Today, NVIDIA officially launched two new modules based on the Thor architecture—the Jetson T3000 and T2000—designed to meet the demand for compact, highly energy-efficient edge AI computing platforms as general-purpose robots and autonomous systems transition from laboratories to mass production. These new additions further expand NVIDIA’s Jetson product line, offering developers flexible choices ranging from entry-level to high-performance options. The Jetson T3000 features an NVIDIA Blackwell GPU and an eight-core Arm Neoverse CPU, equipped with 32GB of LPDDR5X memory providing bandwidth up to 273GB/s, along with support for 25GbE high-speed network connectivity. It delivers 865 FP4 TFLOPS of AI compute power. With dimensions and power consumption approximately half that of the flagship T5000, it achieves performance comparable to the T5000 in multimodal inference tasks such as large language models, vision-language models, vision-language-action models, and world foundation models. For application scenarios requiring functional safety, the IGX T3000 integrates functional safety capabilities while maintaining equivalent performance and can seamlessly run NVIDIA Halos for Robotics full-stack security system, ensuring robot safety around humans. Targeted at broader edge AI systems, the Jetson T2000 provides 400 FP4 TFLOPS of compute power and 16GB of memory, making it suitable for developing intelligent edge devices like visual AI agents, autonomous mobile robots, and industrial robotic arms. This completes NVIDIA’s scalable edge AI product matrix spanning from 70 TOPS to 2000 TFLOPS, capable of covering nearly any edge AI workload. Beyond hardware updates, NVIDIA also introduced Agent Skills for Jetson, which reduces software stack optimization cycles from weeks to days through automated memory optimization, system configuration, and deployment tasks. This feature has already helped several enterprises significantly reduce memory costs: UBTech and Agile Robots saved up to 15GB of memory through software optimizations, enabling migration from 64GB Jetson AGX Orin modules to 32GB variants; smart retail company SandStar saved 4GB of memory, allowing deployment on 8GB Orin NX modules; and smart transportation firm NoTraffic reduced memory usage by 30% on TX2 NX, freeing space for additional AI functions. Regarding foundational models, NVIDIA brings Cosmos 3 Edge—a lightweight world foundation model with four billion parameters—to the Thor platform. Developers can leverage the open Cosmos framework to complete post-training tailored to specific applications within about one day and achieve real-time visual analysis and device-side robotics strategy reasoning on Jetson Thor. Currently, developers can begin development using the Jetson AGX Thor Developer Kit. The simulation mode for T3000 will be released later this month via JetPack 7.2.1, while T2000 simulation modes will become available in subsequent releases. Both new modules are expected to launch commercially in Q1 2027. Numerous ecosystem partners including ADLINK, Advantech, Antai International, among others, have begun providing Thor-based solutions, and software partners offer emulation and migration support for transitioning customers. As physical AI and embodied intelligence move toward mainstream adoption, the new Thor computing platform offers developers a robust, scalable foundation for deploying intelligent humanoid robots and autonomous systems in the real world.