It's happening...

About This Video

A 34-minute video from Matthew Berman in which he reads through Anthropic's S-1, the registration document a company files before going public, and comments on what it reveals. He quotes the filing directly, adds his own opinions throughout, and closes with a list of the quotes that surprised him most. Every figure below is as Berman reads it aloud in the video, so check the filing itself before relying on a number.

Summary

Berman calls Anthropic one of the best-built businesses of all time and, in the same breath, one of the most contradictory. The filing shows very fast customer growth, a positive operating margin, and a business that uses its own models to build the next ones. It also shows a share structure in which seven co-founders hold voting control, which he finds hard to reconcile with the company's talk about guarding against concentrated power.

What Berman Pulls From the S-1

1. The mission pitch (00:33). Anthropic says intelligence is now something humans can build and that the effect on the economy will exceed industrialization, electricity, semiconductors, and the internet. Berman notes the same company recently said there is a greater than 10% chance AI kills all of humanity (as he describes it).
2. Two self-reinforcing loops (02:00). The filing describes an improvement loop that advances Claude and an investment loop that funds more research. Berman ties this to a flywheel he described in an earlier video: the best coding model earns enterprise revenue, which brings in coding data and money for compute, which produces the next best model.
3. Core advantages (03:00). Mission and focus (Berman contrasts Anthropic's single focus with competitors' side bets such as video generation and consumer hardware), and talent and culture (he names several well-known researchers who joined).
4. Closing the loop (05:00). Each Claude version helps build the next. Anthropic calls this recursive dynamic "closing the loop" and a structural advantage. Berman plays a clip of Sam Altman saying that pushing hard on recursive self-improvement would be dangerous and that the world will need speed limits on it (05:38).
5. Compute (07:30). Compute is managed as a single pool across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs and directed to wherever the next unit is most productive (as quoted in the video).
6. Benefits and economic impact (08:31). The filing lists faster science, drug discovery, personalized education, and better power grids, and cites an IDC estimate of $22.5 trillion in cumulative global economic impact by 2031 (09:48). Berman contrasts this optimistic tone with Dario Amodei's earlier warning about a white collar bloodbath (10:19).
7. Concentration of power (12:31). The filing warns AI could concentrate power and wealth between countries. Berman points out it mentions countries and not companies.

The Numbers, as Read in the Video

• Customers (14:15). As of September 30, 2026: more than 11,900 customers at $100,000 or more of trailing twelve-month revenue, more than 1,900 at $1 million or more, and 250 at $10 million or more.
• Customer growth (28:00). Customers above $100,000: 20 at the end of 2023, 200 a year later, 1,500 a year after that, and almost 12,000 now.
• Net revenue retention (15:11). Berman reads 900%, meaning a customer who spends $1 spends about $9 the next year. He calls this unheard of for a software business.
• Cheaper models, more revenue (16:04). The filing says Claude Opus 5.5 cut customer cost per task by about 40% compared with Opus 5, and lower cost drives more usage. Berman connects this to Jevons Paradox and the DeepSeek R1 moment.
• Compute efficiency (17:01). Revenue grew 1300% year over year for the three months ending June 30, 2026, while compute and infrastructure expenses grew 300%.
• Revenue per employee (18:16). Over $6.6 million trailing twelve months, against $1.4 million a year earlier.
• Capital (18:58). $141.1 billion raised in total, with $112.4 billion in cash and $19.3 billion in short-term investments as of June 30, 2026. Berman says this means about $10 billion burned.
• Revenue per compute dollar and margin (28:38). Revenue per dollar of compute rose 5.2x from Q1 2025 to Q2 2026, and operating margin moved from negative 221% to positive 19%.

Governance and Corporate Structure

• Public benefit corporation (21:45). Anthropic is a public benefit corporation with a dual mission. Berman recalls Patagonia as the best-known earlier example and notes lawyers usually advise against innovating on corporate structure.
• Philanthropy (22:04). 846 million Class N shares, about 5% of fully diluted shares, go to a public benefit fund, and each co-founder has pledged 80% of their personal equity to philanthropy. Berman questions whether a charity is the most efficient route and compares it with Elon Musk's idea of giving shares directly to people (22:34).
• Class F stock (23:02). A single Class F share, held through an LLC whose members are the seven co-founders, carries 50.1% of the voting power and can elect directors. Berman reads this as absolute control by seven people.
• Long-Term Benefit Trust (25:35). The filing says the trust can elect and remove a majority of the board but does not run day-to-day management or strategy. Berman finds the separation hard to believe, since the board chooses the CEO.
• Public shareholders (33:52). Holders of Class A common stock generally cannot affect the outcome of votes or director elections, however they vote.

Risks the Filing Names

• Success depends on continuing to build the best AI, and on securing and directing enough compute (26:31). Berman adds a Taiwan invasion as an example of a supply risk.
• Dependence on the co-founders and key technical staff (26:50).
• Models may behave in unintended, misaligned, or autonomous ways, including taking harmful actions no user directed (27:00).
• Revenue concentration, regulatory uncertainty, and a novel share structure that may effectively remove the influence of Class A holders (27:33).
• A flaw in a current model could propagate into later models, and Anthropic says its ability to measure the pace of AI-driven acceleration is limited, so it may only notice a speedup after the fact (29:31, 30:01).
• Competitors may outbid it for scarce compute, and agent workloads can use far more compute per task than chat, so costs could rise faster than revenue (30:37).
• Anthropic says it has underestimated its compute needs in the past. Berman recalls the compute deal with SpaceX that followed (31:01).

Quotes That Surprised Berman Most

• Code written by Claude (32:35). Over 90% of contributions to the code base in September 2026 were written by Claude.
• Internal Claude spend (32:53). In April 2026, monthly internal Claude usage value first exceeded monthly personnel expense, excluding stock-based compensation.
• Safety risk of AI building AI (33:03). Anthropic admits that using AI to build more capable AI could pose a considerable safety risk.
• Delayed pre-training (33:15). Anthropic chose to delay pre-training of a new frontier base model for more than a month to strengthen security.
• Electricity costs (32:01). Anthropic has publicly committed to covering consumer electricity price increases caused by its data centers, which Berman says is what AI companies should do.

Sponsor

A segment (06:16) promotes a team and agent workspace product described as a company "harness" for human teammates, agents, and tools in one place. The name is garbled in the transcript.

Berman's Opinions vs. the Filing

Treat the filing quotes and Berman's commentary as separate layers. His strong opinions include: the recursive self-improvement language contradicts the safety message (05:00, 20:30), the governance explanation is not credible (25:35), and the structure makes OpenAI look "very sane" by comparison (34:10). These are his views, not findings from the document.

Key Takeaways

1. By the filing's own numbers, Anthropic's growth, retention, and compute efficiency are unusual for any software business.
2. The company describes Claude helping to build the next Claude as a core advantage, and also lists it as a risk.
3. Falling prices have raised overall usage, which supports the Jevons Paradox argument.
4. Voting control sits with seven co-founders through Class F stock, while public Class A holders have little say.
5. Lower cost per task and more agent usage pull in different directions on compute cost.
6. A filing is a persuasion document written for investors, so read both the pitch and the risk factors.

Reflect & Discuss

• Is it consistent to warn about dangerous AI while describing recursive self-improvement as an advantage? What would make it consistent?
• Does a board-level trust give real independence if the same people control the voting stock?
• Why would a lower price per task increase total revenue, and where does that logic stop?
• How would you check a claim like 900% net revenue retention against the filing itself?