Mexico, met As thirty billion
dollar hyperiod data centered deal, and a one point three
(27:17):
billion dollar data centered deal in Australia through Stack Infrastructure.
You remember I mentioned those like ten minutes ago if
you if you email me for a fish biscuit, I
haven't gotten. I'm sorry. Anyway, they owned that company and
may I mentioned it earlier. Anyways, Let's keep going to
be clear. Blue Owl pulling out is not the same
as a regular deal. It's a BBC business development company.
(27:39):
Then invest both this money and rallies together various banks
in this case SMBCBMP, Peribus, MUFG, and Gold and Sachs
all part of Stargate, New Mexico, which makes me wonder
why it didn't happen. In fact, it makes me very
much worry about that. Per the Financial time, and I quote,
(28:01):
the private Capitol group had been has been the primary
backer for Oracle's largest data center projects in the US,
investing its own money and raising billions more in debt
to build the facilities. Blue Ol typically sets up a
special purpose vehicle which owns the data center and leases
it to Oracle. Blue Owl is incredibly well connected and
experienced in putting together these kinds of deals and very
likely went to many banks, many banks. It's worked with
(28:24):
banks that have been basically giving them blank checks who
apparently had and I quote concerns about Blue Owl's rising debt,
much of which it had issued them. While rumor suggests
that Blackstone may step in, I need to be clear,
and I've spoken to numerous people in private equity about this.
This is not a consumer mortgage. This is not a
(28:45):
young couple buying a start at home. This is a
giant deal. If it's ten billion dollars, Blue Owl probably
would put in two billion themselves. That'd be like, I
don't know, maybe eight billion dollars of credit. I mean,
pipping in would require billions of dollars in legal logistics
and likely Blackstone talking to the very same banks it
(29:05):
already said no, why are they not doing this? And indeed,
why are things looking shaky? Well remember that thing about
how this data cent would be least to Oracle. Well,
Oracle had a free cash flow of negative thirteen billion
dollars on revenues of sixteen billion dollars, with its most
recent earnings only beating as analyst estimates thanks to the
(29:27):
sale of its two point sixty eight billion dollar steak
in Ampare. And you remember that soft bank Bridge loan.
I mentioned that fifteen billion dollar one Yeah, half of
that went to open AI, half of that went to
buying Ampare, So that's a one off of them. Oracle's
date is exploding, with over a billion dollars in interest
payments in its last quarter alone. It's GPU gross margins
of fourteen percent, which does not mean profitable. Its latest
(29:50):
Nvidio GB two hundred GPUs have a negative one hundred
percent gross margin, and it has two hundred and forty
eight billion dollars in upcoming data center leases is yet
to begin, and thus all for the most part to
handle compute for one customer open Ai, which needs to
raise one hundred billion dollars and then probably much more
(30:10):
to survive. Is that good anyway, that's a good point
to jump off from. Next episode, we're gonna wrap all
this up. It's been a great time reading this for you,
and we're gonna talk about how the data center of
Apocalypse will start and what might come next. Thank you
(30:35):
for listening to Better Offline. The editor and composer of
the Better Offline theme song is Matasowski. You can check
out more of his music and audio projects at Matasowski
dot com, M A T T O. S O w
Ski dot com. You can email me at easy at
Better offline dot com, or visit Better Offline dot com
to find more podcast links and of course my newsletter.
(30:57):
I also really recommend you go to chat dot Where's
your dot am to visit the discord, and go to
our slash Better Offline to check out I'll Reddit. Thank
you so much for listening. Better Offline is a production
of cool Zone Media. For more from cool Zone Media,
visit our website Coolzonemedia dot com or check us out
on the iHeartRadio app, Apple Podcasts or wherever you get
(31:18):
your podcasts
Media. Hi, I'm ed Zetran and welcome to Better Offline.
Speaker 2 (00:19):
That's right, folks, were back for the third part of
our four part series about the terminal stage of the
incitification process, the in shitification of the financial markets. So far,
we've talked about the incertification of the stock market, the
complicity of the analysts and the media, and how venture
capital will be the Canarian the coal mine for any
eventual collapse. Venture capital is both the perpetrator and the
(00:41):
victim of the air bubble, and it's worth exploring how
it came to this point. Now, hang on the gentle voice, folks.
