Every AI company's unprofitable. The

largest customers of AI compute are

unprofitable. Their customers also

unprofitable. OpenAI spent $34 billion

to make 13.07 billion. Their net loss is

just a little under 21 billion. The

amount of revenue to substantiate and

justify the data centers that are

allegedly under construction means that

we need two or three more OpenAI and

anthropic sized customers that do not

exist. Never assume there's a plan.

Anthropic and open AI are people like

well they must have a plan. I don't

think they do. Got a very special

conversation. I am speaking to one of

the most prolific skeptics about AI. I'm

joined today by Ed Zitron author of

Where's Your Ed at Newsletter and the

Better Offline podcast. Ed, welcome to

Monetary Matters.

>> Thanks for having me.

>> Ed, how about you start off and lay out

your view of artificial intelligence and

what the return on investment is for the

enormous sums that are currently being

spent. So trillion plus dollars in the

fact we're still debating the ROI kind

of says everything. I think if this was

a real industry with the kind of TAM

that they've been selling us on for the

past four years, we wouldn't have that

debate. There wouldn't be one. The fact

that we're having it, the fact that you

have people to this day in the year of

our Lord 2026 saying, "Well, AI is

real." That I'd never heard something

real where someone has had to insist

upon that. Economically speaking, the

vast majority of AI companies barely

make more than hund00 million a year in

revenue. Anthropic and Open AI account

for, I think, 89% of all the top. And on

top of that, they're all horribly

unprofitable. They lose billions of

dollars. And the only way that they can

reconcile those is by doing some uh what

I would call wacky accounting. And so,

in my reporting from earlier in the

week, OpenAI spent $34 billion to make

13.07 07 billion and yeah their net loss

is just a little under 21 billion. They

are of they have of course with the FT

given comment that suggests that they

only lost 8 billion. I think that's

laughable. I think that that was uh

booster key jingling. I think that that

was put in there because they have to

find a way to reconcile with the fact

that they told CNBC they lost $8 billion

at the beginning of the year when it's

very clear they didn't. like put aside

whatever fanciful quotes there are they

they lost $21 billion.

>> How much by the way you know has Uber or

Amazon lost when they were losing tons

of money and before they became the

giant behemoth and and how does that

compare to the as you say $21 billion

that OpenAI lost last year?

>> Okay. So Uber I believe burned like $32

billion. They're now kind of like messy

gap profitable. A lot of that was sales

and marketing and a lot of that was

subsidizing rides. But the level of

subsidies they did were just completely

different. The Amazon Web Services

example is the really egregious one.

Between 2003

and 2017, I think it is normalized for

inflation. Amazon Web Services was maybe

5355 billion total. And that's all of

Amazon's capex, not just AWS. Just to

give you some context, if OpenAI closes

all the money they've been promised this

year, they will have raised $122 billion

in the last 6 months. If Anthropic

closes all the money they've been

promised, they will have raised $95

billion in the last 6 months. OpenAI

raised $40 billion last year. Anthropic

raised $16.5 billion. The answer is

incomparable. Like Amazon's capex for

its retail operation on top of AWS was

less than half of what OpenAI has raised

in one round this year. Like it's just

there's no real comparison. And people

making this comparison are looking for a

means of rationalizing the irrational

and of finding a way to not reconcile

with the fact that the two largest

companies, the pretty much the only

demand for AI compute between anthropic

and open AAI,

those two companies don't make sense.

They don't make economic sense. There

the only way to make them make economic

sense is to just ignore your lying eyes.

It sucks because

I I think that the tech industry has

been poisoned with a very specific kind

of ideology. And I think the irony is so

many of them are atheists,

stern hard rationalists, but they have a

quasi religious attachment to artificial

intelligence. Every minor sign is proof

that the great prophecy is true. And

every little point that could possibly

attack a skeptic is considered

unilateral proof that they're wrong. And

it sucks because they're trying to dump

these companies on retail investors who

are going to be the victims of this

theological hype cycle.

>> I think you're exactly right. I mean, it

was a senior executive co-founder of

Google who said that they would rather

go bankrupt than lose an AI. They're

committed.

>> Yeah. And all of that I think comes from

have this theory called the rockcom

bubble which is tech is out of ideas.

They're out. When I say that I don't

mean they literally have none. I mean

they have no hyperrowth ideas. They

don't have a new AWS. They don't have a

new iPhone or a new smartphone. They

don't have a next cloud computing. And

AI was meant to be the panacea. It was

meant to be AI models as in via the API

were meant to be the thing that created

the next generation of startups that

created the next generation of

enterprise bolt-ons so that the software

industry could grow. It was meant to be

the future of consumer software and it

was even meant to be the next AWS in the

form of AIGUs which is why they've sunk

so much money into it. I think that the

hyperscalers Meta put aside I don't

think Meta really has a strategy. I

think it's just like spend money until

Mark gets bored. But Microsoft, Google,

Amazon, I think they saw this as an

opportunity to create the next

generation because they don't have

anything else. There is no other

business line that shown anything close

to the possibility of having the growth

just just talking pure growth. And the

worst thing is is that while they've

seen some growth from AI, it's mostly

because Anthropic and Open AAI spend an

alarming amount of money on compute,

which means that they're basically just

feeding their money back into themselves

and then feeding their own cash flows

into Nvidia or Broadcom or one of the

Taiwanese ODM. So, Honghai, Foxcon, the

companies that build the servers in

Taiwan. So, it's rough. It's very rough

because for me to be wrong, there needs

to be hundreds of billions of dollars of

AI compute demand. Put aside Open AI and

Anthropic. Hundreds of billions of

dollars because there are hundreds over

100 gigawatts of data centers planned.

12.5 to 15 million per megawatt.

We the the compute demand does not

exist. And the only companies that are

spending that much are these

unprofitable AI companies which are

having their gaping holes plugged by

hyperscalers. It's just it's very rough

and I don't see any path for it to to

normalize. It hasn't been normalized and

there are tons of signs things are

actually getting worse. One core uh

claim that you said I just want to

revisit is that the losses experienced

by the two largest labs Anthropica and

open a AI right now especially open AAI

are unlike anything we've seen in

history on an on an annualized basis.

Um, so as you said, if we can get into

the $21 billion in losses in 2025 and

I'll also add something that leads to

another core claim of yours, which is

that, you know, it's no secret that like

Silicon Valley has funded startups that

are initially unprofitable, but the

reason that they some of them work out

and work out tremendously well is

because they're software type businesses

that have low costs and in particular

they have low variable costs. And you've

pointed out that that is not true at all

about open AAI that they are

tremendously expensive not just to

launch and start but to operate and

maintain.

