The Warning That Sold the Stock

$44 trillion. That is the number Anthropic's Jack Clark carried into his meetings with IMF and World Bank policymakers last week, presented as the American GDP the company believes AI could reach by 2030 ^[1]. The number has a provenance worth tracing. It is the extreme scenario in the company's own September economic model — the case that requires recursively self-improving AI — and inside that scenario unemployment runs near 14%, knowledge workers' wages fall 11.5%, and capital's share of GDP climbs from 40% to 54.8% ^[2]. One term was left out of the simulation: the company's own alignment lead puts the odds of AI-caused extinction within a decade above 10%, and that probability never enters the math that produces the $44 trillion ^[2]. Clark's title is Head of Public Benefit. Days before he carried the upside to the lenders, he described the industry to a different audience in different terms.

I think it's Formula One at night without maybe brakes in the cars. — Jack Clark

The same person, in the same week, sold the lenders the upside and the regulators the danger. The title does quiet work throughout: whichever direction the story runs, it is told by the officer charged with the public's interest. The sales document runs both directions on the same pages. Roughly a third of the filing is given over to risk factors, including a warning that the technology could bring harm to humanity at a catastrophic or existential scale ^[3]. Beside that section sit at least $518 billion in compute commitments, about 80% of them non-cancelable, and a $42 billion net loss for 2025 ^[3]. The filing is candid about which side is the constraint.

catastrophic or existential risk to humanity — Anthropic

Set beside the doom section, that sentence does its own work: the company is telling investors that catastrophe is possible, and that the only thing standing between it and more revenue is how much compute it can lock in. Anthropic has spent the year telling the world the industry has a throttle and no brake; in the filing, the throttle is what is for sale and the brake is the section warning how fast it can go. The record is careful about one thing and silent about another. Every structure that follows was built by the company — that much is documented. What any of it was built for is not. The day before the filing, the two big labs issued a joint call for existential-risk regulation in which they pushed to choose their own auditors and name the evaluators — Anthropic named Metr ^[4]. The filing then concentrated control in a Founder LLC whose seven co-founders hold 50.1% of the votes through a single Class F share, beneath a Long-Term Benefit Trust that elects four directors and whose advisers include former Fed Chair Ben Bernanke ^[5]. The filing itself warns investors this arrangement may lead to decisions that conflict with their financial interests ^[5]. On October 14 the company published its founder's manifesto on pacing the frontier, warning that agents could take over the internet within six to twelve months ^[6]. It landed the same day as the investor day for the potential $2 trillion listing, where sentiment split between the revenue numbers and doubts about the build-out ^[7]. Through the summer the same office had insisted the evidence did not support a pause. Two days later the evaluator it had named was inside the building: Metr, embedded with employee-level access on a tiered system Anthropic itself administers, while the company also funds another evaluator, Accenture ^[8]. The company framed the move as a unilateral self-binding. The same week it signed the White House self-policing pledge — a promise the president himself said carries only moral force ^[9] — even as the administration's own default still holds the makers liable for harms ^[10]. The bill gets the deepest read. Anthropic's draft lets the government block a model through exactly one channel: Commerce petitioning a federal judge in Washington on a foreseeable, imminent risk of a billion dollars in damage or fifty deaths ^[11]. It would override state laws, including California's, and bar every other federal agency from halting AI development ^[11]. The company says it has been advocating this framework to lawmakers since June ^[11]. The venue is the detail. The single DC courtroom the bill routes the government's power through is where, in the spring, a judge described the administration's earlier moves against the company as looking like an attempt to cripple it ^[12]. The blocking power is channeled through the one process where the company already won. Three days before listing, the company created its Public Benefit Fund: wholly owned, seeded with 5% of its shares, escalating by 2% for every 50% gain in the company's value above a baseline, with roughly a fifth of shares pledged to public good through founder and employee commitments ^[13]. The public's share of the great wealth is denominated in the company's own stock, administered inside the company, and grows exactly as fast as the share price does ^[13]. Beneath it sits the Founder LLC's 50.1% of the votes ^[5]. Critics note the structure keeps control of the money inside the company ^[13]. The rivals built their own versions at their own price points. Google lobbied at least ten states with model language exempting its own search-integrated chatbots from liability, until Arizona's governor vetoed her state's copy ^[14]. OpenAI asked the public to serve as the insurer of last resort ^[15]. The money follows the same lines: Anthropic's $20 million went to candidates who favor safety rules ^[16], against a PAC backed by OpenAI and Andreessen Horowitz that has raised more than $125 million for lighter regulation ^[16]. Yann LeCun's read on all of it is that the extinction warnings are a marketing campaign meant to achieve regulatory capture, protecting dominant firms from smaller competitors ^[17] — though he runs a competing lab. Neither wall covers the two forces that actually moved the valuation. Above sits the concentration: Microsoft, Alphabet, and Amazon control the compute and distribution layer, and analysts flag that as a drag on the IPO prospects of developers like Anthropic ^[18]. Below sits the cheaper open-weight competition — DeepSeek's faster, leaner release, the same week as the bill, helped knock CoreWeave down 25% in a week and put $3.5 trillion in spending at risk of impairment ^[19]. Every protection in the package is priced in the issuer's own paper — the evaluators the company picked and pays, the pledge no court can enforce, the public's compensation indexed to the warner's own stock, its stake escalating with the share price. The two exposures that moved the valuation arrived in the same window the walls went up, and neither is inside any of them. The catastrophe got a third of the document. The thing that actually moved the market last week sits outside every wall.