In the Loop AI

Last week, Anthropic's S-1 showed a company that was finally making money. Today's amended filing adds the third quarter, and that profit is gone: revenue jumped to nearly $17 billion, up 47% in three months, but Anthropic swung to a $567 million operating loss, and it disclosed $895 billion in compute and infrastructure commitments, about $70 billion of it due in the next 12 months. Ejaaz breaks down why: demand outran the compute Anthropic had booked, so it paid up for $5 billion of short-term compute to keep serving customers.

We also cover the $2 billion donation-matching charge, where the $895 billion is going (SpaceX, Google, Amazon, Broadcom and the neoclouds), whether Anthropic can pay for it, OpenAI's cut-down fundraise and the junk-rating worry, and the bull and bear case for the Anthropic IPO.

==
Timestamps
00:00 Anthropic's Profit Is Gone
01:50 Catching Up on the S-1
04:28 The Q3 Numbers
05:24 Three Reasons for the Loss
06:54 The Restaurant Analogy
07:56 The $2 Billion Donation Match
09:19 The $895 Billion Bill
12:32 Where the Money Goes
14:18 Can Anthropic Pay for It?
16:38 Less Than 0.1% of Budgets
17:48 Is the IPO a Buy or a Sell?
22:14 Final Thoughts

#anthropic #claude #ipo #ai #stocks

==
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

What is In the Loop AI?

Exploring the frontiers of Technology and AI

Last week I went through Anthropic's financials and I was shocked.

Turns out Anthropic makes a lot of money. For every $1 of compute they invest,

they roughly make about $2 of revenue.

Except today they released their Q3 financials and turns out all that profit is gone.

Despite making $17 billion over the last three months, which is up 47% from

the last quarter, Anthropic owes one of the largest compute bills ever, $895 billion.

That is worse than the craziest bear prediction ever. And the worst part is

$70 billion of that is owed over the next 12 months.

So you have to ask the question, how is Anthropic going to pay for all of this?

And it turns out some of the more well-esteemed financial creditors are also

getting worried, assigning Anthropic a potential junk credit rating right before

the biggest IPO that we've seen in history.

Now, it's easy to argue the bear case of this, but what I want to present to

you is something that is maybe slightly more optimistic.

You see, the reason why Anthropic is losing money over the last three months

is because they weren't able to meet customer demand in a suitable fashion.

They had so many people trying to access their models that they didn't have

enough compute to go around.

So they to end up forking out a large premium to access that compute.

Now, without getting too much of the details, the idea here is if Anthropic

is able to secure enough compute over the next 12 to 24 months in these long-term

leases, which capture around $900 billion,

apparently, and demand stays sufficient, they should be able to theoretically

make a decent chunk of money.

That is still a huge amount of risk. And I want to dig into the bull and bear

case of what's going on given these new financials. So let's get into it on

today's episode brought to you by Qualcomm.

So just to catch you up anthropic released their s1 last week which gave us

the entire financials for 2025 up till about end of q2

of this year so we got a peek under the hood as to whether anthropic was making

money and a lot of people's opinions including myself was sort of skeptical

about this i didn't think that they were making as much money as i eventually

then found out plus i thought they were burning way too much cash

turns out the actual opposite was true anthropic had around $132 billion.

On their cash balance sheet, and they'd only burned around $7 to $9 billion.

And most importantly, the money, the revenue that they were generating was actually quite lean.

For every dollar of compute that they would invest, which is arguably the largest

CapEx cost to create these models and serve these models, they were making around $2.08 of revenue.

Inference margin on that is around 85%. So that's crazy for any company,

irrespective of whether it's an AI lab or not. But there was some very fair

criticisms about this, which was Q1 and Q2, everyone was super hyped about AI.

Loads of the Fortune 500 companies were doing something known as token maxing.

So trying to throw clawed models at anything and everything to try and generate

ROI, but that ROI hasn't quite yet been seen.

And so the main argument that was given was, if we see the Q3 results and it

backs up the Q1 and Q2 results, then we'll believe that Anthropic is making money.

The issue is, with the Q3 results that we just got, it shows that Anthropic

is actually making a loss to the tune of around $554 million.

And this ties into the general sentiment

that we're seeing not just with Anthropic, but other major AI labs.

If you look at OpenAI, the rumor has it that the most recent funding round that

they've been trying to raise, $30 billion, had to get slashed to $20 billion.

They were trying to raise money enough to fund the compute and capex spend over

the next couple of months before they eventually IPO, I think,

early 2027 is what the rumor has it.

And they had to slash that. But worse, they had to slash the valuation from

$1.5 trillion, which compared to Anthropik's $2.2 trillion is fairly conservative.

