Anthropic's S-1: the AI lab that became a business

Anthropic's S-1, filed ahead of its IPO, addresses the question hanging over the whole AI boom: is any of this a real business? I walk you through its answers.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Azeem Azhar?

  What we can see in this very interesting S1 is that Anthropic has transitioned from being a lab to a business. And I think that that is the key message that we've learnt in this filing that addresses the elephant in the room, which is, is AI a real business? Are any of the companies real companies?

• 0:24Azeem Azhar?

  Do they have the hallmarks of things that might last or they just fly by night?

• 0:34Azeem Azhar?

  Hello, everyone. I'm Azim Azar. I'm the founder of Exponential View. And this is a live discussion where I really wanted to talk about Anthropic, the AI company, and their S1. Now, the S1 is, in a way, the marketing document that companies file when they're going to go public, that is, offer their shares for sale to ordinary investors. And this discussion is a casual discussion. It is not meant to be investment advice. I'm not giving you any investment advice. You should take your own advice. But what we're going to do is just talk a little bit about what the S1 tells us and what's interesting about it.

• 1:17Azeem Azhar?

  In order to do this, I will take us through a few slides. As many of you may know, we spun up a new team and a new product in Exponential View called our AI Investment Brief. That just allows us to drill into those questions for a very, very specialist audience. So we could be broader for Exponential View readers who, you know, love to explore the landscape of ideas and how the world is changing in very, very broad ways. So what I did was I pulled some material from that investor brief that I think tells the key business story of Anthropic as we learn about it from the S1.

• 2:01Azeem Azhar?

  Anthropic, from a lab to a business. If we start with the nuts and bolts of it, the profit and loss statement. What we now know is that Anthropic, of course, has been growing very, very quickly. $11.6 billion in revenue in the second quarter. That's to the end of June 2026. You can see the growth rate ramp in that purple slide. You know, we reckon, based on our own modelling and press leaks and things that are hinted at the S1, that the next quarter is $16.9 billion of revenue. The same period last year, third quarter 2025, was $1.4 billion in revenue.

• 2:46Azeem Azhar?

  Now, the key thing is that in the second quarter of this year, Anthropic made a small profit. There are all sorts of different types of adjustments we can get into that we talk about in the investment brief. But what you can actually see there is that there is a small profit that they manage in Q2. Small, a billion dollars. That's quite a lot of money for me personally. What I draw attention to is that blue line next to it, which is the computing spend. How much is spent on compute? One of the things people were worried about last year was it was so expensive to train and serve these AI models. Could these companies ever really get profitable? And what you see is Anthropic's revenues are growing far faster than their compute spend.

• 3:29Azeem Azhar?

  What drags them down are staff and sales and other costs. And then a drag on the P and L line because lots of the employees are also paid in equity. And Anthropic is also making equity donations to a public benefit corporation. So that means that on the same basis, we think the company may be slightly loss-making in Q3. But the underlying point is the direction of travel is something that looks increasingly sustainable. We dig into all of this in like 39 charts in the other newsletter, if you're interested. The second question here is, is how real is that customer profile, right?

• 4:12Azeem Azhar?

  There was a lot of concern, quite absurd concern. In reality, given the data that we had, that 80% of Anthropic's revenues came from just a handful of customers. That hasn't been true for a long time, and it certainly isn't true in the S1. And in fact, what you see is a company which has broadened and deepened its customer base. So if you look at our estimates for Q3, the top five customers that Anthropic had used to account for 36% of revenues. And we think that that number is now down to around 21%.

• 4:52Azeem Azhar?

  So in all enterprise businesses, you often have very, very large customers, whales, who weigh you down, especially in the early days before your customer set distributes and broadens. The key thing here is that Anthropic is well, well, well on that track. And five customers for 21% of your revenue, even as all of those customers are spending more than they ever have before, is not a problem of concentrated revenue. And I think it's worth noting also that those top five customers, they account for about $3.5 billion of revenue in the third quarter of 2026, which puts $13 billion of revenue roughly from all of the other customers.

• 5:34Azeem Azhar?

