Jamin Ball expects Anthropic to begin public trading between November 15 and 20.
Anthropic's S-1 shows 85% gross margins, 20% free cash flow margins, and 8% GAAP operating margin on $16.8bn Q3 revenue.
Ball debunks the $108bn convertible notes as debt and the $81bn net loss as mostly convert appreciation.
Anthropic's S-1 shows a near-profitable lab facing a $5tn build-out
Anthropic's S-1 offers the first clear look at a frontier lab's economics, and the numbers are stronger than the headlines suggest. Gross margins run at 85%, only about $7bn of the $141bn raised has been burned, and the alarming $108bn of converts and $81bn loss are accounting artefacts. What comes next is the open question, with roughly $100bn a year of compute commitments and a build-out that would run near 5% of GDP.
Takeaways
Anthropic is almost accidentally profitable. Ball points to 85% gross margins, free cash flow margins above 20%, an 8% GAAP operating margin and $16.8bn of quarterly revenue. Anthropic raised $141bn and burned about $7bn of it.
The scary numbers are accounting artefacts. About $11bn of converts were raised, $5bn have already converted, $6bn remain outstanding and that remainder is now marked at $108bn. That $11bn sits inside the $141bn raised, and $72bn of the $81bn first-half loss is converts appreciating.
Inference efficiency is holding margins up. Compute spent serving models grew 2.3x while revenue grew 14x, roughly a 7x gain in serving efficiency, and only 27% of compute goes to serving. Ball expects margins to fall again as Anthropic reinvests.
Compute commitments run about $100bn a year. $855bn of committed compute deals annualise to about $100bn a year. Ball's maths, 30 gigawatts each for OpenAI and Anthropic at $50bn a gigawatt, comes to about $5tn over five years, near 5% of GDP.
Enterprise token spend has barely started. Half the Fortune 500 uses Anthropic but spends 0.1% of expenses on it, and 80% of revenue comes from models six months old or younger. Stripe's John Collison expects token spend to reach 15 to 20% of expenses.
There is no bid for GPU debt. Data centre debt deals are clearing above 10%, and Ball says credit providers already carry heavy Anthropic and OpenAI exposure through neo clouds such as CoreWeave and Nebius. He sees no dark GPUs and no glut.
Anthropic's path to listing
Anthropic files its S-1 roughly 15 days before the roadshow, with trading expected between 15 and 20 November.
The roadshow runs about a week and a half of pitching to IPO investors.
Labor Day to Thanksgiving is a hot listing window, and Ball says investors would make time for Anthropic.
Bigger companies can raise privately, so they float less of themselves and take less dilution.
The build-out argument
Ball treats the difficulty of building as a natural governor, so a fibre-style overbuild is unlikely and there are no dark GPUs.
swyx says timing the market is impossible and that investors fixate on one of 2,000 data points in an S-1.
They disagree on how much inference efficiency is left and whether frontier models keep their lead over open weights.
Ball expects compute for safety and verification work to be an underrated driver of demand.
Tap a name to go there · tap a line to see why they are linked