December 19, 2025 • 33 mins
In part three of this week's three-part NVIDIA series, Ed Zitron walks you through why there are millions of Blackwell GPUs sitting in warehouses, and why AI’s lack of any profits makes NVIDIA’s future entirely dependent on endless debt and venture capital.
This series took a lot of work, so if you want to support me, why not subscribe to my premium newsletter? Get $10 off a year’s subscription today: https://edzitronswheresyouredatghostio.outpost.pub/public/promo-subscription/p94my1c5ya
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 (00:02):
Media.
Speaker 2 (00:04):
Hi, I'm ed Zitron, and welcome back to Better Offline
And this is our third and final part of our
Better Offline and Vidia special, where we're talking about well,
(00:27):
the shakiness behind its growth and how the company, despite
being on incredibly infirm ground, is definitely not Enron or
Nortail or Wolcom or Lucien or any other dot com
bubble era affirm that imploded under its own way and
well quite dodgy accounting. The thing is, even if Enron
is nothing like them, there's still quite a few causes
for concern, and that's largely driven from the fact that
(00:49):
in Vidio makes the majority of its money selling GPUs
to a handful of customers, and so well, some of
those also look to be on some of their own
incredibly shaky ground. And yeah, I'm talking about Oracle now.
In Vidia's health saying nothing of its growth isn't just
tied to these customers. It's also tied to whether these
customers can actually turn a profit from their capex spending.
(01:12):
And even that's not even certain so due to the
fact that so much money has been piled into building
AI infrastructure and big tech has promised to spend hundreds
of billions of dollars more in the next year. Big
tech has found itself in a bit of a hole.
How big of a hole? Well, by the end of
the year, Microsoft, Amazon, Google, and Meta will have spent
over four hundred billion dollars in capital expenditures, much of
(01:34):
it focused on building AI infrastructure, on top of two
hundred and twenty eight point four billion dollars in capex
in twenty twenty four and around one hundred and forty
eight billion in capital expenditures in twenty twenty three, for
a total of seven hundred and seventy six billion in
the space of three years, and they expect to spend
more than four hundred billion dollars more in twenty twenty six.
(01:56):
Every time I read these numbers, I feel a little crazy.
As a result, based on my own analysis, big tech
needs to make two trillion dollars in brand new, brand
spanking new revenue, specifically from AI by twenty thirty. All
of this was effectively for nothing. Now, I go into
detail about this in the premium newsletter I did on
October thirty first, but I'm going to give you a
(02:17):
short explanation here. First, though, we have to talk about depreciation,
and because I'm lazy, I'm going to quote myself in
that newsletter I just mentioned a couple of seconds ago
at hem. So when Microsoft buyers say one hundred million
dollars worth of GPUs, it immediately comes out of its
capital expenditures, which is when a company uses money to
invest in either buying or upgrading something. It then adds
(02:40):
to its property, plants and Equipment assets PPE for sure,
although some companies list this on their annual and quarterly
financials as property and equipment PPE sits on the balance sheet,
it's an asset as it's stuff for the company that
it owns or as least, GPUs depreciate, meaning they lose
value or over time, and this depreciation is represented on
a balance sheet and the income statement. Essentially, the goal
(03:03):
is to represent the value of an asset that a
company has on the income statement, and we see how
much the assets have declined during the reporting period, whether
that be a year or a quarter or something else,
Whereas the balance sheet shows the cumulative depreciation of every
asset currently in play. Depreciation does two things. And I
know this sounds like a lot, but I'll break it
down for you First, it allows a company to accurately
(03:25):
to an extent, represent the value of things it owns
over their useful life. Secondly, it allows a company to
deduct the value of an asset across said useful life
right up until its eventual removal, versus having to take
a big hit up front. The way this depreciation is
actually calculated can vary. There are several different methods available,
with some allowing for greater deductions at the start of
(03:46):
the term, which is useful for those items that will
experience the biggest drop in value right after buying them
and their initial use. An example you're probably familiar with
is a new car which loses a significant chunk of
its value the moment is driven off a dealership. Long
creation has become a big ugly problem with GPUs specifically
because of that useful life to find either as how
long the thing is able to be run before it
(04:07):
dies or how long before it becomes obsolete, and nobody
seems to be able to come up with a consensus
about how long this should be. In Microsoft's case, the
appreciation for its service is spread over six years, a
convenient change it made in August twenty twenty two, A
few months before the launch of chat GPT and before
it bought a bunch of fucking GPUs. This means that
Microsoft can spread the cost of tens of thousands of
(04:28):
a one hundred GPUs brought in twenty twenty or the
four hundred and fifty thousand, h one hundred gupus it
bought in twenty twenty four across six years, regardless of
whether those are the years they'll be generating revenue or
naturally functioning. Corwy for what It's worth says the same thing,
but largely because it's betting that it'll still be able
to find users for older silicon after its initial contracts
(04:49):
with the companies like OpenAI expire. The problem is is
that aigpus are fairly new concepts, and thus all of
this is pretty much untested ground. Whereas we know how long,
say a truck or a piece of heavy machinery can
last and how long it can deliver value to an organization,
we don't know the same thing about the kind of
data center GPUs that hyperscalers are spending tens of billions
(05:10):
of dollars on each year. Any kind of depreciation schedule
is based on at best assumptions and at worst hope.
Now this is important. The concept of an AI data
center is super new. We maybe saw the first ones
in twenty ninety. In question, it's kind of hard to say,
but even at the scale we're seeing today a Gigawa
(05:31):
data center pretty much brand new, maybe a couple years old.
I don't even think they've even built any but we'll
get to that in a bit. There are a lot
of assumptions at play. There's the assumption that the cards
won't degrade with heavy usage, or the assumption that future
generations of GPUs won't be so powerful and impressive that
they'll render the previous ones more obsolete than expected, kind
of like how the first jet powered planes of the
(05:52):
nineteen fifties did to those manufactured just a decade prior.
The assumption that there will be in fact a market
for Alder cards, than that there'll be a way to
lease them profitably. What if those assumptions are I don't know, wrong.
What if that hope is ultimately irrational. So there's a
quote from the Center for Information Technology Policy framing this problem, well,
(06:13):
that'll link to in the notes. Here is the puzzle.
