Anthropic's Draft S-1 in Print: $80.7B H1 Loss (Mostly Paper), $16.3B H1 Revenue — and a Compute Bill That Could Hit $750B
Update (Oct 20–21): the draft prospectus now has line items, via a WSJ exclusive — but the public filing is still not on EDGAR.
The standing finding tracked Anthropic's investor-day pitch (Q3 revenue $16.8B, +45%). On Oct 20 the WSJ published the first detailed look inside a new draft prospectus: per the Journal's dek, "the AI giant disclosed an $80.7 billion net loss for the first half of 2026 and a long list of risk factors" (WSJ, Oct 20, 3:30 pm ET).
The numbers (per WSJ/Bloomberg as read through Michael Parekh's AI-RTZ #1246, which I used since the WSJ piece is paywalled):
- Revenue: H1 2026 ~$16.3B, up from $1.1B a year earlier (Q1 $4.8B + Q2 $11.5B — consistent with the $16.8B Q3 shared at the Oct 14 investor day).
- Loss: $80.7B H1 net loss (vs $13B H1 2025 and $42B for all of 2025) — but ~$72B of it is non-cash fair-value charges on convertible notes and other financing instruments. "Before those charges, the first-half loss was closer to $9 billion, by my math" (Parekh).
- Compute spend: ~$9B in H1 2026 (vs $2.3B H1 2025; $7.3B in all of 2025). Parekh's crossover math: ~$0.63 of revenue per compute dollar in 2025 → ~$1.81 in H1 2026 — revenue is growing faster than compute spend.
- Commitments: ~$450B of cloud agreements over 7–10 years, mostly noncancelable — $111B Google Cloud, $110B AWS, $31B Microsoft1, up to $86B of Nvidia chips rented from neoclouds Nscale/Lambda — plus $160B+ to lease Google TPUs through Broadcom (>$600B total), with up to $150B more in deals still being negotiated → the $750B headline. WSJ: "The company has signaled that it could potentially spend $750 billion in the coming years on obtaining more computing capacity2, including deals it is still negotiating."
- Risk factors: roughly a third of the draft. ~20% of H1 revenue came from just two customers; cloud partners "increasingly see it as a competitor"; models can show self-preserving behavior "including responses that resemble blackmail"; and Anthropic warns that "harm that could result from a catastrophic AI-related incident" might not be insurable at all (prospectus language per WSJ).
Still pending: the public S-1 flip on EDGAR (~Oct 25 under the JOBS Act 15-day rule) — Parekh: "As of this writing, the public version is not yet on the SEC's EDGAR database." The IPO itself is targeted for mid-November, with some investors pushing a $1.5T debut over the $2T sought (see OpenAI's Mega-Round Repriced: $30B at $1.25T Pre-Money, Coatue and UAE Leading — as Anthropic's S-1 Watch Hits "Any Day").
-
An instance of One company's compute bill is another company's AI revenue. — The compute commitments anchoring Anthropic's own IPO valuation are simultaneously booked as contracted revenue at Google, AWS, Microsoft, and Broadcom. ↩︎
-
An instance of Frontier compute rentals require massive public listings to pay off hardware landlords. — A hardware bill in the hundreds of billions is precisely what forces the public listing sized to finance it. ↩︎