Anthropic's S-1, Day Two: 85% Gross Margins, Inference Costs Falling 5x, and a $63 Billion Broadcom Rescue

Yesterday’s S-1 coverage centered on the headline shock, $16.3 billion of revenue against an $80.7 billion net loss. The second-day analyses, published October 27 by The Information, the Financial Times, Bloomberg, and WSJ, dug into the filing’s unit economics, and what they found is the part that will decide whether the November pricing holds: an 85% gross margin on model serving, inference cost per revenue dollar falling from roughly $2.50 in Q1 2025 to under 50 cents by Q2 2026, 900% net revenue retention, and a financing story behind the compute bill in which Broadcom’s CEO personally rescued a $63 billion debt package in eleventh-hour talks with the banks.

The margin structure, with its asterisks

The Information’s 85% gross margin on running models is, per the filing’s own definitions, a serving-cost number: it excludes model-training costs and the revenue shares paid to distribution partners, and those excluded lines have been growing fast, with partner and distribution costs reportedly climbing from 11% of revenue in 2023 to 32% in 2024 and near 50% by 2025. The FT’s contribution is the trend that matters more than any single margin: compute and infrastructure cost per revenue dollar ran about $2.50 in Q1 2025 and fell below $0.50 by Q2 2026, a five-fold efficiency gain in five quarters, which matches what serving-efficiency work across the industry has claimed and gives it audited numbers. Bloomberg adds the demand-side proof: 900% net revenue retention, with TechCrunch counting 11,900-plus customers spending $100,000 or more annually and around 250 spending $10 million or more. A skeptical read and a bull read both fit these numbers; what changed today is that they are now filed with the SEC rather than leaked.

The $63 billion rescue that almost wasn’t

WSJ’s reporting on the financing is the second-day story with the most drama: Broadcom CEO Hock Tan held eleventh-hour talks and agreed to higher rates plus loan guarantees to secure bank participation in a $63 billion debt package backing Anthropic’s compute leases, with Bloomberg reporting the structure as a special purpose vehicle arranged with Bank of America, Blackstone, Citi, and Morgan Stanley, described as the largest chip-financing transaction to date. The disclosure also quantifies the physical estate this debt sits under: Culpium’s analysis counts just $482 million of first-half capex against roughly $782 billion in lease obligations, and Bloomberg totals $417 billion in future AI spending, numbers that reframe yesterday’s coverage: Anthropic is not a capex company in the Microsoft mold, it is a lease-and-service company whose capital intensity lives off-balance-sheet in Broadcom-, Google-, and NVIDIA-arranged structures, with NVIDIA GPUs at 4.7 GW of Anthropic’s compute against 4.5 GW of Google TPUs, per the filing’s breakdown as reported by Wccftech.

What to watch

Three things follow from the second-day numbers. First, the margin-quality debate: 85% serving gross margin with rising distribution costs is a different business than an 85% all-in margin, and the sell-side models built this week will fight over where training amortization lands. Second, the Broadcom precedent: a chip vendor arranging customer financing at this scale, with the CEO’s personal intervention reported as decisive, ties silicon vendors’ revenue to their customers’ IPO outcomes in a way that has no clean precedent; Broadcom’s own stock now has Anthropic beta. Third, the efficiency slope: if inference cost per revenue dollar keeps falling at anything like the last five quarters’ rate, the path to yesterday’s price war sustaining itself at positive margin becomes plausible, and the bear case built on unsustainable unit economics has to find a different foundation. The IPO is reportedly targeted for the week of November 16; the roadshow that prices it now has its two competing narratives in writing.