Jetson Thor: NVIDIA's Supercomputer Module Powering Physical AI and Next‑Gen Robots
A supercomputer for physical AI
Last week NVIDIA revealed Jetson Thor, a new robotics platform that combines the Jetson AGX Thor Developer Kit and the Jetson T5000 production module. Designed as a compact supercomputer for physical AI, Jetson Thor enables generative reasoning and multimodal sensor processing to run inference and decision making directly at the edge.
Compute performance and efficiency
The platform delivers up to 2,070 FP4 teraflops (TFLOPS) of AI compute through a Blackwell‑based GPU, a roughly 7.5× improvement over the previous Jetson Orin series. That peak performance is available inside a 130 W power envelope, with configurable operation down to 40 W for energy‑sensitive deployments. NVIDIA reports about 3.5× better efficiency compared with Orin, making high‑throughput robotics workloads more practical on mobile and embedded systems.
Architecture and scalable GPU resources
Jetson Thor integrates a 2,560‑core Blackwell GPU with 96 fifth‑generation Tensor Cores and supports Multi‑Instance GPU (MIG). MIG enables partitioning the GPU into multiple independent instances so teams can run parallel workloads or assign dedicated GPU slices to separate tasks. The CPU side is handled by a 14‑core Arm Neoverse‑V3AE processor, featuring 1 MB L2 per core and a 16 MB shared L3 cache for balanced control and planning tasks.
Memory, I/O and real‑time sensor fusion
Memory is provisioned as 128 GB LPDDR5X on a 256‑bit bus, delivering about 273 GB/s of bandwidth. Storage and connectivity options include a 1 TB NVMe M.2 slot, HDMI and DisplayPort outputs, multiple USB ports, Gigabit Ethernet, CAN headers and QSFP28 support for up to four 25 GbE lanes. Those I/O capabilities are crucial for high‑bandwidth, low‑latency multimodal sensor fusion in robotics.
Software stack for robotics and real‑world AI
Jetson Thor ships with NVIDIA’s specialized software tailored to physical AI:
- Isaac (GR00T) for generative reasoning and humanoid control
- Metropolis for vision AI
- Holoscan for real‑time, low‑latency sensor processing and sensor‑over‑Ethernet (Holoscan Sensor Bridge)
This stack allows a single system‑on‑module to execute multimodal workflows combining vision, language and actuation without offloading to external chips.
What is physical AI and why it matters
Physical AI unites perception, reasoning and action planning so robots can simulate possible sequences, anticipate consequences and generate both high‑level plans and low‑level motion policies. With on‑device multimodal inference, robots gain adaptability closer to human reasoning instead of acting as narrowly programmed automata.
Practical applications
Jetson Thor is aimed at scenarios where robust real‑time decision making matters: navigation in unpredictable environments, complex object manipulation, instruction following without repeated retraining, and operations across manufacturing, logistics, healthcare and agriculture.
Availability and pricing
The Jetson AGX Thor Developer Kit is now generally available at $3,499. Jetson T5000 production modules are offered through NVIDIA partners with volume pricing around $2,999 per unit for 1,000‑unit orders. Pre‑orders and partner channels indicate growing availability for both research and commercial robotics projects.
Jetson Thor brings server‑grade, multimodal inference and generative reasoning into a single, power‑bounded module. Its combination of high FP4 TFLOPS, efficient design, broad I/O and a focused software stack positions it as a foundational platform for the next wave of real‑world AI agents.