Anthropic Opens Its Books: $17 Billion in a Quarter, a $23 Billion Loss, and $895 Billion in Compute Promises
*Disclosure: Anthropic, which makes Claude, appears in several sections below, including its own IPO filing.*
On October 29, this blog flagged a mystery: a political resolution quoted an eye-watering loss figure for Anthropic but didn't explain it. Yesterday's amended filing offers an answer, and it landed on a day when Wall Street's appetite for AI debt was being tested everywhere at once.
What Anthropic's Amended Filing Says
According to AI Weekly's summary of CNBC's reporting on Anthropic's amended S-1:
- Revenue: About $17 billion in the third quarter (up roughly 1,127% from a year earlier) and $33.24 billion over nine months.
- Loss: A third-quarter net loss of $23.29 billion, which the filing attributes mainly to fair-value changes on convertible notes and derivatives, not day-to-day operating costs. If that holds up, it helps explain the large loss figures we couldn't decode last week. I haven't read the filing myself.
- Commitments: About $895.4 billion in purchase and lease obligations, “substantially all” tied to compute and infrastructure. The New York Times links it to more than 3.5 gigawatts spread across Amazon, Google, SpaceX (Nvidia), Microsoft and AMD.
The Wall Street Journal reports institutional investors are pushing for an IPO valuation near $1.5 to $1.6 trillion, down from the roughly $2 trillion figure previously floated, ahead of a planned November Nasdaq listing. Separately, Business Insider says Anthropic's internal Claude usage, priced at retail API rates, now exceeds what it spends on employee pay; commenters note that's an accounting quirk rather than real cash outlay.
Software's Reckoning
Three announcements in one day point the same way:
- Squarespace is cutting 19% of staff, about 300 people, describing it as simplifying operations. TechCrunch ties the move to AI “vibe-coding” tools reshaping web building.
- SUSE will cut roughly 10% of its 1,900 employees, with notifications due by November 4. The CEO's memo says automation and AI are fundamentally shifting how software companies work and that demand for some roles has greatly diminished.
- Goldman Sachs has limited Claude Code and GitHub Copilot to its roughly 12,000 engineers, pulling them from bankers and traders who had used them for months, according to Bloomberg. The bank says the tools belong with those “best suited” to use them.
It extends the jobs thread from last week's buyouts and rehiring stories: some firms cut, some re-hire, and some, like Goldman, narrow who may use the tools.
Wall Street Meets the AI Bill
- Meta is sounding out investors on a cross-border bond of at least $20 billion, after negative third-quarter free cash flow and capex guidance of $135 to $145 billion (Bloomberg).
- Oracle filed to sell $10 billion of stock and $10 billion of notes; shares fell more than 6%. Debt stood at $125.34 billion in late August, and credit-default swaps were reportedly pricing junk-level risk.
- Alphabet opened a €6 billion bond sale; orders topped €17.1 billion, days after a reported record one-day market-value drop of about $500 billion.
- Marvell agreed to buy Lumentum in an all-stock deal valued at $115 billion, billed as the largest merger of two public tech companies ever, targeting a 2028 close.
Demand for the bonds is strong. The question is how long it stays that way.
Research Worth Reading
MIT CSAIL researchers report that GPT-6 Astra can read and edit a model's internal weights from a plain-English description. On a tiny 0.8-billion-parameter Qwen model, it traced a looping failure to directions in two layers, removed them by changing 0.016% of parameters, and lifted MATH-500 from 7.9% to 48.1%. These are the authors' numbers on a very small model, and it's not peer reviewed. Separately, a guest essay on Terence Tao's blog urges “egoless mathematics” as AI cracks more proofs, valuing truth over credit.
Odds and Ends
- Meituan's LongCat-2.5 is a 1.6-trillion-parameter model with about 48 billion active, so only about 3% works per token, the Mixture-of-Experts trick we explained on October 30. It was trained on domestically produced Chinese accelerators.
- Russia ordered private data centers to hold 10% of capacity for state use, accessible within 24 hours, following October drone strikes on Yandex and Sberbank data centers, per the Moscow Times.
- Microsoft patent: An application describes live video AI that edits speakers' words and mouth movements. A patent isn't a product, but it lands as new deepfake-safeguard standards start taking effect.
Watch List
- Nov 4: SUSE notifications; Apple/Qwen developer preview.
- This month: Anthropic's IPO pricing; Meta's and Oracle's offerings.
Our Take (Opinion)
Last week's question was whether AI is a bubble. This week's answer is “not yet, but the bills are due.” Revenue is real, commitments are enormous, and investors are now asking for proof that one covers the other.