enesptjazh-Hanszh-Hantfrdeitkoruarhiidtrnlplvithsvhemsukcselrohudafinobn

Anthropic's S-1 shows $2 of revenue per compute dollar and $430 billion it can't cancel

Gil Luria reads Anthropic's S-1 as a profitable war chest. Paul Kedrosky sees $430 billion of locked-in compute and an IPO timed at the peak.

By Published October 27, 2026 7 min read Grounded in the CEOInterviews.AI corpus

Anthropic filed its S-1 on October 26, and within hours two analysts had read the same document and come away with opposite companies. Gil Luria of D.A. Davidson saw a business that sells compute for twice what it pays and holds about $131 billion of cash, so the IPO is "not an act of desperation." Paul Kedrosky of SK Ventures went to the liquidity section and found hundreds of billions of compute contracts that don't touch the income statement yet. Both are right about the numbers. They disagree about which numbers describe the company you would be buying, and that is the whole question for anyone pricing the roadshow.

Anthropic CEO Dario Amodei speaking on stage at TechCrunch Disrupt 2023
Anthropic CEO Dario Amodei at TechCrunch Disrupt 2023. Photo by TechCrunch, Wikimedia Commons, CC BY 2.0.

Luria's read: the margin the whole AI trade needs

Luria went on Bloomberg the afternoon of the filing and led with the number that has been argued over for two years, whether a model lab loses money on every query. His answer is that it doesn't.

Their revenue is twice as much as their total compute costs, including training. And that was unclear from all the leaks, and it is now very clear in an SEC filing.

Gil Luria, October 26, 2026, Bloomberg Tech

The Bloomberg host filled in the rest. Compute ran about 10 times revenue in 2023 and 55% of revenue in the first half of 2026, and the second quarter showed GAAP operating income of $974 million. The headline loss, over $80 billion for six months, is mostly a non-cash charge on convertible notes. Luria's point was that the trend matters beyond Anthropic. If a lab can resell compute at a widening markup, it keeps buying from the hyperscalers, who keep buying chips, who keep buying power and construction.

The whole AI economy depends on that margin. That's why to me that was the highlight of this S-1.

Gil Luria, October 26, 2026, Bloomberg Tech

That is a fair reading of the past eighteen months. Revenue for the first half was about $16.3 billion against $1.2 billion a year earlier, and the Motley Fool's tally puts first-half gross margin at 85%. The question Kedrosky asked is whether the past eighteen months are the right sample.

Page 161 holds the contracts

On Prof G Markets the next morning, host Ed Elson walked Kedrosky through the liquidity and capital resources section. Broadcom at about $161 billion. Google at about $115 billion. Elson's own sum came to more than $800 billion. Yahoo Finance's count is more conservative: about $637 billion across every named deal, and roughly $430 billion in four contracts with Broadcom, Google, Amazon and Microsoft that Anthropic cannot cancel. The filing leaves its own total blank, so every sum you see, including mine, is an outside count. The Broadcom piece is five-year leases on hardware built around Google's TPUs, booked as finance leases.

Oh, 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.

Paul Kedrosky, October 27, 2026, Prof G Markets

Yahoo puts the SpaceX deal at up to $121 billion through 2029, with either side free to leave on 90 days' notice from early 2027. That is bigger than Kedrosky's yearly figure suggests, but it is also the one contract Anthropic can walk away from. The $430 billion it can't is the number to carry into the roadshow.

A Google Tensor Processing Unit 3.0 board
A Google TPU v3 board. Anthropic's largest disclosed commitment is Broadcom leases on TPU-based hardware. Photo by Zinskauf, Wikimedia Commons, CC BY-SA 4.0.

Why the income statement may be the wrong page

Kedrosky's real argument goes beyond the size of the bill. Anthropic today sells largely through Amazon, Google and Microsoft, which host its models and take a cut of marketplace sales, and which also sell competing models. His analogy was Spotify paying a fixed rake to labels that each run their own streaming service. The way out is to own more of the compute, which means capital spending, plus the maintenance spending that follows it every year after.

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.

