0:02

Zone Media. Try

0:06

spinning. That's a good trick. I'm

0:19

at Zetron. This is Better Offline and this is

0:21

the second episode of my two part series

0:23

where I explain how open ai has become a systemic

0:26

risk to the tech industry, even with its

0:28

massive forty billion dollar funding round and bird

0:30

brain benefactor in the form of soft Bank,

0:33

the world's foremost authority in losing money.

0:35

Now, before I continue, shameless request,

0:38

Better Offline has been nominated for a Webbie and I want

0:40

to win this thing. I've linked to it in my Twitter

0:42

and my Blue Sky and if you could vote for me, well, it will be in

0:44

the episode notes two. Moving on

0:47

back at it. Okay, all

0:49

right, open ai now has

0:52

forty billion dollars somehow, right

0:54

great? Right, Well, Dolgier horses. As

0:56

part of its deal with soft Bank, open Ai must

0:59

also convert its bizarre nonprofit structure

1:01

into a for profit entity by December twenty twenty

1:03

five, or it will lose ten billion dollars

1:05

from that forty billion dollars and of funding.

1:08

And just to be clear, by the way, they've only really

1:10

got ten billion dollars of that so far. The rest

1:12

is at the end of the year. Furthermore,

1:15

and the event that open ai fails to convert into

1:17

a for profit company by October twenty twenty

1:19

six, investors in its previous six point

1:21

six billion dollar funding route can claw back their

1:23

investment with a converting into a loan with an

1:25

attached interest rate. Naturally, this

1:28

represents a night mess scenario for the company,

1:30

as it will increase both its costs

1:32

and its outgoings. This is a

1:34

complex situation that almost warrants its

1:36

own podcast, But the long and short

1:38

of it is that open ai would have to effectively

1:40

dissolve itself, start the process of reforming

1:43

an entirely new entity, and distribute

1:45

its assets to other nonprofits or seller

1:47

license them to a for profit company at fair

1:49

market rates which they would not set.

1:52

It would also require valuing open ai as assets,

1:55

which in and of itself would be a difficult task,

1:57

as well as getting past the necessary state regulators,

1:59

the IRA state revenue agencies,

2:01

and the upcoming trial with Elon Musk Well,

2:04

that only adds further problems. I've

2:06

simplified things here, and that's because, as I've

2:08

said, this stuff is a little complex and pretty

2:11

boring. Suffice to say, this isn't

2:13

as simple as liquidating a company and starting a fresh

2:15

or submitting a couple of legal filings. It's

2:17

a long, fraught process and one that will be as

2:20

as has been subject to legal challenges,

2:22

both from open AI's business rivals as well

2:24

as from civil society organizations in California.

2:27

You may have heard the lost monologue. Based

2:29

on discussions with experts in the field of my own research,

2:31

I simply do not know how open ai pulls

2:33

off this by October twenty

2:35

twenty six, and honestly, I'm not sure how

2:38

they do it by the end of this year. It's

2:40

insane. It's a it's

2:42

a really I just

2:45

every time I read this stuff and

2:47

I write out, I'm like, how is nobody else

2:49

reading this and going, what the fuck is going on?

2:51

You See, this is a big problem, this nonprofit

2:54

thing, because open eye really

2:56

has become a systemic risk to the tech industry,

2:58

and anything that increases that is bad

3:00

news for everybody. Open Ai,

3:03

they've become a kind of load bearing company

3:05

for this industry, both as a narrative, as I've discussed

3:07

multiple times as chat GPT is the only large

3:10

language model company with any meaningful

3:12

use base, and also as a financial entity.

3:15

Its ability to meet its obligations and its future

3:17

expansion plans are critical to the future

3:19

health or in some cases survival of multiple

3:21

large companies. And that's before the after effects

3:24

that will affect its customers as a result of any

3:26

kind of financial collapse. The

3:28

parallels to the two thousand and seven the two thousand

3:30

and eight financial crisis are starting

3:33

to become a little worrying. Layman Brothers

3:35

wasn't the largest investment bank in the world, although

3:37

it was pretty big, just like open Ai

3:39

isn't the largest tech company, though again it's

3:41

certainly large in terms of alleged

3:44

valuation and expenditures. Layman

3:46

Brothers collapse sparked a contagient that would

3:48

later spread throughout the entire global financial

3:50

services industry and consequently the global economy.

3:53

Now I can see open AI's failure not

3:55

having as big an effect, but I

3:58

can imagine a systemic effect.

4:00

Still, you have to realize

4:03

that the whole AI trade, the

4:05

narrative, the bubble, it's holding up the economy.

4:08

I think like thirty thirty five percent of

4:10

the US stock market is in the Magnificent

4:12

seven, and all of their bullshit numbers

4:14

right now are held up by this nonsense. And

4:17

like the financial crisis, the impact

4:19

in this case won't be limited to just bankers and insurers.

4:21

It will bleed into everything else. This

4:24

episode is going to be a bit grim. I'm

4:26

not going to lie. I want to lay out the direct

4:28

result of any kind of financial crisis

4:30

at open Ai, because I don't think anybody is taking

4:33

this seriously. Let's start with

4:35

Oracle, who will lose at least a billion dollars

4:37

if open ai doesn't fulfill its obligations

4:40

per the Information. Oracle, which has taken responsibility

4:43

for organizing the construction of the stargate data

4:45

centers with unproven data center build

4:47

a Cruso and I quote the Information here, may

4:49

need to raise more capital to fund its data center

4:51

ambitions. Oracle has signed a fifteen

4:54

year lease with Crusoe, and to quote the Information

4:56

is on the hook for one billion dollars in payments

4:58

to that firm. To further quote information.

5:01

While that's the standard deal length, the unprecedented size

5:03

of the facility Oracle is building for just one customerment

5:05

makes its riskier than the standard cloud data

5:07

center used by lots of interchangeable customers

5:10

with much more predictable needs. According to

5:12

half a dozen people familiar with these types of deals.

