Anthropic has committed at least $518 billion over roughly a decade to cloud services, computing capacity and artificial-intelligence infrastructure, according to a confidential initial-public-offering prospectus reviewed by two news organizations. About 80% of that total is non-cancelable or payable regardless of actual use, Reuters reported Tuesday.

The commitments put a concrete price on a technology strategy that increasingly depends on reserving scarce processors, electricity and data-center capacity years before demand can be known. The Financial Times, which said multiple people familiar with the prospectus confirmed its details, also reported the $518 billion figure. Anthropic has not publicly released the document, leaving investors unable to examine the full wording until the filing becomes public.

The scale matters beyond one company. Frontier AI developers once bought much of their computing through comparatively flexible cloud arrangements. Anthropic’s reported obligations show that leading laboratories are moving toward long-term reservations, dedicated clusters and leased equipment, transferring part of the technology race from model design to physical infrastructure.

Most of the obligation is concentrated among four suppliers

The prospectus assigns at least $111.1 billion to Google, $110 billion to Amazon and $31.4 billion to Microsoft under infrastructure agreements lasting seven to 10 years, according to Reuters. Anthropic also disclosed about $161.2 billion in equipment-lease obligations associated with Broadcom. Those four relationships account for more than $413 billion of the reported total.

The contractual structure is as consequential as the headline number. Anthropic said it must pay Google and Amazon if its spending falls short of minimum commitments, while the Microsoft contract is cancelable only after an uncured material breach. Broadcom-related leases are largely non-cancelable except after default. The arrangement may secure supply in a market where advanced accelerators and powered data-center space remain constrained, but it also converts uncertain future demand into fixed costs.

Public agreements support the scale, though not every term

Previously announced transactions substantiate much of the underlying expansion. Amazon said in April that Anthropic had agreed to spend more than $100 billion over 10 years on its technologies and reserve as much as five gigawatts of capacity using current and future Trainium chips. The Amazon agreement also included an immediate $5 billion investment and the possibility of as much as $20 billion more tied to commercial milestones.

Microsoft separately announced that Anthropic would buy $30 billion of Azure capacity and contract for as much as one gigawatt of additional computing power. That partnership also made Claude models available through Microsoft’s cloud and provided for investments of up to $10 billion from Nvidia and $5 billion from Microsoft.

Anthropic said in April that it had also signed for multiple gigawatts of next-generation tensor-processing-unit capacity from Google and Broadcom, with deployment beginning in 2027. The company described the expansion as its largest computing commitment at that time, with most capacity expected to be located in the United States. The public announcements do not disclose all payment floors or cancellation clauses reported in the prospectus, but they independently establish the extraordinary physical scale of the program.

Compute has become both an advantage and a liability

Anthropic’s premise is that future AI demand will be limited more by access to computing than by customer interest. Reserving capacity early can protect the company from shortages, reduce dependence on spot availability and let engineers optimize models across Amazon Trainium, Google tensor-processing units and Nvidia systems. Using several chip architectures also reduces exposure to a single supplier, at least operationally.

Financially, however, diversification does not eliminate concentration. Google, Amazon and Microsoft serve simultaneously as infrastructure vendors, investors, distributors and developers of competing AI products. That web of relationships can broaden Claude’s market access, but it also leaves Anthropic dependent on counterparties whose commercial interests do not always align with its own. Long-term minimum payments further raise the cost of forecasting error if model demand, pricing or computing efficiency changes faster than expected.

Rapid revenue growth has not yet matched spending

The infrastructure commitments sit against financial results that combine unusually fast growth with unusually large losses. Anthropic generated nearly $4.6 billion in 2025 revenue, about 12 times the prior year, while recording an operating loss of roughly $8.06 billion, according to a separate Reuters review. It spent $7.33 billion on computing and infrastructure during the year, more than half of total operating expenses.

The reported $42 billion net loss requires qualification: roughly $34 billion reflected an accounting charge tied mainly to the rising estimated value of financing liabilities that could convert into shares, rather than cash consumed by operations. Even after separating that charge, the operating loss shows that revenue growth had not covered the cost of building and serving the models. The company held about $20.28 billion in cash, cash equivalents and short-term investments at year-end, far below the face value of its multiyear commitments.

The public filing will determine what investors can verify

The $518 billion figure is a contractual obligation spread across many years, not a single capital expenditure or an amount due immediately. Some potential spending, including reported arrangements for additional Nvidia-based capacity, can be canceled with notice. The total therefore should not be treated as a forecast that Anthropic will write one check or consume every reserved unit of computing at once.

What the disclosure does establish is that the competitive barrier around frontier AI is becoming increasingly physical and financial. Anthropic is wagering that locking in vast capacity will preserve its ability to train and operate more capable models, even if that means accepting payment obligations far beyond its current revenue base. When the prospectus becomes public, investors will be able to test the commitments’ timing, termination rights and accounting treatment. Until then, the independently reported totals and the companies’ own partnership announcements show both the scale of the buildout and the risk embedded in securing it.