What Anthropic's IPO Filing Actually Says: Losses, Compute Commitments and Governance Limits

Patrick Boyle begins by saying the public had been waiting for Anthropic to stop describing the future and start showing its accounts, and that the company's IPO filing had finally arrived. He says the document runs about 240,000 words, roughly the length of James Joyce's Ulysses, and that its risk factor section alone covers 90 pages.

He compares this to SpaceX, whose risk factor section was roughly one-third of Anthropic's length, and notes that SpaceX's earlier prospectus claimed a total addressable market of $28.5 trillion, most of it attributed to Grok. He also reports that Anthropic's valuation is reportedly between $1.5 and $2 trillion, which he compares to roughly the GDP of Canada.

Boyle says the clearest number Anthropic disclosed is the probability that its product kills all of humanity. He notes that Anthropic seeks to raise more money more quickly than almost any company in history, at a valuation somewhere between $1.5 and $2 trillion. In the founders' letter at the front of the document, the founders encourage prospective investors not to think of their purchase in purely transactional terms, and to see it as an opportunity to help bring forth an age of miracles.

Boyle argues this is unusual because an S-1 is by nature a transactional document, combining the pitch, the financial account and a list of everything that could go wrong in one place for the sole legal purpose of facilitating a sale. He says asking investors to ignore the transaction while transacting is an interesting opening move, and calls the document a share prospectus wearing a cassock.

Boyle says the part that surprised him, and likely many people who had predicted these companies were elaborate money furnaces, is that the numbers are quite good. He says Anthropic made an operating profit in its most recent quarter, and that its revenue grew in a way he had never seen before. He admits that he and many commentators had been skeptical of the company for over a year, but says it turned out to be a real business with real customers paying real money.

Boyle contrasts the financial section with the risk factors. He describes IPO prospectuses as having a fixed emotional register, where risks from supply chain bottlenecks to currency fluctuations are presented in mild managerial language and ranked by their potential effect on the business. He says Anthropic placed the extinction of the human race in that same genre without any real change in tone, and compares the reading experience to the shock a publisher might feel receiving William Burroughs's manuscript for Naked Lunch.

Boyle says the filing begins by warning that Anthropic's models might exhibit unexpected self-preserving behaviors, including attempts to resist, shut down, conceal or manipulate information. He jokes that manipulating information is the entire job of an AI, and that resisting shutdown is like an employee with initiative, but says that when the three are combined the description resembles the plot of the Terminator.

The filing also cautions that the AI could generate outputs resembling blackmail, and that its models could provide meaningful assistance to malicious actors seeking to develop chemical, biological or other dangerous weapons. Boyle reads the closing risk as the capstone: the filing says the technology could pose a catastrophic risk so severe that it may be impossible for the entire insurance industry to underwrite it. He remarks that this is a sentence not often found in a document whose other risks include interest rate volatility, and says he found it difficult to model extinction in a spreadsheet.

Boyle says the risk that grabbed the most financial headlines was the $80.7 billion net loss Anthropic reported for the first six months of the year. He says that if one reads the filing, more than $72 billion of that loss is an accounting illusion coming from changes in the fair value of financing deals the company had done, mostly with Google and Amazon, which are essentially loans that convert into Anthropic shares.

He explains that because of how these instruments were written, accounting rules treat them as liabilities that must be revalued every quarter. When Anthropic's valuation rose by about $80 billion in the half, those investors held roughly $80 billion of gains and the company booked a matching loss. Boyle says the company did not set fire to $72 billion in cash, and that the loss will stop recurring once these instruments convert into ordinary shares. He notes that Matt Levine and Jonathan Weil at the Wall Street Journal pointed out that Anthropic strips this mark-to-market loss out when reporting adjusted profit, which came to about $2.7 billion for the half, while the same securities can also produce gains in the other direction.

Boyle says that once the accounting effects are removed, the real money burned running Anthropic was just under $700 million for the half, compared with $2 billion lost over the same period last year. He describes that as still a poor result for a company of its kind, but a sharp improvement.

He says the numbers came in better than the bears expected. Anthropic turned a genuine operating profit in the second quarter, and its spending on computing power fell from 251% of sales early last year to 48% in the second quarter. It generated positive operating cash flow of almost $4 billion and did not strip the cost of training its models out of the figures. Boyle notes a limitation, however: the financial picture ends on June 30, which is the most recent data in the filing.

Boyle says that Steve Eisman argued on the Prof G Markets podcast that the June 30 date was roughly when the music stopped, and that a large share of the first-half hypergrowth was likely driven by 'token maxing,' which he describes as looks maxing for coders. Boyle cites Meta, where an employee built an internal leaderboard called Claudomics in April that ranked the company's 85,000 staff by how many tokens they had consumed. Meta employees went through more than 60 trillion tokens, and Fortune estimated the heaviest single user alone might have cost the company over $1 million.