We haven't seen values this big and alone time. These
the biggest numbers we've ever seen. They're simply tremendous open
airs maybe with eight hundred and thirty billion. Can you
believe that they lose so much money? But folks, we
don't mind. Clamisam Morbourn. They call him Klamiel because everybody
he's going to give them one billion dollars and where
(01:02):
we love him dead. Centers are going to have the
biggest deals we've ever seen, even even if we have
to work with Dario. Sorry to those of you who
don't like that voice. It was only about thirty seconds,
you see right Now, AI startups are big exciting news
for the limited partners funding large language model companies. Things
feel exciting because the value of the assets under management
(01:22):
aum are going up, which is nothing dodgy, per se,
but it's just how vcs value things. And if they're
valuing AI stocks, well that's how their fees are getting paid,
because the thing is they've invested in are worth more money,
and now limited partners are paying them a percentage so
that they can sit in an apartment that costs too
much not see friends because they just sit online posting
for eleven hours a day. Yeah, and then they can
(01:44):
lose all the money in the future. I guess investing
early in open Ai allows a bench capitalist or even
an asset manager like Blackstone, which invested in twenty twenty
four to say it as a big holding and a
big increase in its aum AI stocks, make venture capitalists
who bet on them two years ago, Like genius is
on paper, you got in early on open Ai, Athropic, Cursor, Cognition, Perplexity,
(02:05):
or any other company that loves to burn several dollars
per dollar of revenue. You have a big, beautiful number,
the biggest you've ever seen, and your limited partners to
pay you a feed just to manage it. Venture capital
has not seen valuations like this in a long time,
and on paper it feels a lot like a lot
of vcs got in on companies worth billions of dollars
right on paper, that is, on paper, Cognition is worth
(02:28):
ten point two billion, Perplexit eighteen billion, Cursor twenty nine
point three billion, Lovable six point six billion, Coher six
point eight billion, Repic three billion, Glean seven point two billion.
Massive valuations to companies that all basically do products that
open AI or Anthropic or Amazon or Google or any
number of Chinese companies are already working to clone. They're
(02:48):
all losing lots of money too, and have no path
the profitability of any kind. But right now the numbers
are simply tremendous. Folks. I've heard venture capitalists tell me
that there are times when they have to agree to
invest with little to no information or know that they'll
lose the opportunity to another sucker, I mean, investor. I've
heard venture capitalists say they don't have any insight into finances,
(03:11):
but we love the numbers. Folks, they're simply tremendous. We
love it. Let me tell you how not tremendous they are.
I had a source recently suggests that nobody who invested
pasted open AI's one hundred billion dollar valuation have any
information rates. And what that means is they literally do
not know anything about the company. They don't know anything
other than what the company will tell them. That the
(03:32):
company doesn't have to tell them shit, which means they
don't know the revenues, they don't know the costs. They
only know what new start new ridiculous stat Sarah Friar,
CFO of open ais, farted out. They posted recently that
twenty billion dollars annualized revenue. Yeah. I think that's complete nonsense,
But they want to do that so that you go
open ai made twenty billion dollars in twenty twenty five.
(03:54):
Fuck that, I don't even think they made ten. Now,
venture capitalists would argue, of course that they'd say I'm insane.
They would say they would think me mad, and they
would say that the growth is obviously there, while pointing
to whatever startup has made one hundred million dollars in
annualized recurring revenue, which is eight point three million dollars
a month, by the way, all while not discussing the
(04:16):
underlying operating expenses. The idea I believe is that the
current state of AI spending is only set to increase
the next year, and that will somehow lead to fixing
the margins. I think venture capitalists staunchly refuse to learn
anything other than investing growth and then profit from growth,
even if profiting from growth doesn't appear to be happening anymore.
(04:39):
In reality, venture capital shouldn't have touched LMS with a
shitty stick because the margins were obviously blatantly bad from
the very beginning. We knew open ai would lose five
million dollars in the middle of twenty twenty four. The
same venture capital climate would have fucking panicked, but instead
chose to double dribble and kadruple down. I believe that
massive valuation drawdowns are a certainty for AI companies, and
(05:01):
by that I mean these companies will end up being
revalued and the number will be smaller than the big
number everybody fell in love with. Look, venture capitalists, I
gotta ask you a question, what happens if open AI dies?
