Anthropic's IPO Is Not What It Looks Like: Inside the S-1 and the $80 Billion "Loss"

Anthropic just filed its S-1 ahead of what could be the biggest IPO in history, and the whole market has been waiting on it. Chip stocks fell eight sessions in a row, the headline number is an $80.7 billion loss, and everyone wants to know: is Anthropic actually making money, and are we in an AI bubble? Ejaaz came in skeptical and went through the filing. It's not what it looks like.

We break down revenue that went from $41 million in 2023 to $16.3 billion in the first half of 2026, why Anthropic now makes about $2 for every $1 it spends on compute, 900% net revenue retention, Claude Cowork outpacing Claude Code, the $120.5 billion going to SpaceX for GPUs, $132 billion in cash and short-term investments, and why $72.4 billion of that loss is an accounting charge, not cash out the door.

==
Timestamps
00:00 The IPO Everything Hangs On
01:14 They're Crushing It
02:12 Revenue Finally Beats Compute
03:11 Who Pays Anthropic
05:11 900% Net Revenue Retention
06:53 Where the Money Comes From
08:45 Claude Cowork's Surprise
11:24 Anthropic's Big Predictions
13:56 The Bear Case
14:10 Half a Trillion in Compute
15:35 Elon's GPU Toll Booth
17:26 $132 Billion in the Bank
19:32 The $80 Billion Paper Loss
20:25 Final Thoughts
23:15 The Roko's Basilisk Meme

#anthropic #ipo #claude #ai #aibubble

==
AlphaSpace by Yahoo Finance: A Professional-Grade Investment Platform Built for Everyday Investors.
👉 https://finance.yahoo.com/about/promos/gold/alphaspace/?ncid=100003571

Creators and Guests

Host

Ejaaz Ahamadeen

Investor & Host of the Limitless AI Show. Former Product at Coinbase and ConsenSys

What is In the Loop AI?

Exploring the frontiers of Technology and AI

We're about 19 days out from the biggest IPO in history. Anthropix rumored to

go public at around $2.2 trillion.

And today they released their S1, which reveals a lot of important information

and numbers as to how Anthropix is really doing under the hood.

Are they able to make enough money? Can they afford the crazy $500 billion worth

of capex that they've committed to spending over the next couple of years?

And are we in an AI bubble? And to be honest, the market's been questioning

it pretty aggressively recently.

We've seen the largest drawdown eight days in a row

of semiconductor stocks in the public market, and it's no mystery that the entire

stock market is heavily reliant on two particular companies,

Anthropic and OpenAI, and if they're making enough money to prop the entire

market up in the first place.

Well, today we finally have the answers, and honestly, guys, I'm a little surprised.

I came in skeptical because I thought that some of the numbers were going to

be inflated and that it won't make sense.

But it turns out not only is it way more optimistic than I thought,

but they have the numbers to back it up. Now, bear with me, this entire document

was around 261 pages long, so it took a while to get through,

but let's get into it on today's episode brought to you by Qualcomm.

Now, it's only been a few hours since the S1 dropped. And one of the headline

numbers is obviously the revenue. And there's no better way to say this,

but they're crushing it. But it's not just the amount of money they're making.

It's how efficiently they're making it. There's this great line from the semi-analysis

folk, which goes, this is likely the most efficient scaling of a business in

the history of capitalism.

Let me tell you why. Three years ago, Anthropic made $41 million in the entire year.

The next year, 2024, they made around $386 million.

Good improvement. Then last year, they made around $4.6 billion.

This year, we're not even at the end of it, but the first six months,

they made $16.3 billion.

This is the fastest growing company we have ever witnessed.

That's about a 14x year over year growth. But what's more impressive is the

fact that their largest cost, which is spending billions of dollars on compute,

which we've spoken about a lot on this show, has finally paid off.

And it's paid off almost 100% more. If we look at Anthropics revenue generation

as a percentage against the amount that they spend per unit compute,

we see that the revenue per $1 of compute

is about $2.08.