  Anthropic runs out to several million customers when you think about very small users. Who are those customers outside of the technology companies that we often hear about? Well, over half of the Fortune 500 are paying Anthropic, which is, you know, quite a rapid distribution. And what's interesting is that the average Fortune 500 customer spends about 0.1% of their revenues on AI services through Anthropic. So that's quite a small number relative to the IT budget. It also shows there's lots and lots of room to grow.

• 6:15Azeem Azhar?

  And you can see on the left-hand side of that chart how that changes as you look at the top 10 Fortune 500 customers of Anthropic. And these companies spend about eight times more. So that talks a little bit about the kind of trajectory you might be able to see as companies get more confident in using AI. What I'd note is that if you look at the top 10 companies in many sectors, Anthropic has already got broad coverage within them. And what often happens in markets like this, if you think back to diffusion curves, is that, you know, your earliest customers are always experimenters. You then get to early adopter companies.

• 6:59Azeem Azhar?

  And what early adopters want to do is they want to see use cases that they feel quite comfortable with, but they don't have to be in their own sector. And as you go further down the distribution curve, you get into customers who really want to see what their peers are doing. And this slide, to me, shows that peer activity within particular sectors is a potential area for competitive growth, not just for Anthropic, but for other AI companies. If you look globally, roughly 25, 28% of Anthropic's businesses outside of the US. And the global 2000, about a third of them, 600 out of 2000, are customers of Anthropic.

• 7:41Azeem Azhar?

  So you see that that penetration rate is slightly lower than the Fortune 500. It just is typical that American companies are more proactive in getting involved in new technologies. So just to net out where we stand on the customer base. The customer base is broad and it is deepening. It doesn't look concentrated in any sense. There are some big customers who are billion-dollar accounts, but set against the sea of other customers, it's less of a risk. One thing that I was really keen to understand is, to what extent does all of this investment in new AI models actually drive the AI company's revenue growth?

• 8:25Azeem Azhar?

  And we do get an insight into that. You can look here and you see quarterly revenue growth for Anthropic going back to Q2 '25. It grew 109% in revenue over the previous quarter. 76% from Q3 to Q2, 48% from Q4 to Q3 in 2025. So that growth rate is slowing down in that period. What the company has said is that they see seasonal effects like Christmas affecting their growth rate. You get the launch of Opus 4.5 in late November. We've marked it there. And that was a model that was much better at code development. It was much better at agentic tasks.

• 9:08Azeem Azhar?

  And in the two quarters afterwards, you see a really, really extensive growth rate. The, you know, the famous growth rate that takes things off the chart. So, you know, our hypothesis is that these new models actually do ultimately become important in driving, you know, revenue growth and new customer business. We can start to understand that when we look at where Anthropic's revenues come from a model-by-model basis. When we analysed the S1, we saw an observation that 80% of Q3 revenue

• 9:50Azeem Azhar?

  was from models launched in the previous six months, from the beginning of April through to, you know, through to now. And what that shows is how important the new models are in driving revenue growth and what customers want to buy.

• 10:07Azeem Azhar?

  It also shows that customers can move quickly, at least within model families, into these new models. Now, there are precedents in other markets in technology. I mean, I used to think quite a lot about TSMC. So TSMC is like at the other end of this, because TSMC is making chips. But they regularly, on that cadence that is connected to Moore's Law, have improved process nodes. So this is sort of the next layer of technology. And what you can see is how TSMC is really effective at retiring previous process nodes as new technologies come in and moving their customers with them. And I think analogically that that is similar to what we're seeing here,

• 10:51Azeem Azhar?

  except that the timeframe is very, very much shorter. You know, it's measured in a couple of quarters than in several quarters or in a year or so. I think that this is important because it speaks to, you know, the amount that needs to go into the training of new models. It speaks to how important it will be for Anthropic to talk about recursive loops and speeding up the development of the latest models. And they talk about this in some extent in the S1. It also raises a question, which is, of course, to what extent is their existing revenue

• 11:32Azeem Azhar?

  sensitive to the latest model, wherever that model comes from? And I think that that is really important to recognise, because one of the things that we talk about is that still the majority of Anthropic's revenue comes from people paying for usage rather than these retained contracts. I mean, many of us are probably paying for Claude Pro or Claude Max. You can look at AI Investment Brief and get the actual number, 25 to 30% of revenue is kind of from longer term subscription relationships, but the rest is from this consumption basis, which of course constructs another risk, right?