5:14

In simpler terms, Oracle is building a giant

5:17

data center for one customer, open

5:19

Ai, and has taken on the financial burden

5:21

associated with them. If open Ai fails

5:23

to expand or lacks the capital to actually pay for

5:25

its share of the Stargate data center project, Oracle

5:28

is on the hook for at least a billion dollars, and based

5:30

on the Information's reporting, it's also on the hooked by the

5:32

GPUs for the site. This is me quoting

5:34

them again. Even

5:37

before the Stargate announcement, Oracle and open Ai

5:39

had agreed to expand their Abelene deal from two to

5:41

eight data center buildings, which can hold four hundred

5:43

thousand and VIDIA Blackwell GPUs, adding tens

5:45

of billions of dollars to the cost of the facility. In

5:49

reality, this development will likely cost tens

5:51

of billions of dollars, nineteen billion dollars

5:53

of which is due from open Ai, which does not have

5:55

the money until it receives its second tranche of funding

5:57

in December twenty twenty five from SoftBank,

6:00

and this is contingent partially on their ability

6:02

to convert into a for profit entity, which has mentioned

6:04

is extremely difficult and extremely unlikely.

6:08

It's unclear how many of the Blackwell GPUs

6:10

the Oracle has had to purchase in advance, but

6:12

in the event of any kind of financial collapse,

6:14

open Ai Oracle will likely have to

6:16

toss at least a billion dollars, if not several

6:19

billion dollars, and then we get

6:21

the core with a companies. So

6:23

a company whose expansion is likely driven entirely

6:26

by open Ai now and cannot survive

6:28

without open Ai flilling its obligations

6:30

if it doesn't die. O Anyway, Now,

6:32

I've written and spoken a lot about publicly

6:34

traded AI compute firm Core Weave, and it would

6:36

give me the greatest pleasure in my life. Never think or talk

6:39

about them ever again. Nevertheless,

6:41

I have to. This is my curse.

6:43

This is my curse. Core Weave has become

6:46

my curse every time I think about this. Fuck

6:49

Okay. The Financial Times revealed

6:51

a few weeks ago that core Weave's debt payments could balloon

6:53

to over two point four billion dollars a year by the end

6:56

of twenty twenty five, far outstripping its

6:58

cash reserves, and the Information reported that its

7:00

cash burn would increase to fifteen billion dollars

7:02

in twenty twenty five, as Bird's

7:04

IBO filing. Sixty two percent of core weaves

7:06

twenty twenty four revenue a little under two billion,

7:09

with losses amounting to eight hundred and sixty three

7:11

million was Microsoft Compute, and based

7:13

on the conversations I've had with sources, a good

7:15

amount of this was Microsoft running compute for open

7:17

Ai. Starting October twenty

7:19

twenty five, open Ai will start paying core Weave

7:22

as part of its five year long, twelve billion

7:24

dollar contract, picking up the option that Microsoft

7:26

declined. This is

7:28

not great timing, or maybe it's perfect timing,

7:30

because this is also when core Weave will

7:32

have to start making payments on their massive, stupid,

7:35

multi billion dollar d DTL two point

7:37

zero loan mentioned in previous

7:39

episodes. But really, there's a newsletter if you

7:41

want to you hear me, go mad, You want

7:43

to read me go mad? You read my read my core

7:45

with piece because it really drove me insane.

7:48

Nevertheless, these

7:50

core Weave payments, the ones from

7:52

open Ai to core Weave that October, they're

7:54

pretty much critical to Corewave's future.

7:57

This deal also suggests that open ai will become

7:59

core Weave's life just customer. Microsoft

8:01

had previously committed to spending ten

8:03

billion dollars on core Weave services by the end of

8:05

the decade, but CEO satchly Adella added

8:07

a few months later on a podcast that its

8:10

relationship with core Weave was a one time thing.

8:13

Man, man a really like really

8:15

fucking around there, such a don't love core Weave.

8:17

Assuming Microsoft keeps spending it its previous

8:20

rate, So about one point like sixty

8:23

six percent of two billion dollars whatever

8:25

those I have something that isn't guaranteed.

8:27

By the way, it would still only be half of open

8:29

AI's potential revenue to core Weave. Core

8:32

Weave's expansion at this point is entirely driven

8:34

by open Ai. Seventy seven percent

8:36

of its twenty twenty four revenue came from two customers,

8:38

Microsoft being the largest and yes, I just fucked up

8:40

a number at sixty two percent and using

8:43

core weaves auxiliary compute for open Ai.

8:45

As a result, the future expansion efforts the

8:47

theoretical one point three gigawatts have contracted,

8:50

and by the way, that means it doesn't exist. Compute

8:52

at core Weaver are largely, if not entirely, for

8:54

the benefit of open Ai. In the event

8:56

that open Ai cannot fulfill its obligations,

8:59

core weave will collapse. It's that fucking simple,

9:01

and then the shock waves will ripple further. In

9:04

Video relies on corewey for more than six percent

9:06

of its revenue and corwy's future credit

9:08

worthiness to continue receiving said revenue.

9:11

Well, much of that is dependent on open ai

9:14

continuing to buy services from core Weave now

9:16

and basically this in a comment I received from the legendary

9:18

Gil Luria, Managing director and head of Technology

9:21

research at Analyst DA Davison and Co, I

9:23

quote him, since cor We've bought

9:25

two hundred thousand GPUs last year, and those

9:27

systems are around forty thousand dollars. We believe

9:30

cor We've spent eight billion dollars on in Vidia

9:32

last year. That represents more than six

9:34

percent of Nvidia's revenue in

9:36

twenty twenty four, he said last year, but I've

9:38

just wanted to make it sound better. Core

9:40

We've receives preferential access to Nvidia's

9:42

GPUs, though Nvidia kind of denies

9:45

that and makes up billions of dollars of Nvidia's

9:47

revenue. Corweave then takes those GPUs

9:49

and then they raise debt using the GPU's

9:53

as collateral as well as customer

9:55

contracts. Then they use the money they've

9:57

raised to buy more GPUs from Nvidia. You

9:59

may think that doesn't sound right. I am

10:01

being complete, like this is factual information.

10:04

At this point in video was the anchor for Corwave's

10:06

IPO, and CEO Michael and Trader

10:08

said that the IPO would not have closed

10:10

without Invidia buying two hundred and fifty

10:12

million dollars worth of their shares. Nvidia

10:15

also invested one hundred million dollars in the early

10:17

days of core weaves, and for reasons I cannot

10:19

understand, also agreed to spend one

10:21

point three billion dollars over four years

10:23

two and I quote the information rent its

10:25

own chips from core weave fum

10:28

fact. I can't find a single fucking

10:30

mention of core weave in any of Nvidia's filings.

