Anthropic $895B Compute Commitments
Anthropic
This filing turns Anthropic into a capital allocation story as much as an AI product story. The roughly $895.4B in purchase commitments and lease agreements means Anthropic is locking in the physical capacity needed to train bigger models and run longer agent workloads years before the revenue is earned. These are contractual spending obligations, not historical expenses, and they sit far above the company’s $146.6B of liquidity, so future execution depends on keeping utilization high and revenue compounding into that reserved compute base.1 2 3 4
- The commitment stack is mostly compute. Anthropic disclosed that substantially all of the total is tied to compute and infrastructure, with about $447.5B expected before 2031, $69.8B due in the next 12 months after September 30, 2026, and most of the rest running from 2031 to 2045. That schedule shows a long dated bet on sustained frontier model demand.1 2
- Named counterparties in the filing and related analysis point to a supply chain built across multiple providers instead of one cloud. Reported commitments include at least $114.5B with Google, at least $110B with Amazon, at least $44.9B with Microsoft, about $41.3B with Lambda, up to $44.6B with Nscale, up to $120.5B with SpaceXAI, plus $161.2B of Broadcom linked equipment lease rent.2 5 6
- The operating logic is simple, if Anthropic can keep selling high intensity products like Claude Code, Cowork, API usage, and long running agents, reserved capacity becomes an advantage because customers get faster access to scarce GPUs and TPUs. If demand slows, those same commitments become a utilization problem because unused reserved infrastructure still carries fixed payment obligations.3 4 6
The next phase is a race to turn reserved capacity into dense, daily usage across coding, enterprise workflows, and agent execution. The labs that win will not just have the best models, they will have the best ability to keep giant compute estates full of paying work, which is why Anthropic’s future increasingly looks like a blend of software company, cloud buyer, and industrial planner.3 4 6