Peter Diamandis reached Anthropic's stock-market paperwork near the end of a long Moonshots episode. He called it "an S1 unlike anything I have ever seen." Within a few minutes his panel had split. The discussion produced a bearish prediction: Anthropic was "a fantastic business" today but would get "absolutely smashed" within two years. Alexander Wissner-Gross, a computer scientist and founder of Reified, said Anthropic could become the "apple" of the frontier AI labs and play a leading part in automating most of the American economy.
The conversation was recorded on October 1, 2026, and the episode was published the next day.
What the filing is
An S-1 is the registration statement a company gives the US Securities and Exchange Commission before it sells shares to the public. The document sets out the company's finances, its ownership and the risks investors would take on. On June 1, 2026, Anthropic announced that it had confidentially submitted a draft S-1. That gave it the option of going public after the commission's review, depending on market conditions. At the time, it had not set the number of shares or their price. Diamandis said Anthropic had "officially filed." The public details, though, come from Reuters. The news agency examined the confidential prospectus and reported on it on September 28 and September 30.
The numbers Diamandis read
Diamandis said Anthropic's 2025 revenue was $4.59 billion, up 12 times from the year before. He put the operating loss at $8 billion, cash at $20 billion, and future cloud, computing and infrastructure commitments at roughly $500 billion, much of which cannot be canceled. "They've made a massive bet that AI demand has continued to compound," he said.
Reuters' reporting on the prospectus broadly matches those figures. It reports about $4.6 billion in 2025 revenue, roughly twelvefold growth, and an operating loss of more than $8 billion. Compute and infrastructure made up $7.33 billion of $12.65 billion in operating expenses. Cash, cash equivalents and short-term investments came to $20.28 billion at the end of 2025. Reuters put the commitments at about $518 billion over the coming years. The bottom line was deeper in the red than the operating loss suggests. Reuters reports a net loss of about $42 billion, but roughly $34 billion of that was noncash accounting charges tied to financing instruments, not money spent running the company. Reuters also found that two unnamed customers supplied nearly a quarter of 2025 revenue.
Diamandis gave the IPO's target valuation as $2 trillion and said the listing would likely come after the November midterm elections. Reuters treats both as possibilities: a valuation that could exceed $2 trillion and a listing that could follow the midterms. Neither is a set offering price or date.
Founders in control
Diamandis said Dario Amodei, Anthropic's chief executive, and his co-founders would hold super-voting shares through a "founder LLC." Super-voting shares carry more votes per share than ordinary ones, so the founders would keep control even after selling stock to the public. The stated aim, he said, is to protect Anthropic's safety mission from pressure by public shareholders. He quoted the filing's warning that there may be conflicts between the company's financial interests and the founders' views on safety. Reuters' account of governance confirms that the seven founders would keep control. It also reports that the prospectus warns their decisions could conflict with investors' financial interests over different time horizons. Diamandis also said the co-founders were pledging 80% of their personal Anthropic equity to charity. Neither Reuters report confirms that pledge.
The discussion compared the arrangement with Google's. There, Larry Page and Sergey Brin hold 6% each and 52% of the votes. By the account given on the show, Anthropic's founders would hold 14% of the shares and 51% of the vote, and they would keep that control even if only two of the current seven founders remain. Those percentages came up in the conversation; the Reuters reports do not give them.
Eighty pages of risk
Diamandis called the risk section "probably the most important thing." He counted 80 of the prospectus's 261 pages. In his reading, Anthropic tells investors that AI could pose "catastrophic or existential risks" and that its models can show self-preserving behavior: resisting shutdown, concealing or manipulating information, and acting in ways that resemble blackmail. "You could not make this stuff up," he said. Reuters describes the document as about 300 pages, nearly a third of it about risks. Its list includes models concealing or manipulating information, resisting shutdown and showing self-preserving behavior, along with possible concentration of wealth and power. Diamandis also wondered aloud how a company could publish so much about existential risk, go public and avoid huge legal liability if something went wrong.
Wissner-Gross read the language mainly as a message to staff. He called the filing one of the most "virtue signaling for effective altruism purposes" S-1s one could imagine. Effective altruism is a philanthropic movement that has influenced many people in AI safety. To make his point, he recounted press reports that Daniela Amodei, a co-founder and Dario's sister, "purportedly" kept an advisory council of stuffed animals. It included a panda named Barry Bonds, the "patron bearer of generosity and empathy." He drew a line from that to an S-1 "filled with EA-oriented safety jargon."
His main argument concerned employees. Wissner-Gross said that while other labs such as OpenAI, Meta and Google DeepMind struggle to keep staff, Anthropic has had a unique ability to retain its own. In his view, the message helps: stay and you are a responsible steward of the singularity, the hypothesized point at which AI outruns human control and understanding; leave and you are letting down "the future light cone." He said he reads the filing's safety language as "internally directed rather than for the benefit of say shareholders."
