NVIDIA introduced two mainstream Jetson Thor modules and a substantial Metropolis software update on the same day, and together they sketch the near future of the video analytics appliance.
The hardware
The Jetson Thor T3000 pairs a 1536-core Blackwell GPU with an eight-core Arm Neoverse CPU for up to 865 FP4 TFLOPS, with 32GB of LPDDR5X at 273GB/s and 25 GbE networking, at roughly half the size and power of the existing T5000. The entry T2000 delivers up to 400 FP4 TFLOPS with 16GB. Modules are scheduled for Q1 2027 availability, with emulation support starting now through JetPack 7.2.1.
Those numbers matter because this class of module is what sits inside analytics appliances, smart NVRs, and roadside cabinets. Compute that previously demanded a GPU server now fits in a fanless box on a pole.
The software
The Metropolis update spans VSS Blueprint 3.2, DeepStream 9.1, TAO 7, and a synthetic-data Physical AI Data Factory, more than 80 new capabilities that NVIDIA claims cut vision-AI agent development time by up to 6x. VSS Blueprint builds agents that watch live or recorded video and respond to natural-language questions, with alerting, summarization, and search across camera networks. DeepStream 9.1 handles multi-camera tracking and large-scale ingestion from edge to cloud.
The read
The analytics conversation is shifting from “which detections does the box support” to “what can you ask the system.” Integrators who treat video AI as a line item on a camera quote will find the value moving into software and the appliances that run it.