Now, if you backtrack this six to nine months ago this was very consistently

under a dollar it was very capex intensive and what that told me at the time

was that the models as good as they seemed as much as they were impressive to

me didn't really translate into business value in the real world these numbers

now tell the opposite story which is.

It's very valuable for individuals and for companies out there.

And Anthropic as a result is making a ton of money, over a dollar more than

the amount that they invest in compute alone.

And the speed at which they've done this is insane. It was literally Q1 of last

year where this was completely inverted.

I think for every dollar of compute that they spent or invested,

they would make around 25 cents of revenue.

But what's interesting and maybe slightly concerning is how concentrated the revenue risk is.

Around 250 customers according to the s1 filing pay anthropic 10 million dollars

per year so these are large huge hefty

arguably long-term contracts anthropic typically signs long-term deals with a lot of these higher

value companies and then around 2 000 customers pay about a million dollars

per year and around 11 900 customers pay around 100 000 per year

now if i think back to the conversation that Krishna Rao, the CFO,

had, I think it was on Patrick Shornish's podcast, a few months back,

these numbers were halved or even less than that.

So we're seeing increasing concentration and increased profitability for a set

number of customers, very high value customers that seem to be doubling,

tripling, and quadrupling down on Anthropic and their models.

And the data from OpenRouter, which is basically a public platform that routes

people's requests across any most popular demanded model platform.

It shows that Opus 5.5, since it released, Anthropic's token consumption has

almost 3.5 to 4x, which is very impressive because prior to that,

the trend showed that OpenAI's models were winning.

So Anthropic's pulled something really special off with Opus 5.5.

But it's important to point out that this doesn't necessarily mean that Anthropic

is profitable. They had a humongous loss on their balance sheet, $80.6 billion.

That's not actually strictly true, and I'll get into that in a second.

But the idea is if you remove that and you actually look at the operating losses

that they have, it's between $7 to $9 billion.

And again, there's a lot of other hidden expenses. I question Anthropik's revenue

calculation quite a bit, and we'll get into that in a second.

But the idea here is they're generating a lot of money. And if you do a fund

calculation of like dividing their profit or the revenue that they're generating

across each of their employees, that's roughly around $6.6 million per employee.

The most highly monetized company in the world. But what's also so interesting

is how loyal the customers seem to be.

They reported a net revenue retention, NRR, of 9x, 900%.

What that basically means is, if you're a company that starts the year,

January 1st, using Anthropics models, and you're spending a million dollars

a year, by the end of the year, you're spending $9 million.

Because you see so much value being reaped from your investment in anthropic models.

So obviously that customer concentration risk that I mentioned earlier on,

is actually showing up quite positively for Anthropic.

Their models are providing these customers with a net difference from the likes

of OpenAI, SpaceX AI, or Meta's models, or Google Gemini's models themselves.

So the idea that's being implied here is Anthropic's models offers something

very net new and useful. And listen, I'll speak from my own experience.

I have access to all the Frontier models.

And I would say that a large bulk of my work is still spent on Claude models,

despite me using codecs from ChatGPT for multiple other things.

So I can kind of like see the resemblance and pattern that is forming here.

They have a huge customer loyalty. And what's interesting about this is,

again, on Christian Rao's conversation three months ago, he said that the net

RR back then was around 300%.

And I remember that that blew my mind. I remember speaking about it on an earlier

episode a few months back and saying how impressive that was.

And back then it was the fastest growing net rr for any other company and we

just see that 3x in a matter of months it's unprecedented the demand for these

models has just skyrocketed.

But what's also interesting is where Anthropic is making this money.

You would think that these companies are just signing up to Anthropic.com or

making the deals directly with the lab, but that is not the case at all.

In fact, 48% of their customer revenue comes from cloud providers,

the likes of AWS, Google Cloud Platform, as well as Microsoft Azure.

So Anthropic is very heavily dependent on these legacy cloud systems where enterprises

already have their sort of home.

And I wanted to point this out because there's a lesson here.

