Zvi's deep read of Anthropic's S-1: profitable on cash flow, ~$800B compute commitments

Don't Worry About the Vase (Zvi) · rss · 2026-10-28

Zvi Mowshowitz gives the S-1 a full RTFB treatment with ten takeaways:

Financials

Anthropic is effectively profitable: H1 operating cash flow was +$3.9B; the headline $80B "loss" is mostly a $72B accounting revaluation of convertible notes as the stock rose — notes already de facto converted.
Raised $141B with $131B on hand; bonds rated BB+, one notch below investment grade.
Q3 revenue of $16.8B (only 45% quarter-over-quarter) marked a slowdown; Zvi expects a Q4 rebound with Opus 5.5. 5,300 employees support $35B TTM revenue — over $6.6M per employee.
Compute commitments total roughly $800B: $111B to Google Cloud, $110B to AWS, $31B to Microsoft over 7–10 years, plus up to $86B renting Nvidia chips from Nscale and Lambda through 2033/2034 — largely noncancelable.

Governance & safety

Founders control the company indefinitely via Class F shares and can overrule the Long Term Benefit Trust with employees; the LTBT looks like a governance body, not an AI safety body.
Existential risk appears only in passive voice; alignment is treated as prosaic, downplaying willingness to pay real safety costs.
The core business plan is "closing the loop" — recursive self-improvement — with claims it will lose to rivals without it; "pacing the frontier" needs industry-wide adoption, not just embedded evaluators.

More from AGI Musings