They had to cut that to $1.25 trillion because investors didn't quite believe

that their models were worth that much.

Now, putting that aside and looking at the crazy market dynamics that we've

seen in stocks recently, we've

seen semiconductor stocks draw down around 35 to 40% with near clouds.

And then we've seen Intel, Amazon, Microsoft, Google suffer major stock losses

over the last week or so, despite record earnings.

It kind of gives you the general sentiment that people are getting sort of concerned.

And indeed, like lending rates for all of this humongous CapEx build-out is getting too expensive.

So without getting too much in the financials, what does this mean for Anthropics

IPO, which is rumored to go live in like two weeks?

Well, first, the headline numbers. They made $17 billion over the last three

months. That is up 1,127% from that period last year.

And it's up 47% from just three months ago. So they are still rapidly increasing

their revenue rate, however way you want to look at it. And that means in the

last nine months, they made around $33 billion.

Over the same period last year, they made around $2.5 billion.

So we've seen a 13x increase.

By any means, the claim that Anthropic is still one of the, if not the fastest

growing company that we've ever seen in history still stands.

But the question mark then becomes, is that even sustainable in the first place?

Last quarter, or rather last week when we saw the S1 in Q2, we saw them make

around an operating profit of, I think it was $974 million.

Fast forward to today where we see their Q3 results and we see that they've

actually got an operating loss of around $564 million.

And there are three big reasons as to why they suffered that loss.

Number one starts with $5 billion of short-term compute payments that they had to make.

This wasn't estimated previously in the S1 because they didn't expect the overwhelming

demand that they receive for their model. So let me kind of put something into

perspective here. Anthropic and any other AI company needs to purchase access to compute and,

before they know how much demand they're going to get. Think about it.

If you try and predict the demand that you're going to have for your models,

it's token consumption-based.

What I mean by that is you release a new model and you hope everyone likes it

and you hope everyone wants to consume as much of it as they can.

And that applies for the most expensive model, but it also applies for the most cheapest model.

And if you look at OpenAI, if you look at Anthropic, you realize that not only

do they just release the latest, greatest, most expensive model that is the

most intelligent, they also release much cheaper versions of that either at

the same time or a few weeks later.

And the idea here is it should cover all cost bases and customer types.

If you wanted to code a company or business from scratch, you shouldn't need

to have to spend so much money or use the most intelligent model all the time.

So the idea here is it's very unpredictable to figure out how much demand you're going to get.

So Anthropic invested a lot of money to get access to compute for Q3.

But they underestimated it. And so the idea here is they needed to get emergency

access to compute, of which they had to fork out $5 billion.

Now, to give you an easy analogy of this, imagine that you run,

I don't know, hottest restaurant in town, right? And everyone wants to book

a table and you're booked out day in, day out.

But one night you run out of ingredients. What are you going to do?

Are you going to turn people away and say, sorry, we can't serve you?

Or are you going to run down to the local supermarket and buy the ingredients

off the shelf at retail price, even though it's eating into your revenue margins,

just so that you can serve up the product.

That is effectively what Anthropic had to do. And if I had to guess,

they were probably paying SpaceX, which is their primary short-term lease compute

provider, a very hefty price to get access to this compute.

So just because of this one maneuver, it meant that the compute costing around

48% of their entire revenue, which gave them the $1 in, $2 and eight cents of

revenue out changed to 58%.

So roughly for every dollar of compute spend that Anthropic invested in Q3 resulted in around $1.72.

Still pretty good, but still notably declining. And you want to make sure that

that isn't a continual decline. Now, the second thing that Anthropic is paying

for that is going out of pocket is a little weirder.

It's $2 billion of charitable donations matching for their employees. Now.

I think about two weeks ago, I made an episode on how Anthropik is basically

preparing for mass unemployment because they created this new charitable foundation,

which is committing, I think, upwards of 10%

of the entire company's equity valuation for the public good.

The idea here is they want to distribute their shares and value of the company

that it generates to the wider public because they're worried that people won't

necessarily be able to earn a livable income.

And so they wanted to create some kind of universal basic income or universal

high income to distribute the goods and wealth that AI is going to generate.

This is that first step towards that. And it's going to the tune of $2 billion.

The reason why this is significant is $2 billion on any company's balance sheet is notable, right?

The fact that this is a charitable donation that is eating up against their

margins should be noted.

But it should be kind of considered separately, I suppose, which is the sense

that if Anthropic is likely going to be doing this consistently,

but probably not to that large of an extent, we probably won't see that as a consistent line item.