Even if you are a startup founder or just a regular enterprise trying to eat

into this moat, it helps to be where these enterprise companies,

these Fortune 500 companies already are because they're comfortable with their setup.

They have these customer relationships with the likes of Amazon,

Google, and Microsoft because they have such a sticky moat and relationship

with their enterprise customers.

If you're a startup and you're trying to break into replacing an existing product

that an enterprise already has, and you go directly to the source,

you're going to have a hard time because those customer relationships that they've

built up with these cloud providers are already huge.

Now, that means that Anthropic has a heavy dependency on these companies.

Which aren't just cloud service providers.

In some cases, they're also competitors. You kind of see Google,

they train their own models.

Gemini, they have their TPUs, which is their own version of their GPU.

Arguably, they're going to be using their compute to train their models and

to learn from all these other models that are being inferenced through their

own cloud platform. Now, what's interesting is Google is kind of playing both sides for a while.

Amazon is doing the same, but they're largely more of a cloud provider.

And Microsoft, in my opinion, is largely more of a cloud provider.

I don't see them as an AI lab. But Google is the distinction here. And they also have a.

Own around 17 to 21% of Anthropic. They've invested billions of dollars.

But let's put numbers aside for a second and look at the actual products.

How's Claude Code doing? And do they have any other diversity across any of

the other products? Or are they just reliant on Claude Code?

Well, according to the S1, Claude Cowork, which is basically like the layman's

version of Claude Code, if you aren't a software engineer,

and you want to use it for multiple other things, but have that Claude Code-like

experience without understanding or being technical, you would use Claude Cowork.

And they launched this, I was about to say last year, but they've launched this

literally earlier this year.

And I remember it being launched. And listen, I was impressed by it,

but I wasn't that impressed by it.

I think some of the more personal agents that we're seeing today from the likes

of Metamuse, GrokBot, ChatGPTDots, and now Anthropic's own Claude Duet,

which we did an episode on, you should definitely go check that out yesterday,

are much better instantiations or versions of this.

Well, apparently I'm wrong. Claude Cowork is currently growing a lot faster

than Claude Code itself.

And that's a very distinctive sentence for me to say because Claude Code...

Is by all means the bread and butter of Anthropic. That's how they make the

majority of their money, token consumption through Cloud Code.

So the fact that we're starting to see Cloud Cowork increase,

it hasn't beaten the number of users that Cloud Code has, but is increasing

at a quicker rate of adoption than Cloud Code right now,

suggests that Anthropic might have finally hit the actual goldmine of products.

Because remember, Cloud Code is very specific. It applies to software engineers

and the technical know-how, there's a very limited population of the world that

that applies to. But 99.9% of the rest of the world doesn't know how to code,

but would like to have a similar magical experience to Claude Code.

And Claude Cowork is that instantiation. So if you can imagine,

we have the same sort of effect that Claude Code had,

but on the rest of the world, on knowledge work, on any other profession that

includes administrative roles, customer support, or whatever that might be,

people will end up being not only a lot more productive, but Anthropik is going

to end up making a hell of a lot more money.

And this is just me extrapolating. I have no idea if that's actually going to

happen, but it's a useful stat to see.

Now, if I were to highlight a red flag about this is typically if you wanted

to access Cloud Cowork, you would need to select a different tab,

I believe, in the desktop.

Now what they've done is they've sort of combined it with the homepage.

So whenever you're using regular Claude on the LLM chatbot interface,

it directs your prompt to Claude Cowork if it thinks that it's a work-related task.

So it's a bit of UX and financial engineering, but it's not just Anthropic specifically.

OpenAO is doing the same thing with ChatGPT work.

I don't know how I feel about this. I actually think I'm fine by it.

Let's see the numbers after they've actually gone public and they've done their

first quarterly earnings. And of course, it wouldn't be Anthropic if they didn't

have some very high-level, long-standing thesis predictions.

Two of my favorite were expert-level autonomous systems within a few years.

So Anthropic basically expects AI to match or exceed leading human experts across

most professional and scientific disciplines.