Do you think this will make investors interested in funding
or acquiring other AI startups. How much longer are we
(05:21):
going to do this? When will Venture Capital realize it's
setting itself up for a disaster And what exactly is
the plan? Open AI and anthropic will cuck the lake's
dry like an Nvidia GPU named after Nancy Reagan? How
is this meant to continue and what will be left
when it does? The answer is simple, there won't be
any money for venture capital for a while. Those AI
(05:42):
holdings are going to be worth at best fifty percent
if they retain any value at all. Once one of
these startups die, Once there's a major casualty, a panic
will ensue, sending venture capitalists scrambling to get their holdings
acquired by which I mean sold to someone else, until
there's little or no investor interest left. Why would limited
partners ever trust venture capital after this? Why would anybody?
(06:04):
Because based on the past four years, it doesn't appear
that venture capital is actually good at investing money. It
just got lucky year after year until there were a
few ideas that they could sell for one hundreds of
billions or billions of dollars, And now we're out of them,
we're out. I think we're out. I think there's probably
still some, but I think the era of the unicorn
might be dead. For a while, venture capital believed it
(06:27):
knew better as it turned its back on basic business fundamentals,
starting with Clubhouse, Crypto, the Metaverse, and now Generative AI.
Yet they're far from the only fuck wits on the
Dickhead Express. Because yes, I am going to talk about
data centers. Per Bloomberg, there were at least one hundred
and seventy eight point five billion dollars in data center
credit deals and in the US in twenty twenty five,
(06:49):
rivaling the two hundred and fifteen point four billion dollars
invested in US venture capital in twenty twenty four and
the one hundred and ninety seven point two billion dollars
invested in US venture capital through August seventh to twenty
twenty five, and over one hundred billion dollars more than
the sixty point sixty nine billion dollars of data center
deals done in twenty twenty four. That's that's really worrying.
(07:10):
It's like more than more than more than one hundred
percent more. I'm very worried, and I'm going to tell
you why, using a company called core Weave, core Weave
that I've been actively warning about, warning people about since March.
And I get a little crazy around corwaf because I
have had the core Weave conversation a lot, and I've
seen people straight up just steal the title of my
(07:33):
articles as they come to this realization six to seven
months later. But cor Weave is a mess, and corwy
was obviously always a mess. And everyone has been saying
Corwave's fine because they went public. Core Weave is not fine.
All right, Let's talk about core Weave. Corwave is something
called a neocloud. It's a company that builds data centers
(07:56):
by renting out, and they build data centers they film
full of GPUs and they rent them to people. They're
in something called AI Compume, And as I explained a
few months ago, they do so by building data centers
backed by endless debt. And I quote myself saying, up
a neocloud is expensive. Even if the company in question
already has data centers, as core We've did with its
scryptocurrency mining operation, AI requires completely new data center infrastructure
(08:20):
to house and call the GPUs, and those GPUs also
need paying for, and then there's the other stuff like power, water,
and the other bits of the computer, the CPU, the motherboard,
the memory of the storage, and the housing and the
power supplying all that good stuff. As a result, these
neoclouds are forced to raise billions of dollars in debt,
which they collateralize using the GPUs they already have, along
with contracts from customers which they use to buy more GPUs. Corwei,
(08:44):
for example, has twenty five billion dollars in debt on
an estimated five point three five billion dollars in revenue,
losing hundreds of billions of dollars in a quarter. You
know who also invests in these neoclouds in Video, the
people who make the GPUs in VideA is also one
of Corwey's largest customers, accounting for fifty percent of its
revenue in twenty twenty four, and just signed a deal
(09:05):
to buy six point three billion dollars of any capacity
the care we can't otherwise sell to someone else through
twenty thirty two, an extension of a one point three
billion dollar deal from twenty twenty three reported by The
Information oh And in Video was the anchor investor of
Corweav's IPO, putting in two hundred and fifty million dollars.
