Anthropic’s safety promises meet the terms of its IPO

The AI Power Dossier · Evidence checked through 11 November 2026

Anthropic’s 10 November prospectus amendment describes a company with about $900.9 billion in aggregate spending commitments, substantially all for computing and infrastructure. The same filing proposes a single share that would give a group of founders 50.1% of general voting power. The current charter gives a separate trust the right to elect four of seven authorised seats. The filing identifies three sitting directors elected as its nominees. Anyone buying into the proposed offering has to make sense of both the bill and the division of power. Anthropic Form S-1/A, governance and contractual obligations

Former OpenAI employees founded Anthropic in 2021. Its computing suppliers include investors, and its limits on Claude’s use led to a dispute with the US government.

Anthropic is a Delaware public benefit corporation. Its board is expected to balance shareholder interests with the interests of people materially affected by its conduct and the company’s stated public purpose. Its Long-Term Benefit Trust holds the Class T share. The filing says three of the six sitting directors were elected as its nominees. They were Daniela Amodei, Reed Hastings and Vas Narasimhan. If the offering completes on the proposed terms, the four authorised Class T seats would initially be assigned to Hastings, Christopher Liddell and Narasimhan, with one vacant. Special procedures would apply to Liddell’s designated seat. The trust’s role concerns board appointments. The company handles day-to-day management and strategy. Anthropic S-1/A No. 2, pp. 262–265 · Current charter, Art. VIII(B)(5) · Proposed charter, §§ 7.1 and 8.7

Anthropic says it paid the trust $8,425,213 in the nine months to 30 September 2026 for operations and trustee pay. It consolidates the trust as a variable interest entity, citing power over its economic activities, financial support and rights to financial assets other than Class T stock on termination. Legal control and intervention in trustee decisions remain separate questions. Anthropic 10 November S-1/A, Transactions with the LTBT, p. 296, and Note 1, p. F-18

If the offering completes on the proposed terms, Anthropic would issue a Class F share immediately before closing. The Founder LLC would hold that share and have a majority on matters put to shareholders generally, including the election of directors other than those elected through Class T. A board majority and a majority of general shareholder votes are different powers. The founders’ agreement is described as intended in the filing. An executed agreement was unavailable in the records reviewed. Final offering terms, issuance and actual operation remain Unresolved. Anthropic Form S-1/A, Founder LLC Agreement and description of capital stock

The proposed trust agreement and charter give some coalitions a route to change the arrangements without trustee consent. Before Class F sunsets, those routes require separate approvals from the Class F holder and the Class E and Class A voters, with turnout and approval thresholds. Under the proposed trust agreement, the trust would end if it ceased to hold any Class T shares. The filings establish the rights on paper and leave their effect in a sustained board dispute open. Proposed trust agreement, §§ 4.3–4.5 · Proposed charter, §§ 5.1(dd), 5.1(jj) and 5.5

The $900.9 billion estimate includes specified purchase commitments and leases, some of which have yet to start. The prospectus identifies roughly $186.2 billion subject to exercisable termination rights or contingent conditions. It expects roughly $70.7 billion of the disclosed commitments to be paid in the 12 months after 30 September 2026. These are commitments with different timing and conditions. Cash paid, completed data centres and supplier revenue each require separate evidence. Anthropic Form S-1/A, contractual obligations

The counterparties can play several roles at once. Google supplies computing capacity, has invested in Anthropic and competes in AI. Anthropic disclosed at least about $121.3 billion under its Google hosting agreement through July 2033, subject to the terms and adjustments described in the filing. The filing describes the economic relationships in both directions. Google’s current ownership percentage and the path of particular invested dollars into supplier invoices remain Unresolved. Anthropic Form S-1/A, Google agreements and private placement

In a proposed $26 billion Class N private placement at the assumed $74 IPO price, the filing lists about $10 billion each from Nvidia and Google and $1 billion from Microsoft, among other participants. The placement requires completion of the public offering and other conditions, including shareholder approval. The IPO can proceed without it. The filing’s separate figure of up to $303 million concerns nonbinding interest from certain directors and related entities in public Class A shares. Neither disclosure establishes what has been paid. Anthropic S-1/A No. 2, 10 November 2026, pp. 20–22 and 348

Anthropic says demand has regularly outstripped its available computing power. Booking capacity could help it serve customers, while exposing it if demand changes faster than commitments can be adjusted. Anthropic Form S-1/A, risk factors and infrastructure strategy