10:33

Now buried in that core weaves s one the document

10:36

every company publishes before going public

10:38

was a warning about counterparty credit risk, which

10:40

is when one party provides services or goods to another

10:43

with specific repayment terms and the other party

10:46

doesn't meet their side of the deal. While

10:48

this was written as a theoretical as it

10:50

could, in theoretically speaking,

10:52

come from any company to which core Weave acts as

10:54

a creditor. It only named one open

10:57

Ai now has discussed

10:59

previously. Core Weaver is saying that should

11:02

a customer, any customer, but really

11:04

they mean open Ai failed to pay its

11:06

bills for infrastructure built on their behalf or

11:08

services rendered, it can have a material

11:10

risk of the company. Now. As an aside, the information

11:13

reported that Google and someone's going to email me

11:15

there, so I just want to get ahead of it. The core Weave

11:17

is apparently in advanced talks with Google to

11:19

nd GPUs it Also, it also

11:21

added another thing in this story, just so that I don't

11:23

have to hear from any of you. The Google's potential

11:26

deal with core Weaves significantly

11:28

smaller than their commitments with Microsoft as according

11:31

to one of the people briefed on it, but could potentially

11:33

expand in future years. Do not come to me

11:35

and claim that Google's going to save core Weeve. I'll

11:37

be so mad anyway. Even with

11:40

Google and open AI's money, Carew's continued

11:42

the ability to do business hinges heavily on its

11:44

ability to raise further debt, which I have previously

11:46

called into question a newsletter that gave me madness,

11:49

and its ability to raise future debt

11:51

is to quote the financial times secured

11:53

against it's more than two hundred and

11:55

fifty thousand n vida GPUs and

11:58

its contracts with customers such

12:00

as Microsoft. Now,

12:02

any future debt that core Weave raises

12:04

will be based off of its contract with open Ai. You

12:06

know, the counterparty credit risk threat that represents

12:09

a disproportionate sharef it's revenue I just mentioned,

12:12

and also whatever GPUs they still have left

12:14

that they can get debt on. As a result, a

12:16

chunk of a video's future revenue is dependent on open

12:18

AI's ability to fulfill its obligations the core Weave

12:20

both in its ability to pay them, and they're timing less in doing

12:22

so. If open ai fails, then core Weave

12:25

fails, then that hurts and video Jensen's

12:27

going to have to go. He's going to have to go to a

12:30

cheaper leather jacketarium and

12:32

it gets worse. Open AI's expansion

12:35

is dependent on two unproven startups, one

12:37

of them I just mentioned, who are also

12:39

dependent on open ai to live with Microsoft's

12:42

data center pullback and open AI's intend to become

12:44

independent from redmen. Future data center expansion

12:47

is based on two partners supporting coll Weave

12:49

I know will get there

12:51

and Oracle. Now I'm referring, of course to

12:53

Core Scientific, which is the data center developer

12:56

for core Weave, and of course Crusos,

12:59

the data center develer four Oracle.

13:01

Now, if you were wondering, I can't hint it about this

13:03

earlier how many data center how

13:06

many DIIDENTA centers do you think Cruso's

13:08

ever built, and the answer is none,

13:10

And of course Scientific, how many do you think they've

13:13

built and the answer is also none. These are

13:15

the fucking companies underpending the AI

13:17

boom. I also really

13:20

must explain how difficult it is to build a data

13:22

center, and how said difficult he increases when you're

13:24

building an AI focused one. For example,

13:26

in Video had to delay the launch of its Blackwell GPUs

13:29

because of how finnicky the associated infrastructure,

13:31

so the servers and the cooling and such is for

13:33

customers. This was for customers

13:35

that had already been using GPUs

13:38

and therefore likely knew how to manage the temperatures

13:40

created by them. Also is another reminder

13:42

open eyes on the hook for nineteen billion dollars of funding

13:45

behind Stargate, and neither of

13:47

them have that money. I just want to remind

13:49

you of that, because it costs so much money to

13:51

build a fucking data center. And imagine

13:53

if you didn't have any experience and

13:56

effectively had to learn from scratch, how do you

13:58

think it would be building

14:00

these data centers? Let's

14:03

find out. So let's start in Abilene,

14:05

Texas with Crusoe and the Stargate Data

14:07

Center project. Now. Cruso is a former cryptocurrency

14:10

mining company that has now raised hundreds of millions

14:12

of dollars to build data centers for AI companies,

14:15

starting with a three point four billion dollar data center

14:17

financing deal with asset manager Blue Owl

14:19

Capital. This yet to be completed

14:21

data center has now been leased by Oracle, which

14:23

will allegedly fill it full of GPUs for open

14:25

AI. Despite calling itself and I

14:28

quote the industry's first vertically integrated

14:30

AI infrastructure provider, with the company

14:32

using flared gas as a waste by product

14:34

of oil production to power I infrastructure, Cruso

14:37

does not appear to have built a single AI data

14:39

center and is now being tasked with building one point

14:41

two Gigawatt's had a data

14:43

center capacity for open

14:45

Ai. It's just so fucking Cruso

14:49

is the sole developer and operator of the Abilene

14:51

site, meaning, according to the information that it is in

14:53

charge of contracting with construction contractors

14:56

and data center customers as well as running the data

14:58

center after it is built. Oracle,

15:00

it seems, will be responsible for filling said data

15:02

center with GPUs as mentioned. Nevertheless,

15:05

the project also appears to be behind

15:07

schedule. The Information reported in October

15:09

twenty twenty four that Abilene was meant to have fifty

15:12

thousand of Nvidia's Blackwell AI chips

15:14

in the first quarter of twenty twenty five, and also suggested

15:17

that the site was projected to have a whopping one hundred

15:19

thousand of them by the end of twenty twenty five. Now

15:22

you can join me back here in reality, because

15:24

a report from Bloomberg in March twenty twenty five

15:26

said the Open AI and Oracle we're expected

15:29

to have sixteen thousand available by the summer

15:31

of twenty twenty five, with and a quote open

15:34

Ai and Oracle expecting to deploy sixty

15:36

four thousand then video GB two hundreds

15:38

at the Stargate Data center by

15:40

the end of twenty twenty six. That's

15:44

that's very delayed. That's

15:46

really delayed. Again. How I run

15:49

a PR firm in that I record a podcast, I've

15:51

write a newsletter, I have a book. I'm

15:53

right in. I got all this shit on and I'm the asshole

15:55

who notices this anyway, has discussed

15:57

previously. Open ai needs this capacity very

15:59

bad. According to the information, open

16:02

ai expect stargate to handle three quarters

16:04

of its compute by twenty thirty and these delays

16:06

call into question, at the very least whether this schedule is

16:08

reasonable, or logical or even possible.