Corwave is also one of the largest providers AI compute
(09:25):
in the world, and its business model is indicative of
how most data center companies make money. In fact, i'd
argue that they are probably doing better because they have
the backing of everyone. Now you may think that's a
too big to fail thing. Not really, It just means
that they're kind of the village bicycle of AI compute grossness. Aside,
let me explain how this works. First, core we've signed
(09:46):
contracts such as it's fourteen billion dollar deal with Meto
or it's twenty two point five billion dollar deal with
open Ai before it has the physical infrastructure to actually
service them. It then raises debt using this contract as
collateral orders GPUs from VideA. Those take three months to
arrive and then another three months to install, at which
point monthly client payments begin. And I should also add
(10:07):
that that's kind of taken out of deferred revenue that
the customers already put down. To really simplify this, data
center developers are raising money months up to a year
before they ever expect to make a penny of revenue,
not profit. In fact, I can find no consistent answer
to how long a data center takes to build, and
the answer here is pretty important because that's how the
(10:28):
money is going to get made from the data center's
for building fucking everywhere. You may also notice that monthly
payments in the chart that I linked, which I should
have mentioned, begin at six to thirty months, a curious
and broad blob of time. You see, data centers are
extremely difficult to build, and the concept of an AI
data center is barely a few years old, with the
concepts of hundreds of megawats in one data center campus
(10:49):
entirely made up of aigpus barely two years old jpus,
which means basically everybody building one is doing so for
the first time, and even experienced developers are running into problems.
For example, core Scientific, Corweave's weird partner organization it tried
and failed to buy that is truly different despite the
similar name, has been trying to convert it's Denton, Texas
(11:11):
scriptocurrency mining data center into an AI data center since
November twenty twenty four, specifically so that Corewave can rent
it to Microsoft for Open AI. This hasn't gone well,
and this is when things get weird. The Wall Street
Journal reported a few weeks ago that Denton had been
racked with several months of delays thanks to rainstorms preventing
contractors and pooring contry contrete. I'm going with it. It's contrete.
(11:34):
Now we're calling it contry concrete. I originally, within this
script wrote down that this was a possible thing. I
thought that this was normal. However, I should add something,
this is this may actually be bollocks. So I'm a psycho,
and I went and I looked up whether. I looked
up weather in Abilene, Texas, and it turns out that
(11:57):
there were like eleven days with rain total, and there
was only I think on June third and September twenty second,
there were rainstorms or thunderstorms of any kind. I think
that there are only two days where rain was over
point one inches. The people I've talked to about this
deal are saying it wouldn't be weather, it would actually
just be money. Nevertheless, those are things that are stopping
(12:19):
them building a two hundred and sixty megawat data center,
and that's fucking nothing compared to the Giga. What's the
open Ai claims to want to build. And what this means,
by the way, is that call we've can't actually get
paid by open Ai because per its contract, customers don't
have to stop paying until the computer is actually available.
This is a very important deal to know for literally
(12:40):
any data center development you've ever heard of or seen.
As of cour we've SQ three twenty twenty five earnings,
they're sitting on one point one billion dollars into thirst
revenue that's income for service is not yet rendered are
(13:03):
from nine hundred and fifty one million in Q two
and four hundred and thirty six million in Q one
twenty twenty five. This means deposits have been made, but
the contract is yet to be serviced. Now, I'm also
a curious little critter, so I went and found the
nine hundred and twenty one page two point six billion
dollar DDTL delayed draw Term Loan three point zero loan
(13:23):
agreement between core Weave and banks including Morgan Stanley, Mitsubishi,
UFG and Gold and Sachs, and in doing so I
learned something well. First of all, open ai appears to
have net three sixty payment terms from Corwave, meaning that
it can literally pay a year from invoice. Second, Corweve
is required to maintain something called a contract realization ratio
(13:44):
of point eighty five times, meaning that core Weave has
to make at least eighty five cents of every expected
dollar or it is in default of its loan. This
is important to note because it means that if say
open Ai decides not to pay up in a year,
corwave will be up shit creek without a paddle. Now,
I apolog that suggests that core Weave is an already
up shit creek, or that it might have a paddle
(14:05):
of some sort buried inside in Vidia's latest earnings on
page seventeen, there's a little clue at home, and I
quote in the third quarter of fiscal year twenty twenty six,
we entered into an agreement to guarantee of partner's facility
lease obligations in the event of their default. The agreement
allows our partner to secure a limited availability facility lease
backed by our credit profile in exchange for issuing US warrants.
