# Anthropic S1 Economics and AI Market Uncertainty

**Podcast:** Another Podcast
**Published:** 2026-11-03

## Transcript

Hi, I'm Tony Cowan-Brown.
And I'm Benedict Evans.
And we have the long-awaited Anthropic S1.
Where should we?
90 pages, correct?
Yeah, 900 or something.
Nice and short and sweet.
A Rorschach blot, if that's the term, or a Rorschach test.
What is it that you want to see?
Yes, I like that.
What are you interested in?
What does it show?
There is, I think...
Well, let me kind of break this apart.
First of all, there's the unit economics, which had actually been much leaked and are now confirmed.
Secondly, there's all sorts of operating detail throughout it.
Thirdly, there is all of the future commitments that they have made, both for capacity and now they say that they're going to try and build their own infra as well.
And then there is, okay, Great, but what does this actually tell us about?
What's going to happen next?
And I kind of think that's the most interesting one in that, you know, if I was a CFA, I could sit and go through, you know, every line of the 300 pages of notes to the accounts and say, aha, in fact, I'm going to do a podcast with a financial analyst next week.
And he will probably want to do that.
I think the.
But I think the sort of TLDR of this is that right now, when we have an enormous shortage, and hence you can price all the way up to your customers ROI, they are just about breaking even.
However, and that's grown enormously in the course of this year, which we can talk about.
But meanwhile, we have the cost per...
result dropping by sort of 10x a year on the one side from more CapEx and more efficiency improvements in the models and the chips.
So clearly there's going to be massively more capacity.
And on the other side, agentic ads, pick your number, 10x, 100x, more token usage.
And the usage today is still really, really narrow.
Like there's a billion people who use this a little bit and a few million people who are using this for agentic coding.
And we don't have half a billion or a billion people doing something on the scale of Instagram on this thing.
So it's pretty easy to imagine that some arbitrary point in the future, something happens that means the usage goes up by orders of magnitude as well.
So you've got enormous growth in capacity and enormous growth in usage, neither of which you can really forecast.
I mean, you can, but not seriously.
So you don't know where supply and demand are going to be in two and three or four years' time.
Secondly, the models themselves are more or less commodities.
And certainly if you're deep into this, as again, we've discussed, you can argue about our workload is better than Gemini for this.
But in generality, everyone is looking at using two and three models and you don't have pricing power.
You have pricing power in the sense that an oil company has pricing power.
There is shortage.
I can name my price.
But you don't have pricing power in the sense that people have to buy my oil and not their oil because it's all the same.
And yes, I know different kinds of oil are different, but like.
The point here is, these are pretty fungible systems.
And so you don't know where supply and demand is.
Right now, you don't have differentiation on your own terms.
We don't know if that will change.
We don't know if a winner-takes-all effect will emerge.
This is why people talk about recursive self-improvement.
Maybe that produces a winner-takes-all effect.
Maybe it doesn't.
And we also don't know whether all of those use cases that explode will be things where people are actually using the model as model or whether it's just infrastructure underneath and it's somebody else has built the application.
And so all of which is to say you don't really know the profitability of selling tokens in three years time.
You also don't know the value capture of the model companies themselves.
And so you've got this kind of snapshot of where we are right now, but we have no clue what any of the bits of the algebra of that are going to look like going forward.
Is there anything in the document that you read as, OK, well, this actually makes sense and this is the reality as it is today?
Or does everything feel like a lot of this is guesstimations?
Because they have to be.
Oh, so the numbers in the thing now, I mean, these are all audited numbers.
I mean, the numbers are the numbers.
The point is that you can't really draw straight lines from this and say, well, this is what's going on.
And I think it's very kind of instructive to compare this.
with, as you may have said once, I'm a big fan of comparisons that are unfair but relevant.
And it's kind of instructive to go back and look at the...
What are you going to compare this to?
I'm intrigued.
Well, I was, I mean, you know, unfair but relevant is to go and look at Meta and Google when they are IPO'd.
And so this is where we should talk about unit economics.
So what's been very widely leaked, So first of all, the two things that have been leaked are leaked and or released as these kind of weird annualized monthly revenue numbers.
And so now we have actual monthly, we have actual quarterly numbers.
We have actual revenue, like real revenue numbers.
And you can argue a little bit about whether it was slowing down.
But, you know, September last year, Anthropic did 1.4 billion.
And September this year, it did 16.8 billion.
numbers are growing.
That's great.
People are using this.
People are buying this.
Now, within that, yes, what percentage of that do they pay to their distribution partners?
Fine.
