← All hardware
A Blackwell-powered supercomputer for humanoid robots and physical AI.
Pros
- Massive 128GB memory runs large multimodal/generative models on-robot
- GPU partitioning (MIG) ideal for safety-critical robotics
- Best software stack for physical AI (Isaac, Holoscan, Omniverse)
Cons
- Very expensive at $3,499 for the dev kit
- High power envelope (40-130W) unsuitable for small devices
- Overkill for simple vision/IoT inference tasks
✓ Where it shines / best for
- Humanoid and advanced robotics needing on-device generative/VLA models
- Physical-AI research and multimodal perception at the edge
- Teams deploying large transformer models outside the data center
✕ Not the best fit for
- Cost-sensitive or hobbyist projects
- Simple single-camera vision tasks (overkill)
- Low-power, thermally constrained battery devices
Features
- ✓ Up to 2,070 FP4 TFLOPS (1,035 FP8 TFLOPS) of AI compute - server-class at the edge
- ✓ NVIDIA Blackwell GPU architecture with Multi-Instance GPU (MIG) support
- ✓ 14-core Arm Neoverse-V3AE CPU
- ✓ 128GB LPDDR5x memory (273 GB/s bandwidth) for large multimodal/generative models
- ✓ Over 7.5x higher AI compute and 3.5x better energy efficiency than Jetson AGX Orin
- ✓ 4x 25GbE networking plus a camera offload engine
- ✓ ~130W power envelope; runs JetPack with CUDA and Isaac for physical-AI/VLA models
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| Jetson AGX Thor Developer Kit | $3,499 | one-time | Generally available August 2025; ships with Jetson T5000 module (part 945-14070-0080-000) |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| power | 40-130W |
| memory | 128GB LPDDR5X |
| networking | 4x 25GbE |
| architecture | NVIDIA Blackwell GPU (2560 cores, 96 Tensor cores) + 14-core Arm Neoverse-V3AE |
| ai_performance | 2070 FP4 TFLOPS |
Sponsored
A full review is being generated for this product and will appear here shortly.