The chips at the heart of the infrastructure build out
have a useful lifespan of one to three years due
to rapid technological obsolescence and physical wear, but companies appreciate
them over five or six years. In other words, they
spread out the cost of their massive capital expenditures over
a longer period than the facts warrant what the economist
is referred to as the four trillion dollar accounting puzzle
(06:35):
at the heart of the AIICLOUD. This is why Michael
Burry brought it up recently because spreading out these costs
allowed spigtech to make their net income i e. Their
profits look better in simple terms. By spreading out the
costs over six years rather than three, hyperscalers are able
to reduce the line item that eats into their earnings,
which makes their companies look better to the markets. So
why does this create an artificial time limit? Well, let's
(06:58):
start with a horrible fact. It takes it's two point
five years of construction time in about fifty billion dollars
per gigawa of data center capacity. No matter when the
GPUs for a giga what data center are bought one
way or another, these GPUs are depreciating in value either
through death or reduced efficacy through wear and tear, or
becoming obsolete, which is very likely as in Video is
committed to releasing a new GPU every single year. Newer
(07:21):
generation GPUs, like in videos Blackwell and Verra Reuben require
entirely new data center architecture, meaning that one as to
why they build a brand new data center a retrofit
an old one. Essentially, we have facilities that are being
built around a GPU design or product that may change
in a year or two. Now I hear that in
the Oberon racks that they use for the Blackwells will
(07:42):
be used with some Verra Rubin. But even then there's
going to be an even bigger, more huger Vera Ruben
that comes that might even that I read somewhere that
there might even be like killer what level ones just
in like one hundred killer what ones that this company's insane. Nevertheless,
at some point Wall Street is going to need to
see some sort of return on this investment, and right
now that return is negative dollars. I break it down
(08:03):
on my October thirty first premium piece, but for your sake,
I'll just say it. I estimate the big techniques to
make two dollars for every dollar of CAPEX they've spent,
and this revenue must be new, brand new, As this
CAPEX is only for AI. This CAPEX is useless for
everything else. It does not help it. And no, it
doesn't help that they bolted copilot onto fucking everything that
(08:24):
is not working, and in fact, the Australian Competition Commission
is suing them. Maybe I mentioned that later, but whatever, Meta, Amazon,
Google and Microsoft are already years and hundreds of billions
of dollars in and are yet to see a dollar
of profit, creating a one point two to one trillion
dollar hole just to justify the expenses. So around six
hundred and five billion dollars of CAPEX all told at
(08:44):
the time I calculated it. Much of this CAPEX has
been committed or spent before they've even turned on a
single goddamn GPU. You might argue that there's a scenario
here where say, an A one hundred GPU is useful
past the three or six year shelf life. Even if
that were the case, the average rental price of an
A one hundred is ninety nine cents an hour. This
is a four or five year old GPU, and customers
(09:06):
are paying for it like they would a five year
old piece of hardware. The same fate awaits the H
one hundred, which was released in twenty twenty two but
was still sold in great volume through twenty twenty four,
and I hear the H two hundred of the same
generation is still selling to this day. Every year in
Vidia releases a new GPU, lowering the value of all
the other GPUs in the process, making it harder to
(09:27):
fill in the holes created by all the other GPUs
is capex and costs. This whole time, nobody appears to
have found a way to make a profit, meaning that
the hole created by these GPUs remains unfilled, all while
big tech firms by more GPUs, creating more holes to fill.
So now that you know this, there's a fairly obvious
question to ask, why in the hell are they still
(09:48):
buying GPUs? Well, so, where the fuck are these GPUs going?
So a few weeks ago I wrote a piece Premium
one called the Hater's Guide to Nvidia, and I asked
(10:09):
the basic question in there, where have all the GPUs
that Invidia has sold actually gone? In particular, the six
million Blackwell GPUs that Jensen one keeps banging on about
Now there's little evidence that these are being used in
the volume which they're sold, suggesting that they're either languishing
in the supply chain or being warehoused by hyperscalers, or
even in Nvidia themselves. Now there's the argument that this
could be and this is wanky Nvidian bullshit, this could
(10:32):
actually be two GPUs per GPU sold because there's two
chips on each GPU. Even if that was the case,
three billion GPUs Blackwell, specifically the brand new ones, they're
not in service.
Speaker 1 (10:45):
Now.
Speaker 2 (10:46):
While I'm not going to go and copy paste an
entire premium piece into this script, I am, however, going
to go into detail about what I found. And the
truth is, I can only really see in this inclodes,
looking over like bunches of data center maps, reading hundreds
of press releases, documents, earning statements. I've only been able
to find maybe a couple hundred thousand Blackwell GPUs in existence,
(11:09):
maybe half a million to seven hundred and fifty thousand
if you include the stuff that hasn't even been built yet.
Let's go into it. So Stargate Apilene allegedly four hundred
thousand Blackweld gps A, going there now Oracle CEO Coco,
I should say, Clay McGurk, that's probably not how you
say that. He claimed very recently there were ninety six
thousand of them in stored, so not great. There's theoretically
(11:33):
in one hundred and thirty one thousand Blackwell GPU cluster
owned by Oracle that they announced in March twenty twenty five,
so that should be online. Never five thousand blackweld GPUs
at the University of Texas Austin, which sound like they're online,
more than fifteen hundred in a Lambda data center in Columbus, Ohio.
Those are online. The Department of Energy is still in
development one hundred thousand GPU supercluster as well as ten
(11:56):
thousand in Vidia blackweld GPUs that are expected to be
available in twenty twenty six, and it's Equinox cluster. Really
can't establish how many of those are actually in operation.
Fifty thousand of these blackweld GPUs going into the still
unbuilt musk Run Colossus to supercluster, Corby's largest GB two
hundred Blackwell cluster of two four hundred and ninety six
(12:16):
Blackwell GPUs, tens of thousands of them deployed globally by Microsoft,
including forty six hundred Blackwell Ultra GPUs and two hundred
and sixty thousand of them, these Blackwell GPUs going into
five AI data centers for the South Korean government, and yeah,
I just want to be clear that that is also
fairly recently announced, so probably not even not even built,
(12:38):
let alone powered on. I goan to be honest, I'm
genuinely unable to find one million Blackwell GPUs like inexistence. Now,
some of you might say, oh, there's a bunch of
secret ones. There's a bunch of them. They don't announce
every single one. Here's the thing. Three million of these
fucking things have allegedly been shipped. I can't find a
million of them. And considering everybody always talks about their
(13:01):
GPU purchases, I'm kind of shocked at calm. Now. I
do not know where these six million black Weld GPUs
have gone, but they certainly haven't gone into data centers
that are powered and turned on. In fact, power has
become one of the biggest issues of building these things,
and the fact it's really difficult and maybe impossible to
get the amount of power these things need to the
goddamn data centers. In really simple terms, there isn't enough
(13:22):
power or built data centers for those black world GPUs
to run, in part because the data centers aren't built,
and in part because there isn't enough power for the
ones that are. Microsoft CEO Sacha Nadella recently said in
a podcast that his company and I quote didn't have
the warm shells to plug into, meaning buildings with sufficient power,
and heavily suggested that Microsoft may actually have a bunch
(13:43):
of chips sitting in inventory that they couldn't plug in now.
Just to give you an estimate here, even if we
say three million GPUs, even if we're going with the moonmath,
the Vinvidia, if we're going into the make believe world,
the twisted mind of Jensen Wang, still three million GPUs.
We'll look at still like five or six gigawatts of capacity.
(14:04):
It's not being built. I don't even think two gigawats
of data center capacity have been built. And I swear
to fucking God, if one of you emails me and said,
and Eric is built, and twenty gigawatts or something, power
can get built. Power can get built. You can build power,
getting it to the data center and actually powering the
data center correctly, as in things turn on, everything works,
(14:25):
nothing overloads, nothing blacks out, and the power is consistently done.
Takes what's just months of surveys and scientific stuff and
then years to just get it done. Stargate Appleine only
has two hundred megawatts. They're gonna need over one point
four gigawats just to turn the fucking thing on. I'm
so tired of He's God damn GPUs bo. With all
(14:45):
this said, why pray tell? Is Jensen Huang of Nvidia
saying that he has twenty million Blackwell and Vera Rubin
GPUs ordered through the end of twenty twenty six. Where
are they fucking going? Jensen? Now, I think that number
also includes the six million. And also, to be clear,
I know a lot of you aren't technical, which is awesome.
I love. I want you all to know about this.
(15:06):
You need to know that this is part of the
course within Nvidia in video loves schmushing accountancy things together
and coming up with random numbers. Credit the case could
go our friend of the show for telling me the story.
But during the early twenty twenties, so that there is
twenty twenty two during the Big Crypto rush. In Vidia
classified gaming GPUs that were sold to bitcoin miners as
(15:28):
gaming revenue. They got digged by the SEC. Wasn't fraud,
but just so you know, in video will move shit around.
And I truly do not know where these GPUs are.
I do not know even why anyone is still buying GPUs.
Speaker 1 (15:41):
Now.