• 12:16Azeem Azhar?

  How resilient does that revenue end up being? The compute expense is inordinate, and I'll spend a little bit of time talking about that. One bet that really is at the heart of what Anthropic does is that scale helps. The company is built on these scaling laws of AI, which is that the more compute and data you put in a model, the more performant it gets. And that scaling relationship is held from when it was first identified, you know, Chinchilla and so on, through OpenAI and other companies. But the other aspect of scale, and it's so true for early technology companies,

• 12:57Azeem Azhar?

  which are never profitable, is that you need to show that scaling turns, gets you to a point where you have those economies and effectively your revenues outrun your costs. A large part of Anthropic's costs are in its compute. And what we can see here is that since the first quarter of 2025, and this holds true further back, revenue growth is going much faster than the compute cost growth. Roughly speaking, 31-fold increase in revenues and a six-ish increase in the compute cost. What that shows us is that the scale is starting to pay off, the scale economies are working.

• 13:40Azeem Azhar?

  A quick note. If you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.

• 13:52Azeem Azhar?

  We also try to figure out like what happens with these economics, right? Where does a dollar come in and where does it end up? And there are some really interesting points here. When a dollar comes in, about 48% of all of Anthropic revenues comes from its distributors, the cloud marketplaces. So that is people like AWS, Google Vertex, that's Amazon Bedrock in AWS, Google Vertex, Microsoft Foundry on Azure. So that's, you know, roughly 50% of the business. The rest is sold directly either through subscriptions or through direct access to the API. And in a way, this makes sense for a product like this, because the cloud companies are the ones with long, deep relationships with enterprises.

• 14:36Azeem Azhar?

  So when that dollar comes in, it turns into revenue. There is a cost of revenue, which is effectively the inference computing cost, which gets us to that gross margin, which is pretty tasty at 85%. But let's remember the famous line, "Your margin is my opportunity." So that's an 85% margin that is attractive to someone else to try to steal. Then there are channel fees. So this is the portion of revenue that Anthropic pays to those cloud marketplaces that represent 48% of its sales. And so after channel, you're down at this 68, 69% of the original dollar. We then take out the training and R&D compute,

• 15:16Azeem Azhar?

  which is a really significant number. Krishna Rao, who's Anthropic CFO, said a few months ago, "Investors were going to have to get comfortable with Anthropic maintaining a particular level," by which he meant a high level of training spend. And you start to see that there. So then at the end of it, you've got 28.5%. That's what the revenue that comes through after compute and channel fees, from which you have to pay for all of your other operating expenses. And you therefore get to that adjusted operating income, which is about 9.9% for this first half of the year. And then from that, you need to take away, you need to adjust for stock-based compensation

• 16:01Azeem Azhar?

  and the donations into the Anthropic Foundation, which is going to be a certain portion of the capital table as the company achieves revenue milestones. So that is a waterfall of funds that effectively come through the company. The key point is there is cash flow in there. Now, the compute spend for Anthropic is substantial by every measure, and they acknowledge this in the S1. So here's a direct quote from it. "Anthropic will make forward investments, especially in compute, that may seem irresponsibly large. This comes from our belief on the scaling law, that AI's economic value is on the exponential curve, and the expected near-future economic value of our products will be much larger than the current value."

• 16:44Azeem Azhar?

  And critically, this is their justification for investing more aggressively than a company in a slow-changing technological regime. Now, this is really important. Of course, this makes sense to me. I've studied these technologies for decades. The nature of that exponential curve does mean you end up needing to front-load these types of investments. And I think the hypothetical game that you can play for yourself is that people back in 2024 and 25 were complaining even then about the scale of Anthropic's compute commitments. And what might be interesting would be to sit where we are in 2026 and look back at those compute commitments relative to the revenue that they have supported and say, was that really, you know, out of the

• 17:28Azeem Azhar?