So we should probably put that aside and just focus on the compute costs,

which brings me to the third thing, the estimate of how much they're going to

spend over the next five to 10 years, which is roughly around 895 billion.

We were discussing rumors, I remember, on this show as to what this spend might

be like. And the initial rumors was $500 billion, half a trillion dollars of spend.

That alone, I remember reacting to at the time, was huge.

It is humongous by any means. And I had no idea as to how Anthropic was going

to be able to pay for this.

Now, the worst bear case prediction was around $700 to $750 billion.

Well, it turns out that the figure is way larger. At $895 billion,

you have to ask the question,

has Anthropic kind of lost control of this? So I guess the way to look at this

is let's actually break down how much of that is going to who and how much of

that needs to be paid up front, because the immediate payment over the next

12 months is $70 billion.

Now, the first thing I want to call out is that the total figure is actually not $900 billion.

It's probably more around $1.05 trillion, because there's around $150 billion.

That haven't been priced in yet. These are deals that they have signed in the

early part of October. So Q4 Financials is probably going to go crazy.

But the idea here is it's a very large sum of money. Now, $320 billion of that

is signed contracts where Anthropik is like, listen, I want to get access to

this compute with the likes of SpaceX, with the likes of Google,

with the likes of other neoclouds.

Now, around $415 billion of that are like leases, which is, OK,

if you build the data centers and you're able to kind of deliver that compute,

we will pay you for this service.

Now, I want to point a very obvious thing out, which maybe a lot of people don't understand.

This is Anthropic renting compute. It's not them owning the compute.

Owning the compute would mean that there's like a lot of debt on the balance

sheet. This is them renting it. So it's only dependent on if these companies,

these neoclouds, these hyperscalers, such as SpaceX AI, are able to deliver on the compute.

But a lot of this is still meant to be paid for upfront.

I think the figure is around 40 to 50%, which is gargantuan.

And $70 billion of that is due over the next 12 months and $447 billion is due by 2030.

Now let's address the $70 billion first. That sounds like a big number.

But when I look at that cash balance sheet from last week's S1,

they have around $132 billion.

Okay, so they should be able to cover that. And then if they IPO over the next

couple of weeks, the rumors say that they're trying to raise around $100 billion,

that should cover that with a bit of extra room, certainly, right?

But they keep signing more expensive leases and crazy compute deals.

So the massive bet that Anthropic is taking is

if we sign enough of these long-term compute deals

and we keep generating enough demand for our products, for our models,

and so we generate increasing amounts of revenue quarter upon quarter,

we will end up being able to net a much higher margin for the amount that we're investing.

So it seems like a big deal up front over here, but it's eventually going to

end up in us earning way more profit than if we were to buy the compute in response

to demand, which they did over the last three months and ended up eating into that profit.

So they're trying to work against that scenario actually happening.

Okay, but where is this money going? Well, you have around $121 billion going to SpaceX.

And as you know, Anthropik's paying them around $3.1 billion per month.

Elon Musk is rubbing his hands and smiling gleefully because he's making a ton

of money. It's what led to the first positive revenue earnings of SpaceX a few weeks ago.

Now, another $120 billion is going to Google for access to their TPUs.

Remember, Anthropic has one of the most diverse GPU clusters,

meaning they aren't just reliant on NVIDIA. They're reliant on Google's TPUs,

which is their own versions of GPUs, as well as a number of different ASICs,

which are specialized versions of GPUs.

They're also working on their own chip in partnership with Broadcom.

So the idea here is if they have this diversified stack they should be able

to continue training Claude in as many different ways with as much flexibility

as they can versus someone like.

OpenAI which arguably has a lot more concentration risk on NVIDIA.

Then you have another $110 billion going to Amazon and you have around another

let's call it $100 billion going to say NeoClouds and then the Broadcom deal,

which I mentioned just now, is around $125 billion.

So all big figures concentrated to around, I want to say, four to six companies.

And the circular economy sort of keeps on going there. I want to call that out,

right? Because a lot of these companies are invested in Anthropic.

You've got Amazon and Google, which I think collectively own around 25% of Anthropic pre-IPO.

And then you've got Google owning a significant chunk of SpaceX.

You've got SpaceX kind of in bed with NVIDIA and Jensen Huang securing all the

compute. You've got NVIDIA and Jensen Huang backstopping a lot of the NeoClouds,

such as Nscale, which is providing $45 billion worth of compute to Anthropic.

So again, a lot of circular motion going on.

And the question is, can Anthropic sustainably pay for all of this?

Well, I answered the short-term, near-term thing, right? Which is they have

around $130 billion in their cash balance sheet.