16:12

And I actually really question whether Stargate itself

16:14

is possible at this point. But it can

16:16

get dumber because we're about to talk about Core

16:18

Scientific, and they are core Weave's

16:21

friends. They're the people building data

16:23

centers for core Wave in Denton, Texas.

16:35

Now, as you can probably tell, I've written a great

16:38

deal about core Wave in the past. That got a monologue,

16:40

got a newsletter, and I got a therapy

16:42

bill for it. And specifically I've written

16:44

about their build out partner, Core Scientific,

16:47

a cryptocurrency mining company, yes, another

16:49

one that has exactly one customer

16:51

for its AI data centers, and you'll never guess who

16:53

it is. It's Core Wave. Now

16:55

here's a few fun facts about Core Scientific.

16:58

Core Scientific was bankrupt lotat year. Course

17:00

Scientific has never built an AI data center, and

17:02

its cryptocurrency mining operations were built

17:04

around a six specialist computers

17:07

for mining bitcoin, which led to an analyst

17:09

to tell CNBC that said data

17:11

centers would and I quote need to be buildozed and

17:13

built from the ground up to accommodate AI compute.

17:16

That's the stuff. Course Scientific

17:18

also does not appear to have any meaningful AI

17:20

compute of any kind. It's AI slash

17:22

HPC, which is high performance computing. Revenue

17:25

represents a teeny tiny, teeny little

17:28

percentage of overall revenue, which mostly

17:30

comes from mining crypto, both for itself

17:32

and other parties. Now,

17:35

hearing all of this, would

17:37

you give this company your

17:40

your compute? Would you think these are the people

17:43

that I am going to call to build

17:45

my data centers. If

17:47

you said no, you are smarter than Corewave,

17:49

who has given their entire one point three gigabat

17:51

build out to Core Scientific. Course

17:54

Scientific. Also, it seems they

17:56

seem to be taking on like one point one four

17:58

billion dollars of capital expenditures to build these

18:01

data centers, which, by the way, is not enough money. But

18:03

nevertheless, Corwave has promised to reimburse them at eight

18:05

hundred and ninety nine point three million of these

18:08

costs. This is all from public filings.

18:10

By the way, it's as one clear house course

18:12

Scientific actually intends to do any of this shit.

18:15

While they've taken on a good amount of debt in the

18:17

past five hundred and fifty million dollars in the convertible

18:19

note towards the end of last year, this would be

18:21

more debt than they've ever taken on. It

18:24

Also, as with Crusoe, does not appear to

18:26

have any experienced building AI data centers,

18:28

a point I keep repeating because it's very important.

18:30

These are other companies behind the growth for open AI, except

18:33

unlike Crusoe, Core Scientific

18:35

is a barely functioning, recently bankrupted bitcoin

18:38

minor pretending to be a data center company. Crusoe,

18:40

on the other hand, is possibly also doing

18:43

the same thing, but less egregious about it. Now,

18:46

how important do you think core Weave is to open

18:48

ai? Exactly? Well, that's our semaphore. Core

18:51

Weaver has been one of our earliest and largest

18:53

compute partners, open ai chief Sam Allman

18:55

said in Corwave's roadshow video, adding that Corwave's

18:58

compute power led to the creation of some of

19:00

the models that were best known for Core. We

19:02

figured out how to innovate on hardware, to innovate

19:04

on data center construction, and to delve for results,

19:07

very very quickly did it.

19:10

But even if it did, will it survive

19:12

long term? Going back to the point

19:14

of the contagion. If open ai fails and core

19:16

we fails, so too does course Scientific

19:18

And I don't really fancy Crusoe's

19:21

chances either. But let's take a step

19:23

back for a moment. We've been going so hard,

19:26

haven't we. I've got a genuine question, just

19:28

for the fact finders out there. There's

19:31

a Microsoft book. Open AI's computer's

19:33

revenue now. Up until fairly recently, Microsoft

19:36

has been the entire infrastructure backing open

19:38

ai, but recently to free open ai

19:40

up to work with Oracle and see other people, released

19:43

it from its exclusive cloud compute deal. Nevertheless,

19:45

put the information, open ai still intends to spend

19:48

thirteen billion dollars on compute on Microsoft

19:50

as there this year. What's confusing,

19:52

however, is whether any of this is booked as revenue

19:54

for Microsoft. Microsoft claimed earlier in the year

19:56

that it surpassed thirteen billion dollars in annual recurring

19:59

revenue, by which it means it's

20:01

last month multiplied by twelve by the way, and

20:03

they said it was from ai. Open AI's

20:06

compute costs in twenty twenty four or five billion

20:08

dollars, and that's at are discounted as your

20:10

rate, which on an anualized basis will

20:12

be about four hundred and sixteen million

20:14

dollars in revenue a month for Microsoft. It

20:17

isn't, however, clear whether Microsoft counts

20:19

open ai is computer's money, which is really

20:22

fucking weird. You'd think with all this

20:24

money they're making from this company, they'd be saying

20:27

there was money coming in. It's peculiar.

20:30

I've yet to find a real answer. Now.

