Good. How are you? Thanks for having me.
I'm doing great. Thanks for joining our show. It's an absolute pleasure to have you on with us. I really enjoyed the podcast you did yesterday. I forgot the name with whom you did it though. It was published yesterday. I'm not sure if you guys remember.
Dmitry. And it's apparently very popular, so I'm... but anyways, yes.
Very good stuff. We have pretty big news. Anthropic filed its S-1 for a Nasdaq listing. Firstly, for the audience, could you tell us what an S-1 is and why this is important, and then your general thoughts on Anthropic's S-1 filing?
So, an S-1 is a legal requirement in the United States when you plan to issue stock and sell it to the general public. So, it's a registration you have to make that discloses the nature of your business and the risks as you see them in the business, and then you make it as a filing to the SEC in a repository called EDGAR. And in turn, it should show at least your last 12 months of financials, and then some prospective ideas about how much stock you're likely to sell and some ballpark sense of what you think the future looks like. Usually, it contains a risk section that is a relatively new thing. It didn't exist before about 15 years ago. And the unusual thing, among many things in the Anthropic filing, is that almost a third of the filing is risks, which is about the most unusual thing I think I've ever seen in an S-1 filing. So, it's essentially a legal document required before you can go public.
What are your general thoughts on the S-1 that Anthropic filed?
So, most of it had been leaked previously. So, Reuters and others had carried stories over the last six weeks — Reuters, Bloomberg, and everyone else — in terms of the growth, in terms of the revenue of the company on a quarterly basis and the annualized revenue. The only things that were somewhat unusual, I suppose, was that the losses were narrower than I expected. So, the operating losses came in at around 700-plus million dollars, leaving aside — there's some accounting numbers, and we don't need to get into that if you don't want to — that looked like it was closer to 80 billion, but that had to do with some arcane things with respect to resetting some convertible notes that they were using for their financing. But in terms of actual cash losses, it was like a little over 700 million. I thought it'd be over a billion, but whatever — these were still gigantic numbers.
So, and the main thing looking forward was that, as I think as we were discussing before we came on air, looking forward there's still massive losses in their future, given the nature of the business. It isn't just about what's coming in. It's about the commitments they've made in future for things like compute and everything else, which is in the hundreds of millions of dollars. And so these are what are called uncancelable obligations. So it doesn't really matter what happens to the business. These are commitments that they've made to the various hyperscalers for access to compute to deliver the kinds of things that large language models deliver. And they're stuck with that. I suppose the only other thing that I thought was interesting and unusual — because the way to look at an S-1 isn't just descriptively, it's to always be asking yourself, how is this different from what people expected? Because it's expectations that drive capital markets. If people knew it already, they don't really care that much. So one of the things that was somewhat unusual was that the revenue growth has been slowing down. In the most recent quarter it slowed from over 100% to something like 45 to 50%, which are still very, very large numbers. But again, in terms of justifying the kind of valuation the company proposes in terms of its public issue, which is one and a half to two trillion dollars, the growth rate matters immensely. So that's one thing that's a little unusual, is that the growth rate has declined to about 50% serial quarter over quarter.
And then the other thing that I thought was striking was that the nature of Anthropic's business is changing fairly radically. If you read through the footnotes, it's going from being largely a licensor, renting access to compute — hence the large compute commitments — to a company that's leasing and renting, but also increasingly building its own, or at least managing its own data centers, which is a huge change, which involves much more capital expenditure and changes the risk associated with the business. And just to wrap that up, the reason why they're doing that is because they're in this weird situation. Well, the analogy I make is as if Spotify was buying all the music from the record labels, and then in turn each of the record labels out there had their own streaming service. So your competitors are both your suppliers and competitors. So in the case of Anthropic, their largest distributors, accounting for almost 50% of their sales, are companies like Microsoft, Google. Microsoft and Google have their own large language models, so they act as both distributors and competitors, which is a really unusual situation and creates immense risk for the company. So the company going forward is looking to sort of change the nature of the business and become less dependent on distributors and building out some of its own computer supplies. So those are just some of the things I noticed in the report that I thought were most striking.
So, their competitors are their biggest buyers as well.
