So $2 trillion, given the numbers that we know, you think that's overvalued?
Oh, it's a ridiculous price. As friends of mine were saying this morning, some of the largest hedge funds in the world looking at this, and it's like anyone who thinks that being the buyer at these kinds of prices in a very late-stage IPO of what amounts to a relatively mature company — when you look around the poker table and wonder who the sucker is, it's you. This is not a financing event anymore. What's really going on is people are unloading shares. They're unloading shares on retail investors and on quick-flip institutions who were able to back in and out. So you have to look at it accordingly and realize that what they're really saying is this seems like a good time to get out, and I'm an insider and I want out.
Welcome to Prof G Markets. I'm Ed Elson. It is September 30th. Let's check in on yesterday's market vitals.
The major indices declined as the bond sell-off continued. The yield on 30-year Treasuries hit its highest level since 2002. The 10-year also climbed towards 5.3%. Brent crude fell to around $102 per barrel as the Trump administration ordered an emergency reserve release. And finally, Apple shares fell nearly 3% on reports that the new CEO, John Turnis, is planning to slim down the company. Okay, what else is happening?
Anthropic's IPO prospectus just leaked, and the numbers are staggering. According to Reuters, which obtained the draft of the S-1, the company reported $4.6 billion in revenue last year, up more than 1,000% from the year prior. But they also reported an operating loss of $8 billion and a net loss of $42 billion. Anthropic devoted almost a third of the S-1 to explaining its risk factors. Those included customer concentration, with close to a quarter of revenue last year coming from just two clients, and then, of course, the existential threat to humanity. Still, the company is expected to go public at a record $2 trillion valuation in November, making it the most highly valued IPO of all time. Joining us to discuss Anthropic's financials — or at least what little we know — we are speaking with Paul Kedrosky, managing partner at SK Ventures. Paul, great to see you. Thank you for joining us. I heard some of your laughter as I went through the numbers. We finally got some numbers, and to be clear, these are 2025. But what do you make of them?
They're honestly — most of those numbers in one form or another had already leaked, and we can go through them one at a time. But let's take specifically the one — as a long-ago equity analyst, the one that catches my attention the quickest is customer concentration, because customer concentration is nothing for a small company. You expect a small company to have really high levels of customer concentration, meaning that a small number of customers is a material fraction of revenues. But to see a company at this size where two customers are on the order of 25% of revenues, and there's been some other leaks or whatever you want to call it that's something like six customers or 60% of revenues — these are absolutely, to use the technical term, bananas numbers. And why they matter is because it shows how unusual their earliest customers are. They are so consequential, and they're using it in such unusual ways that they're such a material chunk of revenues. And one of the things you always have to watch with young companies — well, it's strange to say at a company this size — is if they're able to jump across and succeed with later customers who are nothing like the early ones. And specifically in this case, Facebook is a good example. Meta is one of their largest customers. We know that from other data, and they're wildly unprepared — you know, I don't know, Goldman Sachs, pick a standard industrials company or someone else is going to use these tools. So we've got a pretty open question here in terms of two kinds of risk. One is the level of customer concentration, and the second is how representative these early customers are of later customers.
Let's look at the numbers as well, though, that we know from 2025. There's the $42 billion in net losses. To be clear, roughly $34 billion of that was a non-cash charge tied to revaluing their financing instruments. So it's a big asterisk on that number. Still big, but you know, maybe there's a caveat there. But $8 billion in operating losses. So that's the amount of money that they're losing from the day-to-day operations. Granted, it's from last year, but what are your takeaways from that number? And how important is it when valuing this company?
It's hugely important, and this is the point where finance as theater begins, because what's going to happen is they are going to try and characterize these operating losses as really related to something that we shouldn't be worrying our pretty little heads about, which is to say training costs associated with the creation and running of these models. On an operating level, just ignoring training costs — and this number is one OpenAI has leaked and Anthropic has leaked — they now all say that on an inference-only basis they're already cash flow positive in the last couple of quarters. But that's going to be the debate, right? It's this old joke we used to call it when I was in my analyst days: earnings before bad things. If you let me get away with characterizing my earnings before bad things, my earnings look really good. So you have to decide: are these bad things? And they're not so bad because they're fundamental to the business. Are they things that they should be allowed to characterize as something other than operating costs — something you capitalize, for example, like you might with R&D — or are they actually just the day-to-day parts of running the business? I would argue the latter. The training costs are just the day-to-day running of the business. Whenever you're training new models every six to 18 months, that's not something that you're capitalizing out four or five years like a building. That's part of running the business, and it should be reflected in the earnings that we look at to value the business. And to take it away is kind of finance theater. But nevertheless, we're going to see a lot of EBBT — earnings before bad things — coming up here.