20:32

Microsoft sernings do not include an artificial

20:35

intelligence section. No, They're made up of three separate

20:37

segments, Productivity and Business Processes,

20:39

which includes things like LinkedIn, Microsoft

20:41

three sixty five and so on. More Personal

20:43

Computing, which includes Windows and gaming products,

20:46

and then Intelligent Cloud including server products

20:48

and cloud services like A zero, which is likely

20:50

where open AI's computers included, and where

20:53

Microsoft book the revenue from selling access

20:55

to open AI's models but not open

20:58

AI's compute question. As

21:00

a result, it's hard to say specifically where

21:03

open AI's revenue might sit. Even guessing

21:05

intelligent Cloud might not be right, but based

21:07

on an analysis of Microsoft's Intelligent Cloud

21:10

segment from financial year twenty twenty three

21:12

Q one through its most recent earnings, and there

21:14

was a spike in revenue from twenty three

21:17

Q one to twenty four Q

21:19

one. In financial year Q one,

21:21

which ended on September thirtieth, twenty twenty two,

21:23

a month before Chat GPT's launch, the segment

21:26

made twenty point three billion dollars.

21:28

The following year, in FY twenty

21:30

four Q one, it made twenty four point three

21:32

billion dollars, a nineteen point

21:35

seven percent year of a year growth, or roughly

21:37

four billion dollars. This could represent

21:39

the massive increase in training and inference costs

21:41

associate with hosting Chat GPT, and

21:44

they peaked at twenty eight point five billion

21:46

dollars in revenue in the financial year twenty

21:48

four Q four, before dropping dramatically

21:50

to twenty four point one billion dollars in

21:53

financial year twenty five Q one, and raising

21:55

a little twenty five point five billion dollars

21:57

in financial year twenty five Q two.

22:00

I'm so sorry none of this is easy

22:02

to read. This is a plausible explanation.

22:05

Open ai spent twenty twenty three training its

22:07

GPT four to Roho model before transitioning

22:09

to its massive, expensive iryme model, which would

22:11

eventually become GPT four point five, as

22:13

well as its video generating model

22:15

sourer. According to the Wall Street Journal,

22:18

training GPT four point five involved at

22:20

least one training run, costing around half a billion

22:22

dollars in compute costs alone. These

22:24

are huge sums, but it's worth noting a couple of things.

22:27

First, Microsoft licenses open AI's models

22:29

the third parties, so some of this revenue could

22:31

be from other companies using GPT on Azure.

22:34

We've seen lots of companies launch Ai products,

22:37

and not all of them are based on l

22:39

lambs mrieing things. Further, Microsoft

22:42

provides open ai access to a zerr cloud services

22:44

at a discounted rate, as I've mentioned in the past,

22:47

and so there's a giant question mark over open

22:49

AI's actual contribution to the various spikes

22:52

in revenue for Microsoft's Intelligent cloud

22:54

segment, or whether other third parties played a significant

22:56

role. Furthermore, Microsoft's investment

22:58

in open Ai isn't entirely cold hard cash.

23:01

Rather, it's provided the company with credits to be redeemed

23:03

on in see your services, kind of like Chucky

23:06

cheese tokens. I'm not entirely

23:08

sure how this would be represented in accounting

23:10

terms, and if anyone can shed any light on this, please get

23:12

in touch. Would it be noted as revenue

23:14

or something else? Open Ai isn't

23:16

paying Microsoft or are they? Are

23:19

they doing the tech equivalent of redeeming air

23:21

miles or have they spent a gift card of us?

23:23

You're it really isn't obvious, and Microsoft

23:26

is doing some accounting bullshit here and

23:28

not suggesting impropriety, not suggesting

23:30

anything illegal. I'm just saying it's

23:32

insane that they have this company spending

23:35

billions of dollars theoretically

23:37

on their services and it's

23:39

just nowhere. Additionally, while

23:42

equity is often treated as income for tax purposes,

23:44

as is the case when an employee receives

23:46

RSUs as part of their compensation package,

23:49

under the existing open Ai structure, Microsoft

23:51

isn't actually a shareholder, but rather the owner of

23:53

profit sharing units. This is a distinction

23:56

worth noting. These profit sharing units

23:58

are treated as analogous to equity, or

24:00

at least in terms of open AI's ability

24:02

to raise capital, but in practice they aren't the same

24:04

thing. They don't represent ownership in a company

24:06

as directly as, for example, a normal share

24:09

would. They lack the liquidity of a share in

24:11

the upside they provide, namely dividends is

24:13

purely theoretical. Another key difference.

24:15

When a company goes bankrupt and enters liquidation,

24:18

shareholders can potentially receive a share of the

24:20

proceeds after creditors, employees, and so on

24:22

are paid. Well, that often doesn't happen

24:24

as is, as in the liabilities generally,

24:27

they can exceed the assets of the company. In many

24:29

cases it's at least theoretically

24:32

possible, given that profit sharing

24:34

units aren't actual salaries or shares.

24:37

Where does that leave Microsoft? This

24:39

stuff is confusing, and I'm not ashamed to say that I just

24:41

fucked up a word and that complicated

24:43

accounting questions like these are far beyond my understanding.

24:46

If anyone can shed some light, drop me an email,

24:48

buzz me on Twitter or blue sky, hit me up on clerk

24:50

or gorp or post on the better offline

24:53

subware. Someone might take your wallet

24:55

though. Anyway, back on track, I think

24:57

it's worth understanding the scale of the

24:59

open air vortex and how it's distorting the

25:01

tech investment market and why, even

25:03

without having failed, it represents the systemic

25:06

risk. Without open ai, and American

25:08

startup investment is flat, and even with it, less

25:10

startups are receiving investment. Crunch

25:13

based News reported in early April the North American

25:15

startup investments spiked in Q one due to open

25:17

Ai, hitting eighty two billion dollars. Great,

25:20

right, sounds great? This statement

25:22

sadly has a darker undertone. American

25:24

star up investment was actually like forty two

25:26

billion in Q one twenty twenty five when you remove

25:28

the deal, which is appropriate because none of the money

25:31

is actually received by open Ai yet and

25:33

at best only ten billion dollars if it will be received

25:35

before Sember twenty twenty five. This

25:37

quarter also included a three point five billion

25:40

dollar investment in Wormlight competitors

25:42

Aroundthropic run by Warrio Amma Day, making

25:44

the appropriate number of paltry thirty

25:47

nine point five billion dollars. Now,

25:49

this is still an improvement, though a marginal

25:51

one over the thirty seven point five billion dollars raised

25:54

in Q one twenty twenty four. Nevertheless,

25:56

crunch based news also has a far, far darker

25:59

story. Your volume women in American

26:01

startups has begun to collapse, trending downward almost

26:03

every quarter or deal volume isn't

26:05

the direct result of open AI's financial condition.