Speaker 2 (15:41):
AI bulls will tell you that there's this insatiable demand
for AI and these massive amounts of orders are proof
of something or rather, and you know what, I'll give
them that it's proof that people are buying a lot
of GPUs. I just don't know why nobody has made
a profit from AI, and those making revenue aren't really
making that much. Let me give you an example. My
(16:02):
reporting on open ai from November twelfth suggests that the
company I only made four point three two nine billion
dollars in revenue for the end of September, extrapolated from
the twenty percent revenue share that Microsoft receives in the company.
And now some people who write really shit our substacks
have argued with the figures, claiming that they're either delayed
or are not inclusive of the revenue that open Ai
(16:23):
is paid from Microsoft as part of being's ai integration
and sales of open AI's models throughout Microsoft as ure.
So I want to be clear of two things. Some
a deeply bitter person. This is a crual accounting, meaning
that these numbers are revenue booked in the quarter I
reported them. Any comments about quarter long delays or naive approaches,
and you know who I'm fucking talking about, if you're listening,
(16:45):
are incorrect and a riboso. Also, Microsoft's revenue share payments
to open ai kind of pathetic, totally based on documents
reviewed by this newsletter publication whatever you call me, media
entity floating blob in the podcast, for sixty nine point
one million dollars in counting the year Q three, twenty
twenty five. And by the way, the actual number for
(17:08):
that three month period, including all royalties, is about four
point five two seven billion dollars of revenue. I just
want to be clear about something with open Ai. I'm
not saying they're misrepresenting their numbers to anyone. I hope
that open ai is being honest with their revenues. But
if it comes out, I'm right, if it comes out
that it turns out that they've been telling investors completely
(17:28):
different numbers, I'm gonna be absolutely fucking insufferable. I'm going
to bring in I'm going to be playing Tommy Trumpets
a walk around cheering that you can get five minutes
of monologue about that. Also in the same period, Open
Ai spent eight point six seven billion dollars on inference,
which is the process in which an l them creates
its output. This is the biggest company in the generative
(17:49):
AI space, with eight hundred million weekly active users in
the Mandate of Heaven in the eyes of the media, Anthropic,
it's largest competitor, Allegacy, will make eight hundred and thirty
three million dollars in revenue in December tween twenty five,
and based on my estimates, will end up having about
four and a half to five billion dollars if revenue
by the end of the year. Based on my reporting
from October Andthropic spent two point six six billion dollars
(18:10):
on Amazon Web Services through the end of September, meaning
that it, based on my own analysis of reported revenues,
spent one hundred and four percent of its revenue up
to that point just on AWS likely spent as much
on Google Cloud. Now, the reason I'm bringing up these
numbers is these are the champions, the champions of the
AI boom, yet their revenues kind of fucking stink. Wow.
(18:34):
Even if open ai made thirteen billion dollars this year,
even if Anthropic made five billion dollars, okay, wow, so
that's not even twenty billion dollars. That's like nineteen billion
dollars less than Microsoft spent on GPU's and other capex
in the last quarter. That's dogshit. I'm sorry, I'm just
tired of I am tired of humoring this. I'm sure
(18:56):
all of you are too. I find it loathsome that
we have to pretend these people are gifted somehow, they
have shit. Ask businesses that burn billions of dollars, and
you know what. Another thing I'm tired about is everybody
telling this story about Anthropic being more efficient and only
burning two point eight billion dollars this year. Now one
(19:16):
has to ask a question about why this company that's
allegedly reducing costs had to raise thirteen billion dollars in
September twenty twenty five, after raising three point five billion
dollars in March twenty twenty five, after raising four billion
dollars in November twenty twenty four. Am I really meant
to read stories about Anthropic hitting break even in twenty
twenty eight with a straight face, especially as other stories
(19:39):
say that we cash flow positive as soon as twenty
twenty seven. This company's as big a pile of shit
as open Ai. Open Ai raised eighteen point three billion
dollars this year. That's less than two billion dollars more
than Anthropic, who makes a bunch less revenue. Can't believe
I'm defending open Ai. But these companies are the two
(20:00):
two largest ones in the generative of AI space, and
by extension, the two largest consumers of GPU COMPUW Both
companies burn billions of dollars and require an infinite amount
of venture capital to keep them alive. At a time
when the Saudi Public Investment Fund is struggling and the
US venture capital system is set to run out of
cash in the next year and a half, The two
largest sources of actual revenue for selling AI compute are
(20:22):
subsidized by venture capital and debt. What happens if these
sources dry up. They're not paying out of cash flow,
and in all seriousness, who else is buying AI compute?
What are they doing with them? Hyperscalers other than Microsoft,
which chose to stop reporting its AI revenue back in
January when it claimed it made about a billion dollars
a month in revenue, don't disclose anything about their AI revenue,
(20:43):
which in turn means that we have no real idea
of how much real actual money is coming to justify
these GPUs core We've made one point three six billion
dollars in revenue and lost one hundred and ten million
dollars doing so in the last quarter. And if that's
indicative of the kind of actual real demand for AI compume,
I think it's time to start panicking about whether all
of this was for nothing. Corweave has a backlog of
(21:06):
over fifty billion dollars in compute, and twenty two billion
dollars of that is open AI company that burns billions
of dollars a year and lives on venture subsidies. Fourteen
billion dollars of that is Meta, which is yet to
work out how to make any kind of real money
from generative AI. And no, it's generative AI ads are
not the future four or four media. I love you,
but that story was bunk and the rest of it
(21:26):
is likely a mixture of Microsoft and Video, which agreed
to buy six point three billion dollars of any unused
compute from Corewave through twenty thirty two. Should also be clear,
I do pay and subscribe to four or four I
love it. Just the AI ads story was wank, I
love you, I love you, Joe, I love I love
I love the publication. Sorry. I also forgot Google, by
the way, which is renting capacity from Corewave to rent
(21:47):
to open AI, and I'm not shitting you. Oh fuck Sorry.
I also forgot to mention that Corwy's backlog problem stems
from data center construction delays. That and Corewave has fourteen
billion dollars in debt, mostly from buying GPUs, which is
able to raise by using GPUs as collateral, and then
it had contracts and customers willing to pay for it,
such as in Video, who is also selling it the GPUs.
(22:09):
I also left something out of this script, which is
that just the last week, core We've just raised another
two billion dollars of debt. When this all ends, I
am going to be a little insufferable. But let's just
be abundantly clear. Core Weaver has bought all those GPUs
to rent open AI, Microsoft for open Ai, Meta Google
for open Ai, and in Video, which is the company
(22:30):
that benefits from core Weave's continuum ability to buy GPUs? Otherwise,
where's the fucking business? Exactly? Who are the customers, who
are the people renting the GPUs? And what is the
purpose for which they're being rented? How much money is
renting those gps? Can you? Can you tell me? Can
anyone tell me? Can anyone tell me anything? You can
sit and wank and waffle on about the supposed glorious
(22:51):
AI revolution all you want, but where's the goddamn money?
And why exactly are we still buying GPUs? What are
they doing, who are they being rented for what purpose?
And why isn't it creating the kind of revenue that's
actually worth sharing or products that are actually worth using.
Is it because the products suck? Is it because the
revenue sucks? Is it because it's unprofitable to make the revenue?
(23:14):
And why at this point in history do we not
know hundreds of billions of dollars that have made in
Vidia the biggest company on the stock market, and we
still do not know why people buy these fucking things,
nor do we know what they fucking cost imagine if
we sold cars and we didn't have a milesber gallon rating.
I'm serious. That's effectively where we are. Oh God, and
(23:51):
Video is currently making hundreds of billions of dollars in
revenue selling GPUs to companies that either plug them in
and start losing money or I assume put them in
a warehouse for safety. And those companies increasingly a raking
up mountains of debt to do so, and billions more
in long term lease payments. And this brings me to
my core anxiety. Why exactly a company's pre ordering GPUs?
