Reading Anthropic's S1 Like a Value Investor

Recent updates 🗞️

🏦 Fed hikes 25bp to 3.75-4%, its first increase since 2023
📉 10-year Treasury touches 5.33%, its highest level since 2002
🔻 Chip stocks fall 11.4% in a week, the worst since April 2025
🛢️ WTI crude tops $100 as tanker attacks in the Strait of Hormuz hit a record
📚 Reading “How to Make a Few Billion Dollars” by Brad Jacobs
YTD Portfolio Performance: 14.38% YTD

Every few years, a company arrives that forces you to answer an uncomfortable question: what would I have to believe?

Anthropic’s IPO is that company’s moment. When I started this post two weeks ago, press accounts of the draft prospectus described a target valuation of about $2 trillion, more than double the $965 billion it was worth in May. Then came an investor day, a quarter of results, a rough week for AI stocks, and on October 20, the most detailed look yet at the draft prospectus from the Wall Street Journal. Reuters now reports that some investors are discussing something closer to $1.5 trillion. The listing is expected on the Nasdaq in mid-November, before Thanksgiving, and as of this writing the public S-1 still hasn’t appeared on EDGAR.

I’ve spent the last two posts on AI from two angles. In The AI Margin Squeeze, I asked whether foundation models become commodity infrastructure. In Right Thesis, Wrong Leverage, I watched a correct thesis get destroyed by financing. This post is where those two ideas collide, because the Anthropic numbers contain both stories at once.

A caveat before we start: the figures below come from press reports of the draft filing (most recently the Journal’s) and the company’s investor day, and they don’t agree with each other everywhere. I’ve used the ones that at least two sources corroborate and flagged the rest. Check the final public S-1 before you act on any of it.

The growth is real, and it is absurd

Start with what impresses me, because it genuinely does.

• Enterprise & API Dominance: Enterprise and API customers make up roughly 80% of revenue. That is not a consumer fad.
• Unprecedented Slope: Revenue climbed from about $9 billion annualized at the end of 2025 to a current Q3 annualized rate of roughly $67 billion.

I’ll admit this is the part that gets me excited. Revenue went from about $9 billion at the end of 2025 to something like $47 billion by May, by one account. I can’t think of a precedent at this scale. When growth is this steep, valuation arguments stop being about the present and become arguments about how long the curve keeps bending.

When 45% is a disappointment

Here’s the part that made me think. Anthropic grew revenue 45% in a single quarter, and the reaction was disappointment. Q1 to Q2 had been more than 140%, driven by Claude Code. After the October 14 investor day, Reuters reported that some investors were let down by the slowing sequential growth and started discussing a valuation around $1.5 trillion. That’s roughly $500 billion of value repriced on a quarter most companies would frame and hang on the wall.

Do the quick math and you see why. It’s a rough comparison, since run-rates and quarterly totals are measured differently, but Q3 annualized works out to about $67 billion, barely above the $65 billion run-rate reported back in July. At this valuation, the market isn’t paying for growth. It’s paying for the acceleration of growth, and the acceleration just turned.

The $80 billion loss (and why most of it isn’t what it looks like)

The headline number in the Journal’s story is a doozy: a net loss of $80.7 billion in the first half of 2026, one of the largest half-year losses ever reported, and nearly double the $42 billion Anthropic lost in all of 2025.

But read one line further. $72 billion of it came from accounting charges tied to changes in the value of convertible notes and other instruments the company issued to raise money. Charges like this generally grow when a company’s valuation climbs, because those instruments become worth more to the people holding them. So here’s the counterintuitive part: a big chunk of this “loss” is a side effect of Anthropic becoming more valuable, not less.

That doesn’t make the number meaningless. Strip out the $72 billion and there’s still roughly $8.7 billion of loss (my arithmetic) in a half-year that included a quarter with about $559 million of adjusted operating profit. The bridge between those two figures is whatever “adjusted” leaves out, with stock-based compensation the usual suspect. That bridge is the first table I’ll look for in the public S-1.

The more encouraging data point is the ratio underneath. In 2025, Anthropic spent $7.3 billion on compute and infrastructure to generate $4.6 billion of revenue. In the first half of 2026, it spent nearly $9 billion to generate $16.3 billion. Compute costs nearly quadrupled year over year; revenue grew almost fifteenfold. That’s the operating leverage story working in the company’s favor, at least for now.

The price of admission

Now put the valuation next to that revenue. Annualize Q3 and you get roughly $67 billion. At $2 trillion, you are paying about 30x sales. At the $1.5 trillion now being discussed, about 22x. The company’s own forecast, which one report puts near $200 billion of revenue by 2028, would bring that down to roughly 7.5x to 10x. That is a forecast from the seller, so weigh it accordingly.