>> Yes. And one other thing with those glad

you brought up variable costs. That's

actually a big problem both for the

customer and for the AI companies

themselves, the labs. So the information

reported that OpenAI's margins actually

worsened. didn't specify in the piece

how much worse in 2025 because they had

to get lastm minute compute because it's

kind of you don't just say okay I'm

going to buy this much compute and then

if I need more well I can just do make

do you have to expand because inference

on its own is so compute intensive it's

not training that's taken up it's if

they have an influx of customer demand

which doesn't mean money it just means

if some customers have variable if some

customers because if are using codecs

more that will take up more GPUs so they

have to buy lastm minutee compute buying

compute up front much cheaper than

buying it on the spot and especially

because they need so much of it

whoever's selling it to them naturally

has an advantage and can charge them

more so you have all of these variable

costs and also that Silicon Valley I

don't want to call it mythology because

it's happened the losses are just much

smaller like snowflake for example

snowflake is not a profitable company,

but it's not horrifying. It's something

that can keep going. They have some

degree of venture debt. They can keep

plugging away. They will eventually make

it. I don't think Snowflake's going to

be an incredible margin business, but I

think it will get there. This is not

like that. And on top of that, no one

has a path. No one has been able to

explain. We are now approaching Vera

Rubin. We've had Blackwell for a year or

so. Where's the cheapness happening?

Whatever happened to those OpenAI asex?

Anthropic has been using TPUs for years.

They're not working out their costs

either. There is just multiple things

that just don't make sense. And the only

reason it's continued is this kind of

cult like this worship of capital. This

cult-like belief that just if we if we

believe in this enough, all of the bad

stuff will go away.

That that doesn't really work at this

scale. There's not really there's no

government bailout opportunity here.

There's not really any there's at some

point going to be a limit to how much

capital you can sink into this.

Anthropics last round they had like 30

people in or 30 entities invest. It's

just things don't things don't look good

all over the shop and I'm I've really

looked for good signs and I can't find

them.

>> Tell me what you found when you you got

access to Open AI's financials for 2025.

What just how big are the losses? is I

know there's various ways of measuring

them and then there's you know been some

push back from people who are kind of on

the other side of this and I want to get

your your feedback to that.

>> Sure. So revenue was 13.07 07 billion.

Cost of revenue, which is a very

interesting way of referring to costs,

which I'll get to in a minute, $7.5

billion. R&D, $19.1 billion. Sales and

marketing,

$5.73 billion. General and admin, $1.57

billion. So, you get total cost and

expense about $34 billion. The loss from

operations about 20.9 billion, roughly

21 billion. So that sales and marketing

cost is the thing I want people to

really pay attention to because it's

very very strange. So I went and dug

around. Openai did their first major ad

campaign in September 2025.

>> Go on.

>> I'm thinking the thought that um you

probably had when you were you're making

this point of just like yeah that is

sales and marketing. That is a very high

number for sales and marketing.

>> It's more than Coca-Cola spends.

>> Like it's more than Coca-Cola's annual

budget. I know that OpenAI does Facebook

advertising, Reddit advertising, but if

they spent, let's just say, $1.5 billion

dollars on Facebook in a year, they

would be one of the largest advertisers

on Meta, they would be a material

customer that would likely get called

out somewhere. Couldn't find any

evidence of that. What I can however say

is that buried in the back of these

documents is the specific itemized

related cost to Microsoft and SoftBank.

Now, $527 million of their sales and

marketing costs went through Microsoft.

Now, do you think that Microsoft is

doing a bunch of advertising for OpenAI

or is it the free credits that Microsoft

is giving away via the Microsoft for

startups program much like OpenAI for

startups? And on top of that, I think

they're putting free PE free inference

in there. But my biggest evidence by the

way that they're doing that is that I

reported last November that OpenAI spent

8.67 billion dollars on inference on

Microsoft Azure. I reported that and

that was also validated by the FT. But

nevertheless, this sales and marketing

cost I think is really interesting.

Maybe it's where OpenAI puts the cut

that Microsoft takes from selling its

models. Maybe it's the inference cost

that Microsoft charges when they sell

their models through Azure AI. I truly

don't know. But what I can say is based

on my previous reporting is that the

cost of revenue is not all the

inference. It's definitely not. I'm very

confident in my sourcing from November

and also what the hell are they spending

all this sales and marketing on if it's

not that? It's like it's such a weird

category and I sadly don't have a

segment breakdown. I wish I did. So the

some percentage of the 5.7 billion in

sales and marketing you think ought to

actually be attributed to cost of

revenue which would make gross margins

way less a lot lower than they would be.

>> So and to kind of elaborate on that I

think the way they're getting around it

is they're saying their free customers

are in there which is weird because

I I've seen the information suggest that

their margins include say like their

free users. So, who knows? But it's very

clear that this cost is inflated by

something and it's also linearly

increasing with revenue. So, it's kind

of like where is this going? It's either

them giving away credits and then

putting the inference cost in there,

which actually would make some sense or

it's their free users or it's some

Arnold Palmer of the two cuz it's not

advertising. It's they I think I saw a

reporting that said they had 500

salespeople in 2025. Want to be clear,

if you paid each of them a quarter of a

million dollars, that'd be about $125

million, which is uh not even like like

a third of 10 less than third of the 10%

of this cost. It's bizarre. And you kind

of see similarly inflated sales and

marketing costs when you look at Zepu

and Minim Max, the two Chinese AI labs

that went public last year, I think.

Both of which are their cost of revenue

is lower than their revenue, but their

sales and marketing plus their cost of

revenue is not. And it's very

interesting because this is probably

legal under GAP. I'm sure that they

found I'm sure Sarah Fry is a smart

person. I think she probably wouldn't

have something like this unless it was

defensible. But nevertheless,

this is not a company that is

approaching profitability and to

consider their gross margin without

sales and marketing is just kind of

laughable because that much like the R&D

cost is not going away.

>> I think it's a really important point

also as you said that some percentage of

these include Microsoft credits. Now,

when the FT quoted so-called, you know,

so-called people familiar with the

matter who said actually the true losses

we should say are more like 8 billion, I

believe it's the case, correct me if I'm

wrong, that they netted out a ton of not

just stockbased compensation, but cloud

computing credits from Microsoft, which

if that is true, in my opinion, that is

a real cost. Like if I get billions of

dollars worth of free stuff, I think it

it would be honest for me to include

that as a cost because it's you can't

normally just count that as oh my god,

normally yeah, normally I'm going to get

tons of free stuff, you know, that just

that's just not how it works.

>> So the documents that I've seen do not

mention those credits. There is no

mention of credits. They do mention sal

um stockbased compensation though. And

there what's interesting with that is

there's stockbased compensation of I'm

doing this in memory I'm definitely not

reading something of about

let's see uh $6.4 billion but there's

sharebased compensation for compute

provided by a related party of $1.2

billion.

How strange.

How strange are they trading stock for

compute? That's also weird because

that's the thing. The point you just

made is very salient which is if they

are getting compute credits and those

compute credits are lowering their costs

that is still an expense and it's also

not a permanent one. Is Microsoft going

to permanently feed them credits or are

these I think it I think those credits

are the ones left over from the $13

billion that Microsoft invested in 2023.

Semaphore said that was mostly in

credits. So perhaps it's that the

accounting does not say that I also just

I don't I actually think whoever made

that if that person is somewhat

listening go [ __ ] just like seriously

like eight that is not what you lost

dude you lost nearly $21 billion. You

can dance around all you want and be

like well when you move these numbers

around and do this and do that well it's

eight now. The only people you're doing

that for are Twitter posters. Actual

like actual investors will be able to

see through this. It's very silly to do

this because when the S1 comes out,

they're going to have to define things.