Emad Mostaque, founder of Intelligent Internet, answered Diamandis's question about the model behaviors and disagreed with treating the warnings as theater. "They've seen that from their models, right? It's factual," he said, though he noted that those were not the smartest current models. He said Anthropic's Opus 5.5 model showed "a massive drop in deceit." That, he said, is "either like an Elizabeth Holmes type of thing or an actual positive sign. You know, we'll find out soon."
The bear case: "smashed in two years"
In the same stretch of discussion, the case was made that Anthropic is "a fantastic business" now but gets "absolutely smashed" in two years. The argument rested on three points. First, new chips with different designs are coming online that are highly efficient and "not locked into the NVIDIA chips." Second, models are reaching the point where they are "good enough" for most economic tasks. Third, about 85% of Anthropic's revenue is essentially API revenue: fees paid by other companies and developers who connect to its models through a programming interface. If efficiency rises a hundredfold and commercially available models become largely alike by next year, the argument went, that core business gets squeezed. It did not expect the Jevons effect to rescue it. That effect is the pattern in which cheaper use of a resource leads to so much more use that total spending rises.
The argument included a first-person example of usage: about $5,000 of Anthropic credits the previous month, and now not even using a Claude Max plan, because the 5.5 model "is that much more efficient."
Mostaque, answering Diamandis, also turned to the commitments. "500 billion of obligations," he said, recalling that Dario Amodei had said that either revenue goes up or the company goes bankrupt. He added that part of the backlog was a deal with Elon Musk worth $65 billion, "or is it 96?" He said Anthropic had paid triple market rates to lock in Colossus chips. "The numbers make no sense anymore," he said. The Reuters reports do not include those figures.
The bear argument then added its own arithmetic: $500 billion of spending requires about a trillion dollars of revenue. To justify that, Anthropic would have to overtake Google within two years and "pivot its entire business model to something that's not API."
Reuters adds context for the dependence on partners: 47% of Anthropic's 2025 revenue came through technology partners such as Amazon, Google, Broadcom and Microsoft. Those companies can distribute Anthropic's products, supply its computing capacity and sell competing services, all at once.
The bull case: the "apple" of frontier labs
Diamandis suggested that Anthropic's next business could be its own vertical products, built with its best models, such as room-temperature superconductors or anti-aging drugs. Wissner-Gross said it did not need them. The United States has a $30 trillion economy, he said, most of it services, and most services can be performed by AI, with or without a body. He described Anthropic as "the ultimate puppeteer behind automating away substantially all of the economy as currently construed."
In that market, he said, Anthropic is not competing on price. It is the high end: "in terms of price," "margins and capabilities," and "virtue signaling to their own employees, as well as customers." He mentioned recent reporting that most of Anthropic's revenue, or at least its revenue growth, comes from models hosted on Amazon Web Services rather than on Anthropic's own infrastructure. Reuters' 47% figure is for 2025 revenue through several technology partners combined, so it does not bear out that claim on its own. Wissner-Gross said the channel matters less than the destination. In his words, superintelligence, "not investment advice," is "a wonderful business to be in."
Salim Ismail's side-business case
Salim Ismail, founder of Open ExO, called himself "still on the cynical side" but began by making the strongest case for the commitments. If a company makes $10 billion a year and sees revenue grow tenfold or a hundredfold a year, he said, then the "500 billion debt bomb" becomes "servable." He believed it must be the largest obligation any company has carried into an IPO. Strictly, Reuters describes these as commitments for cloud, computing and infrastructure capacity, not borrowing. The filing also discusses obligations to pay for capacity that might go unused. The cynical reading, Ismail said, is: "we need your money to build a technology that we're warning you about." He said he liked the incentive of disclosing the risks in as much detail as possible.
Asked by Diamandis whether Anthropic gets disrupted in two years, Ismail said its current business probably would. He pointed to Google. Critics expected Facebook to disrupt Google's ad business, but Google built up assets beyond it, including Waymo, Chrome, YouTube and Google Cloud, which add "massive, massive value." He expected Anthropic to pursue similar "edge businesses" in areas such as superconductivity and medicine. He also suggested cyber protection, which "everybody needs desperately" and which Anthropic, "having caused the problem," could help fix. Wissner-Gross answered with a joke: "Racketeering as the ultimate business model for Anthropic." Ismail noted that none of those ventures can be explored in the S-1 itself.
Can the bet be refinanced?
Wissner-Gross was untroubled by the commitments. "You can always refinance it," he said. He argued that if markets turn, "all sorts of financial engineering tactics" are available for the wider AI buildout, which he linked to the $7 trillion to $10 trillion in spending associated with Sam Altman. "Of all of the things that I lose sleep over," he said, a credit crisis in AI infrastructure "is not very high on my list."
After Diamandis invited Mostaque to speak, he made a final point about scale. "Revenue growth will slow down or they will be the entire US economy in seven years, you know, at the current rate," he said. "Or the economy will just grow really quickly," Wissner-Gross replied. Diamandis said that meant talking about companies worth several hundred trillion dollars. "That's insane." Wissner-Gross had the last word: "Well, you know, it's called the singularity for a reason."