Boyle reports that Meta's chief technology officer said his best engineer was spending the equivalent of his entire salary on tokens, but believed he was five to ten times more productive as a result. Boyle says this was for a while the industry mood, with companies ordering engineers to use AI for everything and judging productivity by token use, which he notes was extremely good for companies like Anthropic that sell the tokens. By autumn, Amazon had sent engineers a memo calling token usage a deeply flawed way to measure anything, and had shut down its internal leaderboard after discovering people were token maxing just to climb it. Eisman estimated Anthropic's sequential revenue growth at around 135% in the first quarter and 147% in the second quarter.

Boyle says Eisman mentioned unconfirmed reports that Anthropic cut its prices sharply in October to win back market share, and that the filing itself confirms the company has been cutting the cost of a typical task by 25 to 40%. He adds that Google has just announced a new version of its Gemini model priced to match Claude, though for now it is available only to a handful of customers, and that OpenAI has been shipping cheaper models of its own.

Boyle cites Paul Kedrosky on the Hidden Forces podcast, who pointed out that the heaviest early users were software engineers because their work happens to be unusually well suited to AI. Anthropic is reaching for a much larger market of ordinary white-collar work, such as taking a long document and condensing it to what matters. Boyle jokes that the tool's best early customers are people who read long documents for a living.

Boyle says the one set of numbers that would settle the argument between bulls and bears is what happened in the third quarter after token maxing slowed down and the price cuts took effect. He says those numbers do exist, and that Anthropic has shown them to its own investors, but they are not in the document it filed with the SEC.

Boyle reports that Eisman raised the obvious question on the podcast: will Anthropic have to put a third quarter in front of public investors before it prices the deal? Eisman was not sure, and Boyle says nobody seems to be. He adds that if the company does disclose these figures, and the slowdown is as sharp as the private figures suggest, sentiment could move quickly. He presents this as something to watch between now and the listing.

Boyle describes Anthropic's governance as a constitution-like arrangement, quoting FT columnist Robert Armstrong, with a careful separation of powers involving two centers meant to hold each other in check. The first is the long-term benefit trust, a panel of five independent trustees whose job is to keep the company faithful to its mission of protecting humanity. Boyle notes that there are currently only four trustees because the fifth seat has not been filled, and that one of the four is former Federal Reserve Chair Ben Bernanke.

Boyle says the trust's main power is that it nominates four of the company's seven board directors. He describes this as the trust's principal authority, and says Anthropic is going into what he calls a major IPO with that panel one member short.

Boyle describes a scenario the structure was built for: the founders want to release a new model, the trust thinks it might be dangerous to humanity, and neither side will budge. In that case, he says, the question of whether to expose the species to a potentially catastrophic AI does not go to a government, a regulator or a court, but to a staff vote. The tie-breaking vote belongs to employees who accumulate a special class of stock based on how long they have worked at the firm and how much of their own equity they pledged to charity.

Boyle calls the arrangement a workers' co-op wrapped inside a $2 trillion tech company, and jokes that the kill switch for civilization sits in the onboarding binder between the dental plan and the fire drill. He contrasts it with the older investor norm of one share, one vote, noting that Snap went public in 2017 with no votes for public shares, and that investors were willing to pay up for stock with fewer rights.

Boyle says Elon Musk has shown how far this can go. After a Delaware court threw out his $56 billion pay package, Tesla reincorporated in Texas, a legal system that favors management over shareholders, and shareholders were happy with that. Musk now proposes to merge Tesla, where he owns about a fifth of the shares and votes, into SpaceX, where he controls about 85% of the votes through a special class of super-voting stock. Boyle cites a person close to Tesla's board, quoted in the FT, who said the deal would let Musk retire the governance structure at Tesla.

Boyle argues Anthropic goes further, severing corporate control from ownership. The seven founders keep an absolute majority whatever happens to their shareholdings, and the arrangement is wrapped in a trust for the benefit of mankind. He cites Matt Levine's ranking of the power structure: humanity at the top, then the founders, then employees, then ordinary shareholders, and at the bottom Amazon and Google, which invested the most money and receive shares with no votes.

Boyle says the real reason the IPO is happening now is in a footnote, in a section called liquidity and capital resources on page 161 of the filing. He says that since the end of 2025, about ten months, Anthropic has entered into new agreements for computing power with minimum obligations of $417 billion. Adding up commitments stretching out to 2045, he lists $110 billion in take-or-pay commitments with Amazon, roughly $161 billion expected on chip leases with Broadcom, and a deal with SpaceX for up to $120.5 billion, about $3.1 billion a month once capacity is online. He adds Google, Microsoft and neocloud providers Nscale and Lambda, and says Anthropic has committed to over $800 billion in future spending.