(24:14):
What benefit is there in doing so? Blackwell does not
appear to be more efficient in a way that actually
makes anybody a profit. And we're potentially years from seeing
these GPUs in operation in data centers at the scale
they're being shipped, So why is anyone buying more? I
just want to be really specific about something, because I
don't feel like I nailed this down two and a
half years. Fifty billion dollars per gigabot of data centers.
(24:34):
You may be thinking, what black Wells, You'll just shove
them in the old data centers, right, No, they use
these Oberon racks specific new racks. They take a bunch
more power, and they need a bunch of liquid cooling.
You can't just retrofit easily. You have to bulldoze shit
and rebuild well, remove all the housing and then add
HVAC stuff. It's very expensive and takes a long time.
(24:55):
And look, I just don't know what's happening with these GPUs,
and I'm a little bit concerned, And I doubt these
are new customers. They're likely hyperscalers, neo clouds like core Weave,
and resellers like Dell and supermicro, who also both sell
to Gorewave. Because the only companies that can actually afford
to buy GPUs are those with massive amounts of cash
or debt, to the point that even Google, Amazon, Meta,
(25:17):
and Oracle are taken on massive amounts of new debt
or without a plan to make a profit. Oracle is
looking potentially at fifty six billion dollars of debt. It's
completely bonkers. In Video's largest customers are increasingly unable to
afford its GPUs, which appear to be increasing in price
with every subsequent generation. In Video's GPUs are so expensive
that the only way you can buy them is by
(25:38):
already having billions of dollars or being able to raise
billions of dollars, which means in a very real sense
that in Video is dependent not on its customers, but
on its customers credit ratings and financial backers, and the
larger private credit institutions, which I'm eventually going to have
to do a newsletter on and a podcast an because honestly,
every time I read about the private credit situation with
(25:59):
blue out again, I'll begin here in the wa bit
from kill bill it's not good. And to make matters worse,
the key reason that one would buy a GPU is
to either run AI services using it, or rent it
to somebody else to run AI services, and the two
largest parties spending money on these services are open Ai
and Anthropic, both of whom lose billions of dollars and
(26:20):
thus are much like the people buying the GPUs depending
on venture capital and debt. Now remember open ai and
Anthropic both have lines of credit, four billion dollars for
open Ai and two and a half billy for Aanthropic.
In simple terms, in Nvidia's customers rely on debt to
buy as GPUs, and in Vidia's customers customers rely on
debt to pay to rent them. Yeah, it's not great
(26:42):
yet it actually gets worse from there. Who after all
of the biggest customers paying the company's renting GPUs to
sell their AI bottles. That's right, AI startups, all of
which are deeply unprofitable. Cusser. Anthropic's largest customer and now
it's biggest competitor in the AI coding sphere, raised two
point three billion dollars in November after raising nine hundred
million dollars in June. But plex D, one of the
(27:04):
most popular right put that in their quotes, raised two
hundred million dollars in September after raising one hundred million
dollars in July, after seeming to fail to raise half
a billion dollars in May. After raising five hundred million
dollars in December twenty twenty four, Cognition raised four hundred
million dollars in September after raising three hundred million dollars
in March, and Coher raised one hundred million dollars in September,
a month after it raised five hundred million dollars. None
(27:26):
of these companies are profitable, not even close. I read
a story in Newcomer by Tom Datan that said the
cursor cents one hundred percent of its revenue to Anthropic
to pay for its models. Very cool. So I really
want to lay this out for you because it's very
bad when you think about So venture capital is feeding
money to startups. They fund the startups AI startups, and
(27:49):
then they pay either or both open ai or Anthropic
to use their models. Now open Ai and Anthropic need
to serve those models, right, So they then raised venture
capital or there to pay hypersu scalers or neoclouds to
rent Nvidia GPUs. At that point, hyper scalers and neoclouds
then use either debt or exist in cash flow in
the case of Hyperscalers, though not for long, to buy
(28:10):
more in Vidia GPUs. Only one company appears to make
profit here, and it's in Vidia. Well in video and
its resellers like Dell and super Micro, which buy in
Vidia GPUs, put them in service and sell them to
neo clouds like Lamter or Core with at some point
to link in. This debt back chain breaks because very
little cash flow exists to prop it up. At some point,
(28:33):
venture capitalists will be forced to stop funneling money into unprofitable,
unsustainable AI companies which will make those companies unable to
funnel money into the pockets of anthropic and ope and
AI who rent the GPUs will then not be able
to funnel money into the pockets of those buying GPUs,
which will make it harder for those companies to justify
buying GPUs. And at that point some of this comes
(28:56):
to InVideo and Invidia doesn't make so make so much money.
I'm honest, none of Nvidia's success really makes any sense.
Who's buying so many GPUs and where are they going?
Why are in Vidia's inventories increasing? Is it really just
pre buying parts for future orders? Why are their accounts
receivable climbing? And how much product is Nvidia shipping before
(29:16):
it gets paid? While these are both explainable as this
is a big company and this is how companies do business,
and that's true, why do receivables not seem to be
coming down? And how long, realistically can the largest company
on the stock market continue to grow revenues selling assets
that only seem to lose its customers money and don't
(29:37):
seem to even be in use for years. I worry
about it in Nvidia, not because I think there's a
massive scandal, but because so much rights and its success
and its success rights on the back of dwindling amounts
of venture capital, and there because nobody is actually making
money to pay for these GPUs, let alone running them.
In fact, I'm not even saying in video goes tits up.
(29:58):
I want to be clear about that. I think they
may even have another good quarter or two in them.
It really just comes down to how long people are
willing to be stupid and how long Jensen Wong is
able to call up Sachin Adella and Co. At three
in the morning and say, buy one billion dollars of GPUs,
you pig Finnom style baby. But really, I think much
of the US stock market's growth is held up by
(30:20):
how long everybody is willing to be gas lit by
Jensen Wong into believing that they need more GPUs. At
this point, it's barely about AI anymore, as AI revenue
real cash made from selling services run on those GPUs,
doesn't even cover the costs, let alone create the cash
flow necessary to buy more seventy thousand dollars GPUs thousands
at a time. It's not like any actual innovational progress
(30:43):
is driving this bullshit. In any case, the market's crave
are healthy in video has so many hundreds of billions
of dollars of invidious stock sits in the hands of
retail investors and people's four O one ks, and its
endless growth has helped paper over the pallid growth of
the US stock market and by extension, the decay of
the tech industry's ability to innovate. Once this pops, and
(31:04):
it will pop because there's simply not enough money to
do this forever, there must be a referendum on those
that chose to ignore the naked instability of this era
and the endless lies that inflate the AI bubble. I
will be walking around with a gavel. I am going
to be taking heads. I am fucking sick of this era.
And what I'm most sick of is that so few
people are still to this day willing to admit how
(31:27):
bad this is. And I know in the next few
months we're going to get articles some major media outlets
that say, how could we have seen this coming? And
like I said in the previous episode, they could have
fucking looked. All of them could have looked, and they
could have looked a year ago. The incredible support I
get from all of you truly makes this show a
joy to make, even though I've done way too many
retakes on this and apologies to Mattasowski for the noises
(31:49):
I make. But I think in the next few months
we're all going to be validated. It's going to be
the great vindication. But until then, everybody is betting billions
on the eye idea that wily coyote won't look down.
He's gonna have to at some point, won't it. Thank
(32:13):
you 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
Mattasowski dot com, m A T T O. S O
W s ki 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,
(32:34):
my newsletter. I also really recommend you go to chat
dot Where's Youreed dot at to visit the discord, and
go to our slash Better Offline to check out our reddit.
Thank you so much for listening.