(14:28):
The maximum gross exposure is eight hundred and sixty million,
which is reduced as the partner makes payments to the
leaser of a five years. The partner has placed four
hundred and seventy million dollars in escrow and executed an
agreement to sell the data center cloud capacity, mitigating our
default risk credit where credit is due. Eaglide analysts Just
Dario caught this in November, but in core weaves condensed
to consolidated balance sheets, there sits a four hundred and
(14:50):
seventy seven point five million dollars line item under restricted
cash and cash equivalence non current, though this might not
be in Vidia's escrow. This number shifted from six hundred
and seventeen in Q one to three hundred and forty
million in Q two. It lines up all too precisely,
and who else would in video be guaranteeing In any case,
(15:10):
corweav is likely to get the best deals in data
centered debt outside of Oracle. It has top tier financiers
who I'll get to in a little bit, the fallbacking
in Video who is both an investor customer and apparent
financial backstop, and also Core with as a customer of
theirs and the ability to raise debt quickly. Corweb's deals
are likely indicative of how data center financing takes place,
and those top tier financiers it's they may basically been
(15:34):
in every single deal. It's actually it's actually really worrying
when you, I mean, you lay it all out, which
is exactly what I'm going to do. So who's actually
paying for this shit? I went and dug through a
pile of twenty six prominent data center loan deals, including
the proposed thirty eight billion dollar debt package that Oracle
Advantage data Center partners are raising for Stargate Shackelford and Wisconsin,
(15:55):
Stargate Ablene, New Mexico, and of course soft Banks fifty
billion dollar bridge loan, which I included for a reason
that will become obvious shortly, and multiple core wave loans.
I've found a few commonalities, and forgive me, I'm about
to say a lot of names and numbers. You might
want to pull up the newsletter to read along with
it like a sing along. Now, Blue Owl, Blue ol,
We love blue ol folks. They were present in every
(16:17):
single Stargate deal other than the thirty eight billion dollar
deal that's being raised by Advantage that is yet to close.
Blue Hour was also involved in a one point three
billion dollar Australian data set and a debt package by
Virtue of owning Stack infrastructure. Remember that name MUFG MUG
also Mitsubishi UFJ Financial Group is their full name, but
(16:37):
I like saying. MUG was present in seventeen out of
twenty six of the deals, including three separate core weafinancing,
Stargate New Mexico an eighteen billion dollar deal, the alleged
Stargate thirty eight billion dollar deal, SoftBank's bridge loan which
they had to raise by the way, to fund open AI,
and a five billion dollar green loan package for Vantage
Data Centers, who you might have just heard me say. JP.
(16:59):
Morgan Chase was involved in eight deals, and they were
involved in some of the largest. Two cory was October
twenty twenty four financing their third delayed draw Tone loan
and November financing as well the funding behind Stargate, Apple
in the thirty eight billion dollar Oracle deal, and Blue
Owl's acquisition of IPI Partners in twenty twenty four. And yes,
they were also part of soft banks bridge loan. Now,
(17:21):
let's not forget Deutsche Bank, who've never done anything dodgy. Ever,
do not google what Deutsche Bank has done with banks
or particular fellas Deutsche Bank, they were involved in soft
Bank's Bridge loan and three smaller deals data center in Seoul,
Cory's twenty twenty four debt couise, November financing, and a
data center deal in Latin America, as well as a
six hundred and ten million dollar data center project in
(17:42):
Virginia and a billion euro of data center project in
Germany that was invested in within Video BNP Perribus seven deals.
Cory was delayed draw Tone loan three Stargate New Mexico,
Stargate Wisconsin and Texas IPI partners of Blue Owl, that
data center deal in Seoul, and another data center in Chile.
Morgan Stanley. You remember Morgan Stanley, don't you? While they
(18:04):
were in eight of these fucking things. Core's October twenty
twenty four loan loan, three November loans stargeting New Mexico
Stargate with Advantage. They're thirty eight billion dollar one I
just talked about, and one with EQT Edge Connects where
they I don't I don't need to go into detO.
