Infra spend ↓29% YoY

Batched edge inference on Jetson cut cloud GPU hours 38% (4,800 → 2,976/yr); cold storage tiering cut the footprint 35% (180TB → 117TB). Detection accuracy held within 1.5 points.