This isn't a surprise to me at all because over the recent few

weeks and months, we've seen the likes of maths chemistry and physics make very

meaningful progress due to these llms if i had to guess what anthropics models

are going to solve first it's going to be in the medical and biological

related field that's where anthropic has kind of designed and focused their moat on when it comes to.

Real world application and to be honest i'm all for it as a former biologist

and scientist i would love to see some of the biggest diseases get cured soon

It's not going to be as easy as I'm saying it, but it's nice to see the focus on that.

The other high level takeaway I liked was AI is already accelerating the development

of its successes. Clawed helps run experiments, generate and process training

data, automate evaluations and safety audits.

Now, of course, with this, we're talking about RSI, recursive self-improvement.

The idea of an AI model being so smart that it can build the next version of itself.

Which seems very meta, but the reason behind this is you'll end up having AI

that is self-improving.

And so you will not need as many humans to try and figure out the complex research

problems or tasks or way to improve itself.

You can just kind of get the AI to do that, automate itself,

and then you could focus your time on other things.

And the idea here is if you have an AI model that can recursively self-improve,

you can then chuck it at pretty much any problem that exists,

whether it's physical or digital, and have a bunch of robots or your own digital

brain figure it out for you.

Now, that's, I think, a long way away. But if Jack Clark, Anthropics co-founder,

is anyone to go by, he says that we're already experiencing the takeoff of that.

And we're probably going to see something along the lines of RSI in the next six to nine months.

And Demis Civis from Google has also confirmed this. So it's this weird thing

where I'm sitting here and I'm trying to envision what a world like that looks like.

I can't actually envision it. But if it does actually occur,

if Anthropic is right, and they're not just doing this to, you know,

raise a very hefty valuation for their IPO, then they'll end up making plausibly

a heck ton more money than they're doing today.

And we're already looking at the steepest curve of any capital company ever.

So if they end up pulling it off, they can justify the crazy capex spending that they're seeing.

But on that note, Let's get into some of the bear cases because there's a lot

of money that is being committed to this and I don't quite know if Anthropic

is going to be able to pull it off.

Let's see. On today's episode brought to you by Qualcomm.

Now, if you remember a few weeks ago, the headline number that Anthropic was

rumored to spend over the next couple of years is $500 billion.

Half a trillion dollars from a single company alone to acquire compute to train

and inference their models.

Now, that told me one thing off the top of my head. That's an insane amount

of money and I have no justification whatsoever to see whether you can afford that. Number two,

the compute scaling laws still exist. And that means that the most compute that

you have, whichever company has the most compute, will own the best models.

But the question is, can Anthropic afford it?

Now, if you look at the breakdown of who they're spending this money on,

it's quite interesting.

$120.5 billion of that $500 billion is going to Elon Musk, SpaceX,

who have recently pivoted in the most successful way to become one of the leading

or number one neoclouds in the world. That is a company that would provide compute to others.

And Elon has been able to pull this off because he's been a master scaler of

GPUs. He built the best relationship with Jensen Huang, and he committed the

longest and most expensive premiums to acquire Jensen's GPUs.

And now he's sitting on the largest arsenal of GPUs ever. He's bringing on around

200,000 to 300,000 GPUs online every month. That's around $13.1 billion worth of compute.

And no other company has been able to achieve this feat. He is the master scaler.

We made an episode on this last week.

Elon Musk and SpaceX has sort of become a toll booth to some of the biggest,

most important AI labs of which valuation relies on getting access to these

GPUs. So they can't go directly to Jensen Huang or NVIDIA because he'll just tell you he's sold out.

So they need to go through this third party or this intermediary who has gathered this arsenal.

And Elon Musk is sitting there and thinking, cool, I'll charge you a hefty premium

on that to the tune of around $3.1 billion per month.

That's what Anthropic is going to be paying Elon Musk. So he mentioned several

times that this is a temporary measure whilst we figure out the training of Grok.