That's a separate conversation.
But there is enormous willingness of real people to spend real money on this stuff right now.
So that's the first point.
But then you look at the operating economics.
And what you see is fascinating that the cost of sale, which is to say the cost of running the models, which is partly compute and partly people, is they have a gross margin on that of 14.5% of 85%.
So when Google went public, it was more like 87%, 88%.
And when Meta IPO'd in 2011, it was more like 77%, 78%.
And Antropic, 85.
Antropic, 85.
So, I mean, these are all, you know, close enough for government work.
These are all good, close.
It's all kind of the same number within France.
Yeah, so they're spending sort of 15 to 20 percent, all of those companies spending 15 to 20 percent of their revenue on actually serving the existing business.
Then you look at the R&D, which is the cost for Antropic, is the cost of building the next model.
And.
At Google, R&D was 9.5%.
Meta is 10.5%.
Anthropic is 60%, 60.5%.
Sorry, 59.5% in H1.
So plus the two of those together, 75% of revenue is going on running the existing model and building the next model.
And then they say that 80% of Q3 revenue came from models built in the last six months.
So this is why you've had people comparing this to a mining company or a pharma company, in that actually running a gold mine is really, really profitable, but you've got to dig another one and that's really expensive.
And for Anthropik, the gold mine only lasts for six months or eight months or a year or something, and then it becomes irrelevant because the pace of the frontier is going forward so much, you've got to have the newest model.
So then you get questions which are...
How does the ratio of that change?
Because in the second half of last year, Antwopik spent $4.5 billion on training.
And in the first half of this year, they spent $9.7 billion of training.
So they not exactly doubled, but on the way.
And so the funny thing is, if you put those on a chart, basically the...
The fixed cost seems to go up with revenue and the marginal cost seems to stay flat with revenue, which is kind of the wrong way around.
And of course, there's things you don't know within that.
Some of that is discretionary, that Anthropic decides to spend that much money.
Some of that is that that's a competitive pressure, that that's what it takes to build the next model.
Some of that is, you know, obviously the pacing, the frontier.
If we were building the next model every 18 months instead of every six months, then those economics would be transformed.
But putting all of that together, like it comes back to what I said at the beginning, like we really don't know how the economics of this are going to evolve because the economics now look completely different to the first half of last year.
The first half of this year looked completely different to the first half of last year.
And is it fair to say that a big part of this IPO also is about bringing money?
We've talked about this previously, but bringing money that's outside of the loop, because right now the money seems to be going in the circle of the, you know.
the companies Anthropic rents its computers from also the one backing the revenue is a big part of that IPO.
Just like, can we get other people to put money into this?
So, so then there's a whole, so there's a whole second, second conversation, which is the amount of long-term commitments that Anthropic has made to purchase capacity in the future, which is.
Hundreds of billions, trillions of dollars, depending on how you add something over a trillion dollars, depending on how you add it up, most of which is non-cancellable.
OK.
Which is kind of moot, I think.
OK, why?
If your business collapses and you don't need that compute anymore, you're going out of business.
Regardless, yeah.
Regardless.
So whether you file for bankruptcy or you cancel a contract is what, like, you know.
The money's not there, so yeah.
You don't want to be the reasons why you're not paying a kind of secondary to the fact that you no longer need that capacity.
Now, of course, there's another scenario that says, you know, suddenly, you know, the growth collapses and the efficiency gains and they can deliver the same.
The revenue flattens out.
They can deliver the same thing with that 10x less compute every year, but they've still committed to buy all of that compute.
Fine.
So then that's good.
But again, like.
You wouldn't really expect Dario to bet that this business wasn't going to continue growing.
That's all sort of a second point, which is if the growth collapses, fine, then they're screwed.
But they're not worth one and a half or two trillion dollars if the growth collapses anyway.
So like second order problem.
I think then you can get into a lot of other operational questions here.
So they are paying.
Something like, I can't do the numbers in my head, something like a quarter of the revenue is going on distribution, which is mostly distribution partnerships with Amazon and so on.
People who are reselling the business, you'd argue that should be in cost of sale.
That clearly is a marginal cost because it's basically a rev share.
And there's obviously this kind of weird, complicated relationship in that those companies are both their suppliers and their partners and their shareholders and their competitors.
which is the famous circular revenue point.
But, you know, a lot of that stuff, and this is kind of a point I've made previously, sort of looks to me like leverage in that if everything goes well, that will all be fine.
If everything goes badly, then that will just massively magnify the degree to which everything slows down.
And you see that in the non-cancelling contract, cancellable contracts, and in everything else.
