How AI Companies Spending Billions Solve the ‘Buy Too Much or Too Little’ Compute Dilemma

Why Compute Is the Lifeline of AI Companies

Anthropic CFO Krishna Rao describes compute as the “lifeblood” and “canvas” of the business, arguing that insufficient compute prevents serving customers while excess idle capacity can bankrupt a firm.

The “Buy‑Too‑Much or Too‑Little” Dilemma

Rao introduces the “cone of uncertainty”, showing that small month‑to‑month demand fluctuations compound over 1‑2 years into large forecast gaps, rendering linear financial planning ineffective.

From Linear Planning to Scenario‑Based Procurement

Anthropic abandons single‑point forecasts and instead examines a range of outcomes inside the cone, working backwards from target results to set multi‑year compute purchase plans.

Dynamic Allocation Between Customers and R&D

The company allocates a protected compute floor to model development, even if it temporarily hurts customer service, because long‑term model quality is viewed as the primary source of return.

Cross‑Chip Heterogeneous Strategy

Anthropic spreads its workload across three major chip families—GPU, TPU, and its own Trainium—using flexible scheduling to hedge supply risk and to keep the most demanding workloads on the most efficient hardware.

Financial Mechanisms for Long‑Term Supply

To mitigate the extreme scarcity of AI‑grade silicon, Anthropic signs forward compute contracts worth over a hundred billion dollars, locking in capacity for several years.

Jevons Paradox and ROI Measurement

Rao notes that lowering compute prices can paradoxically increase revenue (Jevons paradox) and stresses that ROI should be measured against the ability to sustain a competitive moat rather than simple cost per FLOP.