Anthropic’s S-1 makes clear it’s on the hook for billions in compute. Its customers can walk at any time

Anthropic, a rapidly growing AI company, has revealed in its IPO prospectus that its business model is highly dependent on outside partners for the computing power needed to run its services. The company's revenue has surged 1,300% to $16.28 billion in the first half of 2026, with 80% coming from enterprise customers who pay based on consumption rather than fixed subscriptions.

This growth reflects the increasing demand for AI services, as businesses utilize advanced large language models like Anthropic's Claude products for various tasks, including coding and accounting.

To meet this demand, Anthropic has signed significant infrastructure agreements with major cloud providers such as Amazon, Microsoft, Google, and others. However, the company faces potential challenges if customer usage patterns change or if it has to reduce token prices, which could negatively impact its revenue. Anthropic has committed at least $475 billion in future cloud compute purchases with its partners, including Amazon, Alphabet (Google's parent company), SpaceXAI, Microsoft, and Nscale, with additional deals with Lambda. These commitments amount to as much as $750 billion when considering deals for computing chips.

While Anthropic's revenue growth has helped it keep pace with compute costs, the mismatch between its revenue generation and expected spending creates a balancing act for the company in the coming years. The company aims to grow its customer base and expand revenue to generate cash flows to satisfy its commitments, but it acknowledges that revenue growth may decline and fluctuate over time.

Anthropic has increasingly moved enterprise customers into non-cancelable, multi-year contracts with fixed minimum spends in exchange for discounted token rates, but 75% of this revenue will be recognized within the next year, while its data center spending commitments extend beyond that timeframe.

In addition, Anthropic has underestimated its compute requirements in the past, potentially resulting in periods of excess or underutilized capacity and commitments that do not produce immediate revenue. The company expects to face challenges in aligning compute availability with demand, as customers typically contract for compute 12 to 18 months in advance.