But don't ask Dario to hedge on whether AI is going to work and be a big deal.
Like that's one argument that nobody in Anthropic is taking seriously.
So like you kind of you would expect them to have made those decisions.
I think the challenge in here is nothing in this really tells us how well this is going to work.
Because if they weren't selling tokens at a profit today.
when there's massively more demand than supply, when everyone in the software industry is running around with their hair on fire saying, this is amazing, we want this, how could they not be making money at this level?
If they weren't, they'd be run by idiots.
And so this is another of those sort of rebukes to these people who make these kind of non-falsifiable statements, or in fact, falsifiable statements, like people who spent the last three years saying none of this is going to work and they're not going to raise any money and they won't have any revenue.
And if they do, it won't be very much.
And if they do, they won't make a profit.
Like, here we are.
People are buying, spending a lot of money on this and they're making a profit.
The problem is they're making a profit at today's prices.
And we don't know how any of the algebra in that is going to move in the next few years.
And when we talked about this a couple of days ago, you were talking about exactly what you're saying here.
We don't know the value of what the...
We don't know the value yet of agents.
It seems like our agents are good business.
More people are using the chat feature than actually building agent models.
But the agent model is the bit that is using a lot more compute that's probably a lot more expensive to build.
Is that an interesting even conversation of just we don't know the value of what we're...
what's coming next value of chat versus value of agents one being much easier to build and cheaper to build yeah so there's a there's a general point and a specific point in here i think okay the general point is um agents use way more tokens instead of you asking a question and getting an answer you tell the agent go work this out, go do all these different things, go all this stuff.
Build all this process.
And so they use way more tokens.
Now, tokens do look a bit like talking about bits in like 1998.
Like there's more bits, global data traffic on the internet is growing.
Well, great, but that's not a very close connection to value or utility or adoption.
But then you get a debate, which is okay.
So first of all, agents work for software development and they work inside enterprises.
Is that take a consumer takeoff so mark zuckerberg on the last meta quarterly call said you know we've got unquestionable we clearly have product market fit for a for muse and you kind of look at this and you think wait so hang on let me just pull exactly what the numbers were um so muse has um 10 million people have set up the app of which 25 million people use it on any given day a quarter of those Which app is this?
Mews.
Meta's Mews.
Okay, yep.
And a quarter of those early users, not sure if it's a quarter of the 25% or a quarter of the total, are making e-commerce purchases at an average price of $200.
No mention of how many, no mention of what daily active use is, because 25% users on any given day is not actually daily active users.
He also said he can't scale it until compute costs fall by at least 10x.
And yes, compute costs are falling by 10x a year.
So it's like...
kind of product market fit, but actually you've got like two or three million people using this.
Not quite sure that's product market fit.
But stepping forward from that, first of all, you could say, well, look, if I go and use this in Muse or Gemini or Siri or whatever it is, then that's value that accrues directly to the model builder.
And certainly so Meta thinks they'll be getting commissions on those purchases.
That's their revenue model.
And so that answers a bit of the value capture question.
You could also argue, no, if I want somebody to optimize my financial life, I'm going to do that for a financial services thing.
If I want an agent to work out what flights I want, that's going to be a travel-y thing.
If I want something to work out what products I like, that might be meta, but meta won't also be doing the other things.
You could argue, so in as much as to say, it's not quite clear whether the agent is vertical or horizontal yet.
No.
At anthropic level, there's another problem here, which is that meta sees quite a lot of you and what you're interested in.
Well, let me wind back.
How is it that you would know that the agent could do that for you?
Option one is you think of that, you have the imagination to think that that problem is an agent problem and you ask your preferred agent to do it for you.
Option two is the agent is watching you and says, hey, I've seen your emails about this.
And did you realize that you could save this much money if you did that?
Buy me a coat that I would like.
Well, Meta kind of might be able to answer that question.
Claude probably can't.
And the thing that I'm getting at here is that Apple and Google and Amazon and Meta can all sort of see some of what you're interested in.
They're like the blind man feeling the elephant.
They feel different bits of it.
But they each have a view on some of the stuff you're interested in.
Anthropic and OpenAI, they only know what you do in the agent.
They don't see your email.
They don't see what pictures are you sharing.
They don't see all this other stuff.
So they have much less chance, A, to do the passive.
both passive and proactive, passively watching and then proactively suggesting something that you wouldn't have thought of.
It is much harder for them to do that because they have much less of that context about you.
They also have much less context about what you like.
So you can say to Siri, what was that restaurant that my wife talked about last week?