Anthropic estimates that 5–7% of its revenue in the nine months to 30 September 2026 came from customers it later identified as violating its usage policy or commercial terms. It recognised revenue when it provided services, before restricting or ending access. The estimate depends on detection and review and may change. It measures revenue from those customers. The share attributable to violating activity and the full extent of abuse remain unknown. Anthropic 10 November S-1/A, MD&A p. 151

The Pentagon sought access for all lawful uses. Anthropic says it accepted many defence uses and held two limits, on mass domestic surveillance and fully autonomous weapons. Dario Amodei publicly refused the broader terms on 26 February. The department’s position, recorded in the litigation, was that disagreements over restrictions on use could create national-security risks. Amodei’s 26 February statement · D.C. Circuit opinion, 25 September

On 27 August, a California federal district court granted Anthropic partial summary judgment on specified claims involving the federal directive and a defence procurement action under 10 U.S.C. § 3252. It found unlawful retaliation in the actions before it and ruled for the government on other claims. In September, a divided D.C. Circuit panel denied Anthropic’s challenges to a separate exclusion under 41 U.S.C. § 4713. The majority found the record supported a national-security risk for that exclusion and rejected the asserted causal connection to the company’s protected advocacy. Both decisions are Verified for the distinct actions they reviewed. The prospectus reports notices of appeal of the California orders and a rehearing request in Washington. The current dispositions and effects on future contracts remain Unresolved. California order, 27 August, pp. 58–59 · D.C. Circuit opinion · Anthropic Form S-1/A, legal proceedings

Whether the Pentagon used Claude for mass domestic surveillance of Americans remains unresolved. In a separate September report, Anthropic says Claude helped design a mass interception platform in Mali. The company says an account ban and new detection measures disrupted design work and the platform ran locally with another model. The reported deployment remains independently unverified. Anthropic September 2026 threat report, GTG-50027

A separate episode shows how quickly access can change. Anthropic said it shut off Fable 5 and Mythos 5 worldwide on 12 June after a US directive restricting foreign nationals’ access. The company said it could not reliably verify nationality in real time. It later said the restrictions had lifted and that general Fable access would resume from 1 July, with Mythos still limited. Those posts verify the company’s account. The original directive, its legal basis and any connection to the Pentagon dispute remain Unresolved. Anthropic 12 June statement · Anthropic redeployment statement

Anthropic’s own rules also changed. It rewrote its voluntary safety policy in February, arguing that unilateral pauses could leave rivals with weaker protections setting the pace. The July version still commits to delay development or deployment under specified conditions. In August the company reported pausing some external and internal evaluations and higher-risk pre-release training. These pauses concerned specific activities. Their effect on broader release schedules remains unresolved. Anthropic’s February policy explanation · Responsible Scaling Policy v3.4, 8 July 2026, Appendix A · Anthropic’s pause update, 31 August 2026

METR published two reports from one engagement on 5 November. It estimated that scores on Anthropic’s internal capability benchmark rose since October 2025 at about 1.9 times the earlier rate. How much of the acceleration came from AI assistance remained uncertain. METR broadly agreed that July models posed low direct catastrophic risk through automated AI research, but found the public risk report too certain that an early warning threshold had stayed unmet. METR used internal datasets and interviews without direct access to Anthropic’s systems in that engagement. Anthropic controlled redactions and cancelled the separate check on how much computing capacity it allocated to safety versus capabilities. A 15 October statement describes wider access for six METR staff under a separate arrangement. The extent and results of that access remain unverified in the records reviewed. METR October assessment, 5 November 2026 · METR review of the August risk report, 5 November 2026 · Anthropic evaluator update, 15 October 2026

SecureBio’s separate July review broadly agreed with Anthropic’s assessment of low chemical and biological risk for safeguarded Opus 4.6, while flagging weaknesses in the safeguards. Fable 5, Mythos 5 and Opus 5 fell outside its scope. SecureBio July 2026 review, §§ 1 and 3–4

Anthropic announced two $20 million donations to Public First Action. It says it restricted their use to education and policy work and barred use to influence federal, state or local candidate elections. The grant agreements and recipient spending remain unverified. Anthropic statement, 12 February 2026 · Anthropic statement, 21 July 2026

The underlying long-form investigation was first drafted with an Anthropic model. A later review using a model from another provider checked selected financial, governance and court claims against primary records. The review separates the court actions, attributes issuer claims and treats both METR reports as one engagement. The complete current appeal dockets, executed founders’ agreement and original June directive were unavailable in the reviewed records.

The filing gives no figure for revenue lost because of the Pentagon dispute. Final offering documents, an executed founders’ agreement and a record of a disputed release would test how the trust’s formal powers work. Court and delivery records could clarify legal and financial exposure.