It's official: Anthropic is going public, and boy, are our eyebrows tired from poring over hundreds and hundreds of pages of questionable claims.

You can listen to the latest episode of The Kettle right here on this page, as well as on Spotify, Apple Music, or YouTube. Those platforms also let you subscribe to The Kettle, so you are always notified when the latest episode goes live.

Anthropic is going to need to do a lot to make its massive IPO more than just a flash in the pan, as systems editor Tobias Mann and senior reporter Tom Claburn tell host Brandon Vigliarolo on this week's episode. 

For one, the company has nearly a trillion dollars in compute commitments to pay for, which is going to mean it'll be burning through more cash than a datacenter gas turbine turned up to 11. In addition to its massive spending for the foreseeable future, there's also the fact that much of its computing power relies on customers who are also competitors.

In short, we're skeptical, but you probably already knew that. It's after all, and we're not just taking what the AI industry claims at face value - if it looks like a house of cards, it probably is. 

A lightly edited transcript is below: 

Hey everyone, and welcome to the latest episode of 's Kettle podcast. I'm your host, reporter Brandon Vigliarolo, and this week we're focusing on what I think our contributor Carly Page aptly described as the peak of the AI hype cycle we've been in for the past few years, and that's the public release of Anthropic's initial public offering filing. With me to talk this over is our systems editor Tobias Mann and senior reporter Tom Claburn, both of whom I'm sure have plenty to say about this matter.

Welcome guys.

Good to be here.

Good to be here.

So, I'm sure we've got a lot to cover. 614 pages is a lot to go through. Tobias, I understand you were combing through this thing with a fine-toothed comb. Talk us through the hard numbers before we get into some of the more wobbly parts of this IPO.

Sure. At face value, going into any S-1 filing, it is a kind of mess of risk mitigation, kind of covering your butt. They all pretty much read like a doomsday prophecy.

Right.

The S-1 definitely did that. But on the numbers side of things, the takeaway for the market was kind of interesting because there were some people looking at the numbers and going, "OK, this actually seems reasonable." Their operating losses were only about 700 million dollars.

I think on 16.3 billion net [revenues], right? That was what they reported.

Right.

Now they also mentioned, I think, and I was going to ask you about this, $80.7 billion total losses, but I understand that was an accounting adjustment of some kind. Did they mention in the IPO what that meant specifically?

Not specifically. It's a little fuzzy, and it's always kind of like this.

Of course it is, right?

Anthropic's business is extremely cash heavy, and so they had to go to partners and say, "Hey, we need capital in order to pay for compute to train these models. We can either take it in debt or you could also get repaid in valuation, our equity." And so I believe a large part of that loss is they're trying to write off a lot of the equity that they're going to have to pay off when they do go public. And so I do believe that's the lion's share of that loss. Because it does look huge, but when you think about how companies go public and how they raise capital as startups, it's not really that surprising. It's just a huge number because Anthropic is arguably the largest IPO in US history at this point.

Right. But bottom line is their actual day-to-day business losses were $693 million....It was $1.2 billion, I think, for the first half of the year or same time last year or something like that. So it's gone down.

Yes.

And they're trying to go out at two trillion now? Is that the number?

Well, the number kind of depends on whether they're going to be able to sustain this. I was going to mention for the first time ever, their revenues outpaced their compute expenses two to one. This has never been true in the company's history, and it's one of the reasons that people are feeling more confident about this. The issue that really comes from all of this, and what has gotten missed somewhat in the more rosy interpretations of the S-1, is that they also are signed up to have to repay upwards of 800 billion dollars and counting of compute commitments out through the next two decades or so.

You outlined in your story just a whole host of people they owed billions of dollars to for compute. Like you said, it would be great. That $16.3 billion in [revenue] would be great if those compute costs were going to hold steady, right? But they're not. They're going to keep climbing and there are a lot of bills that are going to be due. So, profitability is relying on making more than that nearly trillion dollars in compute obligations, and that's a big if.

I mean, the big question mark is in terms of what the technology does for inference costs, because I think a lot of this is betting on being able to get the same hardware that is much more efficient down the road and maybe not drop your prices proportionately and profit on the difference. But it remains to be seen whether there will be enough competition to just have the price go do with the increasing cost. And you see that a lot. I think the inference costs have dropped pretty significantly over the year, on average. I mean, it's really hard to calculate inference costs because there are so many different factors and it's really situational depending on the workload and how much caching is going on and things like that. But overall it's a lot cheaper to do inference now than it was a few months ago.