But you can't say that to Claude because Claude doesn't have your messaging, any kind of messaging.
And then Muse has some of your, Meta has some of your messaging.
Google has some, Apple has some.
OpenAI and Antwopic don't have any of that.
They have none of your context.
So there's a bunch of sort of questions around, okay, how do you know what to do with an agent?
Where does the agent sit?
Is it vertical?
Is it horizontal?
What kind of access does it have?
For all of that, that's sort of super tough.
It's harder for Antwopic than for anyone else in the space.
It's interesting.
I'd never thought about it in that lens, yeah.
Yeah, because you're asking it to be proactive on stuff that you're not giving it.
Hence, there's a barrier there regardless.
Yeah.
So I think you can look at agents and say, this is an answer to the blank screen problem.
Firstly, in that you have many more things where you have a task you want solved than you have a question you want answered.
And secondly, that there's scope for these things to be proactive and suggest things.
Where do you ask them?
What are the right levers for that to happen?
These are all the sorts of questions we're used to asking like 97 about the internet and like...
2001 and then again in like 2009 about mobile.
Like how is this going to work?
Who's going to control this?
Where are the levers going to be?
Why isn't Yahoo going to own the whole thing?
Why isn't AOL going to, all the telco going to own the whole thing?
There was this whole time when the telco said, well, we're a smart pipe.
We know all this stuff about you.
And it turned out, no, they didn't.
And they couldn't use it anyway.
So we're at that moment where like you can squint and see this might be the thing.
This might be a thing that people do.
But how is it going to fit together?
We don't know.
Can you plot that back against Anthropic?
No.
Back to my point.
This is a snapshot of the unit economics in the summer of 2026.
That doesn't tell you how any of this is going to work.
It feels like one maybe last question that I know we touched upon a little bit.
But what does Wall Street need to see to keep funding this?
Where do you think their head is at?
Like what, you know, what happens when we run out of capital?
What's next here?
Well, so one of the things I think I've been saying, I can't remember we said it on the last podcast, is that there's sort of two fundamental changes in AI this year.
One of them is the massive product market fit of software development, AI for software development.
And the other is that the CapEx is outgrowing the free cash flows at the hyperscalers, so they can no longer fund this out of their existing businesses.
And so depending on who you ask, you've had something like, where's my slide, something between half a trillion and a trillion dollars of debt issued around AI this year.
There's well over a trillion dollars of leases, well over a trillion dollars of future purchase commitments.
Google, the hyperscalers on one side and then the neoclouds and Oracle on the other side have had to go to the capital markets.
And so that now means that it matters what the capital markets think, which it didn't in 2024.
So that's the primary observation.
Secondly...
It's funny, like I remember when I was an equity analyst, I would talk to the credit analysts and realise that I had no idea what they were talking about.
They were just interested in completely different things.
And looking at a problem or something from completely different angles.
Yeah, but look, yeah, it's a completely different question.
Because as an equity analyst, as an equity investor...
what you want to know is how much is the share price going to go up.
Whereas as a credit investor, what you care about is do they have the cash flows to pay me?
And you don't care how well the business goes.
Or you don't care how well the new thing works out as long as they still have the money.
And so you've got, as a very not equity analyst for 20 years and never a credit analyst observation, is that if you're lending money to Google to build a giant data center, It doesn't, whether that data center produces any return at all is kind of not your problem.
As long as the cash flows from the ad business continue, which for the time being, they clearly are.
Yes, there's a wobble last year, last quarter.
There's almost like a kind of a fun of catch 22, negative catch 22, in that if AI doesn't work, then Google's ad businesses are just fine.
So it's like heads, you lose, tails, I win.
And that's not entirely true, but it's kind of almost true.
But the point is, as a credit investor, you've got a very different attitude to Google, to Oracle, to a NeoCloud, because you've got all of these external cash flows from the ad business that are propping this whole thing up.
So even if the whole AI thing ends up as a smoking hole in the ground, which clearly isn't, but even if it did, you still got all that cash flow to pay your debts.
Whereas for Oracle or for a NeoCloud, if it turns out you can't make any money from an AI data center, well, then you're just going to have to go home.
There's no pivoting from that.
No, I mean, there's this sort of slightly bizarre argument from NVIDIA that, you know, the chips are fungible.
So if this NeoCloud goes out of business, they can sell the chips.
And you think, but if they've gone out of business, it's because you can't pay the interest by selling the chips.
So no one else is going to either.
I mean, explain this to me like I'm five.
Like, I've bought this thing.