Then add the macro backdrop. The 10-year Treasury ended September at 5.29% and touched about 5.33% in early October, its highest level since 2002. It has since eased to around 5.1%, but mostly because investors ran to safety during a brutal week for tech. A risk-free asset paying over 5% is a very different alternative than it was when the last big AI rounds were priced. Long-duration equities, the ones whose cash flows live mostly in the 2030s, are exactly what higher discount rates punish. A $2 trillion price tag is a claim about cash flows a decade away, priced in a world where the bond market is asking for more to wait.

I’m not saying it can’t work. I’m saying the margin for error is thinner than the growth rate makes it feel.

The mood around AI has turned, too. In the week of October 12, the Philadelphia Semiconductor Index fell 11.4%, its worst week since April 2025, and the neoclouds fell even harder. Broadcom’s Hock Tan told a conference in Singapore that many AI investors don’t understand what they’re buying, and compared it to gambling in a casino. Then Gartner estimated that over $3.5 trillion of AI infrastructure spending could face write-downs by 2030. An IPO priced into that tape gets a very different reception than one priced in June.

The number that caught me off guard

Here’s what I kept returning to: $750 billion.

When I first drafted this post, press reports put Anthropic’s infrastructure commitments at about $518 billion. The Journal’s read of the draft prospectus is bigger. The company is on the hook for around $450 billion of cloud computing agreements over the next seven to ten years, the bulk of them non-cancelable.

All in, future infrastructure costs top $600 billion. Deals still under negotiation could add as much as $150 billion more, which is where the $750 billion comes from.
“Regardless of actual usage” is the phrase to underline. It means these dollars are owed whether or not the demand shows up.

And even that isn’t enough. Anthropic says it has sought more capacity than “those partners were willing to provide,” and that they increasingly see it as a competitor. So it’s leasing its own data centers and working with third-party operators, some of which, it warns, have “limited operating history” and may be less reliable than the big clouds.

Let me connect that to last month’s lesson. Aschenbrenner’s fund didn’t die because its thesis was wrong. It died because a long-horizon conviction was financed with short-horizon obligations, at 4x. Anthropic is not a margin-levered hedge fund, and I want to be careful not to push the analogy too far. But the same structural question applies in a different form: operating leverage.

• Fixed costs: hundreds of billions in committed compute, largely non-cancelable.
• Variable revenue: priced per token, in a market where, as Benedict Evans argued, every dynamic we can see points toward commodity pricing over time.

When your costs are fixed and your revenue is a price per unit that may fall, volume has to outrun price. Today it does, spectacularly, and the reported $1.70 of revenue per compute dollar says so. The business is a bet that it keeps doing so for ten years. Right thesis, wrong leverage is not only a hedge fund problem. It is any business whose fixed obligations outlast its pricing power.

The counterargument deserves its due. These commitments are also a moat of sorts: access to scarce compute, locked in early, while others queue. If demand for inference keeps exceeding supply, a locked-in contract is an asset, not a liability. Both readings are plausible, and the difference between them is the entire investment case.

The leverage is now in the bond market

When I started drafting, this was an argument about operating leverage. Then the Financial Times reported on Project North Star, a roughly $60 billion financing to pay for about 1.5 gigawatts of Google’s custom TPUs. The structure is worth slowing down for:

Read that sequence carefully. Between June and October, Anthropic’s revenue kept growing and it turned profitable on an adjusted basis. And yet the price of its riskiest debt went up by about 1.5 percentage points. Part of that is the bond market, since Treasury yields climbed sharply over the same stretch. Part of it is how lenders now feel about AI credit. Either way, a junior lender asking for 10% while Treasuries pay 5% is telling you exactly how much uncertainty it sees in those lease payments.

This is not margin debt, and Anthropic is not a hedge fund. But it’s the same family of risk I wrote about in August: long-lived commitments financed by obligations that don’t care how the thesis is doing in any given quarter.

The bottom line

Munger’s favorite advice, borrowed from the mathematician Jacobi, was “Invert, always invert.” Don’t only ask how Anthropic wins. Ask how an investor in it loses. The answers I keep landing on are not “the technology fails.” They are structural: fixed costs against falling prices, concentrated counterparties, a rich multiple in a rising rate world, and junior lenders asking for 10%.

The technology may be everything its believers say. The business may still be a great one. And the stock can still disappoint, because those are three different bets, and only one of them is priced in.

I’ll be reading the public S-1 the day it lands. What would you have to believe? Let me know in the comments below.

Further reading

• Anthropic S-1 Shows Red-Hot Growth, Risks and Massive Computing Costs by The Wall Street Journal
• Investors temper Anthropic IPO valuation expectations over revenue growth slowdown by Reuters
• Anthropic offers Blackstone a 10% coupon on $9 billion of chip debt, summarizing the Financial Times
• Research firm calls Anthropic the “most ridiculous IPO of 2026” by The Next Web

Cheers 🥂