26:08

The so called revolution created by open ai and

26:10

other generative AI companies technology appears

26:12

to be petering out, and the contagion is starting

26:14

to impact the wider tech sector. It's

26:17

important to understand how bleak things are the future

26:19

of generative AI, wrest and open ai, and open

26:21

AI's future rests on near and possible financial

26:24

requirements. I've done my best

26:26

to make this argument is in as objective

26:28

a tone as possible, regardless of my feelings about

26:30

the bubble and its associated boosters open

26:33

Ai. As I've said before and argued

26:36

countless times in interviews and podcasts

26:39

and newsletters, it's effectively

26:41

the entire generative AI industry, with

26:43

its nearest competitor being less than five percent

26:45

of its five hundred million weekly active

26:47

users Anthropic, Google,

26:50

Microsoft XAI. They're all rounding

26:52

errors in the grand scheme of things. But

26:54

open AI's future is dependent and this

26:57

is not an opinion. This is an objective

26:59

fact on effectively infinite

27:01

resources in many forms. Let's

27:04

start with the financial resources. If open

27:06

ai required forty billion dollars to continue operations

27:08

this year, it's reasonable to believe it will need

27:11

at least another forty billion dollars next year, and

27:13

based on its internal projections, will need at least

27:16

forty billion dollars every single year until

27:18

twenty thirty, when it claims somehow it will be profitable.

27:21

And I quote the information with the completion

27:23

of the Stargate Data Center project, you

27:26

may be wondering, how's that possible?

27:28

Ed? How you think the

27:30

information wrote that down? Fuck no, just

27:32

go Lesson's too busy humiliating people.

27:35

She let go by name on Twitter,

27:37

Ess Golesson. I like the information. I think you're

27:39

a fucking asshole for how you treated your people.

27:41

Say it on my podcast, I say it on Twitter. Anyway.

27:44

Let's keep talking about some of these resources

27:46

that open ai is dealing with, specifically

27:49

the compute resources and expansion. Open

27:52

ai requires more compute resources than anyone

27:54

has ever needed, and will continue to do so in perpetuity.

27:57

Building these resources is now dependent on two

27:59

partners, Course Scientific and Crusoe.

28:01

Though I've never built a data center, as Microsoft

28:03

has materially pulled back on data center development

28:06

and has as aforementioned,

28:08

pulled back on two gigawads of data

28:10

centers, slowed or paused. Of course, some

28:12

of its early stage center products too,

28:15

with TD Cohen's recent analyst

28:17

reports saying that data center pullbacks were and I

28:19

quote them March twenty six, twenty twenty

28:21

five, data center channel checks letter because

28:24

it's so good, driven by the decision

28:26

to not support incremental open Ai training

28:28

workloads, that's the stuff. In

28:31

simpler terms, open ai needs more computer

28:33

at a time when it's lead backer, which has the most GPUs

28:35

in the world, has specifically walked away from building

28:38

it. Even in my most optimistic

28:40

frame of mind, it isn't realistic to believe that Cruso

28:43

or Core Scientific can build the data centers necessary

28:45

for open AI's expansion, even

28:47

if soft Bank and open Ai had the money to invest

28:49

in Stargate today, which they do not. Dollars

28:52

do not change the fabric of reality. Data

28:55

Centers take time to build, requiring concrete

28:57

would steal in other materials to be manufactured

28:59

and placed, and that's after the permitting required

29:01

to get these deals done. Even if that succeeds,

29:04

getting the power necessary is a challenge unto itself,

29:06

to the point that even Oracle, an established

29:08

and storied cloud compute company run by

29:11

a very evil man at one point to quote the

29:13

Information, has less experience than

29:15

its larger arrivals in dealing with the utilities,

29:17

to secure power and working with powerful and demanding

29:19

cloud customers whose plans change frequently.

29:22

A partner like Cruso will Core Scientific simply

29:24

doesn't have the muscle, memory, or domainer expertise

29:26

that Microsoft has when it comes to building and operating

29:28

data centers. As a result, it's hard to imagine,

29:30

even in the best case scenario, that

29:33

they are able to match the hunger for compute the open

29:35

Ai has now. I want to be clear,

29:37

I believe open ai will still continue to use

29:39

Microsoft's compute and even expand further

29:41

into whatever remaining compute Microsoft may

29:44

have. However, there is now a hard limit

29:46

on how much of that there's going to be, both

29:48

literally and what's physically available, and in

29:50

what Microsoft itself will actually allow open

29:52

ai to use, especially given how unprofitable

29:55

GPU compute seems to be based

29:57

on how every single company that isn't

29:59

in vidio lose his money running them. But

30:16

really, and we're coming to the end

30:18

of this, which leads to a question,

30:21

how does all of this end? Last

30:24

week, a truly offensive piece

30:27

of fan fiction framed as a report called

30:29

AI twenty twenty seven went viral, garnering

30:31

press with the Duiskesh podcast and gormles

30:34

childlike wonder from Dope New York

30:36

Times reporter Kevin Rus and reporter

30:39

I think is a fucking stretch. Its

30:41

predictions vaguely suggest a theoretical

30:43

company called open Brain will invent a self

30:46

teaching agent of some sort. It's

30:48

total bullshit, but it captured the handsome minds

30:50

of AI boosters and other people without object

30:52

permanence because it vaguely suggests that somehow,

30:55

large language models and their associated technology

30:57

will become something entirely different. Like

31:00

making predictions like these because the future, especially

31:03

in our current political climate, is utter

31:05

chaos. But I will say that I do not

31:07

see, and I say this with complete objectivity,

31:09

how any of this bullshit continues. I

31:12

want to be extremely blunt with the following

31:15

points, as I feel like both members of the media and

31:17

tech analysts have categorically failed

31:19

to express how ridiculous things have become. I

31:22

will be repeating myself, but it's fucking

31:24

necessary, as I need you to understand

31:26

how untenable things are. Soft

31:29

Bank is putting itself in dire straits simply

31:31

to fund open Ai once this

31:33

deal threatens its credit rating, with soft

31:35

Bank having to take on what will be multiple loans to

31:37

fund this forty billion dollar round, and open

31:40

Ai will need at least another forty billion dollars

31:42

a year later. This is before

31:44

you consider the other nineteen billion dollars

31:46

that soft Bank has agreed to contribute to the data center

31:49

project with Stargate money it does not currently

31:51

have available. Now. Open

31:53

Ai has promised nineteen billion dollars to the Stargate

31:56

Data Center project two and again they

31:58

do not have it, and they need soft Bank to

32:00

give it to them. And again

32:03

I've said it, and I'll say it again. Neither

32:05

of these companies have the money.