It's another fucking data center. Oh and they were also
in SoftBank's Bridge loan, which they needed to to find
(18:25):
open AI. Now here's another name that you're going to
hear in the future, SNBC, Sumi Tomo, Mitsui Banking Corporation,
and I must say the Japanese nyler name of fucking company. Anyway,
they were in seven deals, all notable cor Weeve's DDTL
three point zero and November Financing loans, Stargate, New Mexico,
Stargate Texas, Wisconsin. That's that thirty eight billion dollar one
(18:46):
A data center in Rowan, Maryland, also involving mug TD
Securities and HSBC, as well as data center deals in
Chile and Latin America you've already heard about. Oh and
guess what Soft Banks Bridge loan which they used to
invest in open Ai. There are a lot more data
centered deals than these, but I wanted to tell you
exactly how centralized they are. I also really need you
(19:08):
to know how worrying that is, because literally every part
of this pie is being funded by like eight to
nine banks. If open ai is raising money for a date, well,
they wouldn't raise it would be Oracle or core Weave
or have you raising for a data center. One of
these people is in it or companies. They're not people.
It's really worrying. It's worrying how centralized this is, and
(19:30):
the largest deals, such as the thirty eight billion Stargate Texas,
Wisconsin deal or the eighteen billion dollar Stargate New Mexico
deal both involved Gold and Sachs b and Peter Ribas,
SMBC and fuf MUFG, and all four of those companies
have at some point funded core Weave. In fact, everybody
appears to a funded core Weave at some point. City Bank, Credit, Agricole, Society, General,
(19:52):
Wells Fargo, Carlile, Blackstone, black Rock, Barkley's, Magnatar, and Jeffrey
is the name of few, and if you've been watching
the news recently, I have heard that term term word
jeffries to refer to the company that invested in first
brands and their exposure anyway. Of the forty banks and
(20:25):
financial institutions I research twenty four have at some point
loan to or organized dead for Corwave. Of those institutions, Blackstone,
Deutsche Bank, JP, Morgan Chase, Morgan, Stanley, MUFG, and Wells
Fargo have done so multiple times. Core Weave is a
deeply unprofitable company, saddled with incredible debt and deteriorating margins.
With one of the largest clients paying net three sixty
(20:46):
and as I've said, it is arguably the best financed
data center company in the world with the best chances
of survival. What I'm getting at is that most data
centered deals are likely much worse than the terms that
corwav faces and are likely financed in a similar way
whereas where a client is signed for data center capacity
that doesn't exist, such as when nebuists raised four point
(21:07):
three billion dollars through a share sale and convertible notes
read loans to handle its seventeen point four billion dollar
data center contract with Microsoft and yees what Dobin Sachs
acted as the lead underwriter on the deal, with assistance
from Bank of America, City Group, Morgan Stanley, all three
of which have invested in core Weave. AI. Data centers
(21:27):
are expensive, required debt due to the massive cost of
construction in the cost of GPUs, and all take at
least a year, if not two, to start generating a
single dollar of revenue, at which point they also begin
losing money because it seems that renting out GPUs is
really unprofitable. Didn't think, no, no one, no one fucking
think to check that one Every single major bank and
(21:47):
financial institution has piled hundreds of millions, if not billions
of dollars into building data centers that take forever to
even start generating money, at which point they only seem
to lose it. Were still in Video sells GPUs and
a one near upgrade cycle, meaning that all of those
data centers being built right now are being filled with
Blackwell chips, and by the time they turn on and
Video will be selling its next generation via Ruben chips,
(22:10):
baking them obsolete. Now now now, now, no, no, no.
You've probably heard that Verra Ruben, the next GPU from
in Video, will use the same ranks called Oberon as Blackwell,
which is true to an extent, but won't be true
for long as in Video intends to shift to their
Kuiper racks in twenty twenty seven, hoping to build one
megawat it rax, which involve entire racks full of power supplies,
(22:34):
meaning that all of those data centers you see today,
whenever it is they get built, if they get built,
will be full of racks incompatible with the next generation
of GPUs. This will also decrease the value of the
assets inside the data centers, which will in turn decrease
the value of the assets held by the firms investing
Stargate Aplin, the one invested in by JP Morgan, Blue
Oul and Primary Digital Infrastructure and Society General, the one
(22:55):
that's been heavily delayed and won't be ready until the
end of twenty twenty six of the earliest fall to
the brim to the gills with two year old GB
two hundred Blackwell racks. Hell yeah, baby woo by the
beginning of twenty twenty seven, when this shithole eventually opens,
I should really say potentially opens, because let's be honest,
(23:16):
I'm not confident Stargate Apolene will be obsolete, as will
any and all data centers filled with Blackwell GPUs, as
will any and all data centers being built today. Every
single one takes one to three years and hundreds of
millions or billions of dollars to build, but probably raised
in damp, and every single one faces the same kinds
of construction delays and better yeah, almost all of them
(23:37):
will turn on in roughly the same timeframe. All right, Look, folks,
I gotta admit I don't really get our money works.