Speaker 1 (32:45):
Better Offline is a production of Cool Zone Media. For
more from Cool Zone Media, visit our website. Cool zonemedia
dot com, or check us out on the iHeartRadio app,
Apple Podcasts, or wherever you get your podcasts.
Media.
Speaker 2 (00:04):
Hi, I'm ed Zitron, and welcome back to Better Offline
And this is our third and final part of our
Better Offline and Vidia special, where we're talking about well,
(00:27):
the shakiness behind its growth and how the company, despite
being on incredibly infirm ground, is definitely not Enron or
Nortail or Wolcom or Lucien or any other dot com
bubble era affirm that imploded under its own way and
well quite dodgy accounting. The thing is, even if Enron
is nothing like them, there's still quite a few causes
for concern, and that's largely driven from the fact that
(00:49):
in Vidio makes the majority of its money selling GPUs
to a handful of customers, and so well, some of
those also look to be on some of their own
incredibly shaky ground. And yeah, I'm talking about Oracle now.
In Vidia's health saying nothing of its growth isn't just
tied to these customers. It's also tied to whether these
customers can actually turn a profit from their capex spending.
(01:12):
And even that's not even certain so due to the
fact that so much money has been piled into building
AI infrastructure and big tech has promised to spend hundreds
of billions of dollars more in the next year. Big
tech has found itself in a bit of a hole.
How big of a hole? Well, by the end of
the year, Microsoft, Amazon, Google, and Meta will have spent
over four hundred billion dollars in capital expenditures, much of
(01:34):
it focused on building AI infrastructure, on top of two
hundred and twenty eight point four billion dollars in capex
in twenty twenty four and around one hundred and forty
eight billion in capital expenditures in twenty twenty three, for
a total of seven hundred and seventy six billion in
the space of three years, and they expect to spend
more than four hundred billion dollars more in twenty twenty six.
(01:56):
Every time I read these numbers, I feel a little crazy.
As a result, based on my own analysis, big tech
needs to make two trillion dollars in brand new, brand
spanking new revenue, specifically from AI by twenty thirty. All
of this was effectively for nothing. Now, I go into
detail about this in the premium newsletter I did on
October thirty first, but I'm going to give you a
(02:17):
short explanation here. First, though, we have to talk about depreciation,
and because I'm lazy, I'm going to quote myself in
that newsletter I just mentioned a couple of seconds ago
at hem. So when Microsoft buyers say one hundred million
dollars worth of GPUs, it immediately comes out of its
capital expenditures, which is when a company uses money to
invest in either buying or upgrading something. It then adds
(02:40):
to its property, plants and Equipment assets PPE for sure,
although some companies list this on their annual and quarterly
financials as property and equipment PPE sits on the balance sheet,
it's an asset as it's stuff for the company that
it owns or as least, GPUs depreciate, meaning they lose
value or over time, and this depreciation is represented on
a balance sheet and the income statement. Essentially, the goal
(03:03):
is to represent the value of an asset that a
company has on the income statement, and we see how
much the assets have declined during the reporting period, whether
that be a year or a quarter or something else,
Whereas the balance sheet shows the cumulative depreciation of every
asset currently in play. Depreciation does two things. And I
know this sounds like a lot, but I'll break it
down for you First, it allows a company to accurately
(03:25):
to an extent, represent the value of things it owns
over their useful life. Secondly, it allows a company to
deduct the value of an asset across said useful life
right up until its eventual removal, versus having to take
a big hit up front. The way this depreciation is
actually calculated can vary. There are several different methods available,
with some allowing for greater deductions at the start of
(03:46):
the term, which is useful for those items that will
experience the biggest drop in value right after buying them
and their initial use. An example you're probably familiar with
is a new car which loses a significant chunk of
its value the moment is driven off a dealership. Long
creation has become a big ugly problem with GPUs specifically
because of that useful life to find either as how
long the thing is able to be run before it
(04:07):
dies or how long before it becomes obsolete, and nobody
seems to be able to come up with a consensus
about how long this should be. In Microsoft's case, the
appreciation for its service is spread over six years, a
convenient change it made in August twenty twenty two, A
few months before the launch of chat GPT and before
it bought a bunch of fucking GPUs. This means that
Microsoft can spread the cost of tens of thousands of
(04:28):
a one hundred GPUs brought in twenty twenty or the
four hundred and fifty thousand, h one hundred gupus it
bought in twenty twenty four across six years, regardless of
whether those are the years they'll be generating revenue or
naturally functioning. Corwy for what It's worth says the same thing,
but largely because it's betting that it'll still be able
to find users for older silicon after its initial contracts
(04:49):
with the companies like OpenAI expire. The problem is is
that aigpus are fairly new concepts, and thus all of
this is pretty much untested ground. Whereas we know how long,
say a truck or a piece of heavy machinery can
last and how long it can deliver value to an organization,
we don't know the same thing about the kind of
data center GPUs that hyperscalers are spending tens of billions
(05:10):
of dollars on each year. Any kind of depreciation schedule
is based on at best assumptions and at worst hope.
Now this is important. The concept of an AI data
center is super new. We maybe saw the first ones
in twenty ninety. In question, it's kind of hard to say,
but even at the scale we're seeing today a Gigawa
(05:31):
data center pretty much brand new, maybe a couple years old.
I don't even think they've even built any but we'll
get to that in a bit. There are a lot
of assumptions at play. There's the assumption that the cards
won't degrade with heavy usage, or the assumption that future
generations of GPUs won't be so powerful and impressive that
they'll render the previous ones more obsolete than expected, kind
of like how the first jet powered planes of the
(05:52):
nineteen fifties did to those manufactured just a decade prior.
The assumption that there will be in fact a market
for Alder cards, than that there'll be a way to
lease them profitably. What if those assumptions are I don't know, wrong.
What if that hope is ultimately irrational. So there's a
quote from the Center for Information Technology Policy framing this problem, well,
(06:13):
that'll link to in the notes. Here is the puzzle.
The chips at the heart of the infrastructure build out
have a useful lifespan of one to three years due
to rapid technological obsolescence and physical wear, but companies appreciate
them over five or six years. In other words, they
spread out the cost of their massive capital expenditures over
a longer period than the facts warrant what the economist
is referred to as the four trillion dollar accounting puzzle
(06:35):
at the heart of the AIICLOUD. This is why Michael
Burry brought it up recently because spreading out these costs
allowed spigtech to make their net income i e. Their
profits look better in simple terms. By spreading out the
costs over six years rather than three, hyperscalers are able
to reduce the line item that eats into their earnings,
which makes their companies look better to the markets. So
why does this create an artificial time limit? Well, let's
(06:58):
start with a horrible fact. It takes it's two point
five years of construction time in about fifty billion dollars
per gigawa of data center capacity. No matter when the
GPUs for a giga what data center are bought one
way or another, these GPUs are depreciating in value either
through death or reduced efficacy through wear and tear, or
becoming obsolete, which is very likely as in Video is
committed to releasing a new GPU every single year. Newer
(07:21):
generation GPUs, like in videos Blackwell and Verra Reuben require
entirely new data center architecture, meaning that one as to
why they build a brand new data center a retrofit
an old one. Essentially, we have facilities that are being
built around a GPU design or product that may change
in a year or two. Now I hear that in
the Oberon racks that they use for the Blackwells will
(07:42):
be used with some Verra Rubin. But even then there's
going to be an even bigger, more huger Vera Ruben
that comes that might even that I read somewhere that
there might even be like killer what level ones just
in like one hundred killer what ones that this company's insane. Nevertheless,
at some point Wall Street is going to need to
see some sort of return on this investment, and right
now that return is negative dollars. I break it down
(08:03):
on my October thirty first premium piece, but for your sake,
I'll just say it. I estimate the big techniques to
make two dollars for every dollar of CAPEX they've spent,
and this revenue must be new, brand new, As this
CAPEX is only for AI. This CAPEX is useless for
everything else. It does not help it. And no, it
doesn't help that they bolted copilot onto fucking everything that
(08:24):
is not working, and in fact, the Australian Competition Commission
is suing them. Maybe I mentioned that later, but whatever, Meta, Amazon,
Google and Microsoft are already years and hundreds of billions
of dollars in and are yet to see a dollar
of profit, creating a one point two to one trillion
dollar hole just to justify the expenses. So around six
hundred and five billion dollars of CAPEX all told at
(08:44):
the time I calculated it. Much of this CAPEX has
been committed or spent before they've even turned on a
single goddamn GPU. You might argue that there's a scenario
here where say, an A one hundred GPU is useful
past the three or six year shelf life. Even if
that were the case, the average rental price of an
A one hundred is ninety nine cents an hour. This
is a four or five year old GPU, and customers
(09:06):
are paying for it like they would a five year
old piece of hardware. The same fate awaits the H
one hundred, which was released in twenty twenty two but
was still sold in great volume through twenty twenty four,
and I hear the H two hundred of the same
generation is still selling to this day. Every year in
Vidia releases a new GPU, lowering the value of all
the other GPUs in the process, making it harder to
(09:27):
fill in the holes created by all the other GPUs
is capex and costs. This whole time, nobody appears to
have found a way to make a profit, meaning that
the hole created by these GPUs remains unfilled, all while
big tech firms by more GPUs, creating more holes to fill.
