Agentic inference capacity grew 4.6× while compute rose 2.6×
The non-obvious takeaway is that inference capacity per dollar of compute is rising sharply, so capacity forecasts based on FLOP/s alone are systematically too low. Builders planning agent deployments and enterprises budgeting inference costs should track indices like the IPI rather than chip shipment counts. With Blackwell behind 74% of the gains, Nvidia's per-GPU efficiency is the main lever shaping serving economics globally.