They're going to have to explain what

these things mean at least a little bit

more. And it's just it's frustrating

seeing that quote because the FD does

good reporting, but it was so blatantly

obvious that someone from OpenAI, I

assume, I truly don't know who it is or

where they're from, was going to try and

feed that in there. But the only way in

which OpenAI lost 8 billion is in I

don't know the La La Land of their

accountancy. And I just I think it's

offensive to just about everybody

involved that they would say this

because it's not what happened. It's it

if the FT took that number seriously, it

would be in the headline. It's not. The

thing to focus on is the actual costs.

And here's the thing, they're going to

run out of tokens from from Azure at

some point. And now I also can confirm

it's not in the story. They spent at

least a billion dollars on Oracle last

year renting H100s and H200s I believe

and whatever exists of Abolene. Larry

Ellison isn't giving them any tokens.

That that's going to cost them real

dollars. Coreweave who they paid net

360, they're not taking tokens either.

They're going to need dollars.

And while they might be able to do this

fuzzy moon math for one year or two

years I guess with 2024

2026 they've signed up with Amazon

they've signed up with Corey with

Cerapus with uh Google who is renting

capacity from Coree to sell open AI they

have all of these partners now that they

can't really dance around with credits.

So I think I can't wait for this S1. I

really cannot wait. I just I think it's

very exciting that we're going to see

more of this and I think it's also

laughable that anyone is seriously

saying they lost $8 billion is the COPE

Olympics.

>> In your your piece there there's some

numbers you cite that are even larger

than the $21 billion loss from

operations. So the net loss attributable

to the company of 38.5 billion um and

then something even bigger the 60.51.

I will definitely say those numbers are

inflated by the conversion from a for to

a from a nonprofit to a for-profit,

which is why I'm very much more focused

on the 21 billion number. But something

bizarre is going on with OpenAI's

accounting. Just going to say this

because this net loss attributable to

non-controlling members capital magic

away $3.7 billion worth of costs in

2024. Where did they go?

Where'd they go? I've spoken to a few

accountants since then. They all say the

same thing, which is, yeah, that still

cost them something. They just put it to

a subsidiary. I guess it's very weird.

It's all It's all so bizarre. I've never

seen anything like this before. I wonder

and I want to say I don't know if I'm

right on this if um so when a company

goes public like an IPO or a spa deal

and it has these warrant liabilities and

it's basically what the company owes to

people who own the warrants when the

stock price goes up the warrants that

they owe increase in value. So these on

a gap thing looks like a a and from

accounting sense it is a net loss. Um,

but it's not a true in my view like

economic loss. So I wonder to what

degree it's it's like that. And that's

I'm referring to what the uh 2025

>> the 2024 is in the article though just

to be clear.

>> No, no, I read I read it. I just I just

Yeah.

>> Yeah. But just to be clear, I agree the

the warrant cost those are non-cash.

Like there's I would never I put 38.5

billion in because that's what it says.

But just to be clear, the cash lost is

lower than that. That I've been quite

clear about. The reason I bring up the

2024 one is their loss from operations

in that year was $ 8.7 billion, 8.78

even in 2024, but it became 5.098

billion because well, you know, those

costs go somewhere else. It's very

weird. I thought for a second it could

be cloud credits. That might make sense,

but it's not clear what entity that

could be. Now, things get messier in

2025 because of the revaluation, because

of the the conversion. I get that.

There's nothing we can do. It's you

can't really pass out much from that.

Nevertheless, they did magic away 17.8

billion of costs, which is nice.

I do wonder how this is going to look in

the S1 though, whether it's going to be

the same, how they're going to find ways

to finagle or finesse this. I'm really

don't have much clarity there and the

economics sorry the uh statements don't

really break down anything about like

the exact breakdown of where that might

be going or what entities might be. The

thing to note though that they do and

this is in the article as well is when

they break down that open AAI spent

about $17 billion on Microsoft Azure.

That is that's a that is a large chunk

of Microsoft's revenue that's coming

from a company that's going to run out

of money. That is very like a large part

of their IPO RPOS even 250 billion plus

is coming from OpenAI a company that

cannot afford to pay them. And even in

2025,

astonishing amount of money just really

like Microsoft's revenue is inflated by

OpenAI. That is now fact. That's very

bad. That's very bad indeed because

there was reporting in 2025 that

OpenAI's compute costs were sold at

cost.

That was uh with the information. Now,

that referred to A100 GPUs. So maybe

it's not the same with H100s and B200s

and all that, but happened at one point.

And it's just if this was a real

industry with the kind of things that

were trending in the right direction,

there wouldn't be so many asterisks.

There wouldn't be so many weird things.

You'd be able to say with your whole

chest, "Wow, what a profitable company.

Look how well it's growing. Look at look

at the way look at its triumphant march

towards profitability.

Yeah, it's it's not doing that. And in

fact, I'm not sure what the plan is for

this company given the most recent news

as well.

>> So, Ed, there are people who are

probably bulls in AI believers in I who

may be watching this and say, Ed, if we

were to wind the clock back to November

2025, basically everything you'd said,

particularly about OpenAI, is true. They

were losing an enormous amount of money.

However, in December 2025, the models

got way better, in particular,

Enthropic. And since then the revenue

has seen a huge upsurge. What is your

reaction to that argument?

>> Well, you know what also increases with

OpenAI's revenue? Its costs. D. So fun

fact about OpenAI's revenues as of the

beginning of this year. So Sam Alman

himself said that the uh it was a huge

issue for their customers how expensive

things got. So what happened was at the

beginning of the year everyone was still

subsidized models. So they've subsidized

subscriptions. Then I think in March,

OpenAI started along with Anthropic

moving people to tokenbased billing.

This created a massive burst of revenue.

And now I'm sure you've seen all the

different conversations about ROI. Uber

burned through their entire AI budget in

3 months. I think it was they're putting

caps on their engineers.

Nothing about that changes the thesis at

all. In fact, the one bit of evidence we

have most strongly is the costs increase

with the revenue. The more the more they

burn, the more they make, the more they

burn. That's been happening consistently

across the board. People will say,

"Well, look, the cost of goods isn't

going up. Sales and marketing increased

over 400% year-over-year. It's the

fastest growing category in this

company." funnily enough, OpenAI has

been giving away $1,000 of codeex API

credits to anyone, anyone who has a

business. Anthropic is doing the same

thing with Claude Code, by the way. So,

even if they had a burst of revenue,

they also had a burst of cost. And now

they're going to have customers who are

already pulling back on spend. Also, the

Wall Street Journal reported that OpenAI

is considering, and I quote, drastic

price cuts.

That's not something you do when

customers are showing anything other

than an intent to churn. And I think

again, come back to it. If this was

working, you'd be able to point to it

and just say it's working. They would

they wouldn't have to do all this kind

of uh three card monty stuff with the

finances. And I think that I actually

think they're going to end subsidized

subscriptions. I think that that we're

getting to a point where there's just no

e economic point other than marketing,

which again makes me wonder what that

sales and marketing cost is. It's just

it's very it's frustrating arguing

against this because there's a great

deal of

there's a great deal of the arguments

against me that are just nu and it's

like come on mate like you can you can

only say Nvidia super cycle and so many

times before that actually has to

happen.

>> What about Anthropic?