Probably they'll be fine for the next $70 billion that's owed over the next 12 months.

But what about the remaining $810 billion over the next couple of years?

Well, that's the bet that Anthropic is making, and there's no easy way to say this.

They are betting that the continual investment in compute, in training better

models, in achieving recursive self-improvement, aka the model that is so far

exceeding of human intelligence that it is able to automate a bunch of things

and generate so much wealth for humanity,

that model, if they are able to build that successfully, then they're going

to generate endless amounts of money.

That is a big, risky, and some might say far-fetched bet.

Now, the bare case is they aren't able to build that in any short time span.

And we're already seeing kind of these AI model labs restricting the release of their AI models.

Rumor has it that Anthropik's been sitting on Fable 5.5, which is the codename

of their next big model drop.

OpenAI has been doing the same thing because it hasn't been able to sufficiently

satisfy alignment procedures, which means that if you release the model right

now, it would go out and do a bunch of crazy shit like cyber hacks,

bioweapons, exploits, and stuff like that.

So the idea is there's this push and pull. Regulations are getting involved.

And most importantly, data centers are getting harder to stand up.

I mentioned on an episode last week that there's hundreds of millions of dollars

worth of GPUs collecting dust in warehouses because we haven't got the power

or the energy grid or just the technical expertise to stand these up.

So the idea here is things might potentially slow down, which means that Anthropic

makes less revenue, which means that they aren't able to meet their compute

bills, which means that this whole thing, this house of cards ends up collapsing.

Now, I want to point out that this is a very big risk to take and there are

great arguments to argue why Anthropic won't be able to pull this off and great

arguments to say that they will. It's my inclination to.

If we're able to continue this acceleration, and we're seeing it kind of recently

on the periphery where you've got internal open-air models solving a lot of

hard, complex challenges in math and physics.

You've got Anthropik models doing the same in medicine.

Imagine a world where these models are able to do so at scale to many other people right now.

You look at the spending budgets of Anthropik as well, right?

Where you see that it was less than 0.1% of the Fortune 500 customers that Anthropik has, which is.

Quite a lot of them, have allocated that much to Claude Models. So less than 0.1%.

That is a tiny amount.

Imagine if that scales up. By the way, what that 0.1% looks like is around 11,900

individuals or companies paying around $100,000 a year.

And then you have around, I think it's like $200,000 to $300,000 paying $10 million per year.

Now, that was about a 600% jump up from that same evaluation from three to five months earlier.

So can you imagine what that looks like, I don't know, a year,

a year and a half from now?

You can imagine that that 0.1% budget spend is going to increase to,

let's say, 1% or conservatively 5% of a company's spend, which I don't think is too crazy.

And if a lot of that accrues to Anthropik, that's going to be a large amount of revenue increase.

So that's the optimistic take. The negative take is they won't to be able to

pull this off and we're going to grind to a halt because we can't get the power

to supply any of these gpus in the first place which i guess brings us to

the final part of this episode which is the bull and bear case for anthropics

ipo is this a buy is this a sell,

is this something to be optimistic about or something to be concerned about

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

My first point is the question has sort of changed from last week.

Last week, the quick question was, does AI make money?

It was answered pretty explicitly. Yes, Anthropic is making a heck ton of money,

and they're doing so at a crazy growing rate.

But now the question is, given this operating loss, and given the explanation

of why they're making that loss, aka a lot of compute expense,

can the demand come first before Anthropic has to pay the compute bill.

In other words, can they generate enough demand for their clawed models before

they have to pay the compute bill?

So they need to make enough money to be able to pay it off before that bill arrives.

Otherwise, you know, they run out of money. And it reminds me of something that

Dario mentioned earlier this year, which he said, and he was kind of boasting, I guess, in a way,

where I think it was, I forgot what podcast it was, but he basically said,

I pride myself and Anthropics self on the ability to not go all in on compute.

And he said this at the start of the because he said that is quite a costly

expense and we have saved ourselves a lot of money and damage because we weren't able to do that.

Of course, he ate his words a few weeks later, a few months later,

when it was very much proven that scaling laws for training these models were

very much intact, and that he needed way more compute in order to catch up.

And that's, in fact, what allowed OpenAI, who was back then behind on models,

allowed them to catch up to clawed models at the same proficiency because they

invested so much in compute upfront.

So what seemed like a pretty bad decision for OpenAI three months ago ended

up becoming a really good decision back then in January, February time.

So Anthropic, in the same vein, is making a very similar bet.

And the answer to that will only be seen over the next couple of quarters.

And it's something to definitely track.