32:08

The money is not there, and

32:11

open Ai needs Stargate to get

32:13

built to grow much further. I

32:16

see no way in which open ai can

32:18

continue to raise money at this rate, even

32:20

if open Ai somehow actually

32:22

receives the forty billion dollars it's been

32:24

promised, which will require it to become

32:26

a for profit entity, which I don't think it can fucking

32:29

do. While it could theoretically stretch

32:31

that forty billion dollars to the last multiple years,

32:33

projections say it will burn three hundred and twenty billion

32:35

dollars in the next five years, or

32:38

more likely, I can't see a realistic

32:40

way in which open ai gets the resources it needs

32:42

to survive. It will need this insane

32:45

streak of good fortune, the kind of which you only really

32:47

hear about in Greek poems or

32:49

JoJo's Bizarre Adventure. You know, the more cultured

32:51

choice, but let's go through them. Somehow

32:54

soft Bank gets the resources and loses the constraints

32:56

required to bankroll this company forever.

32:59

The world's wealthiest entities, those sovereign

33:01

wealth funds mentioned in the last episode, sounds

33:03

and so on, they pick up the slack until open Ai

33:05

receive they reach profitability,

33:09

which is a huge assumption. It's

33:11

also assuming that open ai will have enough of these

33:14

megawealthy benefactors to provide it with the three hundred

33:16

and twenty billion dollars they need to reach profitability,

33:18

which it won't. There'll also need Cruso and

33:20

Core Scientific to turn out to be

33:22

really good at building AI infrastructure,

33:24

which they've never done before, which

33:27

is that's very possible, I'm sure. And

33:30

then Microsoft will then walk back its walk

33:32

back on building UAI infrastructure

33:34

and recommit to tens of billions of dollars

33:37

of CAPEX, specifically on AI data centers,

33:39

and also will give it to open Ai. And

33:41

then, of course Stargate's construction happens

33:43

faster than expected and there are no supply chain

33:45

issues in terms of labor, building materials, GPUs

33:48

and so on. Now I don't

33:50

know, I haven't checked the news in the last three weeks,

33:53

but is there anything going on that might increase the

33:55

costs of materials? Probably

33:58

not. Anyway, if those things happen, I'll

34:01

eat quo. I'm not particularly worried.

34:06

In the present conditions. Open ai is on course to

34:08

run out of money or run out of compute

34:10

capacity, and it's unclear which will

34:12

happen first. But what is clear is

34:14

it's time to wake up. Even

34:17

in a hysterical bubble where everybody

34:19

is agreeing that this is the future, open ai is

34:22

currently requiring more money in more

34:24

compute than is reasonable to acquire. Nobody

34:28

nowhere, ever, anywhere,

34:31

has ever raised as much money as open ai

34:33

needs to. And based on the sheer amount

34:35

of difficulty that soft Bank is having raising

34:38

the funds to meet the lower tranche, the ten

34:40

billion dollar one of its commitment, it may

34:42

not actually be possible for this company

34:44

to continue, even with the extremely

34:47

preferential payment terms months long

34:49

deferred payments for example that open

34:51

ai probably has. At some point someone will

34:53

need a dollar. I'll give

34:55

Sam Ortman some fucking credit. He's found

34:58

many partners the shoulder the burden of the rock economics

35:00

of open Ai. With Microsoft, Oracle,

35:02

Crusoe and Core We've handling the upfront costs

35:04

of building the infrastructure, and SoftBank finding

35:07

the investors for its monstrous stupid round,

35:09

and the tech media mostly handling marketing for

35:11

him, which is really nice, great job everybody.

35:14

He is. However, overleveraged. Open

35:17

Ai has never been forced to stand on

35:19

its own two feet or focus on efficiency,

35:21

and I believe the constant enabling of this

35:24

ugly nonsensical burn rate has doomed

35:26

this company. Open Ai has acted

35:28

like it'll always have more money. In compute, And

35:32

that's kind of because everyone's acted as

35:34

that would be the case. No one's really called sam

35:36

Altman out on his bullshit. There are some people,

35:39

but really no one in the mainstream media has bothered.

35:42

Really, Sam Altman has been enabled.

35:45

Open Ai, by the way, cannot just make things

35:47

cheaper at this point, because the money has always been

35:49

there to make things more expensive, as has

35:51

the compute to make larger and larger language

35:53

models that burn billions of dollars a year. This

35:57

company is not built to reduce its footprint

35:59

in any way, nor is it built for a future

36:01

in which it wouldn't have access to infinite

36:03

resources. Worse still, investors

36:05

in the media have run cover for the fact that these models

36:07

don't really do much more than they did a year ago,

36:09

and for the overall diminishing returns of large

36:12

language models writ large. Now, I've

36:14

had many people attack my work about open

36:16

ai, but none of them, not one of

36:18

them. Nobody has provided

36:21

me any real counterpoint to the underlying

36:23

economic argument I've made since July of last

36:25

year, the open ai is unsustainable.

36:28

Now this is likely because there really isn't one

36:30

other than open ai will continue to raise

36:32

more money than anybody has ever raised in history

36:35

imperpetuity and will somehow turn

36:37

the least profitable company of all time into

36:39

a profitable company. This is not

36:42

a rational argument. It's a religious one.

