Goldman Sachs: AI Needs $0.67 per Million Tokens for 15% ROIC

Goldman Sachs turns its $1.73T Phase-2 capex target and 15% ROIC threshold into concrete AI numbers: hyperscalers must clear a blended ~$0.67 per million tokens or ~$4.30 per GPU-hour. DeepSeek-style pricing at $0.15–$0.30 input and $0.60–$1.20 output generates only low-single-digit or negative implied ROIC, so the build-out fails to cover depreciation and running costs. That gap between massive AI spending and collapsing token prices drives the AI industry impact on future infrastructure returns.