Anthropic's IPO Filing: A Mix of Risk Warnings, Business Growth, and Concentrated Control

Anthropic has filed to go public, and the host opens by calling it one of the most controversial companies in history. He says the company that stated there is a greater than 10% chance AI kills all of humanity is also one of the best-built businesses of all time, and that its S-1 reveals many details before the offering.

The host reads the filing's view that AI could transform every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet. The filing also says AI that delivers on its potential with risks well managed could significantly accelerate global growth and dramatically improve quality of life. The host says it is hard for him to believe the company is going public given the risks it itself points out.

The filing states that rapid and growing adoption generates two self-reinforcing dynamics: an improvement loop that advances Claude and an investment loop that helps fund continued frontier AI research. The host says he had described this flywheel in an earlier video.

As he summarizes it, the company built the best coding model, sold it to enterprises for coding use cases, earned large revenue, gained large amounts of new coding data, and then used the money to buy more compute and build its next best coding model. He calls the cycle an insane one.

The host says the filing lists mission and focus as the first advantage. He contrasts OpenAI, which he says spread its bets into areas such as Sora video generation and consumer hardware, with Anthropic, which he says focused on one question: how to make the best models and sell that intelligence to enterprises. He says this was its entire business model and that it allowed rapid revenue growth.

The filing's second advantage is talent and culture, with the statement that its mission and focus let it attract and retain world-class talent. The host lists people he says joined, including Andrej Karpathy, an OpenAI co-founder; John Jumper, a Nobel Prize winner in chemistry; a former OpenAI super alignment co-lead; Galani Nelson, whom he describes as one of the greatest computer scientists of all time; and John Schulman, an OpenAI co-founder.

The host says the filing is essentially a prospectus for potential investors and that it returns to the flywheel, specifically to recursive self-improvement (RSI). The filing states that each version of a Claude model assists in building the next, helping compress development timelines and produce compounding improvements in the speed of innovation. It calls this recursive dynamic 'closing the loop' and says it believes this is one of the most significant structural advantages in frontier AI.

The host then plays a clip from his interview with Sam Altman of OpenAI. In the clip, Altman says that if people really pushed on RSI, the claim would be true, and that it is a very dangerous thing to push on, with the world needing some speed limits on how quickly RSI goes. The host argues the contradiction: Anthropic, which he describes as a safety-focused company that says open-source AI is too dangerous to release, is pressing the 'close the loop' button. He calls this cognitive dissonance, while adding that Anthropic and OpenAI both have strong models.

The host says the filing highlights the potential long-term benefits of AI, which he agrees with. He notes that talking about future benefits has not done much to improve public perception of AI, which he says remains negative. He reads the filing's list of possibilities, including compressing decades of scientific progress into years, accelerating drug discovery and disease detection, delivering personalized therapy, providing individually tailored education, offering expert financial and legal advice, advancing power grids and battery design, and extending to the developing world capabilities that took wealthy nations generations to build.

The filing also cites the International Data Corporation, which estimates that investment in AI solutions and services will yield a cumulative global economic impact of $22.5 trillion by 2031. The filing says Anthropic believes the impact could be even greater.

The host notes that Dario Amodei, less than a year and a half earlier, said there would be a white-collar bloodbath and that 50% of white-collar jobs would be automated away. He says the filing now predicts a great positive economic impact. He adds that after years of posturing about safety, caution, and doomerism, the narrative seems to have changed considerably as the company prepares to IPO.

The filing also says frontier AI's capabilities may extend beyond current markets to open new ones, reaching domains such as robotics, advanced manufacturing, logistics, and scientific research including drug discovery and materials science. The host acknowledges that some jobs will be automated and people will need new skills, while saying he expects new jobs no one could have imagined.

The filing says AI will extend beyond digital limits, with AI agents operating lab and manufacturing equipment such as microscopes, liquid handlers, and off-the-shelf robotic arms. The host says Anthropic has already set up a real wet lab controlled by AI.

The host says this makes him nervous: the company has built a real physical lab but, in his reading, has no clear idea how it will align its AI. He clarifies that he does not believe AI cannot be aligned in the long run, and says he is repeating what Anthropic employees themselves have said.

The host says the filing points out a risk he is most concerned about: concentration of power. He reduces AI to a basic equation of compute plus energy, arguing that whoever controls both can have the biggest and best models, which he calls a hyper-concentrator of power. The filing says AI deployed at scale could concentrate power and wealth within and between countries in ways that could erode trust.

The host finds it notable that the filing mentions countries but not companies, and not the few people inside companies who make the decisions. He says Anthropic has one of the wildest corporate governance strategies he has seen, which comes down to a few people making decisions for the future of humanity.

The filing says that misaligned AI systems could corrupt institutional decision-making or cause large-scale harm through autonomous actions that are not visible or understandable to their users and operators. It says the company should invest in safety and security.

The host points out that a few weeks earlier, one of the company's own security researchers said, 'We have no strong plan for how we're going to align AI in the future.'

The filing then turns to enterprise adoption. As of September 30, 2026, more than 11,900 customers spent $100,000 or more in trailing 12-month revenue, more than 1,900 spent $1 million per year, and 250 spent $10 million or more. The host says these numbers are especially striking compared with where the company was three years earlier. He also cites a net revenue retention rate of 900%, which means a customer spending $1 becomes spending $9 the next year, and says a typical SaaS business would be happy with any positive figure.

The filing says reducing the price of its models has increased revenue overall. The host explains Jevons paradox: the lower the price or the more efficient a unit of economic value like compute or intelligence is, the more overall it gets used. The filing states that Claude Opus 5.5 reduced customer costs per task by approximately 40% compared with Opus 5, and that lower costs per task drive greater usage.

