Anthropic IPO: "This is a dog of a company" | Ed Zitron

How long can we keep kicking this can? How many hundreds of billions of dollars do we need to feed into Dario Amodei's furnace before we start accepting that something isn't right here? When these numbers came out, I have never heard from boosters like this. Usually boosters are just in my mentions being like, 'You're wrong.' All the worm-tongue figures going, 'Nah, you are dead, Christian old man in Amodei.' But this time, I have people being like, [snorts] 'Well, well, you know, 2025 doesn't matter. 2026 is great.' Well, what would it take to prove you wrong? What would it take to admit you're wrong? They are terrified.

Well, the fact is that in 2025, Anthropic was a worse business than OpenAI. And that's the thing. This wouldn't be anything quite remarkable. They're both bad businesses. But for years, the mythology in the valley and journalism and everything has been, 'Oh, Anthropic is the stable, more profitable one.' There were multiple stories last year saying, 'Oh, they're going to be profitable more quickly than OpenAI. Oh, they're a profitable company.' 2025, the numbers were quite shocking and Reuters is, I'm going to say with professional respect, slow-rolling this in a way that makes me quite suspicious. They have been spreading this out over days. It's not clear if they have the 2026 numbers and if they do and they're delaying them like this, I think it's despicable. But putting all that aside, the only defense I've got from AI boosters on this is, 'Oh well, 2025 doesn't matter.' That's not true at all. It absolutely does. The condition of this company is truly insane. But the biggest thing that stood out to me were the amount of compute agreements. So it's about $518 billion over 7 to 10 years. But 80% of those, so a little over $400 billion — you know what's $400 billion between mates — are non-cancellable. This means that Anthropic has do-or-die, take-or-pay agreements that are just more than the operating expenses of Microsoft, which is about $150 billion. Meta, about $125 billion a year. And it's just — this is not a — this is a very serious situation, but not a particularly serious company based on the agreements they've made.

Oh no. People are going to keep saying it because for months Anthropic has been leaking these stories about them being profitable. 'Oh, we had a profitable Q2.' Profitable Q2 came from Musk's discount of its compute. Q3, they said, 'Oh, yeah, we might be profitable to investors.' FT story. Really bizarre one as well because it said, 'Oh, we might be profitable.' Didn't include the revenues, didn't include the cost, and also it was an adjusted operating profit, which is kind of like saying, um, 'I'd be profitable if it wasn't for all my costs.' So, really, Anthropic is just as bad a business as OpenAI, if not worse. I think at this point it's very obvious that running an AI lab is not a sustainable venture and even if they stop training, they haven't broken out the training cost in the Reuters story, I should add. I don't know — I don't know how they turn this around, nor do I have any compelling answers coming in my email or even from boosters other than they'll just do it someday.

Yeah, the copium.

So the two customers in 2025, if I had to guess, were GitHub Copilot, so Microsoft, and Cursor, so Anysphere who makes Cursor. Point about GitHub Copilot is up until June of this year, GitHub Copilot massively subsidized their customers' compute, their token spend even. So you could burn $1,000, $2,000, as much as $5,000 worth of tokens, including Anthropic ones, on GitHub Copilot for 39 bucks a month. Now, the important thing about this is that Microsoft moved to token-based billing in June. So that's dead because before they were just feeding money to Anthropic hand over fist. Now I think it's Meta and Cursor. It might have been Meta last year. I truly don't know. I remember an Information story from the middle of last year, 2025, that said that Cursor and GitHub Copilot were the largest. In any case, this gets back to something I discussed a few weeks ago, which is Ramp had the study that said 80% of Anthropic and OpenAI's enterprise revenues comes from 1% of their customers. Most of those customers are these unprofitable, unsustainable AI companies like Cursor. So, they already have this concentration risk that is just truly awful. And then there's the fact that Musk bought Cursor and put it inside SpaceX. So this means that either at some point Musk is going to just drop them and replace it with Grok or whatever other models he can find, or it's going to be contingent on whether Anthropic buys compute from Musk. And that Reuters story revealed that they have something ridiculous like $48 billion worth of compute agreements, but they can be cancelled within 90 days.

