# Anthropic’s future compute costs dwarf its current losses

Markdown copy of https://realalpha.app/s/anthropics-future-compute-costs-dwarf-its-current-losses/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Paul Kedrosky](https://realalpha.app/people/paul-kedrosky/) · Managing Partner, SK Ventures · Invests in science-driven startups and founded HSBC James Capel’s technology equity research practice.
- Host: [Ed Elson](https://realalpha.app/people/ed-elson/)
- Show: Prof G Markets
- Episode: [Anthropic's $800 Billion Problem Is Hiding In The Fine Print](https://realalpha.app/v/ilkjZXwb8dU/), Oct 27, 2026
- Moment: 8:16 to 16:46 of the episode (9 min), playing on https://realalpha.app/s/anthropics-future-compute-costs-dwarf-its-current-losses/; the episode from there: https://realalpha.app/v/ilkjZXwb8dU/?t=496

## Quote

> “the preponderance of the commitments are in future, they're massive, and **they're all non-cancellable**.”
> — [Paul Kedrosky](https://realalpha.app/people/paul-kedrosky/) at [10:40](https://realalpha.app/v/ilkjZXwb8dU/?t=640) · Prof G Markets

## Takeaways

1. Nearly $700 billion in compute commitments remain even without SpaceX.
2. Most commitments are non-cancellable, limiting Anthropic’s flexibility.
3. Going direct means upfront and ongoing data-center investment.

## About this segment

Paul Kedrosky says Anthropic’s huge, mostly non-cancellable compute commitments and shift toward building its own infrastructure could change its cost structure. That makes today’s operating losses a poor guide to future margins.

From “Anthropic's $800 Billion Problem Is Hiding In The Fine Print”, Oct 27, 2026. [Watch the full episode](https://realalpha.app/v/ilkjZXwb8dU/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[8:29](https://realalpha.app/v/ilkjZXwb8dU/?t=509) and i get to the the liquidity and capital resources section this is like page 161 it's it's way further down and they go through some of their future compute agreements and they they tell they show us that they have an agreement with broadcom with 161 billion dollars and google 115 $15 billion. And you add it all up and you realize that they are about to spend on a timeline that they don't tell you, but they're about to spend more than $800 billion on compute. And none of that is included in the income statement.

[9:05](https://realalpha.app/v/ilkjZXwb8dU/?t=545) And so I'm looking at the income statement now and I'm thinking, does this actually tell me anything? Because it seems as though they've left all of their compute costs at the bottom as a footnote, not because they're trying to pull a fast one, I don't think, but because those costs haven't actually been reflected in their income statement yet because they're about to happen. So what do you make of that part?

[9:28](https://realalpha.app/v/ilkjZXwb8dU/?t=568) I think that's exactly right. And this is a deep structural problem that I think Moody's pointed to a few weeks ago in terms of the way that all of these compute commitments trickle back to and are consolidated in the financial statements because, well, they're non-cash obviously now because they're future commitments, but they don't even show up necessarily in the form of orthodox debt because these are future commitments that aren't actually debt instruments.

[9:51](https://realalpha.app/v/ilkjZXwb8dU/?t=591) They're more like a lease that's being discounted back to their present value today. And they're all obviously highly consequential and investors would need to think about it as you consider what the actual cash flow nature of the business looks like going forward. Because in particular, it's not just the size of the commitments, they're all non-cancellable, right? This is really important. This is not something where I say, you know what, it's not working out the way we wanted.

[10:14](https://realalpha.app/v/ilkjZXwb8dU/?t=614) We're not gonna have that commitment. No, these are all non-cancellable capital commitments with the only exception being that there's a growing capital commitment in terms of doing compute at SpaceX, which I think works out to about $3 billion a year. So much smaller in the great scheme of things than these other non-cancellable compute commitments. And that, I think, is cancelable on a 90-day basis, 90-day notice, which speaks more to what SpaceX is trying to do, I think, than anything else, that they're desperate to have this cash flow coming back in because they, in turn, consolidate it as growing AI revenues when it's really not, obviously.

[10:46](https://realalpha.app/v/ilkjZXwb8dU/?t=646) This is just compute you're providing to someone else on a rental basis. So it's important, I think, to remember that the preponderance of the commitments are in future, they're massive, and they're all non-cancellable. And even if you were to get rid of SpaceX, you're close to $700 billion in compute spending. That's exactly right. It's immaterial in terms of the great scheme of things. And I think that's by far the most important factor here that they're looking at. And in turn, those compute commitments are even more important because in a sense, you're working through distributors. And they say this in the S1 itself. And so they're distributing their large language models through competitors, right?

[11:24](https://realalpha.app/v/ilkjZXwb8dU/?t=684) Microsoft is one of the prime examples, Google, and so Amazon. And this is a really unusual relationship because you're distributing through companies who also sell on your behalf, but also take a rake off the top as they do it. So my analogy is, this is a little like as if Spotify, which has a rake from the music companies of about 30% of their revenues, but instead of it being 30% of revenues, it was a fixed percentage.