I'm not economist, but I do know that supply and
demand has an effect on pricing. What do you believe
happens to the price of renting at Blackwell GPU when
all of these data centers come on? Do you think
(24:00):
it will it will mean that the price goes up
or down. Also, while we're on the subject, what do
you think happens if there isn't sufficient demand? And demand
is a real question, by the way. Right now, open
ai makes up a large chunk of the global sale
of compute, at least eight point six seven billion dollars
of a zero revenue which is Microsoft's cloud platform through
(24:21):
September twenty twenty five, and they're part of about twenty
two point five billion dollars of Corway's backlog, thirty eight
billion dollars of Amazon's backlog, and so on and so forth,
and made based on my reporting from last year, just
over four point five billion dollars through the end of
September twenty twenty five. Open ai can't afford to pay anyone,
and nowhere is that more obvious than when it negotiated
(24:42):
year long payment terms of core with Otherwise, when you
remove the contracts signed by Hyperscale as an open Ai,
which I do not believe has the money to pay anybody.
Based on my analysis, there was less than a billion
dollars of AI compute revenue in twenty twenty five, or
zero point five eighty three percent of the money spent
on day center credit deals in the US. Is that good?
(25:05):
Hyperscalar revenue is also immediately questionable with Microsoft's steel with
Nebeus per their six K filing set to default in
the event that Nebius cannot provide the capacity it's sold
of its unfinished Fineland, New Jersey data center, which is
being built by Data One, a company that's never built
an AI data center, with a CEO that had his
LinkedIn location set to United Arab Emirates as a very
(25:28):
recently and indeed was in Arabic until very recently as well,
and they have funding from a concrete firm that is
also a vendor on the project. I also believe that
Microsoft is setting Nebeus up to fail based on discussions
with sources direct with direct knowledge of plans for the
violent New Jersey data center. Nebius has agreed to timelines
that involve having eighteen thousand video B two hundred and
(25:48):
B three hundred GPUs up and running by the end
of January for a total of fifty megawats, with another
eighteen thousand, b three hundreds due by the end of May.
On speaking with experts in the field about how viable
he is, the plans are two laugh and one to
me to fuck off. If Nebius fails to build the capacity,
Microsoft can walk away, much like open Ai can walk
away from Stargate in the event that Oracle fails to
(26:09):
build it on time, as reported by Anissa Gardisi of
The Information in April twenty twenty five, and I believe
that this is the case for literally any data center
provider that's building a data center for any signed up tenant.
This is another layer of risk to data center developers
that nobody bothers to fucking discuss, because everybody loves seeing
these big, beautiful numbers, Except the numbers might have become
a little too beautiful for sun. On December seventeenth, the
(26:32):
Financial Times report that the Blue Ou Capitol had pulled
out of the ten billion dollar Stargate Michigan data center project,
citing and I quote concerns about its rising debt and
artificial intelligence spending to quote the ft again. Blue Oul
had been discussions with Lender's and Oracle about investing in
the planned one gig what data center being built to
serve open AI in Saline Township, Michigan. What debt, you
(26:55):
may ask? While Blue Owl, formerly the loosest legs in
data center financing, was in always six hundred million dollar deal,
the seven hundred and fifty million dollar deal as well
for its planned Virginia data center with Teresa Technology Parks,
a four billion dollar corps with have data center project
in Pennsylvania, Stargate Abilene, Stargate Mexico, met As thirty billion
dollar hyperiod data centered deal, and a one point three
(27:17):
billion dollar data centered deal in Australia through Stack Infrastructure.
You remember I mentioned those like ten minutes ago if
you if you email me for a fish biscuit, I
haven't gotten. I'm sorry. Anyway, they owned that company and
may I mentioned it earlier. Anyways, Let's keep going to
be clear. Blue Owl pulling out is not the same
as a regular deal. It's a BBC business development company.
