Anthropic’s IPO prospectus has laid bare the scale of its computing ambitions, revealing commitments of between $542 billion and $567 billion to future compute and infrastructure as it competes for capacity with the world’s biggest technology companies.
The company said access to specialised computing was essential to developing, training and deploying its frontier models. But it warned that the market was “intensely competitive”, with rivals including some of the most heavily capitalised technology groups in the world.
Anthropic signed agreements worth $417 billion with various providers on 31 December 2025. Those arrangements begin between 2026 and 2027 and run for periods of between one and 20 years. A further $125 billion to $150 billion of capacity is due to begin between 2028 and 2031.
Compute was Anthropic’s biggest operating expense in the second quarter of 2026, costing $5.6 billion. The company generated $11.6 billion in revenue over the same period, giving it just over $2 in revenue for every dollar spent on compute.
Anthropic’s direct data-centre strategy
The company is sourcing capacity mainly from Amazon, Google and Microsoft, while agreements with Elon Musk-owned SpaceX provide access to as much as $120.5 billion in compute through 2029. The capacity is powered by Nvidia graphics processors owned by SpaceX.
Anthropic has also developed what it calls a “proprietary orchestration layer” to decide which workloads should run on different chips and cloud providers. The system covers AWS Trainium, Google’s tensor processing units, Nvidia graphics processors and AMD graphics processors.
Anthropic said the technology had taken “years of significant investment to build”. It is intended to control costs and reduce the company’s reliance on any single supplier.
However, the prospectus acknowledged that its relationships with Amazon, Google and Microsoft were “multifaceted and complex”. The three companies are both suppliers and competitors, and may prioritise their own computing requirements or gain insight into Anthropic’s operations and capacity needs.
All three have also made significant investments in Anthropic and use its models as customers, adding to the overlap between the companies.
To reduce its dependence on those competitors, Anthropic is developing a strategy known as Anthropic Direct. Under the plan, third-party partners will secure sites, power and facilities and run data centres, while Anthropic leases the computing capacity it requires.
The company began entering purchase commitments and lease agreements for the initiative around 30 September 2026. Its five-year rent payments are expected to total $125.2 billion, excluding interest, and Anthropic said Direct would account for an increasing proportion of its computing investments over time.
Broadcom is Anthropic’s largest infrastructure partner in 2026, with up to $42 billion in potential financing. Google follows with $20 billion and Amazon with $15 billion.
The strategy will require substantial spending before the company receives a rapid return on its investment. Anthropic warned that it expected to incur “significant cash outflows” that could harm its liquidity, financial condition and results.
Powering the planned data centres presents a further risk. The company expects “substantial energy demands” and said connections could be delayed by permitting issues, legal challenges and supply-chain disruption.
It also warned of possible shortages of chips and data-centre equipment. If its infrastructure plans suffer delays or cost overruns, or the facilities fail to meet its changing requirements, Anthropic said it might not achieve the expected benefits of its investment, with potential consequences for its business and financial performance.