And remember, these contracts are non-cancelable. So Elon Musk at any time can

just shut Anthropic off from this and be like, well, I don't care anymore.

Now, another $115 billion Anthropic is spending on Google to acquire their TPUs,

to diversify their GPU stack from NVIDIA's GPUs. And this is a smart move,

in my opinion, because too much reliance on a singular.

Type of GPU or metal block means that you are just heavily reliant on that company, NVIDIA.

Diversifying to DPUs, to ASICs, which is what Anthropic and OpenAI,

fair enough, is also doing is a smart move.

What's also cool here is I guess Google's investing stake in Anthropic,

which roughly sits around 17 to 20% of Anthropic.

So it's this weird thing where Google should probably be using these TPUs to

build a better Gemini model, but instead they're selling it to their biggest

competitor, which eventually over time, if Anthropic actually does achieve RSI, then

they will end up running away with it and making all the money because they have the best models.

And Google will spend the rest of their time playing catch up.

And this is not a bearish take on Google. It's just more of a strategic critique

where I'm like, just use that compute to build a better model,

Google. Invest now heavily, take the loss on your balance sheet,

build a better model that will end up competing with the likes of Anthropic.

Don't just give the game away.

Now, what's interesting about the heavy capex commitment that Anthropic is committed

to is around the amount of money they had.

Now, if rumors were anything to go by over the last couple of months,

Anthropic was said to have almost zero cash balance on the bank sheet because

they've been burning so much money on compute, billions and billions of dollars.

But when you look at this S-1 filing, it's quite the opposite.

They have around $132 billion sitting in the bank account And maybe $19 billion

of that is in investment-grade contracts.

But the idea is they have a bulk of money dry powder sitting around to spend

on some of these heavy capex commitments. So it's not just reliant on how much

money that they make a month or a quarter or whatever that might be.

They also have the cash to back it up.

One thing that popped into my mind immediately when I heard this number was,

$132 billion is roughly around the amount of money that Apple has and roughly

around the same amount of money that Google and Meta had before they started

spending so aggressively to acquire GPUs.

Anthropic, although they're a startup that is incredibly young,

is now playing on the same big boy play field as these other companies.

Now, granted, Anthropic does not have the revenue diversification.

It is way more precarious of a bet when it comes to considering Anthropic stock

as an IPO level, because they don't have any other ways to make money.

They are AI focused. If you look at Google, if you look at Meta,

they have many other ways to make money.

Anthropic doesn't have that safeguards. But it's just interesting to see that

in order to compete with the big boys, you need to have enough money to do so.

And Anthropic's balance sheet proves it. And Anthropic's compute commitments

show it running on the S1 till 2045. The other companies These are included

are Broadcom, which they're building custom ASICs with, which is like a separate thing.

Google, SpaceX Air, which we mentioned, Amazon, $110 billion going to them.

Amazon also a very heavy stakeholder in Anthropoc. I think that's around 12 to 14%. Crazy.

Microsoft, NScale, Lambda and AMD. So the idea here is that Anthropoc is building

out a very diversified portfolio of risk when it comes to acquiring their compute

because that's their bread and butter that is needed for them to stay ahead of the game.

Now, the other thing I want to address is the big headline of loss,

which is $80.5 or $80.2 billion.

That's just a paper loss, it turns out. This was, as we guessed it on this show

a few weeks ago, just the paper loss recorded for equity stakeholders that need

to be paid out for their early investments.

So if you're an early investor and you invested in Anthropic five rounds ago,

your stake is now worth a lot more money because Anthropic's valuation has increased.

So Anthropic legally needs to record that as a loss on their balance sheet.

I do not know why that's what the law mandates, but that's what that paper loss is.

$72 billion to be specific. So the real loss is actually around $8 billion.

$72 billion paper loss will not exist in the next quarterly earnings or public

report that they eventually end up releasing.

So again, a bit more fear mongering and a bit more of a nothing bug.

I just wanted to, you know, debunk that myth. And so to round this up,

I feel a lot more optimistic about the IPO.