>> I think they're in the same place. So,

Anthropic, funny company. Wall Street

Journal story came out about a month ago

that said, "Oh, they're profitable in

this quarter." And it was because Elon

Musk sold them Colossus's Compute and

gave them a discount for the exact two

months that they were profitable. And

they were profitable by like a couple

hundred million. It's a $1.25 billion a

month contract. The math is pretty

obvious. But Anthropic has the same

things as OpenAI. their revenue

increases, their cost of goods sold

increases, and their sales and marketing

increases. It's the same thing. I

genuinely think sales and marketing is

the the slop trough to put the cost that

they don't want to put on the top line.

I think it should still be considered

the cost. I don't think that that

changes very much, but it exists only to

beguile the easily beguiled to convince

the people that want to be convinced

that this is all going well. The problem

is is also they both of them have

actually increased their sales and

marketing costs. On top of this,

Anthropic has one of the most aggressive

influencer campaigns I've ever seen.

Both have increased their ad spend to an

indeterminate level. I see them all the

time on my subreddit. So, I think

Anthropic is in much the same position.

And Anthrop sorry, OpenAI was doing

those drastic price cuts because they

believe anthropic will do the same. So

it's AI needs to keep accelerating and

it's already kind of slowing down and I

think OpenAI and Anthropic are basically

the same company. I think they run they

run in similar ways. I think Dario Amade

and Sam Wman are similarly faguous

people. I don't think either of them

believe in very much. Both of them are

terrifying BS. At least Jensen Huang's

funny. At least he's At least when he

gets mad about stuff, you're kind of

like, "Oh, oh, oh, is he gonna is he

going to smack this college student?" At

least Hawk Tan has a kind of like

emanating aura from him that's

terrifying and horrible. You have to

hear goddamn Sam. Everyone Sam and Daru

Ammedday, they all say the same thing.

They're all doing the same song and

dance. You'll notice that both of them

brought up recursive self-improvement

recently because they're both giving up

on coming up with ideas. They're like,

"Our idea is the machine will come up

with the idea." And of course, jingle

jingle to the AI boosters. AI that

trains itself. We have no proof that

this will happen, but yay, we can repeat

this and be part of the club. It's a

shame because only so much can be done

on hype and hope.

>> What do you make of the claims of ARR

for anthropic from $9 billion annualized

recurring revenue to 14 billion in 2026

to what a month ago they said it was 42

44 billion or you know over over 40? I

>> believe it was 47

>> in this in their series H announcement.

So the information sync streamer update

over there fantastic reporter reported

that the way anthropic calculates its

annualized run rate which is not ARR. I

made this I've made this mistake too.

ARR refers to annual recurring revenue

with stable contracts. But putting all

that aside annualized they calculated by

taking that day's subscribers

and multiplying them by 12 and the last

four weeks of API spend and tsing it by

13. So the problem with this is API

calls and model spend is not a recurring

expense.

It's it's not perhaps it's something you

can extrapolate from. Perhaps it's

something you can say yeah we got this

much. Perhaps you have contracts with

people that say they have to spend X

amount in a month in a year even or in

within the three-month period. I'm not

party to their contracts. But yeah, that

number can be manipulated real easy. For

example, Axios's Madison Mills reported

that 500 mill someone spent $500 million

in the space of a month on Claude on

because they did not set up well

somebody I mean an enterprise company

spent that because they didn't set up

spend controls. If you use that times 13

mathematics that's $6.5 billion in

annual run rate for a cost that will

never come back. we are they were

measuring based on the token maxing era

that is coming to an end. This is the

problem with run rate. It's a it's a

problematic measurement and it's very

weird and kind of telling that the only

two hyperscalers who have ever talked

about AI revenue are Microsoft and

Amazon and both of them only use run

rate because run rate is a snapshot. You

can have a $47 billion run rate, then

API calls drop and suddenly you don't

have the same thing. But because that's

out there, people will think, "Wow,

they're going to make $47 billion in the

year." When I think the Wall Street

Journal reported they made $4.6 billion

in Q1 and they're on course to make a

little over $10 billion in Q2, which are

large amounts of money. They are not 47

billion. In fact, that what they would

have to they would they would have to

keep growing at a remarkable rate that

is not going to happen to get to 47

billion in the year. In fact, there are

plenty of signs that things are slowing

down. If they're talking cost cuts, then

they absolutely are. They've seen some

there's something they saw that's

genuinely bothered them.

>> Shout out Madison Mill. I uh I know her

and that's that's great to hear you.

>> No, she she's f she is fantastic. She's

awesome. So tell us about the push back

from to token maxing and what do you

think that means for real real revenues

how you measure them

>> to set the scene the only way for these

companies to grow and like I said $1.1

trillion in compute commitments for open

and anthropic combined I think that's

through 2013 for them to actually make

good on that to actually pay their

contracts they have to keep growing and

the only way that that can happen is

through selling direct access to the

models to enterprises because they're

not growing that based on consumer

subscriptions or even enterprise

subscriptions. You're not you're not

doing that kind of rapid growth which

will mean I think OpenAI projects to be

at $284 billion in revenue by 2030.

Anthropic at 174 I think by 2029.

That's not happening unless they can

charge just everyone on tokenbased

billing. On top of that, they need

everybody to be spending more and more

and more. They need to get more

customers and those customers need to

massively increase. The problem is is

that everyone is being all the

executives of all these companies have

been saying use AI as much as possible

without making sure that there's a

measurable return on investment and

there isn't. In fact, it's quite

difficult to measure the actual cost of

an AI task. You because across different

models, different prompts, different

harnesses if I mean the jump from 4.7 to

4.8 with Opus changed how the models did

stuff and you pay regardless of whether

they screw up or not. So, all these

organizations went token crazy. Uber

burned through their entire budget in

three months. uh Zillow, who I reported

on, burned through their entire cursor

budget through uh for the year by the

end of May. And now everyone's pulling

back. Everyone's doing limits. I think

Brex I reported out I think there's 1.5

or 2K per engineer

and then like five bucks a week for

non-engineers. Uber's 1,500 limit for

engineers. Uh I hear data bricks is

still letting people go nuts bananas,

but good luck on that oie. Uh, I think

that what's happening is organizations

are pretty poor at measuring

productivity in general, except when

they've had to measure it before, it

wasn't coupled with a multi-million

dollar monthly cost. It wasn't suddenly

this massive aberrative cost explosion.

And I think what they're doing now is

they've gone from no cost controls to

cost controls. And I think those cost

controls get more. I think they start

crushing a little bit because you've got

open- source models coming

also just it's hard to measure the ROI.

So say you cut from a million dollars to

500 grand a month. How do you know that

500 grand a month's good? How are you

measuring that? Some people are saying

lines of code that's an insane way to

measure how good that's it's just not a

good measurement of productivity. It's

just a measurement of how much you've

shipped. Uh, Prest, same deal. With

Zillow, who I reported on a few weeks

ago, I think it was something like they

increased the amount of review hours for

human beings by like thousands and

thousands of hours a month. They just

added instead of replacing humans, they

just added more labor for their other

humans.

And so, it's it's a situation where I

don't see how that transforms into the

kind of rocket ship growth they need.

And I must be clear, however you feel

about the current state of these AI

companies and their revenues, what

they're doing today is nowhere near

close to what they need, they need to be

rocking within a couple years, 15 20

billion a month. They need to that that

needs to happen because if it doesn't,

they cannot afford their compute. And I

don't think they can raise enough money

and I don't think they're magically

becoming profitable. So it's how does

that work? People are going to say, "Oh,

custom silicon hasn't happened. Just

it's not happening." Which when when's

that going to happen? How long do I have

to wait?