So now that you know this, there's a fairly obvious
question to ask, why in the hell are they still
(09:48):
buying GPUs? Well, so, where the fuck are these GPUs going?
So a few weeks ago I wrote a piece Premium
one called the Hater's Guide to Nvidia, and I asked
(10:09):
the basic question in there, where have all the GPUs
that Invidia has sold actually gone? In particular, the six
million Blackwell GPUs that Jensen one keeps banging on about
Now there's little evidence that these are being used in
the volume which they're sold, suggesting that they're either languishing
in the supply chain or being warehoused by hyperscalers, or
even in Nvidia themselves. Now there's the argument that this
could be and this is wanky Nvidian bullshit, this could
(10:32):
actually be two GPUs per GPU sold because there's two
chips on each GPU. Even if that was the case,
three billion GPUs Blackwell, specifically the brand new ones, they're
not in service.
Speaker 1 (10:45):
Now.
Speaker 2 (10:46):
While I'm not going to go and copy paste an
entire premium piece into this script, I am, however, going
to go into detail about what I found. And the
truth is, I can only really see in this inclodes,
looking over like bunches of data center maps, reading hundreds
of press releases, documents, earning statements. I've only been able
to find maybe a couple hundred thousand Blackwell GPUs in existence,
(11:09):
maybe half a million to seven hundred and fifty thousand
if you include the stuff that hasn't even been built yet.
Let's go into it. So Stargate Apilene allegedly four hundred
thousand Blackweld gps A, going there now Oracle CEO Coco,
I should say, Clay McGurk, that's probably not how you
say that. He claimed very recently there were ninety six
thousand of them in stored, so not great. There's theoretically
(11:33):
in one hundred and thirty one thousand Blackwell GPU cluster
owned by Oracle that they announced in March twenty twenty five,
so that should be online. Never five thousand blackweld GPUs
at the University of Texas Austin, which sound like they're online,
more than fifteen hundred in a Lambda data center in Columbus, Ohio.
Those are online. The Department of Energy is still in
development one hundred thousand GPU supercluster as well as ten
(11:56):
thousand in Vidia blackweld GPUs that are expected to be
available in twenty twenty six, and it's Equinox cluster. Really
can't establish how many of those are actually in operation.
Fifty thousand of these blackweld GPUs going into the still
unbuilt musk Run Colossus to supercluster, Corby's largest GB two
hundred Blackwell cluster of two four hundred and ninety six
(12:16):
Blackwell GPUs, tens of thousands of them deployed globally by Microsoft,
including forty six hundred Blackwell Ultra GPUs and two hundred
and sixty thousand of them, these Blackwell GPUs going into
five AI data centers for the South Korean government, and yeah,
I just want to be clear that that is also
fairly recently announced, so probably not even not even built,
(12:38):
let alone powered on. I goan to be honest, I'm
genuinely unable to find one million Blackwell GPUs like inexistence. Now,
some of you might say, oh, there's a bunch of
secret ones. There's a bunch of them. They don't announce
every single one. Here's the thing. Three million of these
fucking things have allegedly been shipped. I can't find a
million of them. And considering everybody always talks about their
(13:01):
GPU purchases, I'm kind of shocked at calm. Now. I
do not know where these six million black Weld GPUs
have gone, but they certainly haven't gone into data centers
that are powered and turned on. In fact, power has
become one of the biggest issues of building these things,
and the fact it's really difficult and maybe impossible to
get the amount of power these things need to the
goddamn data centers. In really simple terms, there isn't enough
(13:22):
power or built data centers for those black world GPUs
to run, in part because the data centers aren't built,
and in part because there isn't enough power for the
ones that are. Microsoft CEO Sacha Nadella recently said in
a podcast that his company and I quote didn't have
the warm shells to plug into, meaning buildings with sufficient power,
and heavily suggested that Microsoft may actually have a bunch
(13:43):
of chips sitting in inventory that they couldn't plug in now.
Just to give you an estimate here, even if we
say three million GPUs, even if we're going with the moonmath,
the Vinvidia, if we're going into the make believe world,
the twisted mind of Jensen Wang, still three million GPUs.
We'll look at still like five or six gigawatts of capacity.
(14:04):
It's not being built. I don't even think two gigawats
of data center capacity have been built. And I swear
to fucking God, if one of you emails me and said,
and Eric is built, and twenty gigawatts or something, power
can get built. Power can get built. You can build power,
getting it to the data center and actually powering the
data center correctly, as in things turn on, everything works,
(14:25):
nothing overloads, nothing blacks out, and the power is consistently done.
Takes what's just months of surveys and scientific stuff and
then years to just get it done. Stargate Appleine only
has two hundred megawatts. They're gonna need over one point
four gigawats just to turn the fucking thing on. I'm
so tired of He's God damn GPUs bo. With all
(14:45):
this said, why pray tell? Is Jensen Huang of Nvidia
saying that he has twenty million Blackwell and Vera Rubin
GPUs ordered through the end of twenty twenty six. Where
are they fucking going? Jensen? Now, I think that number
also includes the six million. And also, to be clear,
I know a lot of you aren't technical, which is awesome.
I love. I want you all to know about this.
(15:06):
You need to know that this is part of the
course within Nvidia in video loves schmushing accountancy things together
and coming up with random numbers. Credit the case could
go our friend of the show for telling me the story.
But during the early twenty twenties, so that there is
twenty twenty two during the Big Crypto rush. In Vidia
classified gaming GPUs that were sold to bitcoin miners as
(15:28):
gaming revenue. They got digged by the SEC. Wasn't fraud,
but just so you know, in video will move shit around.
And I truly do not know where these GPUs are.
I do not know even why anyone is still buying GPUs.
Speaker 1 (15:41):
Now.
Speaker 2 (15:41):
AI bulls will tell you that there's this insatiable demand
for AI and these massive amounts of orders are proof
of something or rather, and you know what, I'll give
them that it's proof that people are buying a lot
of GPUs. I just don't know why nobody has made
a profit from AI, and those making revenue aren't really
making that much. Let me give you an example. My
(16:02):
reporting on open ai from November twelfth suggests that the
company I only made four point three two nine billion
dollars in revenue for the end of September, extrapolated from
the twenty percent revenue share that Microsoft receives in the company.