I actually think the best metric to track for this is what I mentioned earlier,

the number of companies that are spending around $100,000 to $10 million per year.

Are there more companies like that every single quarter?

If that number keeps going up, then that tells us a very obvious thing,

which is there's insatiable demand for this thing and the enterprise customer

base, which is Anthropix bread and butter, is increasing in the way that you want it to.

Now, the second thing to point out is, I guess, in favor of both the bull and

bear case. On the bullish side of things, you could argue that.

Yes, Anthropic made an operating loss, but it was because they had so much demand

that they had to pay out of pocket for more expensive ingredients,

if you want to use my restaurant analogy earlier.

The bad case is Anthropic didn't even have a cushion or figure out a way to

kind of satiate this demand or anticipate this demand, and that might end up with them blowing up.

And like, listen, this is a high stakes game that Anthropic is playing.

They don't have alternative revenue earnings from any other parts of a business.

They are an AI-focused company, and they are betting the entire house on it.

If you compare them to Google, Meta, Microsoft, they have more flexibility.

Google has search revenue, which I know is going down slightly,

but they're making billions of dollars from that.

And they have so many other product lines. You've got Meta with the same scenario as well.

And so they should be able to kind of keep this expense afloat.

Anthropic does not have the same sort of flexibility, neither does OpenAI.

And the other thing I want to tack on to that is market sentiment is probably

at its lowest that it's been since this entire AI hype started,

I don't know, a few years ago.

So the idea here is, have we reached the breaking point where no matter how

much good news a company can sort of surface, it's still going to be bearish

from a public investment side of things? At the end of the day,

Anthropic needs to raise $100 billion in roughly around two weeks.

Will they be able to do it in the light of their second biggest competitor,

OpenAI, not even being able to raise $30 billion?

And we heard from Elon Musk last week that he even struggled to raise the $40

billion that he needed to buy more compute. And he has a very obvious business

line where he just sells it at a premium.

So it's obvious that the market is questioning whether this thing has reached its breaking point.

But on the positive side, so

little of the world has actually been penetrated by AI in the first place.

I think you could feasibly estimate that if they remain the top two AI labs

in the world, they'll be able to capture a humongous amount of this revenue

that would be able to pay for a lot of this cost.

Listen, when I step back from all of this, I want to point out that I am probably

one of the bigger or biggest AI bulls out there.

But even I am a little skeptical as to how much money is going into this thing.

And I want to acknowledge that it is a humongous bet.

Now, I naturally lean on the side of thinking, if you are a long-term investor,

then this is something that is obvious for you.

Directionally, AI is going to go out there and change so many things.

Even if you're not a massive believer in centralized labs, if you're a believer

in open source or whatever that might be, the idea here is simple.

AI is going to improve in intelligence and it's going to decrease in cost,

which means that eventually it's going to be able to run on any and all of our devices.

We're going to have personal agents that we can interact with at any single

point in time that can do myriad different things for us. It's not just going

to be work-related, it's going to be leisure-related as well.

And so in that world, are you going to be net-net more productive,

net-net more successful, and make more money? What are your odds of that?

Or are you going to lose more money and consume many of these products.

It could go either way. I tend to think that this technology is going to generate

an enormous amount of wealth, and we need to be tentative in how we're spending

and make sure we don't over leverage.

But we do need to fund the build out. It is now more important than ever.

But equally so, if you are maybe more short term minded, and you're concerned

about the inflated valuations of a lot of these NeoClouds, the amount of money

that's being spent on computer, most importantly, and this is,

very understandable and valid, that a lot of these GPUs and compute aren't able

to actually come online as fast as we wanted to. And so there's naturally going to be a slowdown.

Then that's fair. And you should sort of invest and trade accordingly.

Again, like none of this is financial advice. It's just my kind of thoughts on this.

I'm more long-term minded. I know a lot of our recent episodes have been market-oriented,

but hey, the market's sort of calling for it.

Stocks are down pretty massively. And I want to inject a bit of optimism,

I guess, into this entire thing. We are going to see one of the biggest IPOs,

whether you like it or not, in a few weeks' time, if I had to guess.

My bet is a lot of the numbers that we are speculating to be bad a year from

now is actually going to be a lot more surprising.

I think it's going to move towards the upside. And if I had to bet where those

numbers of Fortune 500 companies, number one, spending a larger percentage of

their budget, certainly greater than 0.1%, but also two, spending tens of millions

of dollars versus just $10 million

a year on Anthropik models or just AI models in general,

I will bet you that it's going to be much, much higher a year from now.

And yeah, that's it. Thank you for listening to today's episode brought to you by Qualcomm.