[11:48](https://realalpha.app/v/ilkjZXwb8dU/?t=708) It doesn't matter how much you sell. Not only that, all of your competitors also have music streaming services, and they're perfectly happy to kick you out. So this is the deep fragility in the nature of the distribution business. And I used the Spotify example because it shows you how a company that's been penalized for this kind of relationship is actually in a much more favorable position than Anthropic is.

[12:07](https://realalpha.app/v/ilkjZXwb8dU/?t=727) So this brings us to what I think is the most illuminating part of the filing, which is that we learned, and a lot of people said, oh my gosh, this is so great, that Anthropic is not spending on CapEx, which has obviously been a huge expenditure for the hyperscalers, i.e. building their own compute building their own data centers and everyone said look this works because they don't have this capex issue but they admit to the problem that you're describing which is that they are leasing compute from their competitors primarily which is probably going to be an issue later down the line and so i'd like to just quote read you this quote from the s1 our increasing efforts to build out our compute and infrastructure require substantial capital expenditures and are changing the cash flow profile of our business whereas the compute

[13:01](https://realalpha.app/v/ilkjZXwb8dU/?t=781) purchased through our csp agreements is generally paid as the compute becomes available to us we expect an increasing portion of our compute and infrastructure investment to take the form of capital expenditures and other commitments including lease obligations that require significant upfront payments well in advance of the related compute becoming operational. That tells me they're about to build data centers, and they're about to spend a shit ton on them.

[13:27](https://realalpha.app/v/ilkjZXwb8dU/?t=807) Is that right? Right. That's exactly right. Shit ton being the technical term here. So this is also a fancy way of saying, or maybe a relatively opaque way of saying, that what you see in the past in terms of the structure of the business won't be representative of what it looks like in the future. That we can't maintain these leasehold, these lease relationships with our competitors in terms of passing product through.

[13:46](https://realalpha.app/v/ilkjZXwb8dU/?t=826) So we're going to be forced to make duplicate capital investments so that we're not beholden to them, both in terms of strategically, but also in terms of this rake off the top in terms of what we have to give up from what's called marketplace sales, where we're actually selling products through the marketplaces on these other services. And so that's what you're saying. The nature of what we spend in future not only is going to continue, but it's going to accelerate in terms of the CapEx because we're in this strategic quandary of being reliant on our largest competitors.

[14:13](https://realalpha.app/v/ilkjZXwb8dU/?t=853) Which tells me that the income statement is actually irrelevant, right? I mean, if that's not even close to a reflection of what this business is about to look like, why does it matter? Yeah. And I think that's the thing that people are missing is that, in a sense, we all know that income statements and financial statements in general are backwards looking documents. But this is backwards looking at a particularly sharp inflection. And the sharp inflection is from a business that was largely relying on distributors to a business that's increasingly going to be trying to go direct. This is the transition that's underway. And in going direct, it has two consequences. One is massive current capex for doing that.

[14:53](https://realalpha.app/v/ilkjZXwb8dU/?t=893) And second, and this is the thing people miss all the time, and kudos to Michael Burry and others who are pointing this out, is that it creates maintenance capex. You don't have to spend it just once. You have to maintain these things. And so you have a future capital obligation that's not just the amount you spend to build, but also the amount you have to spend to maintain these very expensive facilities.

[15:11](https://realalpha.app/v/ilkjZXwb8dU/?t=911) But they're in this Hobbesian choice. They have no choice but to do that because otherwise it's to be reliant on their largest competitors who could, you know, like Microsoft did back in its heyday, you know, cut off their air supply at any moment. So here's a little conspiracy theory that I'd like to see if you think is perhaps correct. Is it possible that this is why they wanted to go public now? Is it possible that they wanted to take this one moment in time where the operating profitability looks pretty good right before it's about to look quite bad?

[15:49](https://realalpha.app/v/ilkjZXwb8dU/?t=949) Because that seems like that is what is about to happen. They're at this inflection point. No, no, no, no, no. Listen, I was arguing that earlier today multiple times with people. So I think that's exactly what's going on. This is as good as it looks from the standpoint of the business as currently structured. And now you go through this inflection that makes it much more CapEx intensive. So you're essentially changing jockeys and changing the underlying structure of what's happening.

[16:10](https://realalpha.app/v/ilkjZXwb8dU/?t=970) And so I think this was probably their, and they know this, this is the best time to do it because the future business is going to be very different. They're essentially reduplicating AWS, Amazon Web Services, in both directions, building up and building down in terms of both the client-customer facing as well as the cloud-facing stuff. It doesn't mean it's impossible. It could work. But it does mean that the build…

[16:32](https://realalpha.app/v/ilkjZXwb8dU/?t=992) the way the business is characterized in this current statement is not representative of the