Anthropic’s IPO filing might be the most fascinating business document of the year. It’s a growth story and a governance thriller in one. I just watched this AI professional, Matthew Berman, go through the whole S1, and I couldn’t stop taking notes.
The short version: the numbers are jaw-dropping, and the fine print is even more surprising.
🔄 A flywheel that feeds itself
The expert says Anthropic’s whole business runs on one loop. Better models win more enterprise customers. More customers bring more revenue and more “feedback,” which really means data. That money buys more compute, and the extra compute builds the next model.
Anthropic calls the newest piece of this “closing the loop.” Each Claude version helps build the next one. That’s recursive self-improvement (RSI), meaning AI that keeps improving itself. The creator thinks it’s striking that a safety-first company is pitching RSI to investors. Sam Altman recently told him that pushing hard on RSI is “a very dangerous thing.”
Here are the three parts of the filing that stood out most.
📈 The growth numbers are unreal
- Customers paying $100K+ a year went from 20 at the end of 2023 to almost 12,000 by September 2026.
- More than 1,900 customers spend $1M+ a year, and 250 spend $10M+.
- Net revenue retention is 900%, so every $1 a customer spends turns into $9 the next year. For a typical SaaS company, anything above 100% is a good sign.
- In the three months ending June 30, 2026, revenue grew 1,300% year over year, while compute and infrastructure costs grew 300%.
- Operating margin went from -221% in Q1 2025 to +19%.
- Revenue per employee hit $6.6M, up from $1.4M a year earlier.
The author also highlights Jevons paradox: when something gets cheaper, people use more of it. Claude Opus 5.5 cut cost per task by about 40% compared to Opus 5, and overall revenue still went up.
🤖 Claude is building Anthropic
- Claude wrote over 90% of the contributions to Anthropic’s codebase in September 2026.
- In April 2026, the monthly value of Anthropic’s own Claude usage passed its monthly payroll, not counting stock compensation.
- Anthropic runs its compute as one shared pool across AWS Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. It sends that compute wherever it’s most useful at the moment.
The risk section is just as honest. It says a flaw in today’s models could carry into future versions. It also admits Anthropic may only notice a big jump in progress “after it has occurred.”
🏛️ Seven people hold the keys
This is the part that shocked the creator most.
- One special Class F share gives the seven co-founders 50.1% of total voting power, plus the power to elect the board.
- People who buy Class A shares “generally will not be able to affect the outcome” of votes, no matter how they vote.
- Anthropic is putting 846M shares (5%) into its Public Benefit Fund, and each co-founder has pledged 80% of their equity to charity.
The filing says the founders don’t run day-to-day operations. But they control the board, and the board picks the CEO. The author calls the setup “almost like a monarchy.”
There’s one move he openly praises, though. Anthropic has promised to cover any electricity price increases its data centers cause for nearby consumers, including grid upgrades.
💡 My take
I was blown away by how much Anthropic has built on so little cash burn. It has raised $141.1B and still holds $112.4B in cash plus $19.3B in short-term investments. That works out to roughly $10B burned to build all of this. Still, that voting structure is worth thinking hard about before buying a single share.
Watch the full video for the author’s complete breakdown, including the Sam Altman clip and the full list of risk factors. It’s well worth your time.