Yeah, it seems like the training costs will be something that — who knows, maybe that'll be stripped out because, as you point out, they're saying that at least on an adjusted basis they are profitable right now. But we don't know what they're stripping out. Maybe they're stripping out the training costs.
We don't know what they're stripping out. It's training costs. I can — I will — I'll take that bet all day long that that's what they're doing.
Stripping out the training costs.
The other thing that supposedly they are stripping out, or there's a question about if they're stripping it out, are these revenue-sharing agreements, because of course Anthropic owes a significant share of its revenue to Amazon, as an example. And it sounds like — I mean, we know that when they're looking at their gross margins, they're not including that in their calculation.
Less material than the training costs, but still consequential. I agree.
Still consequential. How profitable? To be clear, we don't really know yet because we're only getting leaks. But if you had to make a guess at the profitability — we'll just focus on Anthropic right now — in 2026, what would you say the profitability picture probably looks like given all of those questions?
Well, again, absent earnings before bad things — taking away training costs and some of the revenue-share commitments — I doubt it would be hard for me to believe that they're not showing some cash flow positive as it currently stands on an inference-only basis. What those numbers look like, we've seen numbers suggested that it could be as high as a couple of billion dollars, some positive cash flow just from inference alone. But that's purely speculation, and they haven't released it, so we don't know. But it wouldn't surprise me. But it's more than dwarfed by the business of running the business, which is to say training costs and some of the other, I'll say, more circular revenues. If you back those out and you take those things out, the business looks entirely different. So then the question becomes: are you trying to tell me that training costs are not a part of the business going forward? Because I've sometimes argued that the first frontier model company to stop training models and just do inference is probably going to win, because Wall Street will reward you for cutting costs and generating huge amounts of cash flow. So you can't have it both ways. Either training costs are integral to the business, or they're the thing you're going to cut so that you can be fantastically cash flow positive going forward. So I just think they're trying to have it both ways, and Wall Street's going to give them a wakeup call on that.
Let's assume that they are stripping out the training costs. That's part of the bad things, and they're reporting the earnings before that bad thing. How bad, in your view, of an accounting gymnastic move do you consider that to be? I mean, the ultimate sort of accounting mismanagement that people cite often is WeWork, where they invented this community-adjusted EBITDA, which a lot of people said was the most ridiculous, and we know how that went. It didn't work. It was a disaster of an IPO. If that is what they're doing — they're taking out the training costs and saying we're profitable — how bad is that in your mind? How does it compare to, say, community-adjusted?
So I think the community-adjusted earnings was, frankly, a fraudulent measure. It was an attempt to hide the fundamental broken economics of the business. I think there is a cash-flow business here, but it requires far less money spent on training. So is it fraudulent? No. Is it poor accounting? Yes. Should it be supported by the auditors? No. Should Wall Street punish them for it? Yes. That would be the way I come at it.
Yeah. So, $2 trillion, given the numbers that we know, you think that's overvalued?
Well, it's a ridiculous price. As friends of mine were saying this morning, some of the largest hedge funds in the world looking at this, and it's like anyone who thinks that being the buyer at these kinds of prices in a very late-stage IPO of what amounts to a relatively mature company — when you look around the poker table and wonder who the sucker is, it's you. Not because it's a bad business, just because what's happening is this is not a financing event anymore. They're not raising money for anything. I think what's really going on is people are unloading shares. They're unloading shares on retail investors and on quick-flip institutions who are able to back in and out. So you have to look at it accordingly and realize that what they're really saying is this seems like a good time to get out, and I'm an insider and I want out.
Is your view that if they stopped training — and just so everyone knows the difference, I mean, the training is the building the models, it's creating these advanced frontier models; the inference is just running them, just operating them. Is it your view that — because this is something that I do hear from AI people — that don't worry, if we just flip the training switch and just say, okay, we're not going to spend money on this anymore, then we have a great business. Is that your view that if you do that, if you get rid of the training costs, you just focus on inference, then actually these are sustainable business models that work?
No, it's a trap I'm laying out for them. It's actually catastrophic for them if they do that. So what happens when they do that is they then become basically solar panel manufacturers who are trying to compete with China, and China crushes them with cheaper power and vastly larger industrial token production than these frontier companies could ever cope with. So they're caught in a terrible place where, to protect their so-called moat, they have to spend profligate amounts on training. But if they don't spend on training and now it looks like they could be cash flow positive on inference, well, now they're into sort of industrial token production, no different than industrial photovoltaics or industrial EVs, and now you're up against this colossus called China, who wins that game over and over again and is already setting the floor in terms of token prices. So you're in an impossible situation: if you don't keep training, you have no moat; if you do stop training, you're crushed by the industrial production of tokens coming out of China.
