Anthropic's S-1 Filing Reveals $602 Billion in Compute Commitments and Profitability Risks

This roundup covers Anthropic's IPO disclosures and other AI developments. Anthropic published an IPO prospectus yesterday, revealing financial results, extensive compute commitments and risks to future profitability. The S-1 shows that, as of June, Anthropic earned more than two dollars of revenue per dollar spent on compute. Anthropic reported a 19% adjusted operating profit margin for the latest disclosed quarter, excluding stock-based compensation, non-cash charitable contributions and payments settling authors' copyright claims. The filing shows a 50% increase in revenue per dollar of compute expense between the first and second quarters of 2026. Anthropic's operations produced $3.9 billion in free cash flow during the first half of 2026.

Anthropic warns that expenses for substantial additional compute will precede the full associated revenue, potentially reducing operating margin and adjusted operating margin during those periods. Anthropic also warns that price competition could materially reduce margins or undermine the company's ability to become or remain profitable. Anthropic spent $5.6 billion on compute and infrastructure during the six months through June 2026. Anthropic's commitments include $45 billion for Microsoft data center capacity. Anthropic has committed $110 billion to Amazon between now and 2036. Anthropic has committed $121 billion to SpaceX. Anthropic has committed $115 billion to Google. Anthropic also has additional commitments of $125 billion to $150 billion. Anthropic has committed $45 billion to Nscale and, separately, $41 billion to Lambda. These capacity commitments total at least $602 billion to $627 billion. Anthropic's chip spending includes at least $20 billion with AMD and a $161 billion leasing agreement with Broadcom, bringing the combined total to as much as $808 billion.

Jeremy Kahn questions whether revenue can cover these costs, noting that Anthropic's 1,300% revenue growth between the first halves of 2025 and 2026 almost certainly cannot persist. Anthropic's customers can readily switch models, while non-cancellable compute contracts fix most of the company's costs. The five members of Anthropic's Long-Term Benefit Trust can appoint a majority of the corporate board. Anthropic's seven cofounders collectively hold 50.1% of the voting power for electing the remaining directors through Founders LLC. Ordinary investors receive non-voting Class A shares, while major backers Amazon and Google receive non-voting Class N shares. Employees holding Class E voting stock can vote on certain changes to the trust's operation. Kahn argues that this restriction on shareholder voting contrasts with Anthropic's support for democracy and efforts to gather diverse views for Claude's written constitution.

Anthropic's disclosed risks include authoritarian surveillance and repression, malicious use of AI for cyberattacks, possible US government nationalization and potentially catastrophic or existential harm from inadequately safeguarded advanced AI. Anthropic says the company is committed to helping free societies counter authoritarian misuse and support freedom and democracy. Anthropic omitted third-quarter 2026 financial results from the S-1 despite briefing existing investors on those results at an October 14 investor day. Social-media critics allege that Anthropic withheld those results because the figures present a weaker picture of financial health and growth. Anthropic disclosed that revenue for the three months ended September 30 was 86 times the corresponding 2024 figure, implying about $16.8 billion and matching a leaked investor-day figure. That implied revenue represents approximately 45% quarterly growth, compared with 147% between the first and second quarters of 2026.

Social-media critics argue that Anthropic's use of convertible notes for some computing costs obscures operating economics and makes adjusted results and cash flow look stronger. Anthropic issued most of those notes to Amazon and Google, which may also have discounted compute prices. The S-1 does not provide an exact payment schedule, describing obligations across 2026 to 2033 or 2036 and, elsewhere, commitments lasting one to 20 years. Analysts want a detailed schedule to assess whether revenue can meet the obligations. Anthropic's shares are expected to begin trading between November 9 and Thanksgiving.

The Washington Post reported that Cathedral, a cybersecurity startup founded by former DOGE staffers, used AI agents to access sensitive Anthropic systems in under 12 hours. According to that account, Cathedral obtained internal meeting transcripts, compensation and job-offer information, and Slack discussions about advanced models through connected third-party business software. Cathedral operated under Anthropic's bug-bounty program rather than penetrating the company's core AI systems directly. Anthropic said access was confined to limited sales, HR and IT systems, excluded customer conversations, model code and model weights, and was quickly closed; Anthropic accused Cathedral of overstating national-security implications.

The roundup also covers OpenAI's claims of mathematical, computer-science and physics breakthroughs that still require scientific scrutiny, OpenAI's release of GPT-6.1 Terra at prices the company says are up to 50% lower per token than GPT-5.6 Terra and GPT-6.1 Sol, and Menlo Ventures partner Deedy Das's figures showing that discovered cybersecurity vulnerabilities tripled over the past six months.