It's like, imagine I'm an Uber driver and I've...
bought a car with a loan, but it turns out I can't get enough Uber rides to pay the loan.
And the person who's given me the loans, who sold the car says, well, that's okay, because somebody else could drive the Uber and pay off the loan.
No, that's not why I can't pay off the loan.
If I can't pay off the loan, it's because there aren't enough Uber rides.
So they're not going to be able to do it either.
You know, there's a correlated downside here, put it like that, which is why...
Everybody pushed back massively on, I can't remember which neocloud it was that said they should be AAA rated because we can sell the chips and the chips are gold-plated collateral.
Well, yeah, but no.
I think part of what's been going on, and again, I'm not going to do financial markets commentary really, but part of what's been going on in the last couple of weeks is a question of leverage.
in that everything, take everything that I've just said about how uncertain the future of Anthropic is, you know, that's a variability from like a $100 billion company to a $5 trillion company making these numbers up, depending on where you think that's going to go.
The same thing applies to Amazon, Google, Meta, Oracle.
The same thing applies to NeoCloud.
It's the same thing by extension applies to NVIDIA and to all the chip companies.
You've got a very wide range of possibilities.
You've got very, very elevated valuation ratios.
I mean, not relative to the dot-com bubble, but the numbers have gone up a lot.
And then you are hunting for signal.
And so anything that looks like it might be the, oh, no, it's not that, it's this, gets jumped on, which is why you get volatility.
And instead of writing my column for the week, and it occurred to me, the sort of phrasing is like, you know, on the market scale, it's kind of bipolar.
On the one hand, everything's great looking at the numbers.
Look at all these numbers going up.
And on the other hand, everything's terrible.
Look at all these numbers going up.
Yeah, and it comes down to the thing that we always talk about.
What's the value here?
What are we building?
Depending on which numbers you're looking at.
I mean, there's a very, very old stock market saying that there's no such thing as a bad company, only the wrong price.
And back to the beginning, OK, the revenue on this thing is exploding.
Really don't know what's going to happen to the pricing power around that.
We know the demand is going to go up massively.
We also know the supply is going up massively.
We do not know what the balance of those is going to look like.
We all do not know whether you're going to have differentiation in your own terms.
We do not know whether you're going to have pricing power.
We do not know whether you're going to build something on top of this that everyone has to buy from you rather than from anybody else.
We don't know what's going to happen to, you know.
the underlying science behind all of this, if all of this just kind of slows down and it turns out that the models flatten out next summer, then that produces a different outcome to if the models keep growing for another two years, three years, five years, 15 years, like we all become pets.
To your point, we just don't know.
Yeah, now try and value that.
Now, a long time ago, I started my career as a telecoms analyst and I remember chatting to an analyst at a bank, a big bank that no longer exists.
Deliberately, that could be three or four companies.
And he said that his model for Colt Telecom, which was a European fibre carrier, had a section entitled value I can't capture in the DCF.
Meaning my DCF says that the price is 50 quid.
The share price is 400 quid.
My target price is 450 quid.
Fair enough.
There's value I can't capture in the DCF.
That's not quite what's going on here.
The problem is it's like it's a different problem, which is you just don't know what growth rate to put into your DCF.
You can't know.
Yeah, yeah.
It's not someone hasn't done their job.
It's just it's impossible right now for us to calculate that.
Yeah, which is a friend of mine that is a semi-sanalyst at one of the...
world's biggest fund manager is a portfolio manager, used to be a semis analyst.
He said, I'm really glad I'm not a semis analyst anymore.
How can I predict what SK Hynix's revenue is going to be in three years?
I can draw straight lines on a chart.
I can draw straight lines on the chart.
I can draw a straight line on the chart around Anthropic.
I can model.
It's like, as I used to say, it's like trying to model broadband in 1997.
There's four or five lines.
Shut your eyes for a minute and think, well, what would you do?
Well, you can imagine there's four, five, six lines you'd make in the spreadsheet.
Like there's going to be more people on the internet.
They're going to be using it more.
Telcos are going to deploy broadband.
Some will get video working.
Okay, could you broadcast what internet bandwidth was going to be in 2005?
Of course not.
Yeah, this is why I'm no longer an activist.
Now the truth comes out.
That only took 28 minutes.
Well, I had no interest in share prices, no interest in talking to clients.
And I would look at questions like this and saying, how can I possibly forecast this?
That feels good.
That feels like a good price end.
Yeah.
None of the above is investing advice.
That feels very important.
We'll make sure to repeat that at the top of the podcast as well.
Speak to you soon.
Yes.
Speak to you soon.