Brian Reed: Hey. You doing okay? Because I've been staring at this thing for two hours and I'm still not sure I've processed it.
Brian Reed: Anthropic filed with the SEC. October 26th. Class A shares, Nasdaq, ticker ANT. That part's settled. That actually happened.
Eliza Ward: Sixteen point three billion dollars in revenue. First half of this year alone.
Brian Reed: Right — and then you turn the page.
Eliza Ward: Eighty point seven billion GAAP net loss.
Brian Reed: For the same six months. I mean — I had to reread it. Same period. $16 billion in, $80 billion loss. Goldman Sachs and Morgan Stanley are the lead underwriters on this. They've looked at these numbers. And I still don't know how to hold both of those figures at once.
Eliza Ward: Wait — and no share count. No offering price. No listing date. They filed and gave us almost none of the IPO mechanics.
Brian Reed: Right — but the part that doesn't fit is, they give you a number to make it feel better. One point six billion in adjusted operating income. Non-GAAP. And the adjustment is — okay, two point two billion in stock comp, a hundred and twenty-six million in charitable contributions. That's your two point three billion of explained gap. And then the GAAP loss is eighty point seven billion. So where does the other eighty billion go?
Eliza Ward: That's the thing the filing doesn't reconstruct cleanly.
Brian Reed: Think about it like a restaurant. The chef tells you they made ten dollars profit on every plate they served. That's true. But they haven't counted the lease, the equipment loans, or the equity they handed the chef on day one. That's the whole story in one sentence — the adjusted number is real, and it's also incomplete in a way that matters enormously.
Eliza Ward: And Denny Fish at Janus Henderson said exactly that — key details are still missing. That's not retail skepticism, that's an institutional portfolio manager saying the filing doesn't close the loop.
Brian Reed: Is that the market-moving read though? Because the number circulating — unofficially — is a one point eight to two trillion dollar valuation. With no share count filed. No price. Nothing the SEC has actually confirmed.
Eliza Ward: No, that's — wait, that's the hype layer. Those valuation estimates have no anchor. Anthropic hasn't filed a share count or a price range, so any number floating around right now is someone's back-of-envelope math on sixteen point three billion in revenue and a growth rate that's — I mean, thirteen to fourteen hundred percent year-over-year is real, but it doesn't tell you what the equity's worth if you can't explain the eighty billion dollar hole.
Brian Reed: No, I don't buy that the gap is just noise.
Eliza Ward: It's not noise — that's actually the argument about whether frontier AI economics work at all. Like, the adjusted number says almost breakeven. The GAAP number says the structure underneath is — well, we don't fully know yet, and that's the honest answer. The filing exists to answer that question and it hasn't, cleanly.
Brian Reed: So the thing to watch is whether Goldman and Morgan Stanley's roadshow fills in what the S-1 doesn't. Because right now the retail enthusiasm and Denny Fish's caution are reading the same document and landing in completely different places.
Eliza Ward: That retail-versus-institutional split — okay, but the number that's actually doing the most work in the bull case right now is the 900% net revenue retention. That's the one.
Brian Reed: Right — and that's the take I want to push back on. Because analysts on X, Ben Bajarin included, are citing that number as proof Claude customers are locked in. Enterprise stickiness. The software story.
Eliza Ward: But the S-1 says all major contracts are terminable.
Brian Reed: Terminable. And — hang on — roughly 80% of revenue is consumption-based token fees. So if a customer just... reduces usage on a Tuesday, revenue drops immediately. There's no contract holding them.
Eliza Ward: And then the concentration piece. Twenty percent of first-half revenue from two customers. Two. The 900% retention number tells you nothing about what happens if one of those two sends a termination notice.
Brian Reed: Picture a pension fund's retirement-savings manager. It's a Tuesday morning, the contract renewal on one of those two accounts doesn't come through. The 900% NRR — that's a trailing metric from September 30th. It literally cannot tell you what next quarter looks like.