  ordinary? But it's important, because this is a really, really large number. And it shows up with this tension, this duration tension. The bulk of Anthropic's customer revenue is, you know, relationships that are paid by usage. So Anthropic roughly has visibility on $10, $11 billion of revenue, 12 weeks of revenue. That's what customers have contracted to buy. But on the other hand, Anthropic has contracted for more than 12 years of compute costs, $587 billion. And they do acknowledge this in the S1. "Our revenue is primarily consumption-based, and much of it is not subject to long-term

• 18:12Azeem Azhar?

  contractual commitments, while our compute infrastructure generally extend over multiple years." And so this is the thing that I think will make a lot of people uncomfortable. And naturally, it is uncomfortable. You've got this short-term revenue with, you know, limited contractual visibility against long-term, highly contracted, very large compute obligations. The critical thing to note, those compute obligations don't hit tomorrow. They run out over the course of 10-plus years in a ramp that, you know, we don't, no one has really good visibility on. But it's important to recognise that there is this, you know, this mismatch. So the question, I suppose, is, you know, can they make, can they pay

• 18:57Azeem Azhar?

  for the compute obligations? The signed computes that they have for 2028 would require revenue to grow at four times today's run rate. So effectively, the, if you look at the purple line, that is what we reckon in the named contracts are the, the sort of the revenue that's, that's, you know, required at today's conversion rate. But of course, to get comfortable, you need to make more than that. And that's the green line. And it's effectively saying that.

• 19:36Azeem Azhar?

  You know, if we model Anthropic's revenue retention diminishing over the course of a few years, this is ultimately where their revenue would get. It's about four times today's run rate, and it would be sufficient to pay for the compute obligations. We dig into this in a lot more detail, of course, in Investment Brief, and you could take a look at it there. The question, I think, is how realistic or not are those numbers? And I think off the basis of what we've seen in the S1 and what we've seen about the breadth and depth of customers, what we've seen about net revenue retention being 900%, how we have seen, quite importantly, that even during

• 20:18Azeem Azhar?

  the price cuts of the Opus model over the last few months, Anthropic actually made more money per dollar, for every dollar it spent on compute than it did before. That path is plausible. It is definitely not impossible. Now Anthropic as well, of course, is kind of this quirky company, the way in which it talks about its risks and the future of humanity. And in the S1, they don't depart from that. In fact, they say, we have a departure from profit maximization. We are committed to standing up for our values at critical points, even when this is costly for us commercially. And they give examples of where they've done that, where they've done that with

• 21:00Azeem Azhar?

  some of the advanced models like Mythos and Fable, where they didn't release them, they put more safeguards around them. They give an example that they won't accept advertising because they think it creates the wrong kind of incentives between the company operating the AI and consumers like us. So that is definitely a departure. I suppose if you're a steely eyed capitalist, you'd want to keep an eye on that if you were considering getting involved in this in this company. But they do talk about it, and it gets reflected, I think, in the way in which the stock in the company is organized. And there are a few kind of curious things about it. So one is

• 21:40Azeem Azhar?

  that there is a special class of shares, the class F shares, which have 50.1% of the votes, and they reflect the 9% of the shares held by the seven co-founders. Now, there are some little rinky dinky details here about about what happens if co-founders leave the business and do those rights fall away. It's not uncommon in in tech companies. Facebook was a good example, where, or Meta, as it's now called, where founders don't just exert a lot of moral authority over the board of the company, but structure their voting rights so that they have a kind of voting control. And again, this is something that we see within Anthropic, you know, right now. The

• 22:24Azeem Azhar?

  other thing, of course, I talked about is that there is this foundation for for public good, which will continually get equity from the business as the business grows. You know, as a mechanism, I don't think that's a bad mechanism. I actually talked about mechanisms like that in my previous book, The Exponential Age, when I was looking at the increasing returns to scale of technology companies. Of course, Anthropic didn't exist at the time, and trying to figure out how you can manage those increasing returns to scale without it all turning into a win for shareholders of a particular type. So I think it's kind of an intriguing idea. You know, Anthropic itself recognizes this sort of, or claims there's this odd moment that that the

• 23:06Azeem Azhar?