And now some people who write really shit our substacks
have argued with the figures, claiming that they're either delayed
or are not inclusive of the revenue that open Ai
(16:23):
is paid from Microsoft as part of being's ai integration
and sales of open AI's models throughout Microsoft as ure.
So I want to be clear of two things. Some
a deeply bitter person. This is a crual accounting, meaning
that these numbers are revenue booked in the quarter I
reported them. Any comments about quarter long delays or naive approaches,
and you know who I'm fucking talking about, if you're listening,
(16:45):
are incorrect and a riboso. Also, Microsoft's revenue share payments
to open ai kind of pathetic, totally based on documents
reviewed by this newsletter publication whatever you call me, media
entity floating blob in the podcast, for sixty nine point
one million dollars in counting the year Q three, twenty
twenty five. And by the way, the actual number for
(17:08):
that three month period, including all royalties, is about four
point five two seven billion dollars of revenue. I just
want to be clear about something with open Ai. I'm
not saying they're misrepresenting their numbers to anyone. I hope
that open ai is being honest with their revenues. But
if it comes out, I'm right, if it comes out
that it turns out that they've been telling investors completely
(17:28):
different numbers, I'm gonna be absolutely fucking insufferable. I'm going
to bring in I'm going to be playing Tommy Trumpets
a walk around cheering that you can get five minutes
of monologue about that. Also in the same period, Open
Ai spent eight point six seven billion dollars on inference,
which is the process in which an l them creates
its output. This is the biggest company in the generative
(17:49):
AI space, with eight hundred million weekly active users in
the Mandate of Heaven in the eyes of the media, Anthropic,
it's largest competitor, Allegacy, will make eight hundred and thirty
three million dollars in revenue in December tween twenty five,
and based on my estimates, will end up having about
four and a half to five billion dollars if revenue
by the end of the year. Based on my reporting
from October Andthropic spent two point six six billion dollars
(18:10):
on Amazon Web Services through the end of September, meaning
that it, based on my own analysis of reported revenues,
spent one hundred and four percent of its revenue up
to that point just on AWS likely spent as much
on Google Cloud. Now, the reason I'm bringing up these
numbers is these are the champions, the champions of the
AI boom, yet their revenues kind of fucking stink. Wow.
(18:34):
Even if open ai made thirteen billion dollars this year,
even if Anthropic made five billion dollars, okay, wow, so
that's not even twenty billion dollars. That's like nineteen billion
dollars less than Microsoft spent on GPU's and other capex
in the last quarter. That's dogshit. I'm sorry, I'm just
tired of I am tired of humoring this. I'm sure
(18:56):
all of you are too. I find it loathsome that
we have to pretend these people are gifted somehow, they
have shit. Ask businesses that burn billions of dollars, and
you know what. Another thing I'm tired about is everybody
telling this story about Anthropic being more efficient and only
burning two point eight billion dollars this year. Now one
(19:16):
has to ask a question about why this company that's
allegedly reducing costs had to raise thirteen billion dollars in
September twenty twenty five, after raising three point five billion
dollars in March twenty twenty five, after raising four billion
dollars in November twenty twenty four. Am I really meant
to read stories about Anthropic hitting break even in twenty
twenty eight with a straight face, especially as other stories
(19:39):
say that we cash flow positive as soon as twenty
twenty seven. This company's as big a pile of shit
as open Ai. Open Ai raised eighteen point three billion
dollars this year. That's less than two billion dollars more
than Anthropic, who makes a bunch less revenue. Can't believe
I'm defending open Ai. But these companies are the two
(20:00):
two largest ones in the generative of AI space, and
by extension, the two largest consumers of GPU COMPUW Both
companies burn billions of dollars and require an infinite amount
of venture capital to keep them alive. At a time
when the Saudi Public Investment Fund is struggling and the
US venture capital system is set to run out of
cash in the next year and a half, The two
largest sources of actual revenue for selling AI compute are
(20:22):
subsidized by venture capital and debt. What happens if these
sources dry up. They're not paying out of cash flow,
and in all seriousness, who else is buying AI compute?
What are they doing with them? Hyperscalers other than Microsoft,
which chose to stop reporting its AI revenue back in
January when it claimed it made about a billion dollars
a month in revenue, don't disclose anything about their AI revenue,
(20:43):
which in turn means that we have no real idea
of how much real actual money is coming to justify
these GPUs core We've made one point three six billion
dollars in revenue and lost one hundred and ten million
dollars doing so in the last quarter. And if that's
indicative of the kind of actual real demand for AI compume,
I think it's time to start panicking about whether all
of this was for nothing. Corweave has a backlog of
(21:06):
over fifty billion dollars in compute, and twenty two billion
dollars of that is open AI company that burns billions
of dollars a year and lives on venture subsidies. Fourteen
billion dollars of that is Meta, which is yet to
work out how to make any kind of real money
from generative AI. And no, it's generative AI ads are
not the future four or four media. I love you,
but that story was bunk and the rest of it
(21:26):
is likely a mixture of Microsoft and Video, which agreed
to buy six point three billion dollars of any unused
compute from Corewave through twenty thirty two. Should also be clear,
I do pay and subscribe to four or four I
love it. Just the AI ads story was wank, I
love you, I love you, Joe, I love I love
I love the publication. Sorry. I also forgot Google, by
the way, which is renting capacity from Corewave to rent
(21:47):
to open AI, and I'm not shitting you. Oh fuck Sorry.
I also forgot to mention that Corwy's backlog problem stems
from data center construction delays. That and Corewave has fourteen
billion dollars in debt, mostly from buying GPUs, which is
able to raise by using GPUs as collateral, and then
it had contracts and customers willing to pay for it,
such as in Video, who is also selling it the GPUs.
(22:09):
I also left something out of this script, which is
that just the last week, core We've just raised another
two billion dollars of debt. When this all ends, I
am going to be a little insufferable. But let's just
be abundantly clear. Core Weaver has bought all those GPUs
to rent open AI, Microsoft for open Ai, Meta Google
for open Ai, and in Video, which is the company
(22:30):
that benefits from core Weave's continuum ability to buy GPUs? Otherwise,
where's the fucking business? Exactly? Who are the customers, who
are the people renting the GPUs? And what is the
purpose for which they're being rented? How much money is
renting those gps? Can you? Can you tell me? Can
anyone tell me? Can anyone tell me anything? You can
sit and wank and waffle on about the supposed glorious
(22:51):
AI revolution all you want, but where's the goddamn money?
And why exactly are we still buying GPUs? What are
they doing, who are they being rented for what purpose?
And why isn't it creating the kind of revenue that's
actually worth sharing or products that are actually worth using.
Is it because the products suck? Is it because the
revenue sucks? Is it because it's unprofitable to make the revenue?
(23:14):
And why at this point in history do we not
know hundreds of billions of dollars that have made in
Vidia the biggest company on the stock market, and we
still do not know why people buy these fucking things,
nor do we know what they fucking cost imagine if
we sold cars and we didn't have a milesber gallon rating.
I'm serious. That's effectively where we are. Oh God, and
(23:51):
Video is currently making hundreds of billions of dollars in
revenue selling GPUs to companies that either plug them in
and start losing money or I assume put them in
a warehouse for safety. And those companies increasingly a raking
up mountains of debt to do so, and billions more
in long term lease payments. And this brings me to
my core anxiety. Why exactly a company's pre ordering GPUs?
(24:14):
What benefit is there in doing so? Blackwell does not
appear to be more efficient in a way that actually
makes anybody a profit. And we're potentially years from seeing
these GPUs in operation in data centers at the scale
they're being shipped, So why is anyone buying more? I
just want to be really specific about something, because I
don't feel like I nailed this down two and a
half years. Fifty billion dollars per gigabot of data centers.
