Anthropic expects to spend at least $518 billion on AI infrastructure over roughly a decade, according to a confidential IPO prospectus reviewed by Reuters.

The commitments cover cloud computing, specialised AI hardware, and infrastructure agreements with companies including Amazon, Google, Microsoft, and Broadcom. Reuters reported that around 80% of the commitments are non-cancellable or require payment regardless of usage.

According to Reuters, Anthropic’s commitments include $111.1 billion with Google, $110 billion with Amazon, and $31.4 billion with Microsoft. The filing also lists about $161.2 billion in largely non-cancellable equipment lease obligations involving Broadcom. Reuters separately reported that Anthropic spent more than $7 billion on compute and infrastructure in 2025, while its operating loss exceeded $8 billion.

Anthropic secures gigawatts of compute capacity

Amazon remains Anthropic’s primary cloud provider and training partner. The companies expanded their agreement in April this year, with Anthropic committing more than $100 billion over 10 years to AWS technologies and securing up to 5GW of additional computing capacity. Anthropic said nearly 1GW of additional Trainium2 and Trainium3 capacity would be available by the end of 2026.

Anthropic said it was already using more than one million Trainium2 chips to train and serve Claude when the agreement was announced. AWS and Anthropic also developed Project Rainier, a Trainium-based compute cluster that Anthropic uses to train and run its models. The expanded AWS agreement covers current and future generations of Trainium processors.

In October 2025, Anthropic said it planned to expand its use of Google Cloud technologies to as many as one million TPUs. The agreement was worth tens of billions of dollars and was expected to provide well over 1GW of capacity during 2026. In April this year, Anthropic expanded that relationship through an agreement with Google and Broadcom for next-generation TPU capacity expected to begin coming online in 2027. Anthropic later described that agreement as covering 5GW of capacity.

Under a separate agreement announced in November 2025, Anthropic committed to purchase $30 billion of Azure compute capacity. It also agreed to contract up to 1GW of additional capacity using Nvidia Grace Blackwell and Vera Rubin systems.

Building that capacity requires expenditure before customers begin using it. Amazon CEO Andy Jassy said AWS typically spends on land, power, buildings, chips, servers, and networking equipment six to 24 months before it can begin billing customers for the resulting capacity.

The agreements also give Anthropic access to different processor architectures. The company uses AWS Trainium, Google TPUs, and Nvidia GPUs for training and inference workloads. Anthropic says this allows it to match workloads with the hardware best suited to them.

Amazon said its custom chip business, including Trainium, Graviton, and Nitro, exceeded a $20 billion annual revenue run rate in early 2026. Jassy said Amazon expects Trainium, at scale, to reduce annual capital expenditure by tens of billions of dollars and provide an operating margin advantage of several percentage points compared with relying on third-party chips for inference.

Anthropic also works with AWS subsidiary Annapurna Labs on optimising future generations of Trainium for Claude workloads. Its separate partnership with Nvidia includes joint design and engineering work to optimise Anthropic models for Nvidia hardware and future Nvidia architectures for Anthropic workloads.

Cloud suppliers invest in Anthropic

Amazon has invested billions of dollars in Anthropic alongside its role as the company’s primary cloud provider and training partner. Under the partnership announced in November 2025, Nvidia committed to invest up to $10 billion in Anthropic and Microsoft up to $5 billion, alongside Anthropic’s $30 billion Azure compute commitment.

The cloud providers also distribute Claude to enterprise customers through Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Reuters reported that 47% of Anthropic’s 2025 sales passed through Amazon and Google’s cloud platforms. Anthropic identified its dependence on large technology partners as a business risk in its confidential filing.

Microsoft’s financial disclosures also show the cost of expanding AI infrastructure. Microsoft reported $59.3 billion in Microsoft Cloud revenue in its fiscal fourth quarter of 2026, up 27% year over year. Commercial remaining performance obligations – revenue contracted but not yet recognised – reached $678 billion.

The company said Microsoft Cloud gross margin percentage fell year over year partly because of continued investment in AI infrastructure and increased product usage. Nearly 90% of full-year Microsoft Cloud revenue came from customers outside frontier model companies.

Anthropic’s own revenue grew alongside its infrastructure commitments. The company reported an annualised revenue run rate of more than $30 billion in April this year, up from around $9 billion at the end of 2025. More than 1,000 business customers were each spending more than $1 million on an annualised basis at the time. The company subsequently said its annualised revenue run rate crossed $47 billion in May.

Anthropic cited capacity constraints when announcing its expanded AWS agreement in April. It said rapid consumer growth had affected reliability and performance for free, Pro, Max, and Team users, particularly during peak hours. Anthropic said additional AWS capacity would begin arriving within months.

Anthropic has also sought capacity outside the three largest cloud providers. In May this year, it announced an agreement with SpaceX giving it access to all compute capacity at the Colossus 1 data centre, with more than 300MW of capacity and over 220,000 Nvidia GPUs expected to become available that month.

(Photo by Brecht Corbeel)