36:44

It's a call for faith. And it's disgusting

36:47

to see well paid reporters

36:50

with one hundred and fifty thousand

36:53

subscribers to the newsletters and a

36:55

really shitty podcast with a major news

36:58

outlet constantly just ignore them share

37:00

and I see no greater payal horse of the apocalypse

37:03

than Microsoft's material pullback on data

37:05

centers. Well, the argument might be that Microsoft

37:07

wants open ai to have an independent future. That's

37:10

fucking laughable when you consider Microsoft's

37:12

deeply monopolistic tendencies, and for that

37:14

matter, it owns a massive proportion of

37:16

open AI's pseudoequity. At one point,

37:18

Microsoft's portion was valued at forty nine

37:20

percent, and while additional fundraising has likely

37:23

diluted Microsoft's steak, it still owns

37:25

a massive portion of what is, at

37:27

the very least, if

37:29

you believe any of this nonsense, the most

37:31

valuable private startup of all time. And

37:34

we're supposed to believe that Microsoft's pullback,

37:36

which limits open AI's access to infrastructure

37:38

it needs to train its and run its models, and

37:40

thus is mentioned represents an existential threat

37:42

to the company. You meant to believe that this is because

37:44

of some sort of paternal desire to see

37:46

open ai leave childhood behind

37:49

to spread its wings and enter the real world.

37:51

Are you fucking stupid? Sorry?

37:55

I shouldn't be calling people stupid. I shouldn't.

37:58

I really shouldn't. But I am more

38:00

likely Microsoft got would have needed out of open Ai,

38:03

which has reached the limit of the models that can develop,

38:05

and which Microsoft, by the way, already owns

38:07

the ip of due to their twenty nineteen funning

38:09

round. There's probably no reason for Microsoft

38:11

to make any further significant investments other than

38:13

just kind of throwing a little cash in there, and then I

38:16

imagine some sort of tax dodge. I'm just guessing. It's

38:19

also important to note that absolutely nobody

38:21

other than Nvidia is making any money from generative

38:23

AI. Core Weave loses billions of dollars,

38:26

open Ai loses billions of dollars, Anthropic

38:28

loses billions of dollars. And I can't find a single fucking

38:30

company providing generative AI powered software

38:32

that's actually making a profit. The

38:35

only company is even close to doing so Are

38:37

Consultancy is providing services to drain and create

38:40

data for models like Churing and Scale AI, and

38:42

Scale isn't even fucking profitable now.

38:44

The knock on effects of open AI's collapse will

38:47

be wide ranging. Neither core Weave nor

38:49

Crusoe will have tenants for their massive, unsustainable

38:51

operations. An oracle will have nobody

38:53

to sell compute to because they've

38:55

leased that thing for fifteen fucking years, or one

38:57

customer who else is going to take that

39:00

anyway. Cor we will likely collapse under the weight of

39:02

its abominable debt anyway, which will lead to a six

39:05

seven percent or more revenue drop for in video or

39:07

a time when revenue growth has already begun

39:09

to slow. On a philosophical

39:12

level, too, open AI's health is what keeps

39:14

this industry alive. Open ai

39:16

has truly the only meaningful user base

39:18

in generative AI, and this entire hype cycle

39:21

has been driven by its success. Meaning any

39:23

deterioration or collapse of open ai will

39:25

tell the market what I've been saying for over a year.

39:27

The generative AI is not the next hype of growth

39:29

market and it's underlying economics do not

39:31

make sense. But

39:34

look, I'm

39:36

not saying this to be a hater. I'm

39:38

not saying this to be right. This

39:40

stuff has driven me insane, but

39:43

I'm not doing it to be a pundit,

39:45

to be a skeptic, to be a cynic, to be someone

39:48

that hates because I want to hate. And I hate

39:50

them not because I think people

39:52

like me because I hate them. I hate them because I have brainworms.

39:55

I have something wrong with me inside

39:58

my brain that tells I have

40:00

to be like this, and I have to look at these things

40:03

and I have to try and find

40:05

what's going on, otherwise I will be driven

40:07

mad, which is why I'll say if

40:09

something changes, if I'm wrong somehow,

40:12

I promise you I will tell

40:14

you exactly how, exactly why,

40:17

and what mistakes I made to come to the

40:19

conclusions I have in this episode

40:21

and the episodes before. But

40:24

I don't believe that my peers in the media will do the

40:26

same when this collapses. But I promise

40:28

you that they will be held accountable

40:30

because all of this abominable waste

40:33

could have been avoided. Large

40:35

language models are not on their own

40:37

the problem. The tools capable of some

40:39

outcomes, doing some things, But the problem, ultimately

40:42

are the extrapolations made about their abilities

40:44

and the unnecessary drive to make them larger,

40:47

even if said largeness never really amounted

40:49

to much. Everything that

40:51

I'm describing is the result of a tech industry,

40:54

including media and analysts, that refuses

40:56

to do business with reality, trafficking

40:58

in the ideas and ideolo, celebrating

41:01

victories that have yet to take place, applauding

41:03

those who have yet to create the things that they're

41:05

talking about, cheering on men lying

41:07

about what's possible so that they can continue

41:10

to burn billions of dollars and increase their wealth and

41:12

influence for barely any fucking reason.

41:15

I understand why others might not

41:18

have said what I've said. What

41:20

I am describing is a systemic failure,

41:22

one at a scale here too unseen, one

41:25

that has involved so many rich and powerful

41:27

and influential people agreeing to ignore reality,

41:30

and that'll have crushing impacts for the wider

41:32

tech ecosystem when it happens. Don't

41:36

say I didn't warn you. Thank

41:45

you for listening to Better Offline. The editor

41:47

and composer of the Better Offline theme song is Matasowski.

41:50

You can check out more of his music and audio projects

41:53

at Mattasowski dot com m

41:55

A T. T O S O w

41:58

Ski dot com. You

42:00

can email me at easy at better offline dot

42:02

com or visit better offline dot com to find

42:04

more podcast links and of course, my newsletter.

42:07

I also really recommend you go to chat

42:09

dot where'soead dot at to visit the discord,

42:11

and go to our slash Better Offline to check

42:14

out I'll Reddit. Thank you so much

42:16

for listening. Better Offline

42:18

is a production of cool Zone Media. For more

42:20

from cool Zone Media, visit our website

42:22

cool Zonemedia dot com, or check

42:25

us out on the iHeartRadio app, Apple Podcasts,

42:27

or wherever you get your podcasts.