That's right. It is a really unusual situation to find yourself in. I mean, think of it in like Spotify — the record labels take 30% of Spotify's revenues off the top, right? This is part of the reason why Spotify has sort of an unusual business model. But this is even worse because not only are they taking it off the top, as you said initially, these compute commitments are locked in. So even if Anthropic's sales decline, the commitments do not decline with it, unlike Spotify's royalties, as an example, to the music labels. So if for whatever reason something were to change in the business, they're still on the hook for these hundreds of billions of dollars in capital commitments. And at the same time, they're also selling through their competitors who are just as happy to sell their own products. So this is a really unusual situation. And one of the reasons why the S-1 is such a striking document is because it shows that the company's at a kind of inflection where it's going from largely relying on selling through its competitors to ramping up costs dramatically as it tries to go, let's say, more direct, if you will, and be less reliant on them. And so in a weird way, the document is a reflection of mostly where Anthropic has been, not where it's going.
So the S-1 shows approximately 542 to 567 billion in computer obligations. These obligations — are they contractually obligated to pay those no matter what happens throughout the 2030s?
They are uncancelable obligations. So other than in the event of what's called Chapter 11 restructuring, where you actually are bankrupt and a court decides what obligations have to be paid, those must be paid. So these are the floor, if you will, in terms of what their future capital obligations look like. So exactly right.
And these are obligations for data centers, GPUs, etc. Although...
These are largely already compute commitments, meaning that I'm telling Microsoft and I'm telling Google and I'm telling others, here's how much I'm already planning to pay you for you to give me access to compute for training and for inference. What they will spend on data centers and what they will spend building out comes on top of that.
Mhm. Can you explain the convertible notes and how this actually affects the losses shown for, I think, the year? Because apparently it's a markup to make up for the valuation, if I'm not mistaken, and it's a fake number when it comes to the actual...
So there's two big, let's call them, accounting numbers. I won't call them fake numbers. They're not real cash. So there's two big ones in the report. One is this issue with respect to the conversion of the notes, and the second is with respect to what's called SBCs, stock-based compensation. So both of these are accounting artifacts. They're not cash per se. They're just charges that get reflected in your balance sheet and your cash flows because of the result of what you have to do to reconcile it for... This is real accounting wonkery for what's called generally accepted accounting principles, GAAP. So to reconcile your statements to GAAP, you need to make these revisions. And so specifically with respect to the notes, one of the things that's happening is much of the early financing of the company was done through a mechanism called a convertible note, which is a note that gives you the right, but not the obligation, to convert at some future price and then some future date if you believe in the company's prospects and you want to participate in what's happening right now, which is the planned IPO. And so these things are typically sold at a discount and have mechanisms that allow you to do the exercise. So the way to think about it is that so many of the early note holders are converting to equity that that creates an accounting cost reflecting the transition of the note from debt to equity. There's no specific cash changing hands. It's mostly an accounting artifact, but having changed from being a note holder to an equity holder incurs an accounting cost, but not a cash cost. But it's massive, and the way to think about it is there's so many people who want to convert to equity because they're hoping to, you know, bail out after the IPO, right? They want to sell as quickly as possible, and you can't do that as a note holder. It's a reflection of the amount of demand there is from the early investors to sell the stock post-IPO.
And if you exclude that, what are their actual losses approximately?
But just over $700 million. So if you exclude 70-odd billion related to the note conversions, and you exclude, I think it's two and a half billion in stock-based compensation, exclude both of those, you're left with about 800 million. The perversity, of course, is there's still other weird things in the S-1. Like, for example, the company at the end of the period, the most recent period, had almost $112 billion in cash sitting in various forms in notes, overnight deposits, and other things, which is, I mean, a towering amount of cash to have sitting on your balance sheet. And to give you a sense, that cash was generating about a billion dollars in interest income, which to put it in context would make them — were they a bank — they'd be like a regional bank in the United States. That's how large their deposits are that are sitting on the balance sheet in preparation for upcoming spending based on the money that they've been fronted by largely credit providers.
So just to finish that point, I suppose I should say, so the billion dollars in income is actually larger than — the interest income is actually larger than the losses in the quarter, which is just astonishing.
Mhm. You mentioned at the beginning that their growth rates declined. Do you have any idea why that might be?
It's not why it declined, and I think — I mean, we can speculate, and they didn't provide any information about why that might be. And as a matter of fact, the only reason we know that it declined was because the Q3 numbers, which were masked in the report in the S-1, were leaked previously by I think it was Reuters or possibly Bloomberg. So we know what the numbers were, so you can back out the growth rate by adding back the numbers that we know from a previous leak that are the Q3 numbers. And so as a result, we can see this decline in terms of the growth rate. My own guess is that one of the many dynamics going on here is there's this — the early users of Anthropic were very code-centric. It was largely software engineers who were willing to pay very high prices. And we heard about things like token maxing and all these other phenomena. And if you've been following over the last few months, the whole idea of token maxing and encouraging employees to use as many tokens as possible has largely collapsed. And people are putting in very sharp guards because there were cases where companies accidentally spent five and ten million dollars without planning to and blew their entire budgets in months that was supposed to be for the year. And which is great for Anthropic, but not a particularly good situation to find yourself in as a public company unless you're generating something that's insanely valuable, and that was obviously not the case. So there's been — Microsoft, Meta and others have all cut back dramatically on their employees' ability to use these products, sometimes by as much as half. So my guess is in large part given that the dependence of Anthropic on a small number of customers, most of whom are using it for coding, and the sharp cutbacks in terms of what they're willing to spend on it, what you're seeing is a reflection of that.