Paul Kedrosky, October 27, 2026, Prof G Markets

Elson then floated what he called a conspiracy theory: that Anthropic chose to list at the one moment its operating line looks best. Kedrosky didn't push back.

This is as good as it looks from the standpoint of the business as currently structured.

Paul Kedrosky, October 27, 2026, Prof G Markets

He named two things that would prove him wrong. Compute intensity per token has to keep collapsing, and revenue growth has to reaccelerate. Bloomberg reported third-quarter revenue of $16.8 billion against $11.6 billion in the second, 45% sequential growth but a smaller dollar add than the quarter before. Luria pinned that on enterprise share falling from near 100% as OpenAI and open-weight models caught up. Kedrosky put it more bluntly: a company needing a growth reacceleration at the moment it lists "just blows the top of my head off."

Seven founders, one charter, and no vote for you

The part of the filing Luria found hardest to price has nothing to do with compute. Anthropic is a public benefit corporation, and Bloomberg's reading is that its seven founders keep 50.1% of the vote. Most dual-class companies give you one or two people to model. This one gives you a committee.

Here we will have to wonder what those seven people agree on, and whether there will be some dynamics there of them disagreeing on things and therefore changing the trajectory of the company.

Gil Luria, October 26, 2026, Bloomberg Tech

He added that the filing says outright Anthropic won't always maximize value and doesn't want its shares trading above what the company is worth. I have not seen that sentence in a prospectus before. CEO Dario Amodei spent September arguing that the industry should slow down, and the risk factors include catastrophic outcomes for humanity. Investors have bought founder control many times. They have not often bought a founder group that tells them in writing that the share price comes second.

The tape didn't care

None of this slowed the market down on the day. The Nasdaq 100 rose 1.4% and chips and neoclouds traded green. Dan Ives put first-day appetite at $1.5 to $1.7 trillion and said he would buy.

It's the third inning. It's, you know, it's called 15% through this AI revolution.

Dan Ives, October 26, 2026, CNBC International

Ed Yardeni went further on CNBC the same day, and his version doubles as a reason to worry. He expects failures among neoclouds after Elon Musk's warning of "guppy-level death" for more than 200 of them, but he thinks the big players will catch the small ones.

You know, Anthropic is going to go public, they'll have a lot of cash, and if they need to help out a neocloud, they'll do it.

Ed Yardeni, October 26, 2026, CNBC Television

Read that next to Kedrosky. If the IPO cash becomes a backstop for Nscale or Lambda, two of the GPU suppliers named in the filing, then the $131 billion war chest and the compute bill are the same money. Luria made a related point on BNN Bloomberg. Capital "is starting to seize up right now," he said, except for the investment-grade hyperscalers. Anthropic is raising while the window is open, and Meta is doing the same.

OpenAI's price makes Anthropic look reasonable, on loose math

The other yardstick out this week was OpenAI's reported round of up to $30 billion at a $1.25 trillion pre-money valuation, down from the $1.4 trillion talked about in September. Jason Calacanis divided it by an annualized run rate of about $50 billion and called the result fair.

Anyway, it's 30 times topline for an incredibly high-growing business. I think they have actually right-sized the valuation to a certain extent.

Jason Calacanis, October 26, 2026, This Week in Startups

His division is off, since $1.25 trillion over $50 billion is 25 times, and he admitted he didn't know how the run rate was calculated. But the instinct is the one the market is using. Price the labs on revenue and the compute contracts sit off to the side, which is exactly the habit Kedrosky is warning against.

So which S-1 is the real one?

I think Kedrosky has the better argument about the future and Luria the better argument about the present, and the roadshow will decide which one sets the price. The things to watch are concrete. Luria expects fourth-quarter and 2027 guidance to come during the roadshow, with pricing around November 19 or 20. Luria said $100 to $150 billion of 2027 revenue would impress him. The four locked contracts cost about $66 billion a year if you spread them evenly. Revenue in that range lets the margin Luria likes absorb the commitments Kedrosky fears. If growth keeps fading, the next filings will show the compute bill instead of the 85% gross margin. Either way, an investor buying at $1.5 trillion gets the bill and no vote on it.