  company is in. I think most of us agree that right now the world feels different in some way. And so they do give this warning, or this acknowledgement, or this indication to to investors, which is not something I've regularly seen in S1s. We encourage you, as a prospective owner of our stock, not to think of your purchase in purely transactional terms. We hope you will see it instead as an opportunity to be part of something greater. The advent of machine intelligence is one of the seminal moments in human history. So, you know, there is this kind of grand positioning, seminal moments in human history. You know, one could also argue, given all the warnings, perhaps the last moment in human history. Of course, as you know, as you know,

• 23:49Azeem Azhar?

  people who read Exponential View, I don't think it is that. But I do think it it sort of ties a bow on this quite intriguing, you know, somewhat complicated thing that is now a business, not just a startup. I'm very happy to take a few questions if things have cropped up. Yeah, William, what can we, implications can we read from Anthropic using so much compute for training and R and D versus inference? Yeah, this is important. They're roughly spending about a quarter their compute on inference and 75% on on training. There's a few things going on there. One is that

• 24:30Azeem Azhar?

  it's the newest models that create the value. They argue that in the S1, that exponential curve, and you need to stay at the front of it. We've also shown it in our own analysis that the vast bulk of the revenue in Q3 has come from models that were only released starting in April this year. So there is an obsolescence that is pretty rapid, and it's a curve they'll need to to get in front of, although I do think those dynamics may slightly change as they get more ordinary companies. I think the second thing is that it is a lever that gives Anthropic and the managers there a lever they can play around with as they run the business, which is, you know, they can

• 25:15Azeem Azhar?

  always flex that a few percent one way or another to meet customer demand if they need to, or to to do other things, to sprint in other areas. It's quite difficult with a business like this at all to imagine that there is a natural steady state for the amount of, you know, compute that you will will ever need. And I think that that is probably the case for for this business that, you know, certainly is as far out as we can see for a few years, that they will continue to need compute. You know, remember that customers will be doing much more inference. And Krishna Rao, the CFO, has said, well, there's going to be a level of training we're not going to fall below.

• 26:00Azeem Azhar?

  And how much more inference will customers do? That we have this analysis that's worth taking a look at, which is that essentially enterprise customers who use Claude Code use 20 times as much of Anthropic's product as enterprise customers who don't use Claude Code. So you you definitely, and I'm sure many of us have had that experience, which is as we get better with Claude Code, or CoWork, or building large agentic flows, or even as some of my team have agentic loops, your consumption of tokens increases really significantly. David asks, access to capital seems like the way the race is being played. You either have it or you don't. If you do invest

• 26:42Azeem Azhar?

  for a place on the board, compute is really, really important. And compute is is clearly the way in which the value gets created and the work gets done beyond the individual humans who are sort of, you know, doing some of that work. I do think that, and we we saw this within Anthropic's S1, that data still matters for making a difference. And they've talked a little bit about how many customers and how there's a growing desire for zero data retention. But if you've only got your own data, but you don't have the compute, I think that that is that is not very helpful. And about three years ago, my colleague Nathan Warren and I, we teamed up with

• 27:24Azeem Azhar?

  BCG, which is a consultancy, to write a forecast saying that CEOs and boards need to really think about compute as a strategic input into their businesses. And we were saying that at the time, I think back in 2024, that boards were not thinking in those terms, and they absolutely needed to think in those terms. I don't think this necessarily closes the door on small, highly capable models that will run on on less compute, and for those to have a, you know, a great business

• 28:03Azeem Azhar?

  around them. But I would think that the, you know, for the general use case, these generalist, highly capable models that have this loop turning on them are now on a track to certainly being able to to fend off many competitors for the next few years at least, whereas, and you know, you then have space for specialist models, whether whether it's, you know, models that do formal verification, which I think is going to be more and more important to make software robust, or whether you have models that are better at things like uncertainty quantification, which I think becomes

• 28:44Azeem Azhar?

  important in many, you know, many use cases. But I wouldn't, I wouldn't say that there's an immediate way to outwit the frontier models with, you know, some bit of magic, although I wouldn't rule it out. Thank you so much for joining us today. We will, of course, get back into the habit of doing more of these. You know, please do share Exponential View with your friends. And if you are interested in Investment Brief, you can find that at investmentbrief.ai and you can sign up there.