Apart from this industry's phenomenon where your biggest customers are also your biggest competitors, could you give our viewers a quick breakdown of how the industry is structured right now and what makes it so quote-unquote bubbly?
Sure. So there's a whole bunch of different characteristics of it. One of the ones to think about is the one you just described, the nature of this sort of interlocking and often called circular relationships among the various players where competitors are also suppliers, but they're also often investors. And a great example, and the one I kind of point to most often, is the role of Nvidia in this greater sort of AI ecosystem. Nvidia is kind of like — if you want to think about it in Federal Reserve terms or European Central Bank terms — Nvidia is kind of the central bank of AI. So what they're doing is essentially extending credit to what it thinks of as its best clients so they in turn can use that money to buy Nvidia products. So Nvidia does this across the board, they'll extend credit to Anthropic, they'll extend credit across for companies to buy GPUs. And obviously these GPUs are one of the core and most important and most expensive parts of current generation AI, they're as much as 65% of the cost of building a new data center. So this is a very significant component and it really drives the performance. So this circularity comes in, and what makes this unusual is this notion that not only do you have customers and suppliers going in both directions, oftentimes Nvidia can also be an investor. So I might be an investor in one of these companies, and in turn exchanging them credit, but they use it to buy my products. So it's really difficult to unwind all the circularity and see what the real size of the business actually is as a result. And so this is a huge distortion. We've seen this many times in the past. We saw this during the fiber optic bubble. We saw it to a degree in dot-com. And if you want to go way back in history, you saw some similar phenomena even back in the 1920s during rural electrification in the US and in the 19th century on railroads. It's a very typical phenomenon here as people act in what they see as their best interest and extend credit to buyers who in turn use it to buy my products. And of course, the problem is that creates an incredible amount of fragility in the entire process because were that to change, were that to stop, or I in any way to become a less aggressive lender, that everything gets repriced because the business itself changes, which has an impact right down the chain on public companies all the way down to construction, venture-backed companies, which is, you know, where I mostly live. Everything down the stack gets repriced because right now everything is priced on the basis of what it looks like the actual market size is, which is in turn distorted by the circularity.
What percent of revenue share would you say in companies like OpenAI and Anthropic is non-institutional and retail, let's say from users who have a $200 plan with Claude, like maybe you as well? What percent of that would you say is...
So, right now, we actually have a pretty good idea of that. So about 80% of Anthropic's revenues are API-centric. So that's not typically individuals. That's much more likely to be applications that have some aspect of Anthropic embedded in it. So the calls are being made through the application programming interface, often at huge volumes, but either in apps or in coding tools. So that's about 80% of their revenues. The remaining 20% is subscription-centric. So these are the kinds of plans you're talking about. We don't know the actual split in terms of the enterprise versus consumer, but a reasonable guess is it's probably skewed at least 80/20 towards enterprises versus consumers. So consumers are maybe 20% of the 20% that is not the API-centric — of the subscription piece. So consumers are a relatively small piece of this, and I'm trying to — I'll exclude like small developer shops and things like that. But it looks from the numbers in the S-1, that's what it looks like.
What elements of this quote-unquote bubble are you most critical about yourself?