>> You're say people would say custom

silicon is going to make it cheaper.

>> Yeah, they've been saying it for years.

I mean, another weird thing as well,

back in October last year, Broadcom and

Open AI said that they were going to do

10 gigawatts of AI data centers

together. I don't think OpenAI's ordered

a single chip from them yet. What

happened there? That's weird.

That's really That was just a weird AMD

said they were going to do six gigawatts

with Open AI. Just never happened. SK

Highix and Samsung said they were going

to sell them 900,000 wafers of RAM a

month. That also didn't happen. It's

almost as if lots of this isn't real.

Ed, what do you make of Agentic AI and

the claim that in the future a lot of

work is going to be done by AI agents?

So you know if 5 years ago you know in a

job of a lawyer or a tax accounted

profession you know you know would be

completely done by a human that in the

near future you are going to have

computers who are working for that

person sending emails buying things from

maybe from other agents u um doing work

you know setting all the numbers aside

which we've talked about do you think

that fundamentally that vision is is

right or wrong

>> I think it's wrong I think the

everything you're talking about there

involves multiple different

deterministic functions. Large language

models cannot do those. You cannot rely

on a large language model to replicably

do something. You change a hardness, you

change a model, you mess up a prompt, or

it just misreads a prompt because it

hallucinates, which is mathematically

certain. That's not going to happen.

Anything with money, that's not going to

happen either. I think I've seen six

different companies say that agents can

buy with them. I can't find evidence of

a single dollar being spent. I've heard

people doing it with their open claw.

Again, no evidence that's actually

happened. On top of the incredible

expense of doing this, which is not

going anywhere. It's also just not

happening. We're not getting science.

We're not getting science that's I've

talked to people who use Harvey or Lora

or what have you and I it sounds like

they're a rapper for large language

models. It sounds like it's just you can

feed law stuff in and we've got some

custom pro prompt engineering that makes

it do stuff. Perhaps that's useful,

perhaps it's not. I've yet to speak to a

person who's super excited about it or

even likes it. The people that I hear

going on about those products are always

goddamn partners. They're partners at

the top of the law firm who aren't doing

the grunt work that associates do.

Accountancy. Yeah, just no on that one.

just don't think that I think that

there's probably a small crop of people

that get something out of chat GPT

looking at a PDF. I don't think that

scales to a business that replaces

people. And also just Agentic AI is this

it's like the beginning of PeeWee's big

adventure. It's the breakfast machine.

It's all these cobbled together

deterministic scripts to try and not

these goddamn models into doing

something. do doing something never

seems to leave the realm of coding with

any seriousness. So when people say,

"Well, in the future, Agentic AI," I'm

just kind of like, "I'm sorry. I'm not

going to fill in the gaps for these

companies." And I mean this to the

boosters who might listen. You're being

conned. You're debasing yourself because

it's one thing. You can believe this

will happen, but the fundamental proof

is not there. And the fundamental proof

not being there means that to basically

say that this is going to work out just

involves saying don't believe your lying

eyes. You are doing the exact thing you

are being propagandized to.

And I I think it's because there are

people who are like emotionally invested

in the success of this. I actually don't

think the majority of them are

financially invested. And it sucks

because these people are mocks. They

have been they have been conned. You can

be I have no problem with anyone being

excited about LLMs. I truly don't. My

problem is the way they're being sold,

the way they're being lied about. The

way that people are talking about these

things is actively misleading. You want

to do ondevice stuff, go nuts. I I think

ondevice is the future of this stuff. I

think if it hangs out, which I am still

think is an open question, I think it

becomes this very specialist ondevice

software engineer tool, which could be

kind of cool in a decade. That could be

interesting. Good for them. I think that

the era of GPUbased LLMs will die. And I

and even if it doesn't even if it

doesn't for a while at least the amount

of revenue to substantiate and justify

the data centers that are allegedly

under construction means that we need

two or three more open AI and anthropic

sized customers that do not exist. I

mean just I don't mean that as like

anything about the efficacy of LMS or

anything. I just mean on a raw spend

level. We need hundreds of billions of

dollars of AI compute revenue by 2030.

And we need it to exist because if not,

we're going to have theoretical millions

of H200s, GB200s.

Well, I guess it' be MVL 72s, but

nevertheless, we're going to have racks

of these things that are sittingow. And

I also think the customers are all

unprofitable AI companies. So yeah, I I

don't hate on this stuff because I'm

like, "Yay, good. I get to be angry at

something. I get to criticize

something." It's because I think people

are being misled. I even think that the

booster types are being misled. I feel

bad for them on some level because it's

like, why are you angry at me? Be angry

at the companies for being so dodgy.

Like that's if if you don't like what

I'm saying, be angry at them. Be angry

at them for not giving you better

evidence. Because that's that's the

thing. If if it wasn't going if if these

companies weren't dodgy, why do they act

so dodgy?

>> Earlier you said that the you know the

bull market in compute the fact that

demand for comput seems off the charts

that that is wrong. Why do you say that?

>> Because the majority of AI compute

revenue is either anthropic, open AI or

meta or hyperscalers giving compute to m

uh anthropic and open AI. So, let's go

through them. So, Meta, I just want to

be clear, does not have an AI strategy.

Mark Zuckerberg cannot be fired. Yeah.

It's just like talking over what Meta is

doing is just like talking about a

friend with a drug problem. Like, it's

just like you, Mark, you got you got to

get off the compute, mate. You got to

get clean. But let's talk about

anthropic for example. $330 billion of

the remaining performance obligations

for Microsoft, Google, and Amazon are

anthropic. just that's like they're not

selling very much compute to anyone

else. Coreweave's revenue is principally

either Microsoft for open AI, Nvidia's

backs stop, a $6.3 billion back stop, or

Google for OpenAI or Anthropic or Meta,

of course. Iron is being hired by

Microsoft at Nvidia. What would they

what could Microsoft possibly be doing

with that compute other than selling it

to OpenAI? uh Cipher is I think it's

Cipher Mining and Terraolf both had

loans backstock by Google to build data

centers for fluid stack for Anthropic.

Um let's see uh Nebius $17 billion deal

with Microsoft. Why is that compute

going to be used? I've already confirmed

it's going to be used for open AI. This

is the story across the board. When you

peel away the non-open AI,

non-anthropic, non-meta compute revenue,

it's like a billion or two. Like just

because most people don't need that much

inference or training. Just put aside my

bare case for a second, just on very on

a very basic level, the demand and

natural need for AI compute is not that

high. Like it's just not there. And

people's arguments might be, well, we'll

use more of this in the future. The ROI

conversation is actually pushing back

against that. It kind of suggests that

actually people are people are kind of

hesitant with the costs. Well, they

could move to open source models. Great.

Those things seem to use less compute.

So, what we going to do with all those

data centers? And no one really has a

compelling answer to this. And the thing

is the largest customers of AI compute

are unprofitable. their customers who

are the ones pushing them as in the AI

startups who use those APIs also

unprofitable. Every AI company's

unprofitable.