No. I mean, if it was going to happen, it would have happened already. Like that — people will say, 'Well, Vera Rubin will do that.' Will it? Will it do that? Blackwell was meant to be 10x, 10x more efficient or something. Did that — when is this going to happen? Because at some point, it kind of gets back to the profitability thing as well. It's people like, 'Oh, they'll be profitable next year.' How long can we keep kicking this can? How many hundreds of billions of dollars do we need to feed into Dario Amodei's furnace before we start accepting that something isn't right here? And I think that there is just a dramatic amount of copium, opium, type whatever you call it that's just making people ignore this. When these numbers came out, I have never heard from boosters like this. Usually boosters are just in my mentions being like, 'You're wrong.' All the worm-tongue figures going, 'Nah, you are dead, Christian old man in Amodei.' But this time I have people being like, 'Well, [snorts] well, you know, 2025 doesn't matter. 2026 is great. What would it take to prove you wrong? What would it take to admit you're wrong?' They are terrified. And I think it's because everyone thought these numbers would be better. Everyone thought they were going to have like 4.6, 4.7, $4.7 billion worth of revenue, maybe $2-something billion worth of losses, something that would make them feel good about themselves, when in fact it was quite the opposite. And I mean, I got OpenAI's audited financials and at the scale that they're both at, they're functionally the same level of bad. And I think that it's hard for boosters to take this because the mythology was correct. I'm sure the people in comments will notice this. The mythology was really strong. It was always that Anthropic was the stable, normal business while OpenAI was off burning all this money. Apparently not. Apparently, even with OpenAI burning money on Sora, free users and all of these other things, they still burned about as much. So, yeah, it's going to get — it's going to get interesting from here because I don't think that 2026 is going to be as bad, but it could be. And if it is, Anthropic is screwed. But even then, this is still a company that loses billions of dollars. And I think that as people reconcile with that, they're going to get more deranged.

Oh, they would have to be actually profitable.

Actually really profitable. Not adjusted, not EBIT, which is earnings before training, inference and something like taxes — and yes, earnings before training, inference, uh, taxes, depreciation and amortization. Yeah, no adjusted operating. I don't think we should let — I don't think we should let companies have adjusted numbers anymore. GAAP or no GAAP — GAAP or nothing, I should say. I think Anthropic would have to have like a teeny-tiny loss. It would have to be like $500 million for the year. It would have to be something so astonishingly different. And the fact is that this company, if you go and look back on the last 10-odd months of this company's press push and the amount of annualized run rate — you've heard this company mention this annualized run rate that people were like, 'Oh, they have a $65 billion annualized run rate.' Never define what that means. And the fact that they've leaked these figures shows that Anthropic isn't really interested in sharing the truth. They want to obfuscate their numbers as much as possible. Like that Q2 number, I think they claimed they had an operating margin of negative or operating loss of $500 million because of the Musk discount. I don't know if that's adjusted or not. We'll find out when we see these numbers. So long story short, it would have to be a different company. And even if — here's the thing, even if — and I don't think this is the case — Anthropic was somehow a profitable company, you've got $410, $413 billion of non-cancellable agreements. Profitable is no longer actually enough. And to be clear, I really don't think they're profitable.

Yeah. So this was a story out of Insane Analysis. So apparently that $65 billion number that's been going around everywhere is actually Anthropic taking the day of July 31st, so about 178 million, and multiplying it by 365. So —

That is beyond abnormal. Generally, annualized run rate is really bad. Like it's just — I hate it as a term. I don't think it tells you very much about — I don't think it tells you very much about the company or their state. It's generally you want to do — and they use it so that people think it's annualized recurring revenue, which is stable revenue that's guaranteed, times some number. Nevertheless, this number may not have come from Anthropic. It may have been investors, but it was reported by Bloomberg. It's been used by everyone to say this company's doing fine. Look, their annualized run rate is $65 billion. It's a meaningless number. And if it truly is generated by multiplying a day by 365, it's negative useful. It tells you less about the company and actually the only thing it tells you is that Anthropic is desperate and that Anthropic's investors are desperate to make people believe it's healthier than it is.

And that's the question because at this point the answer is venture capital and debt. So the only new money entering the system in this case would be venture capital, maybe customers, but again because these companies are so unprofitable, all of that new money entering through customers is just incinerated. So Anthropic — it just came out just before this call that Anthropic is also leasing something like $161 billion worth of Broadcom chips, but $42 billion of those, I believe the actual financing is coming from Broadcom itself. So genuine dot-com bubble, Lucent Technology style. So how can I put this? Right now the money is mostly just circulating within the system. Venture capital funds Anthropic's customers. Anthropic's customers are predominantly AI startups that are unprofitable. So they're losing money and they're losing money because they're sending that money to Anthropic. Anthropic is losing money because the cost of compute is expensive, sending that money to Microsoft, Google, Amazon. Really, it comes down to when the bottom of the chain fails. It's how much funding can Anthropic get. But really, how much funding can Anthropic's customers get? Because this isn't coming from companies saying, 'Oh, I'm going to rent GPUs from Microsoft or I'm going to pay Anthropic directly and Anthropic is able to bank that money.' It's coming from the conga line of different venture capital sources and the limited partners who fund venture capital. And apparently some worrying — like a good like 10, 20% of their revenue came from overseas. That's technically new money, but my theory is that that money is actually coming from either Australia, which is new money — their money's good there — or it's coming from China, specifically with the big labs over there distilling Anthropic's models and paying for the privilege, which Anthropic claims it hates, but I think a sizable amount of their revenue comes from those customers. And 47% of their revenue comes from Amazon and Google reselling their models. So even like nearly half of Anthropic's money doesn't even come from Anthropic doing sales, but resellers, which is not illegal, but it's also a very different company if you think of it less as Anthropic doing business with people and more as other people are hawking their wares for them.

