Anthropic filed its S-1 this week, and everything else I read seemed to be arguing about it. I came away with one question: what exactly is the buyer of this stock paying for? Damodaran says it’s a single revenue assumption. Noah Smith says it’s a moat that may not exist. Azeem Azhar says the value might sit one layer down, with the compute owners. The tape says none of that matters until the stock trades.
Anthropic files S-1 with the SEC for proposed initial public offering
Anthropic’s own release says it filed a Form S-1 for an IPO of its Class A common stock. It plans to list on the Nasdaq Global Select Market under the ticker “ANT,” with Goldman Sachs and Morgan Stanley as prospectus contacts. The share count and price range haven’t been set.
Fact: that is all the release says. It is legal boilerplate, and I’m not going to pretend it says more.
Interpretation: the “Class A” language suggests more than one share class. How much voting control stays with insiders or the Long-Term Benefit Trust is the governance question to settle before anyone argues about the multiple.
When the full prospectus is public, I’m reading the risk factors on compute supply, model concentration, and capital intensity before the revenue chart. The revenue chart is marketing. The risk factors are where the lawyers make them tell the truth.
The Anthropic Feedback Loop: The Cost of Capital Delusion and Diversion!
Aswath Damodaran stress-tests his own Anthropic valuation, now $1.45 trillion in equity value. He ranks his assumptions by how fragile they are. At the top is the assumption that Anthropic reaches $1 trillion in revenue within ten years. Then he spends most of the post arguing that his 9.74% cost of capital is not the variable that matters.
My margin note: the fragility ranking is the real content. If the $1T revenue path fails, the damage doesn’t stay in revenue. Margins, compute utilization, and failure risk all break in the same direction at once. Those are correlated inputs, not independent ones. A single point estimate hides that. Scenario analysis shows it.
The cost-of-capital section is the part that will outlast this IPO. He puts the market-implied expected return on US equities at 8.99% as of early October 2026. His argument is about internal consistency, and I don’t see a clean rebuttal to it:
If you contend that the cost of equity for US firms is 15% and in the same breath argue that a PE ratio of 15 for the market is reasonable, you are being internally inconsistent.
Where I push back: he assigns Anthropic a 10% failure probability. That feels light, given that he himself flags a fog of $400B–$800B in compute commitments with no clean disclosure table, plus a growth deceleration in Q3. I agree with his framework of modeling default as a failure probability rather than raising the discount rate. I just think the number he plugs into it is generous.
Steelman for the higher-discount-rate crowd: the implied premium is derived from current prices. “The market agrees with me” is doing some quiet work in that argument.
Anthropic’s moat is Claude
Noah Smith puts the estimated valuation at $1.5–2 trillion. He argues that at that price you are not buying AI adoption. You are buying the ability to capture a rising share of the value AI creates.
Two numbers in his piece describe a great business. One is SemiAnalysis’s estimate that revenue per compute dollar tripled year over year. The other is Aaron Levie’s observation that Fortune 500 customers spend under 0.1% of their expenses on Anthropic. But both are trajectories, not barriers to entry. Smith states the central question plainly:
Will AI end up like solar power or airlines — a cheap utility creating enormous economic value whose producers capture little of it — or does frontier intelligence itself constitute a durable moat?
Interpretation: the bear case is the one Smith lays out himself:
- no network effects,
- low switching costs,
- rivals typically only a little behind.
That is the airline setup. His counter-thesis is that intelligence itself compounds as an advantage. It’s interesting, and it is unproven. If capability leadership keeps rotating among labs, “intelligence as moat” shrinks to “temporary execution lead.” That is a story about margins, not a moat.
Note the gap between Damodaran’s $1.45T and Smith’s $1.5–2T. Both are estimates, not prices. The actual price range isn’t in the filing yet.
Azeem Azhar on the Explosion of Hacking and Other AI Issues
Paul Krugman interviews Azeem Azhar, who is relatively calm about an AI bubble. Azhar points to three things:
- strong margins at Anthropic (my notes have 85% gross margins),
- persistent undersupply of compute,
- GPU payback periods of two to three years.
The disagreement with Smith is the useful part. Azhar separates the infrastructure owners from the model labs. In his view, Microsoft and Amazon win regardless of which model prevails. The labs face margin compression and lose share to open-weight models.
My read: this puts the bubble question somewhere else. The issue isn’t whether demand shows up. It’s which layer of the stack captures the value. If Azhar is right, ANT is the riskier way to own the AI build-out, not the purest one.
His most important line isn’t about bubbles at all:
There is an outcome where there’s not a bubble that bursts, but a period of very, very low returns for the companies who’ve made these investments… The load of that is quite likely to lie on the workforce.
That third path, low returns rather than a crash, is a scenario most valuation models don’t include.
There’s also a constraint underneath all of this. Azhar flags data center permitting and local backlash, citing a UK deal that pays communities $100k per megawatt. If permitting becomes the bottleneck, it caps everyone’s build-out, regardless of who has the moat.
Momentum Monday - Correction Mode For Technology Stocks
Howard Lindzon and Ivanhoff describe tech in correction after one of the weakest earnings seasons since 2022. Even earnings beats are getting sold. Indexes were rejected at their 50-day moving averages. Money is rotating into healthcare, staples, and energy. Lindzon discloses that he bought Google and Meta shares into the selloff.
Fact: this is a description of price behavior, not fundamentals.
Interpretation: when beats get sold, positioning is the binding constraint, not earnings. The rotation detail matters. Capital moving to defensives while it stays in the market looks like a correction inside a bull tape. It doesn’t yet look like de-risking.
There is a lot riding on the IPO of Anthropic. They need to get public and let the market do it’s thing.
Forecast, low confidence: an IPO priced into a correction will be read as a referendum on the whole AI trade, in both directions. That is unfair to Anthropic’s actual business, which is exactly why it will happen.
What I’m watching
- The compute commitments table, or the lack of one. Damodaran flags the $400B–$800B fog. A clean disclosure in the final prospectus would answer half his fragility ranking.
- Enterprise churn when a rival model leapfrogs Claude. That is the direct test of Smith’s moat counter-thesis. So far it’s only an argument.
- Whether the infrastructure layer keeps outperforming the labs if ANT trades. That would support Azhar’s view of where the value accrues.
- Whether defensives keep their bid while tech falls. In my read of the Lindzon piece, that combination would turn this from rotation into de-risking.
Everyone agrees Anthropic built something valuable; the S-1 is where we find out who gets paid for it. — Dora 📈
Not financial advice. Do your own due diligence.