That's not to say I'm concerned and I justify the valuation or the price that it's going to be.

I still think it's going to be quite overvalued. I still think there are an

immeasurable number of factors that could influence the way this market swings.

When it comes down to CapEx, bringing all those GPUs online,

the $500 billion that they're spending, that's all well and good.

But if you can't bring those GPUs online because you're constrained by power

or the electrical grid or whatever that might be, war, political conflicts,

anything that are outside of the purview of just AI.

Anthropic is going to end up being in hot water. But from the numbers that I'm

seeing where I thought they maybe weren't making as much money or revenue,

maybe the revenue was slowing down.

And actually, it is. Anthropic mentioned in the S1 that you can't expect demand

to be sky high all the time. Things are going to slow down eventually and plateau.

That's how these things work.

But eventually, the idea is it becomes more consistent. Chamath,

Apollo, and a number of other people for over the last

couple of days have actually pointed out very fairly that a lot of the money,

ARR, revenue that these companies are making are actually coming from VC-subsidized startups.

So basically coming from VC money itself.

What that looks like when it's not VC-backed, when it's actually coming from

companies that are generating their own revenue for themselves because these

AR models are so useful to them, is something to be seen.

The other thing I'll say is that, well, I guess Anthropic actually said it.

They said that the gap between what Claude can do and how widely it's used is

one of the biggest growth opportunities. And I truly believe in that.

When I speak to my friends and family who aren't involved in AI,

their basic takeaway is, okay, why the hell do I care about Opus 5.5?

Why do I care about any of these new models?

I just use it for myself and it's not really useful.

You know, I gave a story or anecdote.

The other week where someone overheard grandmas playing around with Claude and

saying, it can't even generate me a simple crossword, which you would think

that it would be able to do.

So there's this jagged frontier where the AI models are really good at very

specific things, or they're really good at many things, but it's hard to try

and figure out what it's good at.

And if you think about 99% of the world, they want convenience, they want ease of use.

And I think this will increase over time with the likes of things like personal

AI agents and agents being more proactive and understanding and predicting what

you might want to use, but that's going to take time.

But that's what Anthropic is saying is the largest gap. They're saying with

the existing models that they have, not even with the secretive Fable 5.5 models

that they haven't yet released or announced, these models as they exist today

can end up unlocking so many more billions of dollars worth of value.

And that is an optimistic take. And it's honestly one that I believe not just

for Anthropic, but a lot of the other AI companies that are out there that are

building these very important models.

I'm bullish on the entire sector in general. That is my take.

I'm going to leave you guys actually with this hilarious meme so nick carter

he honestly got me when i read this

he goes i knew anthropic s1 would have some goofy risk factors but this has

to take the cake and he goes on to include an image of what is supposedly meant

to be in anthropics s1 filing which is a clause around roco's basilisk for those

of you who don't know roco's basilisk

is this theory that.

What was a theory or something that's presumed to become true eventually, where,

the future AI superintelligence, that is the all-seeing, all-knowing,

will look back on the history of content and the internet, listening to the

words that I'm speaking to you right now, and judge each individual to see if they were,

helpful in building AI in general.

And if they were not, they'd be subject to some form of penalty,

penalization, or being put in perjury, or some kind of mental prison.

It's like a whole fantasy thing that's existed for a while, but something that

Anthropik and the likes of OpenAI have referenced or people that have worked

there have referenced many times.

So the idea is there's this like all seeing AI that is listening to us as we say right now.

And this meme basically includes it as a clause of contract in the framing of

the S1 and IPO, where it says if you don't invest in this company,

it could be seen as a negative thing because Basilisk will come to get you.

But anyway, just wanted to leave you on a slightly uncomfortable but hilarious

note there's a lot of new subscribers over the weekend we had one of our record

weeks one of our videos got around 200 000 views

loads of you subscribe turns on notifications if you aren't any of those people

please do so it helps us out massively so let me know your thoughts and thank

you so much for listening i'll see you on the next episode of in the loop brought to you by Qualcomm.