So this entire industry's revenue

appears to be a test of how long venture

capital and debt can hold it up. And now

Broadcom is backto stopping $30 billion

of a $35 billion deal for and god this

thing's so stupid

to borrow money that goes into a joint

venture that buys TPUs from Broadcom who

then sells them to Google who then rents

them to Anthropic and Anthropic pays the

lease on them.

So again, gets back to the larger point

of this industry wasn't dodgy. Why is it

acting dodgy? Like these are not the

things that you do when there's real

demand or where there's tangible demand.

And I mean you've what base 10 raised $

1.5 billion. If the demand existed at

the scale that should theoretically be

be happening, B would be worth way more

than that because the inference demand

would be so obvious. There are no real

signs that this demand is coming either

because I don't know who else is there.

Who else needs it other than Mark

Zuckerberg?

>> Tell me about Meta's AI strategy. I when

I say I don't understand it, I'm not

trying to insult it, you know, by by way

of being polite. I literally don't

understand it. And you know, I asked

Gemini about Meta's AI strategy. And

other than, oh, we're going to make our

ads more effective, which of course, you

know, they lit literally, one of the

things they said is a health application

that's going to charge $7.99 per month,

and that's not going to justify.

>> So, one thing I like to my my advice I

give basically anyone is never assume

there's a plan. Like that's like

anthropic and open AI, people like,

well, they must have a plan. I don't

think they do. I think they thought,

which I kind of I don't like it, but I

get it's like, let's throw as much money

at this and it should work out. It

didn't, but I see the strategy with

Meta. Mark Zuckerberg is a capricious

man. He moves from idea to idea. He

moved on from the metaverse. He changed

the name of the company to Meta. And

then a year later was like, sorry, two

years ago was like, "Nah, mate. I'm

going to do AI now." But the thing is,

no one's really had much of an AI

strategy. Google and Microsoft and

Amazon kind of had an obvious one.

Amazon built infrastructure. Microsoft

build infrastructure and bolt AI

services that people hate onto other

products. Google, same deal. And Google

had TPUs already. I think Google's

probably the best positioned if only

because so much of their silicon is

their own. Don't know if it's as good,

but anyway. Meta, on the other hand,

well, their first strategy was to buy as

many H100s and H200s as they can find,

then put them in data centers. Then they

put them in their gem model, their

generative advertising model, which lots

of people have you tried to extrapolate

to say, well, that means that Meta is

using all these GPUs for ad targeting

kind of, but not in a way that I think

is drastically increasing profits

because I don't know if it was, we're a

year or two into this. They got all

these bloody GPUs. Wouldn't their growth

be astronomical? It wouldn't be it would

be like 50% year-over-year because the

power of AI. What I think happens is

they got incremental improvements out of

there like a couple percentage points

here and there. There are various blogs

about GEM that suggest that that's the

case but when you actually look at them

it's like okay they had a lift in

engagement here and here how does that

actually where's that end up on the cash

flow statement? Like where where's the

money from that? What I think happened

was Mark Zuckerberg saw everyone else

doing something and decided to do it

too. He bought all the compute and he

went great now we'll train our own model

and so they did actually something

pretty cool which was Llama the open-

source model they then realized wait

crap this is open source we can't charge

people for that so they actually the

information reported this last year they

went around the major hyperscalers being

like hey can you give us money for llama

and then the hyperscalers said no it's

free like you you made this free why why

would we pay you for that and so meta

spent spent $14 billion to bring in

Alexander Wang who then decided to do

internal models which are an

indeterminate level of good. I have

heard from a source that they are doing

so they have this plan to do an open

claw style thing called hatch.

>> Okay.

>> And they're going to do like a pendant.

The reason I'm listing all these things

out likely listeners are going to say,

"Wow, that sounds like a bunch of

disconnected ideas just kind of thrown

together. That is Matt's AI strategy."

because that's meta. This is actually

all in line with the company's history.

This is a company that has not had a new

idea since Facebook.

Instagram, they bought stories, they

stole from Snapchat. Reals, they stole

from Tik Tok. Every idea they've had,

they've stolen. All of this is to say is

that Meta doesn't have an AI strategy.

They've reorganized their AI department

four times, maybe more. And so

everyone's just kind of chasing their

tail at this point there. It's so

strange. I think Meta will be the last

man standing in AI. I think that

>> I don't think they have any reason to

stop. I think that Zuck has at this

point overcommitted

many times over the scale AI

acquisition, whatever you call it,

insane acquisition. Just what are you

doing? Mark Zuckerberg can't be fired

also. So there's not really any pressure

the board can put on him if the market

kills the stock if just something

massive changes maybe. But my greater

theory that I mentioned earlier is when

these companies stop doing AI which is

the reason that they put so much money

in and the reason that they have not

stopped yet is because once they stop

doing AI the markets will ask a very

reasonable question which is great.

What's next? What how are you going to

grow further? They don't have anything.

Quantum isn't going to do it. Robotics

isn't going to do it. These are quantum

is they have quantum. It works, but it's

not really a product yet, so to speak.

Robotics is

I no mark. Mark, walk away from the

robot. Amazon tried Amazon already tried

robotics. They're not going to do that.

I guess Microsoft will raise prices

again. That's what they do every few

years anyway. But yeah, no one has

anything more. So that's why they

haven't given up. That's why they're not

That's why they're so steadfastly

dedicating themselves to the graveyard

smash of spending a trillion dollars a

year on this with no return on

investment for them.

>> How much cash do does Enthropic and

OpenAI have left? Um I think you you had

numbers in your your recent piece on

exclusive on Open AI. Um but I don't

know if that was before or after that

the fund raise. And can you also share

your views on the ability or willingness

of the venture capital community as well

as like corporate VCs like Nvidia,

Google or whatever um as well as the

general public if there's a or

institutional investors if there's an

IPO to to invest in these companies like

what do you think the odds are of you

are right on the fundamentals but over

the next 18 months like $150 billion is

raised uh to to keep on funding these

you know as you say money losing

operations. First and foremost, I think

venture capital is running at its limit.

I think that's why you have a bunch of

private equity firms who got involved. I

think you even had a a private credit

fund. Like I think you've had you've

started moving into the big asset

managers. As far as the cash position

goes, even they had 22 billionish in

cash. More than that, they had like 50

billion total in assets. The information

just reported that they had $73 billion

in cash and other things. So, it's kind

of hard to pass that out. The thing is,

they could raise that money. It could go

on a little longer, but at some point,

OpenAI is going to have to pay $300

billion to Larry Ellison over three five

years. At some point, Anthropic is going

to have to pay part of that $330

billion. They're not going to be able to

just keep scraping along on cloud

credits or venture capital subsidies.

And also, if they go public, it's going

to be if they want to do equity dumps,

they can. But in the state of these

companies, I don't know if that's a good

idea. Nor do I think that the bond

market's going to be very helpful. Maybe

they do one or two bond sales just

people are really stupid. But even then,

it's like we have not seen a company

with this frightful level of economics

go public. SpaceX is a piss poor

company, but at least they have business

lines like Starlink, which makes more

money, and SpaceX, which blows up

rockets and has government contracts,

and also XD Everything app, which

generates non-conensual porn. Like, it's

they have money losing operations, but

the thing losing the money the most is

AI. Open AI, Anthropic, they are just

AI, and they have businesses that are

increasingly commoditized.