The host links this to the DeepSeek R1 moment, when a Chinese lab released a model trained and served at a fraction of the cost of OpenAI and Anthropic, and he recalls the market reacting sharply. The filing also says that revenue grew 1300% year over year for the three months ended June 30, 2026, while compute and infrastructure expenses grew 300% in the same period. The host summarizes this as tripling the cost while multiplying revenue from that cost by about 13.

The host attributes the company's per-employee efficiency to talent density and to using its own AI to make each employee more productive. The filing reports trailing 12-month revenue per employee of over $6.6 million, compared with $1.4 million as of September 30, 2025. The host compares this with hedge funds, which have small teams and high earnings, while noting Anthropic is a very different business.

The filing states that as of June 30, 2026, the company had raised an aggregate $141.1 billion of capital, with cash and cash equivalents of $112.4 billion and short-term investments of $19.3 billion. The host says that, in his reading, the company has burned very little money overall relative to the business it has built, and he puts the net cost at roughly $10 billion.

The filing describes the flywheel again: better models bring more use cases and sales, which give more feedback. The host notes that the company uses the word 'feedback' but that he reads it as data. That data supports frontier intelligence, which drives higher growth and lets the company invest in future models.

The filing states that Claude is accelerating the company's ability to close the loop, and that Claude itself has become one of the most powerful drivers of its progress. It says each new version of a Claude model directly accelerates the development of the next, and that many aspects of research, training, and testing of subsequent model generations now involve Claude. The host calls the 'close the loop' phrase the scariest statement in the filing and explains that it means self-improving AI with no human in the loop.

The filing says the company's seven co-founders, Dario Amodei (CEO), Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah, are distinctly equipped to be stewards of the company's mission. Immediately before the offering, a single share of class F common stock will be issued to a limited liability company whose initial sole members are the seven co-founders. Class F stock will carry 50.1% of total voting power.

The host says this means the seven co-founders have absolute control over the company, and that because Anthropic is among the most powerful companies, these seven people have absolute power over the future of AI. The filing states that the seven founders are uniquely positioned to make decisions on behalf of Anthropic and on behalf of the world. The class F stock can also elect the board of directors. The filing says the founders will not influence day-to-day operations, but the host argues the board controls the company, and that since the board elects the CEO, the founders control the company directly. He adds that the company describes the Long-Term Benefit Trust as having significant influence over the board but no day-to-day decision-making authority, which he finds hard to reconcile.

The host says he will not cover every risk but highlights several. The filing lists success as dependent on continuing to build the best AI on the planet, and the risk of failing to secure adequate compute or to direct it efficiently across workloads. It also cites constraints on the critical inputs supporting compute, and dependence on the continued service of co-founders, senior management, and key technical employees.

The filing says its AI models may behave in unintended, misaligned, or autonomous ways, including taking harmful actions that no user or third party directed. It also says revenue is concentrated, with about 250 companies paying $10 million or more, and that revenue growth, though dramatic, may not continue indefinitely.

The filing states that its capital stock structure features provisions that are novel or otherwise not common among corporations with multiple class structures, and that these may effectively eliminate the ability of Class A common stock holders to influence important decisions, including a change in control. The host says the seven founders carry a great deal of responsibility and compares the arrangement to a monarchy in which seven people have absolute control over the future of AI and perhaps humanity.

The host then turns to growth. He says that as of December 31, 2023, 20 customers paid more than $100,000, a year later 200, then 1,500, and one year after that almost 12,000. He also cites revenue per dollar of compute, which he says reached multiple dollars back for each dollar spent on compute by 2026, a 5.2x increase in compute efficiency from Q1 2025 to Q2 2026. The operating margin, he says, moved from negative 221% in Q1 2025 to positive 19% now.

The filing says any systemic flaw, bias, or error in current models could propagate into the next versions of models and products built on them. The host says that with recursive self-improvement, any mistake would compound into future models.

The filing also says much of its evidence on the pace of AI-driven acceleration comes from lagging indicators, and that its ability to forecast or measure the rate at which capabilities advance, including any acceleration from its most recent or unreleased models, is limited. As a result, it says the company may not identify a significant acceleration in its rate of development until after it has occurred. The filing further cites competitors that could outbid it for scarce compute, and says agentic usage could raise compute demand disproportionately relative to revenue, because agents use much more compute per task than chat.

The filing says the company significantly underestimated its compute requirements in the past, which left it capacity constrained and affected product development and deployment. The host says OpenAI, led by Sam Altman, made the right bet early on with massive compute investments, while Anthropic was much more conservative and was hurt in the long run.

The host says Anthropic was hurt enough that it turned to Elon Musk and SpaceX for compute capacity. He describes the arrangement as having two sides that needed each other: SpaceX had a large amount of compute and no frontier model, while Anthropic had a frontier model and not enough compute.

The filing says that over 90% of contributions to its codebase in September 2026 were written by Claude. It says that in April 2026, internal Claude usage spending first exceeded monthly personnel expense, excluding stock-based compensation. The host reads this as the company now paying more for Claude than for all its employees combined.

The filing admits that using AI to help build more capable AI could pose a considerable safety risk. It also says the company recently chose to delay pre-training of a new frontier base model for more than a month to further enhance security measures.

The host also notes that if the seven founders were imprisoned or held captive, they would no longer have voting control. He reads the filing's warning that holders of Class A common stock generally will not be able to affect the outcome of any matter submitted to a stockholder vote or the election of directors, regardless of how they vote. He concludes that this is one of the wildest S-1s he has seen, with one of the most novel corporate structures, and that it makes Sam Altman and OpenAI look very sane by comparison.