Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.471]

What It Covers

Anthropic CFO Krishna Rao explains how the company manages compute as its core strategic resource, covering the allocation framework across training, internal use, and customer demand, the economics behind frontier model pricing, the $9B to $30B ARR growth in one quarter, and why enterprise returns to frontier intelligence continue accelerating rather than plateauing.

Key Questions Answered

• •Compute Allocation Framework: Anthropic runs daily meetings to allocate compute across three buckets: model training, internal acceleration, and customer inference. A non-negotiable floor protects model development spend even when customer demand spikes. The company uses three chip platforms—AWS Trainium, Google TPUs, and NVIDIA GPUs—fungibly, switching workloads between morning inference runs and evening training jobs on the same hardware to maximize utilization.
• •Cone of Uncertainty Planning: Rather than point-estimate forecasting, Anthropic models a range of exponential growth scenarios over 12–24 months and builds compute procurement flexibility into contracts. Small differences in weekly growth rates compound into vastly different compute requirements. Contracts include flexibility clauses, and multi-platform chip capability allows rapid reallocation when actual demand diverges from projections at either end of the range.
• •Returns to Frontier Intelligence: Enterprise customers consistently upgrade to the newest model immediately upon release because frontier models unlock entirely new use cases, not just incremental improvements. Anthropic's net dollar retention rate exceeds 500% annualized. Revenue grew from $9B to $30B run rate in a single quarter, driven by model capability leaps enabling longer-horizon agentic tasks, faster completion times, and expanded enterprise workflows beyond coding.
• •Jevons Paradox in Model Pricing: When Anthropic reduced Opus-class pricing, consumption increased far beyond what price elasticity alone would predict. Customers had been forcing Opus-level problems into cheaper Sonnet models due to cost. Lowering the price unlocked latent demand and allowed customers to build Opus into production workflows. Pricing stability across model generations matters: when Opus 4.6 launched with no price change, customers slotted it directly into existing integrations.
• •Finance Team AI Deployment: Anthropic's finance team uses Claude to produce statutory financial statements across all legal entities, with human review as a final check. An internal tool called AntStats, combined with a library of over 70 finance-specific Claude skills, generates monthly financial reviews that are 90–95% complete before human review. Weekly revenue and compute utilization reports that previously took hours now take 30 minutes, shifting team time toward strategic analysis.
• •Platform vs. Application Strategy: Anthropic builds primarily horizontal platform infrastructure—prompt caching, agents SDK, managed agents, computer use—and enters vertical application layers only when it can demonstrate capabilities ahead of the market or show the ecosystem what is possible. Claude Code was built because the company had visibility into upcoming model capabilities that would make autonomous coding viable. Vertical products are launched in partnership with ecosystem players rather than in competition with them.