So, I think they can raise more money. I

don't think that's impossible. I think

that the scale of their last raises

tells me that they are running up

against the limits. The fact that Google

had to do an equity sale 85 billion I

think it was proves that the credit

markets are starting to run dry. There

was an NT story a few months ago, maybe

a month or two ago where they said that

banks were afraid they were choking on

data center debt. I don't think that

those same banks are going to be

particularly excited about loaning money

to these companies. They've given them

some lines of credit, but we're talking

they're based on their fundraising

history, these companies are going to

need $150 billion within the next 6 to

12 months and also they're going to go

public and they're going to have to

raise that money with everyone knowing

they're dirty business. So, not really

sure how that works out, but they've

also just made so many compute

commitments. And though Anthropic and

OpenAI both project that they will be

profitable by the end of 2030, no proof

as to how that's happening. no actual

evidence other than, oh yeah, we just

didn't pay that. We were allowed not to

pay a cost. They don't they don't have a

plan. And also, it's very obvious that

training is not going away. This is the

thing that no one wants. Everyone hears

training and they're like, "Oh, it's a

temporary cost. They could just stop

training when

when because it keeps going up. Does

that stop going up?" Because the basics

of machine learning are that model drift

happens. You have to constantly update

these models. I think through the middle

of last year, Joe Biden was still

president according to chat GPT. Like

these things need constant updates and

it's not just pre-training, it's

post-training, it's specialist training

data. It's a ton of investment that has

to go into this just to keep them going.

And also, they have to produce new

models because the ones right now aren't

doing enough. They're not providing the

ROI that justifies their current costs.

They need to keep noodling at this. And

I think the markets are going to

eventually ask, "How long do you need?

Do you actually have a plan? What is it?

Can we see it? What do you mean you left

it at home?" Like that kind of thing.

And I want to talk about the public

markets exposure to AI. Like they're

probably some people saying, "Okay, I

don't love AI at all, but you know, I'm

I'm diversified. I'm invested in the S&P

500." I've talked to lenders who have

confidence that they're lending against

basically the credit of the hyperscaler.

So the biggest companies in the S&P 500,

in other words, they're confident that

these companies have entered into lease

commitments of hundreds of billions of

dollars that are going to appear as

costs that I don't think many people are

considering that.

>> During the dotcom bubble, Lucent

Technologies, they had a $2 billion deal

with Winstar, a company that only ever

lost money. $2 billion where they did a

circular financing thing. Now, Winstar

ran out of money and actually ran out of

money because of the cost of that loan.

It's very possible to sign a huge deal

and then just not get paid. That happens

many times. But the fundamental thing is

is yeah, you're you're betting against

you're betting that the hyperscalers

will make like the cipher mining deal

for example, the terowolf one backed by

Google, that Broadcom deal, that

Broadcom deal with Anthropic. Anthropic

still has to make those payments and if

they don't, Broadcom will have to. I

think that we are yet to see a real test

of any of these situations because the

data centers are taking so long to

build. I think there's the slower that

happens, the longer it will take to have

that test. But I think that the

fundamental problem is that hyperscalers

can only cosign these like student loans

or student credit cards so many times

before it starts to affect their balance

sheets, before it starts to affect

investor considerations of those

companies. And I think that once

anthropic and open AAI go public and I

think it's very possible one or both of

them do I think that that will become a

much more serious issue because there

will be you will have to mention your

exposure to these companies is a risk

and I think it's also a real risk when

you are one of the people that's keeping

them alive and there comes a time when I

genuinely think just there is not enough

money if the I think I read some stat

where it's like 90 by by next year I 98%

of all hyperscala cash flows going into

capex like they're going to have to take

on debt. The markets do not like the

debt. The whole reason you invest in one

of the magnificent 7 is cuz they're

cashri asset light. Now they're

just plumbed full of these bloody GPUs.

These bloody GPUs that aren't useful for

anything outside of AI. And the only

reason they've been given this

affordance is because they've been

relative like they've their other

businesses have kept growing and because

the markets are invested in by people

with the brains of dogs at times

people are just like well number keep

going up that must be AI

that must be AI is doing that even

though they won't tell us how much money

they're making from AI even though they

offiscate that in every way shape or

form. Well, you know, I that's good

enough for me. That stops being as fun

when you get burdened with all of this

debt and they're going to have more and

more, Dan. It's going to get worse and

worse because they're also not making a

profit from Mayi. At some it's just a

question of when the markets eventually

care and also how desperate the

hyperscalers get.

>> What is going on between Enthropic and

the US government with the the commerce

secretary saying you can't use these

models. you have to exclude them to

foreign nationals. So, basically,

Anthropic has taken the fable model

away. Um, which is a, you know,

moderately dumbed down version of of of

mythos. What's going on here? What

should we take away? What are you taking

away from this?

>> Let's go back to April. So, in April,

Anthropic said with Project Glass Wing,

oh, we've made this big scary model

called Mythos, and it has these powerful

cyber security things, and it can find

vulnerabilities in all sorts of things

up and down, side to side. Since then,

it's come out that the system card

mostly overstated things. They don't

include how many false positives there

might be. So, and also it's good at

finding vulnerabilities, doesn't really

exploit them, isn't really clear what

all the freaking out was. Nevertheless,

they said, "This is too dangerous. We're

only going to give it to 15

organizations." About a month later,

they were like, "Actually, it's going to

be it's going to be 150 organizations."

And then couple weeks ago, they said,

"Well, actually, we're going to release

something called Fable 5. Fable 5 is a

mythos class model with guard rails. So

you can't use it for biological stuff.

You can't use it for cyber security. So

eventually

because guess what? Here's what happens

when you tell software engineers they

can't do something. Their first thing

they try is to do it immediately. They

were just like I will break. There's a

guy called um was it ply the liberator

on Twitter as well. He jailbroke it as

well. That guy that guy really loves

jailbreaking [ __ ] Putting all that

aside, an Amazon research group and then

this message then went through Andy

Jasse, the CEO of Amazon, reported it to

the commerce secretary that there was a

jailbreak. The Howard Lnik went to

Anthropic and Anthropic said, "It's not

a big deal." The government said, "You

need to fix this jailbreak." Anthropic

said, "Uh, we're not going to. It's not

a big deal." Government said, "We're

going to add export controls. No non- US

citizens inside or outside of America

can use this." And so Anthropic went the

only way we can comply with this is to

just take Mythos and um Fable offline.

Now there's some back and forth and it

seems that there's a degree of something

with the argument of the Department of

Defense from a few months ago. There's

clearly bad blood. They clearly want

them to kiss the ring. They're claiming

they're working on a framework. But what

this is is the consequence of lying for

years of just saying our models are big

and scary and they're going to destroy

everyone and they're going to take every

job and it's big and scary. Mythos is

too powerful to launch other than the

fact we're launching it. It's not safe

enough for anyone to use other than JP

Morgan, Goldman Sachs, and multiple

other organizations and also 150 of them

in 15 different countries. Otherwise,

it's not safe at all. I've spoken to

people that use Mythos. They're like,

"Hey,

>> it's just like I've spoken to they're

like, "Yeah, it was able to find some

vulnerabilities. It found a bunch that

weren't actually vulnerabilities, too,

and many that weren't even executable.