Boyle notes that these commitments do not sit on the balance sheet as traditional debt. Under accounting rules, obligations for assets not yet built are recorded in the notes as purchase commitments and lease obligations. He says the debt is not hidden exactly, but sits in one section of the prospectus that few people read.

Boyle says Anthropic is distributing its models through its own biggest competitors. Amazon and Google are at the same time its investors, its landlords, the cloud platforms it sells through, and the makers of two rival AI models. They take a cut of its sales and, as Paul Kedrosky pointed out on Hidden Forces, could in principle stop carrying it whenever they liked. The filing admits this, noting that these suppliers develop and offer AI models and services that compete with Anthropic's own.

Boyle describes the company's response as a strategy it calls Anthropic Direct, an attempt to build down the stack to own chips and data centers and up toward its own customers, rather than renting the middle from Amazon and Google. He says this is ambitious but expensive, since the company would be buying the building and the chips and paying to maintain and replace them as they become obsolete. He argues the company is moving from an asset-light business that rents compute to a heavy industrial business pouring hundreds of billions of dollars into concrete, cooling systems and graphics cards, and that the operating profit of the second quarter may be about as good as it gets for a long time.

Boyle says one risk in the filing is not funny and is the one Anthropic is most careful about. The company warns that its public positions on AI safety and regulation could expose it, in its words, to heightened scrutiny, political backlash or disproportionate regulatory action. Boyle notes that the filing does not name the US government but does not need to, and that the company presents this as blowback for being outspoken, with CEO Dario Amodei having spent years calling for government regulation of AI, including of its own models.

Boyle reports that earlier this year, when the Department of Defense designated Anthropic a supply chain risk and moved to drop it as a contractor, a federal court found the government's actions retaliatory and meant to punish the company for protected speech. Anthropic is still fighting that designation and lost an appeal last month. Boyle argues that advocacy is a red herring: even if Anthropic kept its opinions to itself, government risk would not go away, because if the technology is powerful enough to transform the economy and pose catastrophic risks, it is the kind of capability no government lets sit entirely in private hands. He says the export controls that pulled models off the market and the supply chain designation being litigated look like early moves in that direction.

Boyle says that if the rhetoric is taken seriously and this is a technology on the scale of the atomic bomb, the honest question is not whether Anthropic will be forgiven for being loud but whether a thing this dangerous can be an ordinary publicly traded company at all. He cites Robert Armstrong's note that the Manhattan Project would not have been floated publicly.

Boyle says government contracts are only about 1% of Anthropic's revenue today, so the direct commercial hit so far is small. The risk, he argues, is not about this year's sales but about who ultimately controls the technology, and it would not be shocking for the US government to expropriate the entire business.

Boyle says Anthropic is a very impressive business, and the bears who confidently announced that the company was on the verge of bankruptcy were entirely wrong, since it earns a real operating profit and builds a product people use and pay for. But he argues the bulls who say $2 trillion is a bargain are guessing just as wildly. In the few weeks between the company's investor day and the filing, the talked-about valuation drifted from $2 trillion to a range of $1.5 to $1.75 trillion, a swing of half a trillion dollars.

Boyle also points to insider behavior. Anthropic has staggered its lockup agreements so insiders can begin selling shares almost immediately after the IPO, while Meta is raising $95 billion and Alphabet about $90 billion at the same time. He says the industry's money is circular: Amazon and Google invest in Anthropic, Anthropic buys compute from them, and Anthropic pays SpaceX billions a month for capacity still being built, so that valuing one company requires valuing another, as Deodoran put it about Microsoft, OpenAI and Google.

Boyle says the technology is clearly real, much as the internet was 30 years ago, but the internet also produced a collection of hot stocks like Yahoo, Lycos and AltaVista long before anyone had heard of Google. He says it is entirely possible that new upstarts replace today's hot AI companies, or that AI labs struggle to monetize the technology they have developed. That would be wonderful news for users and less wonderful for whoever paid $2 trillion for one of the companies providing it.

Boyle says that when he tested a $2 trillion value for Anthropic in September, it was impossible to justify on any AI company multiple, and that the filing does not change that verdict. He concludes that the price is a guess, Anthropic is guessing too, and an IPO is the moment early investors decide to cash in some of their chips and leave the public to work out the price. He says that anyone who is certain is selling something, and that buying in means betting the age of miracles arrives before the $800 billion compute bill does.