Anthropic IPO Filing, September 2026: $518 Billion in AI Infrastructure Commitments, Largely Non-Cancelable

Anthropic's confidential S-1 reveals $518 billion in AI infrastructure commitments — the bulk structured as non-cancelable contracts — positioning the safety-focused lab as the anchor tenant of a capital buildout spanning multiple continents and a growing web of counterparties.

When Anthropic filed its confidential S-1 with the SEC in September 2026, the headline figure was impossible to ignore: $518 billion in AI infrastructure commitments, assembled in roughly eleven months and predominantly structured as non-cancelable contracts. That single disclosure reframes Anthropic not merely as a frontier AI laboratory but as the nucleus of an unprecedented capital stack. The speed of accumulation is itself notable — one outlet reported that the Akamai Technologies agreement pushed Anthropic's cumulative compute commitments past $517 billion within eleven months.

The commitments span a widening web of counterparties. Akamai announced a seven-year, $11.6 billion compute agreement, with extension options that could lift the total to approximately $20 billion — a deal that routes Anthropic compute through non-hyperscaler infrastructure at a scale that Akamai describes as supporting growing demand. Reports from multiple outlets place Anthropic in negotiations to lease one gigawatt of data center capacity from Apollo Global Management, a transaction that would give the company direct access to power and capacity at a scale associated with hyperscalers. TeraWulf, migrating from crypto mining, signed a 20-year capacity lease; Riot Platforms disclosed a $9.1 billion agreement; Lambda reportedly struck a $35 billion deal. In Australia, Anthropic anchored a $32 billion Queensland data center project and agreed to lease the first phase of the $10.4 billion Zerra DC facility — one of the largest single AI infrastructure commitments in the region, though one that has drawn attention for its reliance on coal-fired power and its interaction with federal renewable energy regulations. One outlet, citing the confidential filing, reported that a computing-capacity arrangement with SpaceX could reach up to $84.5 billion through 2029, roughly double a previously disclosed figure; that claim rests on a single source and warrants caution until corroborated.

What distinguishes this buildout from a standard hyperscaler capex cycle is the contractual architecture. Reuters and The Globe and Mail, both citing the IPO filing directly, reported that the bulk of the $518 billion is non-cancelable — meaning Anthropic's ability to exit these obligations is structurally constrained regardless of how demand evolves. That design reflects a deliberate strategic bet: by locking in capacity across geographies and providers simultaneously, Anthropic secures supply while binding partners to its growth trajectory. The Nscale IPO, testing investor appetite for an AI cloud provider that lists Anthropic as a key customer, illustrates how this model distributes financial exposure outward — infrastructure partners absorb capital risk in exchange for long-dated, contracted revenue. Broadcom, reportedly in line to become Anthropic's largest custom chip customer, would gain a concentrated inference-semiconductor relationship reflecting the AI lab's significant commitment to custom inference silicon.

The IPO itself remains the most consequential near-term event. Multiple outlets in September 2026 placed a target valuation near $2 trillion — which, if achieved, would represent the largest public offering on record. A subsequent report indicated Anthropic had postponed the listing from October to November 2026 to allow third-quarter results to be incorporated, suggesting management is deliberately managing investor expectations rather than rushing to market. The filing's own language around existential risk — reported directly from the confidential document — adds an unusual dimension: a company whose business model is predicated on frontier AI development, publicly acknowledging that the technology it is commercializing poses risks of that magnitude. The company's policy chief was separately reported to have argued that AI safety cannot rely on an honor code and requires enforceable mechanisms, a position that sits in some tension with the scale of the infrastructure race the same filing simultaneously describes.

For investors and infrastructure counterparties alike, the opportunities and risks are structurally asymmetric. On the opportunity side, non-cancelable contracts with Anthropic represent visible, long-duration revenue for data center operators, power-adjacent plays, and non-hyperscaler compute providers that have long struggled to attract anchor tenants at this scale — the Riot Platforms and TeraWulf agreements are the clearest examples of that dynamic playing out in practice. The Australia commitments open a new inference geography, with $32 billion in aggregate that spans both data center development and a meaningful sovereign policy dimension. On the risk side, the non-cancelable structure that insulates infrastructure partners from demand softness is the precise mirror image of Anthropic's own exposure: if model usage, enterprise revenue, or the IPO valuation falls short, the company carries fixed obligations it cannot unwind by renegotiating. Three signals are worth watching closely: whether the November IPO proceeds at or near the reported $2 trillion target, a figure that embeds extraordinary forward revenue assumptions; whether the Apollo gigawatt lease closes and on what terms, which would clarify how far Anthropic is prepared to go in building direct infrastructure outside the hyperscaler ecosystem; and how regulators respond to the antitrust proceedings filed in September 2026 alleging that Anthropic, OpenAI, Google, and SpaceX's AI division agreed to constrain the pace of AI development — a claim that, if sustained, would challenge the strategic rationale underlying the entire $518 billion commitment structure.