What What were we meant to do with

that?" But anthropic scaremongering

because this has been since GPT2 when

Dario Amade still worked at OpenAI.

They've been doing this thing of oh it's

so scary. Oh the models are so scary.

Then they took a model and they sold it

literally saying it's too scary but now

we're going to release it for some

reason. And what do you know? Someone

took it seriously. It's what one of

those well well if isn't the actions of

my the consequences of my actions. It's

just frustrating because there are some

people like it's just proof that it's

too powerful. No, it's not. No, it's the

stop it this kayfabe nonsense. Why are

we doing why are we pretending? Silicon

Valley was built on this kind of

meritocratic

rugged

um rugged like pragmatism and re

realism, rationalism. And it's like,

yeah, but the moment one thing comes

along, they're all like they're talking

about it like they they saw Jesus in a

cup of coffee. They're reading the tea

leaves. They're doing tarot card decks.

Like they become everyone becomes a

goddamn mystic when an LLM's involved.

But that's what happened. They scared

people. They sold something on fear and

then people, the government in this

case, acted like somebody would if they

were scared of something. I think it

could be genuinely really bad for AI

development. I actually think it sets up

a really terrible precedent for pretty

much everything now. I think it's bad

for the software industry. They ne Cal

Newport computer scientist was just New

York Times called it doom trolling. I

think it is I think that these labs

because they realize they can't sell the

software based on today that they have

to do this and I think it's good they

face consequences. I think it sets a

horrible horrible precedent for how the

government is going to deal with tech

going forward and blaming Dario Amade is

necessary because he is responsible him

Sam Orman did a bit of we're scared but

Amade is the number one carnival barker

scaremonger he's a he is a problem he is

an actual problem and both of them are

genuinely bad for the tech industry but

Dario Amade he really sees himself as

some Jobsian

socialite uh kind of like a elder

statesman type when he's just kind of an

oath. He doesn't have you know you're

getting you know things are bad. You

know you're an oath when Sam Alman is

politicking better than you.

>> Do you think the AI bubble pops this

year 2027 or 2028?

>> I think 2027's the safe bet. I think

2026 could be possible if SpaceX starts

tanking for example. If SpaceX, it's

been kind of trundling down. I don't

know when this runs. Probably embarrass

myself and it will be back up. Uh, but

if SpaceX could not transform into a

meme stock like Tesla, I think that

might make the OpenAI IPO a little bit

more dangerous. But there's also the

chance that just the money starts

running out. The data centers stop

getting built that like there's enough

situations and also Nvidia's got two,

three more earnings calls. if their

guidance doesn't make the markets rock

hard every three months. People get

people get like if you what read the

headlines before Nvidia's earnings it's

always like people have been like

okay okay Jensen keep me alive here okay

please don't mess this up Jensen and

because Nvidia has done the circular

financing to keep this inflated they set

these unrealistic expectations Nvidia

better bloody hope they have a trillion

dollars of sales through 2027 because if

they don't think the markets will fall

apart it really comes down to the fact

that because not real revenue because

real revenues from these companies are

not and actual ROI is not what's making

the AI bubble inflate. It's going to

come down to a vibe shift and it's

already begun.

>> Ed, what do you think is the greatest

misconception by the AI bulls or as you

said the AI boosters that we haven't

talked about so far?

>> The average AI booster has this belief

that I'm doing this because I just hate

I hate I hate progress. I actually think

the biggest misconception they have is

that AI is progress. That AI is a

progressive thing when what it actually

is is a flattening of everything. It is

an averaging out of everything. It is a

technology that's not sold on what it

does today, but what it might do in the

future. Every conversation happens in

the future tense. You can't talk about

AI without someone saying, "Well, it

will." I genuinely think that AI bulls

are conflating a semiconductor bubble

and these massive sales that they're

seeing because of all of this debt

fueling the AI capex bubble. I think

they see that and think that that is

demand for AI. And what that is is a

demand for speculative debt. It's a

demand for private credit funds to find

more yield. When you actually go and

look at the like people using AI stats,

it's always like yeah, when you ask a

CEO like it's the best thing ever. When

you ask a worker, it's like it's fine.

It's all right. But I think that the AI

industry has successfully co-opted a lot

of people that conflate technological

progress with stock values

who have taken this era of LLMs to mean

that all AI is going to grow

exponentially. And I think these people

are mocks. I think they're being used by

the companies because the companies

treat them, I mean this for every

booster, they treat them with contempt.

How else do you describe what they're

doing? You can't get a straight answer

out of any of these companies. They

don't want to give you the direct story.

They don't want they move stuff around

their balance sheets to try and make

things look good. They give weird quotes

to the FT and that is contemptuous

towards their fans. If I was an AI

booster, I would want better evidence

than this. I would be genuinely like if

I had to do the the bull case, I would

want better. I would I would be here's

the thing. I would be asking Sam Orman,

Dario Amade, Boris Churnney, all of

them. Hey, this is worrying. What is

your answer? I won't be getting mad at

me. I'd be getting mad at the fact that

I wouldn't have fundamentally sound

information because that's what keeps

happening. It's like, you know what? You

like LLMs? Fine. Good for you. Enjoy. I

don't like them. I think that they do

bad things to the world, but you want to

do that, fine. It's software. Who gives

a [ __ ] But when it comes down to, oh,

if you get in the way of this, you're

against the future. God, no. I love the

computer. I think I genuinely, in fact,

maybe that's the biggest misconception.

I love the computer. I grew up online. I

have great affection for software. I

think the computer has made me a better

person. It's given me so much value. But

I think that many AI balls and I

actually kind of meet on that level. I

think they too have a debt of gratitude

to software. That's not what this is.

It's not what it is. It's the It's an

aberration of software. It is a draining

of Silicon Valley's value. It is an

intellectual bubble that quashes

dissent, that intentionally pits people

against each other, that makes people

angry for stepping out of line in a way

that resembles cult mentality. And

that's not an insult to the people

involved. I consider AI balls largely

manipulated by the companies uh who want

their craving community as we all do as

human beings. And it's frustrating. It's

frustrating because one of us is going

to be right. If I'm wrong, I'm

committing to explaining why. But a lot

of the demands of me often come down to

people that don't want to actually

engage with my work, which I get. I

wouldn't want to read something that

pissed me off either.

I think it's just a level of at the end

of this, we're going to know who's

right. There's plenty of evidence I'm

right. There's nothing wrong with

admitting you're wrong. I will admit I'm

wrong when I'm wrong. I'm sure I'll be

wrong in the future. And I it frustrates

me because

the bulls I think will end up being

wrong and they will have been wrong

after investing probably not a ton of

money but a ton of emotional and

intellectual energy into something that

flattens the experience that takes

attention and money away from actual

innovation and ultimately just makes a

couple of other guys really rich. Maybe

they aspire to be them. It's not worth

it.

>> Ed, thanks so much for coming on

Monetary Matters. If people want to uh

learn more about about your thoughts,

you have you write prolifically at

where's your ed uh newsletter which

people should check that out as well as

the better offline podcast. Thank you

everyone for watching. Please leave a

rating and review for monetary matters

on Apple podcast and Spotify and

subscribe to the Monetary Matters

YouTube channel.

Thank you. Just close the door.