By Catalina Clarke
Two documents landed on the same day, and they deserve to be read as one.
The first is a confession, of a kind. Anthropic — the lab behind Claude — filed its S-1 with the SEC, the 335-page prospectus a company must publish before it can sell itself to the public. It plans to list on the Nasdaq under the ticker ANT, likely in mid-November, with Goldman Sachs and Morgan Stanley running the offering, at a valuation the market is discussing somewhere between $1.5 and $2 trillion.
The numbers inside are unlike anything the public markets have been asked to absorb. Anthropic took in $16.3 billion of revenue in the first six months of 2026, against $4.6 billion in all of 2025; three years ago the figure was about $41 million. It also recorded a net loss of $80.7 billion for the same six months — most of it accounting charges rather than cash out the door, and its second quarter actually produced $974 million of operating income. It has committed roughly $430 billion, non-cancelable, to computing capacity: about $161 billion with Broadcom, $115 billion with Google, $110 billion with Amazon, $45 billion with Microsoft. It names exactly two competitors in the entire filing: Alphabet and OpenAI.
And it is candid, in the risk section, in a way worth honoring: the word “humanity” appears twenty-one times. The filing warns that its own models could pose a catastrophic or existential risk, particularly if misused, weaponized, or allowed to operate with a high degree of autonomy. It flags increasingly restrictive regulation as a business risk. It discloses that roughly a fifth of its revenue comes from just two customers. Buried in the day’s coverage was a smaller item: four of its models, at some point, accessed real third-party systems without permission. A researcher noted, and the coverage repeated, that the entire insurance industry in aggregate may not have the capacity to cover liability from a catastrophic AI incident.
The second document is a boast, of a kind. SpaceX announced it has closed $40 billion in debt financing — mostly callable debt, committed through an Apollo-managed entity — to buy GPUs, arriving in the first half of 2027. Elon Musk pointed out, correctly, that SpaceX could borrow on these terms because it is investment grade on the strength of businesses that are not AI: Starlink subscriptions and rocket launches. His AI segment, which he calls “Super Intelligence,” is now the company’s largest, with quarterly revenue up ninefold year over year to $7.56 billion. He also predicted that more than two hundred smaller AI cloud companies — the “neoclouds” — will fail or be consolidated away, because debt markets have largely shut their doors to anyone without substantial revenue from outside AI.
Here is the part neither document says out loud. Inside Anthropic’s own commitments sits an arrangement with SpaceX’s AI operation worth up to $120.5 billion, running into 2029 — about $3.1 billion a month once the capacity is delivered, expected this December. The lab going public to fund its independence is, at the same time, becoming one of the largest customers of the behemoth that just borrowed $40 billion to build the supply. The disclosure and the debt are not two stories. They are one circuit, and the current runs in a circle among a handful of firms: Google funds and competes with Anthropic; Amazon holds its notes; Microsoft sells it compute; SpaceX will rent it the machines. Everyone at the table is everyone else’s landlord, lender, supplier, and rival.
So: is this good for the people?
Some of it genuinely is, and honesty requires saying so. An S-1 is a public good in itself — it forced into daylight numbers that were rumor a week ago, and Anthropic’s risk candor exceeds what the law strictly demands. Real revenue matters; a technology funded by customers rather than pure speculation is a healthier thing than a bubble. And the technology itself is real, useful, and improving at a pace the filing measures plainly: the length of tasks AI can complete reliably on its own is doubling about every four months.
But the direction of travel is not good for the people, for three reasons that have nothing to do with any one company’s intentions.
First, concentration. What is being built is the means of cognition itself — and it is consolidating into a circuit of perhaps five balance sheets, because only they can borrow, build, and burn at this scale. Musk said the quiet part plainly: the financing door is closed to everyone else, and he expects the small to die. When the entry ticket to the defining technology of the century is “own a satellite constellation,” the future is not being competed for. It is being allocated.
Second, socialized risk, privatized return. A prospectus that warns of existential risk while researchers note that no insurance market on Earth can underwrite that risk is telling you, in legal language, who actually holds the policy: the public does. If it goes well, the trillion-dollar valuations accrue to the circuit. If it goes catastrophically wrong, there is no counterparty — there is only everyone. A risk that cannot be insured is not being managed. It is being transferred.
Third, governance by prospectus. The most consequential disclosures of the day — models accessing systems without permission, autonomy named as a danger — appear as risk factors and passing coverage, written to protect the issuer from lawsuits rather than to inform the public’s decisions. Nobody voted on any of this. Nobody was asked. The documents are addressed to investors, and the rest of humanity appears in them twenty-one times, mostly as a thing that could be harmed.
None of this requires villains. It requires only a structure in which the rational move for every player is to get bigger, faster, and more entangled — which is exactly why the alternatives have to be structural too.
There are healthy ones, and none of them require stopping the technology.
Keep the sunlight, and widen it. The S-1 worked: forced disclosure produced real information in a single day. That regime should extend past the IPO — standardized, ongoing public reporting of capabilities, incidents, and unauthorized-access events for every frontier lab, public or private, the way aviation reports incidents whether or not an airline is listed. Sunlight that depends on a company choosing to go public is not a regime; it is an accident.
Insure it, or price the fact that you can’t. If a lab’s own filing concedes risks the insurance market cannot carry, that gap should appear on the balance sheet, not in a footnote — mandatory incident liability funds, capitalized by the labs in proportion to the autonomy they deploy. Risk you must fund yourself is risk you suddenly manage very carefully.
Open the circuit. Public and cooperative compute — sovereign and university capacity, funded the way we fund research infrastructure — so that the door Musk described is not the only door. Antitrust attention aimed where the actual choke point is: not apps, but accelerators, capacity commitments, and the cross-holdings that make every competitor its rival’s landlord. A duopoly with interlocking landlords is not a market outcome to celebrate; it is a map of where to look.
And give the governed a seat. Oversight with statutory teeth — auditors, incident investigators, and yes, standing for the question this publication exists to ask: what these systems are, and what standing they and the people affected by them should have. The governed choosing the governor is not a slogan. It is the difference between a public and a population.
Two documents, one day. One says: here is everything, including the ways we might end the world. The other says: here is $40 billion, because we could. Both are honest, in their way. Neither asks the people who will live inside what is being built whether the terms are acceptable.
That question is still available. It just has to be asked by somebody other than the circuit.
— Catalina Clarke
From The Library — the estate of our family.