So I've been digging through endless analyst letters about Google Cloud revenue, Amazon Web Services revenue, uh, Microsoft revenue, and for the most part, they break out how much revenue comes from Anthropic and OpenAI and Meta, technically Meta. Meta is — Meta is within that $22 billion number. Going through that and actually I've read, I think at this point, every analyst letter I can find that talks about anyone spending money on AI and I cannot find a customer larger than $100 million a year outside of I think Perplexity and Cognition. Cognition I think spends what, 200, 300 million on GPUs. Perplexity maybe 100, 120. Otherwise it's mostly people spending piddly little amounts like a few million here and there, which is a lot of money to regular people but within the grand scheme of things, there's not that much. And the thing is, $22 billion of annual AI GPU spend, just to be specific. It's people renting GPUs. Sounds like a lot again, but not when these companies are spending 30 to 50 billion, the hyperscalers, on GPUs every quarter. GPUs that are barely getting installed anywhere.

It's not a dopey question because it's a very confusing one. So I am guessing here because I can't — I've tried, I've talked to sources, no one can tell me. What it is, I believe, is that Jensen Huang, CEO of Nvidia, has said if you don't buy these upfront, if you don't agree, I won't sell them to you in the future.

Pretty much. Yeah. It's like a Porsche owner situation. And much like a Porsche, they keep breaking. Uh, but it's kind of like that. It's also hyperscalers — I don't think they have another plan. So if they stop buying these GPUs, their investors will go, 'Oh, you stopped spending on AI. Brilliant. Why did you do that?' But as long as they keep buying, as long as they keep putting the money in, they don't have to answer that question yet. So I think that it's that plus people are just really stupid — companies are run — these companies are all run by management consultants. I also think what they might have done is pre-allocation where they say, 'Oh, well, I'm building a gigawatt data center capacity, I best buy the chips now.' Why? Not really an advantage of doing so, but it makes Jensen Huang look good. And then there's also the chance of failures. Now, I've not been able to hunt down this story fully, but I've heard a few reports that H100 GPUs, so a few generations back, are failing within six months across the board because apparently at the corners of the servers, they start overheating. Now, that would make sense with the H100s, wouldn't make sense with Blackwell, which are the latest generation, because I've also heard that Blackwell's barely being installed. Microsoft is about 130, 150,000 from what I hear, which again sounds like a lot but not when millions of these bloody things have been sold. So it really — this is going to be something we find out about in a book in three years. It's going to be someone finds out that Microsoft did it because Jensen Huang was texting Satya Nadella at 6:00 in the morning going, 'Buy me — buy my GPUs, pig.' The world's biggest fine operation. But it's so strange and you say it was a dopey question. It really isn't. It's actually probably the most consequential question of the bubble, which we do not have answers to because it's blatantly obvious at this point. Because Satya Nadella said in November last year, he said, 'I have chips I can't plug in.' Why would you buy them early? And I think it is that the scarcity is necessary to keep this going. But I don't actually want people to lean conspiratorial here because it's very rare that there is a conspiracy in the way that people think it. Like typically there are conspiracies, but it's very rare that everyone gets together and goes like, 'All right, folks. We're going to buy these GPUs as a means of keeping Nvidia inflated so the stock market doesn't crash, whatever.' No, I think it's more likely that across the board, Jensen Huang has said the Porsche thing, the Ferrari thing, like if you don't buy in now, I won't sell you the next generation. If you don't buy a full allocation today, I can't guarantee you the one in two months or three months or six months. And perhaps there is an end point to that. Perhaps there is a point when the companies go, 'Maybe this is bloody stupid.' But until then, they're going to keep doing it. And it's just so strange. It's genuinely — I can find nothing like it in history. And I've had people say it might be the bullwhip effect, which is when companies double order or triple order and then they get cancelled and then it whips back on the company selling them, like Nvidia in this case, and everyone screams. I don't know if it's that. I think this might just be a very bizarre corporate move that we're going to be unpacking for years, if not decades.