Eliza Ward: And — wait, there's actually a harder version of this. The S-1 disclosed that 5 to 7% of revenue in the first nine months of 2026 came from customers who were later found to be violating usage policy. Enforcement stepped up August-September. Those contracts terminated. That's retention that went to zero, and the 900% figure doesn't capture what that cost forward.
Brian Reed: So the 900% could be reflecting momentum off better Claude models, not actual lock-in. If the next release doesn't move the needle the same way — does that number hold? We genuinely don't know.
Eliza Ward: We don't. And the part that comes later is going to make this capital structure question a lot sharper — because there's $637 billion in compute commitments sitting against those terminable contracts, and I haven't heard anyone explain how those two things fit together.
Brian Reed: And that $637 billion number — I mean, where does that even come from? Pages 171 to 172 of the S-1. Someone aggregated the named contracts. Google, Amazon, Microsoft, SpaceXAI, Broadcom, Nscale, Lambda. All of them. Add it up and you get $637 billion in compute commitments.
Eliza Ward: And Goldman and Morgan Stanley have not independently confirmed whether those are binding obligations or capacity reservations with escape clauses. That's — that's the unanswered question right there.
Brian Reed: Because those are completely different things.
Eliza Ward: Completely. If it's capacity Amazon will hold for Anthropic if Anthropic wants it — that's one capital story. If it's $637 billion of guaranteed infrastructure payments — that's, wait, that's a structurally different company. And customer contracts on the other side are terminable. Month-to-month consumption fees, mostly. So the mismatch is — commitments potentially running six, seven years, customers who can walk on a Tuesday.
Brian Reed: Let me make that concrete. Say it's a CFO at a large financial services firm, mid-morning, reviewing Claude API costs against their quarterly budget. Usage-based billing. They trim it by 40%. Anthropic's Q2 revenue per dollar of compute was $2.08 — that margin is thin enough that a 40% usage cut from a top-tier customer doesn't just hurt, it potentially flips the infrastructure math.
Eliza Ward: Against $5.6 billion in compute spend in that one quarter alone. Yeah. SemiAnalysis and Jamin Ball at Altimeter flagged improving inference revenue-per-megawatt efficiency as bullish — but the forward trajectory isn't quantified. So we don't actually know if that $2.08 ratio improves or compresses as they scale into the commitments.
Brian Reed: And then the existential risk disclosures. Ninety-plus pages of risk factors. Including — actual catastrophic and existential AI risk language. In a prospectus. Has that ever happened?
Eliza Ward: No. That's literally unprecedented for an S-1. The SEC is receiving a document where the issuer explicitly warns that its own product could pose civilizational-scale harm. And you can read that two ways — either it's radical transparency, which is what the public benefit pledge announced October 18th is trying to signal structurally, or it's an admission that no market pricing mechanism actually knows how to value this risk.
Brian Reed: And that's the live tension that the roadshow doesn't resolve. The compute commitment structure and the existential risk language sit in the same document, and the thing to watch is whether the SEC asks Anthropic to clarify which of those $637 billion line items are binding before ANT gets a listing date.
Eliza Ward: That's the question the SEC review is going to sit with. No listing date disclosed. The process starts now and we don't know how long it runs.
Brian Reed: And — I mean, ANT exists as a ticker. The Nasdaq Global Select Market has a slot reserved. The price does not exist yet. And that price, whenever it comes, is going to be the first moment markets are actually asked to put a number on — I mean, how do you price a company whose own prospectus warns the product might cause civilizational harm? What is that risk worth? There's no comparable. Nothing to anchor to.
Eliza Ward: No. And I don't think there's an answer yet. The offering price is the only number that settles anything — and it doesn't exist.
Brian Reed: The real test is when a major customer terminates, or the $2.08 compute ratio moves, or Claude plateaus. Then investors find out which financial story they actually bought. The GAAP one or the adjusted one.
Eliza Ward: And we genuinely don't know. That's the honest state of it.