Per token, yeah.

But again, I feel like if you're betting this $2 trillion IPO on the fact that we're going to keep getting better and cheaper at inference, right? I mean, I know Moore's Law is kind of a thing of the past, but at this point, what if we run up against another compute bottleneck and inference stops getting cheaper before they need it to?

Obviously any IPO prospectus is one big document about risk, right? Like, "Hey, you might make a ton of money investing in this company, or we might go complete belly up and you lose everything," right? That's always a possibility. 

Right. And they still haven't demonstrated that the use case for AI is wide enough at volume that they claim it is. I mean, in certain sectors it's pretty obvious now there's going to be AI for a long time. Coding, everyone kind of agrees that works and we know how to manage that. And then there will be adjacent fields to that. But once you start getting into more general office work and regular business tasks that are not super-compute heavy, I don't know. It's a good question. We'll see if people are going to use it that broadly, but I just see it being overly optimistic.

The big question I have right now is: Is there enough user demand to support even a $200 billion a year revenue stream? Because even going off just commitments that they can't get out of made in the first half of this year, that's $417 billion running out through about 2033 to 2036. 

I think the upper end of your estimate was a total, including newer commitments, of about $807 billion. At the lower end, $782 billion at the lower end. That's not that much of a difference when you're looking at hundreds of billions of dollars.

But if they can't make a profit with a two to one revenue-to-compute ratio — if they have to make twice as much in revenue in order to pay for their compute and still not break even — you extrapolate this out and they're going to have to be making $200 billion plus, assuming no additional compute contracts are made during that period over the next year, which is definitely not going to happen. They've already said they're going to keep making compute contracts. So this number just keeps getting more absurd the further out you go, unless inference magically gets dirt cheap, which we haven't seen.

Right. And it could cap. You know, I think also you made a really good point in your story, Tobias, when you mentioned that some of their biggest customers are also some of their biggest competitors. They're leasing space from hyperscalers, they're making deals with people to host their compute. But these are also companies that are operating their own AI models. And you might end up in a situation where they're like an MVNO, which, if anyone's not familiar with a mobile virtual network operator, is cell phone companies that are not T-Mobile, Verizon, or AT&T. Anyone who's not those big three in the United States is operating a virtual network that piggybacks on someone else's infrastructure. And when that infrastructure is busy, you get less priority. So we could see them run into a similar situation there, I imagine.

There's an interesting trend that another piece I'm working on later today touches on. Some of the AI gateways like OpenRouter are now starting to move away from just focusing on price metering and looking towards things like fastest time to tokens and all the quality of service metrics in terms of the way that they price and deliver the models, because inference pricing has become very commoditized in terms of what they're dealing with. They're not necessarily serving all the frontier models at very high prices. They're dealing with a lot of open weight models. And if you just focus on the lower end of the market, there's not a lot of margin there. So it makes sense to go, "Well, maybe if we deliver this faster or have better reliability or service, we can move up the price and get a little profit." And I think everyone in the industry is starting to realize that the music is stopping and they have to take a seat at some point and make the business model work. So there's going to be a lot of pressure on that. And that's also showing up in things like the licensing of models. A lot of those open weight models have seen their licenses get more restrictive over time as companies realize, "Well, maybe we don't want someone competing with us with this thing that we've made." So I think there's a lot..

That reminds me of Mistral just releasing the open weight version of Large 4, "Le Chonk". And the licensing model basically said if you intend to use this for any down-market purposes, you've got to sign a fairly restrictive agreement with us so that you're not just using this open weight model without us benefiting from it, right?

A lot of the big Chinese models, like Kimi K3, I believe were among the first to move back in this direction, basically saying, "Hey, if you're just going to deploy this internally, we don't care." If this is just for your enterprise, fine, but if you're going to be selling an API service like OpenRouter around it, you gotta cut us in. So there is definitely that. And Anthropic in their S-1 – again, every S-1 reads like a doomsday prophecy. And this S-1 of the things that they raise: open models are a threat on multiple levels, one of them being that a lot of their customers are startups sensitive to economic downturn. So those customers will likely be the first to jump ship to cheaper alternatives that are maybe not as good, but they can afford if the economy runs into any issues.

I wonder if that'll happen or not, right?

