Capacity and cost model for a state-wide real-time video analytics estate. Seven detections share one detector pass; the model solves streams-per-GPU against decode, SM and VRAM limits, decides per site whether inference belongs at the edge or in the region, and sizes the event, media and data tiers around it.
Model basis. Decode capacity is megapixels/second across the GPU's NVDEC engines. Detector capacity is TensorRT throughput at 640 × 640 scaled by (640/res)^1.85 and by precision; secondary heads scale from an L4 baseline by a per-GPU compute index. VRAM is surfaces × W × H × 1.5 B (NV12) per stream plus engine and batch buffers. Streams per GPU is the minimum of the decode, SM and VRAM limits at the chosen ceiling. Site RTTs are typical terrestrial figures to Mumbai, not measured. ILL tariffs, edge hardware costs and the Maharashtra outline are approximate and editable — the map is schematic, not survey-grade. Cloud rates are mid-2026 list prices scaled by region; G4 / RTX PRO 6000 is estimated. A design envelope, not a quote.