(27:39):
Then invest both this money and rallies together various banks
in this case SMBCBMP, Peribus, MUFG, and Gold and Sachs
all part of Stargate, New Mexico, which makes me wonder
why it didn't happen. In fact, it makes me very
much worry about that. Per the Financial time, and I quote,
(28:01):
the private Capitol group had been has been the primary
backer for Oracle's largest data center projects in the US,
investing its own money and raising billions more in debt
to build the facilities. Blue Ol typically sets up a
special purpose vehicle which owns the data center and leases
it to Oracle. Blue Owl is incredibly well connected and
experienced in putting together these kinds of deals and very
likely went to many banks, many banks. It's worked with
(28:24):
banks that have been basically giving them blank checks who
apparently had and I quote concerns about Blue Owl's rising debt,
much of which it had issued them. While rumor suggests
that Blackstone may step in, I need to be clear,
and I've spoken to numerous people in private equity about this.
This is not a consumer mortgage. This is not a
(28:45):
young couple buying a start at home. This is a
giant deal. If it's ten billion dollars, Blue Owl probably
would put in two billion themselves. That'd be like, I
don't know, maybe eight billion dollars of credit. I mean,
pipping in would require billions of dollars in legal logistics
and likely Blackstone talking to the very same banks it
(29:05):
already said no, why are they not doing this? And indeed,
why are things looking shaky? Well remember that thing about
how this data cent would be least to Oracle. Well,
Oracle had a free cash flow of negative thirteen billion
dollars on revenues of sixteen billion dollars, with its most
recent earnings only beating as analyst estimates thanks to the
(29:27):
sale of its two point sixty eight billion dollar steak
in Ampare. And you remember that soft bank Bridge loan.
I mentioned that fifteen billion dollar one Yeah, half of
that went to open AI, half of that went to
buying Ampare, So that's a one off of them. Oracle's
date is exploding, with over a billion dollars in interest
payments in its last quarter alone. It's GPU gross margins
of fourteen percent, which does not mean profitable. Its latest
(29:50):
Nvidio GB two hundred GPUs have a negative one hundred
percent gross margin, and it has two hundred and forty
eight billion dollars in upcoming data center leases is yet
to begin, and thus all for the most part to
handle compute for one customer open Ai, which needs to
raise one hundred billion dollars and then probably much more
(30:10):
to survive. Is that good anyway, that's a good point
to jump off from. Next episode, we're gonna wrap all
this up. It's been a great time reading this for you,
and we're gonna talk about how the data center of
Apocalypse will start and what might come next. Thank you
(30:35):
for listening to Better Offline. The editor and composer of
the Better Offline theme song is Matasowski. You can check
out more of his music and audio projects at Matasowski
dot com, M A T T O. S O w
Ski dot com. You can email me at easy at
Better offline dot com, or visit Better Offline dot com
to find more podcast links and of course my newsletter.
(30:57):
I also really recommend you go to chat dot Where's
your dot am to visit the discord, and go to
our slash Better Offline to check out I'll Reddit. Thank
you so much for listening. Better Offline is a production
of cool Zone Media. For more from cool Zone Media,
visit our website Coolzonemedia dot com or check us out
on the iHeartRadio app, Apple Podcasts or wherever you get
(31:18):
your podcasts
Host
Ed Zitron
Popular Podcasts
Two Guys, Five Rings: Matt, Bowen & The Olympics
Two Guys (Bowen Yang and Matt Rogers). Five Rings (you know, from the Olympics logo). One essential podcast for the 2026 Milan-Cortina Winter Olympics. Bowen Yang (SNL, Wicked) and Matt Rogers (Palm Royale, No Good Deed) of Las Culturistas are back for a second season of Two Guys, Five Rings, a collaboration with NBC Sports and iHeartRadio. In this 15-episode event, Bowen and Matt discuss the top storylines, obsess over Italian culture, and find out what really goes on in the Olympic Village.
iHeartOlympics: The Latest
Listen to the latest news from the 2026 Winter Olympics.
Milan Cortina Winter Olympics
The 2026 Winter Olympics in Milan Cortina are here and have everyone talking. iHeartPodcasts is buzzing with content in honor of the XXV Winter Olympics We’re bringing you episodes from a variety of iHeartPodcast shows to help you keep up with the action. Follow Milan Cortina Winter Olympics so you don’t miss any coverage of the 2026 Winter Olympics, and if you like what you hear, be sure to follow each Podcast in the feed for more great content from iHeartPodcasts.