Was it Elon Musk who was tweeting about the end of the neoclouds recently? With financing getting tighter, a lot of the neoclouds living on borrowed money are going to find it harder to sustain their loans and are not going to survive this. So there will be a shakeout at some point. Does it hit the big companies? I don't know. All the cloud providers look pretty happy because they get business regardless. Even if Google's AI models are not first in people's minds, they're still serving all of this stuff.

We've said that on multiple episodes, and in multiple articles we've written, right? Basically, if the AI industry does implode, Google's going to be fine, Amazon's going to be fine, Microsoft's going to be fine; these guys have other businesses. Anthropic and OpenAI, not so much. That's all they do. And in many cases they're renting compute space from these other providers. So they might take a bit of a hit there, but by and large, they're going to be okay. They've got compute, they'll turn it around, they'll find other people to use it. Which is interesting, from your article, Tobias, you kind of summarize that the entire pitch this IPO makes is that bigger is better, right? But as we were just talking about, there are a lot of reasons to say that's probably not the case. It makes me wonder if they're still living in the token-maxxing era, whereas I just wrote a story this week about AWS telling its engineers to back off, stop using so many tokens, be more judicious, use AI well, don't just use it. And Microsoft said this months ago in August – they were saying tokenmaxxing doesn't factor into how they want to do business. So you've got two very big consumers of AI basically saying, "Y'all need to slow it down and be more purposeful." So, bigger isn't necessarily...

Anthropic's business is predicated on the idea that more compute means more revenue. However, the disconnect is: Can revenue scale fast enough? Is there enough user base? Is there enough token demand to support the scale they need to cash the checks they've already written? And it really looks like a house of cards. They warn about this in the S-1 – they have to – that if their revenue can't keep up with their compute commitments, they could default. And that default could turn into a cascade where if they can't pay one bill, they can't get enough compute to generate revenue necessary to pay the next bill, and it all falls apart.

If it happens to Anthropic, the other AI frontier labs are not far behind, because that's going to be a symptom of a larger problem. And that's that people are basically saying what a lot of non-AI companies or companies that do AI along with other things, like AWS, are saying: We don't need these massive frontier models for the vast majority of business purposes. They're not necessary.

Right. A lot of the enterprise AI products you see now are about trying to figure out ways to reduce that cost, too.You have some sort of gateway or harness system that looks at your jobs and routes them to an appropriately cheap model so that most of the time you're not using top-tier intelligence; maybe you're using a decision model, a regular classifier, or something like Gemma to do language translation. You don't always have to use that expensive model, and that price pressure is going to hit them, too.

To Tom's point, Cadence Systems, which does a lot of electronic design automation and circuit board layout, just rolled out a new product line called Maestro that's built around Nvidia's Nemotron open-weights models. The idea is that they had been using proprietary APIs to power their agents that assist engineers with PCB layout and chip design, but those agents generate a ton of data that could be used to train and fine-tune open weight models. So their latest pitch is, "Hey, we know our customers like this, but it's getting expensive to continue using these proprietary models from OpenAI, Anthropic, or Google. So we're going to give you a way to take an inexpensive model, use the proprietary model to generate training data that we can use to fine-tune this open model, and then you can decide how much of the work you want done on the expensive proprietary models versus the cheap open models. You can deploy those models on-prem, in our cloud, or on your preferred cloud. Just use our agent platform and pay the subscription fee."

I can't imagine most businesses are going to fork over all this money to proprietary model providers when they don't need those resources. It really makes me wonder how much of this IPO is forecasting the bursting of the bubble. I feel sometimes like we talk about the AI bubble bursting like doomsday prophecies: "It's happening now!" But there have been more signs recently. Chinese AI stocks have been hammered recently. There have been recent earnings reports from American companies that have not looked  Yeah.

Google was hammered; they lost like $500 billion in stock value based on the fact that their search was underperforming

That's that one I wasn't able to think of, that minor little blip, right? 

AI overviews not monetizing as well. So there's a lot, yeah.

A problem they created for themselves.

Google creating its own problems? What?

I would agree that Anthropic's survival depends fairly heavily on pushing the frontier forward. In order to do that, it needs a disproportionate amount of compute that it may not be able to afford on a continuing basis, especially if Microsoft, Google, and all of these cloud CSPs that provide compute for them and are among their biggest customers decide, "We don't really need cloud for this. We've reached a point where our own open-weights models or small proprietary models are good enough for drafting emails or generating code bases for our customers."

So they're going to start buying fewer tokens from Anthropic or OpenAI when they eventually IPO. They're likely to go through very similar challenges because they're also getting most of their compute from companies like Oracle and Microsoft at this point.