I hardly know where to start. So I'll just say historically when you have something that becomes this large and begins distorting the economy in ways that this does. And one of the reasons why I first got interested in this episode was that — and I called it AI CapEx eating the economy — was this idea that over the last, say, eight quarters, it's varied from quarter to quarter, but AI CapEx has been anything from 25 to 60% of GDP growth in the United States, which is a staggering figure and was largely misunderstood at the time because people thought, oh, you know, tariffs are working, look how resilient the economy is, or you could pick — you could sort of use what behavioral economists call motivated reasoning. I could just come up with a reason that pleased me on why the economy is doing so well despite all these other headwinds out there. And so what it was was this idea that AI CapEx was so large, it's actually distorting the economy. It's something like 30%, 25% of global trade growth. It's something like two and a half percent of global GDP. It's probably — these are all just outrageously large numbers. But the reason why it matters is because it creates these huge distortions that are in turn funded by circularity and debt. So as a result, we don't know what the reality is of the global economy. All we know is that every time this has happened in the past, the unwind has been fairly catastrophic. So in the railroads' case, it led to three rolling recessions. In the '20s, it led to the crash of '29 and the Great Depression. In the fiber optic bubble, we had a similar crash. In the dot-com, we had a similar crash because of this idea that increasingly all of this gets built on each other. And what's particularly unusual about this episode that I find really striking is that most of this is happening in private markets. So if you think back to dot-com and telecom, we were looking at companies like JDS Uniphase and Ciena and whoever else who had these bubbly valuations of like 100 times earnings and so on. And the rejoinder you often get in talking about this moment is people say, 'Show me the bubbly companies in the public markets. They don't seem to be any. Look, the valuations aren't that bad. Look, Microsoft's not that expensive,' and so on. And of course, the reality is the number of publicly listed stocks in the United States has halved over the last 25 years. Private markets are the new public markets, which makes it more opaque and makes it more difficult to actually untangle. And so if you look at just say SpaceX pre-IPO, Anthropic, and OpenAI, the sum of those three was say $6 trillion in potential future public market valuation, all at ridiculous multiples of whether you want to look at sales or earnings. And to put that in context, the nominal value of the Nasdaq at the peak of the dot-com bubble was about $6 trillion. So we had an entire Nasdaq of bubbly stocks, if you will, sitting inside of three companies, all in private markets waiting to go public, that were collectively the size of the entire Nasdaq dot-com bubble. So these are the things that I look at and say, 'It's actually — the problem is this bubble is in a strange way stealthier, more opaque, and more fragile because so much of it was happening, if you will, in secret, in private markets.' And these are some of the things that make me more nervous.
The last one I'll point to is that the largest bubbles and the most consequential bubbles in US history usually had to do with real estate, or usually had to do with loose credit, easy access to money, or it had to do with some new technology, you know, like the railroads or the dot-com, or it had to do with government policy where the government encouraged you to take decisions that didn't really make economic sense. And we saw that in the global financial crisis. What's unusual about this bubble is it has all four elements. There's a great technology story. AI is probably one of the most important technologies, certainly of the last 100 years. There's a real estate component. Many people are speculating about real estate because that allows you to build data centers, and in turn this becomes kind of a real estate project. There's a government component, which is government sees this as an existential crisis: 'We must win or China will,' that kind of rhetoric. And there's a loose credit component. So what's unusual about this and makes it probably the most important bubble of the last 100 years is it sits at the intersection of all four of those, which has never happened before.
Yeah, I honestly think the quote-unquote arms race between the US and China is such a strong factor, and it's something so new that we've never seen before, and the...
Yeah, absolutely. It creates this race dynamic, right? Where we saw this when Dario Amodei, the founder and CEO or co-CEO of Anthropic, was proposing this idea of — which was a terrible sort of term — but anyway, this idea of pacing the frontier, which sounds vaguely like you're running the Boston Marathon, but it's not. It was this idea that we would slow down development and manage risk better. But of course, the immediate response was, well, now we're into a race dynamic inside of a prisoner's dilemma. So if I... prisoner's dilemma meaning that if we all cooperate, we get a great outcome. But if anyone defects, they get better outcomes and we all get hurt. This is the dynamic, and so the rhetoric was, sure, OpenAI, SpaceX, Anthropic, we can all agree to pace the frontier, but if China doesn't agree, they win and we look like chumps, we look like fools. This was the idea. And so, of course, as soon as that dynamic kicks in, then everyone just further accelerates in the race. The race dynamic just goes faster because it's, well, I can't lose, so I must race further ahead. And that's the existential battle where there's this idea that we can't allow — at least at a government policy level is the argument, and certainly in the Trump administration — that you can't allow China to win. And that's what locks in not just progress, but risk taking. And that's been increasingly the piece of this that's become most dangerous. And I put out something the other day showing that more critical errors in software have been found in the last eight months than in the history of what are called CVEs, critical errors in software, in the prior 15, 20 years, which is absolutely insane. If you're inside of a large organization, it's like being peppered by, you know, sort of drones all day long. They're all coming in and you have to fix things, and I'm so far behind. And it's only a matter of time because you have to be lucky every time as a defender, whereas the attacker only has to be lucky once. And the numbers are absolutely staggering, but that's driven in large part not just by the improvement in the technology, but by the improvement in the technology as a direct byproduct of this race dynamic.
Did you see the PHP vulnerability that they discovered?
Right. I mean, that was part of what triggered me on this. Yeah. So there's a good example.