(24:34):
You may be thinking, what black Wells, You'll just shove
them in the old data centers, right, No, they use
these Oberon racks specific new racks. They take a bunch
more power, and they need a bunch of liquid cooling.
You can't just retrofit easily. You have to bulldoze shit
and rebuild well, remove all the housing and then add
HVAC stuff. It's very expensive and takes a long time.
(24:55):
And look, I just don't know what's happening with these GPUs,
and I'm a little bit concerned, And I doubt these
are new customers. They're likely hyperscalers, neo clouds like core Weave,
and resellers like Dell and supermicro, who also both sell
to Gorewave. Because the only companies that can actually afford
to buy GPUs are those with massive amounts of cash
or debt, to the point that even Google, Amazon, Meta,
(25:17):
and Oracle are taken on massive amounts of new debt
or without a plan to make a profit. Oracle is
looking potentially at fifty six billion dollars of debt. It's
completely bonkers. In Video's largest customers are increasingly unable to
afford its GPUs, which appear to be increasing in price
with every subsequent generation. In Video's GPUs are so expensive
that the only way you can buy them is by
(25:38):
already having billions of dollars or being able to raise
billions of dollars, which means in a very real sense
that in Video is dependent not on its customers, but
on its customers credit ratings and financial backers, and the
larger private credit institutions, which I'm eventually going to have
to do a newsletter on and a podcast an because honestly,
every time I read about the private credit situation with
(25:59):
blue out again, I'll begin here in the wa bit
from kill bill it's not good. And to make matters worse,
the key reason that one would buy a GPU is
to either run AI services using it, or rent it
to somebody else to run AI services, and the two
largest parties spending money on these services are open Ai
and Anthropic, both of whom lose billions of dollars and
(26:20):
thus are much like the people buying the GPUs depending
on venture capital and debt. Now remember open ai and
Anthropic both have lines of credit, four billion dollars for
open Ai and two and a half billy for Aanthropic.
In simple terms, in Nvidia's customers rely on debt to
buy as GPUs, and in Vidia's customers customers rely on
debt to pay to rent them. Yeah, it's not great
(26:42):
yet it actually gets worse from there. Who after all
of the biggest customers paying the company's renting GPUs to
sell their AI bottles. That's right, AI startups, all of
which are deeply unprofitable. Cusser. Anthropic's largest customer and now
it's biggest competitor in the AI coding sphere, raised two
point three billion dollars in November after raising nine hundred
million dollars in June. But plex D, one of the
(27:04):
most popular right put that in their quotes, raised two
hundred million dollars in September after raising one hundred million
dollars in July, after seeming to fail to raise half
a billion dollars in May. After raising five hundred million
dollars in December twenty twenty four, Cognition raised four hundred
million dollars in September after raising three hundred million dollars
in March, and Coher raised one hundred million dollars in September,
a month after it raised five hundred million dollars. None
(27:26):
of these companies are profitable, not even close. I read
a story in Newcomer by Tom Datan that said the
cursor cents one hundred percent of its revenue to Anthropic
to pay for its models. Very cool. So I really
want to lay this out for you because it's very
bad when you think about So venture capital is feeding
money to startups. They fund the startups AI startups, and
(27:49):
then they pay either or both open ai or Anthropic
to use their models. Now open Ai and Anthropic need
to serve those models, right, So they then raised venture
capital or there to pay hypersu scalers or neoclouds to
rent Nvidia GPUs. At that point, hyper scalers and neoclouds
then use either debt or exist in cash flow in
the case of Hyperscalers, though not for long, to buy
(28:10):
more in Vidia GPUs. Only one company appears to make
profit here, and it's in Vidia. Well in video and
its resellers like Dell and super Micro, which buy in
Vidia GPUs, put them in service and sell them to
neo clouds like Lamter or Core with at some point
to link in. This debt back chain breaks because very
little cash flow exists to prop it up. At some point,
(28:33):
venture capitalists will be forced to stop funneling money into unprofitable,
unsustainable AI companies which will make those companies unable to
funnel money into the pockets of anthropic and ope and
AI who rent the GPUs will then not be able
to funnel money into the pockets of those buying GPUs,
which will make it harder for those companies to justify
buying GPUs. And at that point some of this comes
(28:56):
to InVideo and Invidia doesn't make so make so much money.
I'm honest, none of Nvidia's success really makes any sense.
Who's buying so many GPUs and where are they going?
Why are in Vidia's inventories increasing? Is it really just
pre buying parts for future orders? Why are their accounts
receivable climbing? And how much product is Nvidia shipping before
(29:16):
it gets paid? While these are both explainable as this
is a big company and this is how companies do business,
and that's true, why do receivables not seem to be
coming down? And how long, realistically can the largest company
on the stock market continue to grow revenues selling assets
that only seem to lose its customers money and don't
(29:37):
seem to even be in use for years. I worry
about it in Nvidia, not because I think there's a
massive scandal, but because so much rights and its success
and its success rights on the back of dwindling amounts
of venture capital, and there because nobody is actually making
money to pay for these GPUs, let alone running them.
In fact, I'm not even saying in video goes tits up.
(29:58):
I want to be clear about that. I think they
may even have another good quarter or two in them.
It really just comes down to how long people are
willing to be stupid and how long Jensen Wong is
able to call up Sachin Adella and Co. At three
in the morning and say, buy one billion dollars of GPUs,
you pig Finnom style baby. But really, I think much
of the US stock market's growth is held up by
(30:20):
how long everybody is willing to be gas lit by
Jensen Wong into believing that they need more GPUs. At
this point, it's barely about AI anymore, as AI revenue
real cash made from selling services run on those GPUs,
doesn't even cover the costs, let alone create the cash
flow necessary to buy more seventy thousand dollars GPUs thousands
at a time. It's not like any actual innovational progress
(30:43):
is driving this bullshit. In any case, the market's crave
are healthy in video has so many hundreds of billions
of dollars of invidious stock sits in the hands of
retail investors and people's four O one ks, and its
endless growth has helped paper over the pallid growth of
the US stock market and by extension, the decay of
the tech industry's ability to innovate. Once this pops, and
(31:04):
it will pop because there's simply not enough money to
do this forever, there must be a referendum on those
that chose to ignore the naked instability of this era
and the endless lies that inflate the AI bubble. I
will be walking around with a gavel. I am going
to be taking heads. I am fucking sick of this era.
And what I'm most sick of is that so few
people are still to this day willing to admit how
(31:27):
bad this is. And I know in the next few
months we're going to get articles some major media outlets
that say, how could we have seen this coming? And
like I said in the previous episode, they could have
fucking looked. All of them could have looked, and they
could have looked a year ago. The incredible support I
get from all of you truly makes this show a
joy to make, even though I've done way too many
retakes on this and apologies to Mattasowski for the noises
(31:49):
I make. But I think in the next few months
we're all going to be validated. It's going to be
the great vindication. But until then, everybody is betting billions
on the eye idea that wily coyote won't look down.
He's gonna have to at some point, won't it. Thank
(32:13):
you 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
Mattasowski dot com, m A T T O. S O
W s ki 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,
(32:34):
my newsletter. I also really recommend you go to chat
dot Where's Youreed dot at to visit the discord, and
go to our slash Better Offline to check out our reddit.
Thank you so much for listening.
Speaker 1 (32:45):
Better Offline is a production of Cool Zone Media. For
more from Cool Zone Media, visit our website. Cool zonemedia
dot com, or check us out on the iHeartRadio app,
Apple Podcasts, or wherever you get your podcasts.
Host
Ed Zitron
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 Bobby Bones Show
Listen to 'The Bobby Bones Show' by downloading the daily full replay.