Which was one of the questions I wanted to ask as we wrapped up: Do you think OpenAI is going to be in any better position once their IPO comes out? Because neither of them owns a lot of their compute, and OpenAI's attempts to build its own compute seem to have fizzled out a bit. Could they be in a better position or is it going to be more of the same?

I don't know with OpenAI whether they're going to be able to overcome the same risks that Anthropic has. It does feel like Anthropic is sidestepping a lot of hard questions by pushing forward with their IPO now.

With Anthropic going public this year, if OpenAI goes public next year, I don't think they can avoid the scrutiny, because the world is going to have woken up to the risk factors associated with these giant model houses. And I think they're going to face a much harder time with this, even before they get there.

OpenAI has the distribution of ChatGPT, which is demonstrating itself to be a useful tool that people like to use. But that's not where their monetization gold mine is going to be, unless they become the Google of AI search, and Google is already trying to get to that space, too. So I don't know how they're going to make it into the enterprise business and get the margins they need to pay back all the money they've taken.

The point being that OpenAI's revenue streams may be sufficiently different to insulate them from this. They have been trying to position themselves as an AI advertising company as another revenue stream, which Anthropic has flatly rejected. So it's possible that OpenAI's business model is sufficiently different by next year.

12 months is a long time in this industry. If Altman takes them public this time next year, the conversation might be very different. Having said that, Wall Street's eyes are going to be open to a lot of the risk factors in a way that Anthropic is sort of sidestepping right now by being the first really big model house to go public.

Which is honestly kind of surprising to see OpenAI press the brakes while Anthropic presses the gas. I'm assuming they saw an opportunity to get out the gate first and avoid answering the hard questions. Because OpenAI basically said, "Hey, we're going to pause our IPO while we assess these risks," given they've been in the doghouse regarding AI behaving badly this year, giving Anthropic that opportunity to scoot past and not have to deal with these things.

I think the other tell with Anthropic is actually that two-to-one compute ratio, because that's the first time that's ever happened. And it will be very telling if in their first or second public quarter that just disappears. So the timing might be related to the fact that they can now claim their cash flows are on a positive trajectory towards being cash flow positive and moving towards profitability. OpenAI delaying their IPO, I suspect, is largely because they need to get to a similar point. Otherwise, people are going to look at this and go, "This has no possibility of ever being profitable." At least with Anthropic's S-1, halfway through 2026, it looks like they could be profitable. Whether that's true in in their first public quarter is another matter until

To close us out here, I think this statement from your article, Tobias: "The only way to justify this kind of spend and risk is if you believe the upside is enormous, like the biggest economic transformation in the history of humankind enormous." And that raises my final question for you both: Is it going to be worth it? And what's the chance that Anthropic is going to flame out spectacularly in the next year or two?

Tom, I'll let you take the first crack at that one.

Ha ha ha.

Yeah. I mean, I don't know how spectacular it will be. I think it's gotten to the point where it's sort of too big to fail, but I could certainly see some desperation moves where one of these big AI companies strikes a deal with Apple for distribution, because they really need the kind of money that Google was paying Apple for, or sorry, that Google is paying Apple for for distribution. And I can see something like that happening because they really need revenue. I don't really see it blowing up because it's everybody. So they're going to figure out a way to do the least harm if things go downhill. And that may mean parting out part of their businesses. Who knows?

I would tend to agree with Tom. I think about what happened with the dot-com boom as an indicator. The situation is quite a bit different because the AI industry has wormed its way so deep that it's not the surface-level situation we saw with the dot-com boom. There was a ton of fiber laid during the dot-com boom that didn't get utilized until much more recently, but it did get utilized. So I think AI is probably bigger than the internet and going to go down in the history books as the next industrial revolution, in much the same way the internet felt like that.

After the dot-com boom, it was big, but it wasn't as big as people prescribed the value to. It took another 10 years for it to actually become the thing everybody wanted it to become. It eventually rose to that valuation, but it took much longer than the market could sustain. And I think we're probably in for that.

So in other words, don't buy during the IPO. Don't buy high and hope to sell higher, because you might have 10 years to wait before you get back to that IPO valuation or above. Who knows? Maybe it'll go faster. No matter what happens, there's a good chance we're going to be right here to talk about it because this sure can fill a lot of podcasts and articles. Digital ink is going to be spilled on this topic for years to come. We hope you'll join us to hear all about it on the Kettle. Thanks for listening. ®