NVIDIA’s top Jetson developer kit. Blackwell GPU, 128GB LPDDR5X, 273 GB/s, up to 2070 FP4 TFLOPS at 40–130 W. Built for physical AI and robotics; also runs large local models.
Street price (USD)
$3,499–$5,499
Memory
128GB LPDDR5X
Bandwidth
273 GB/s
Chip
Jetson T5000 (Blackwell)
Catch
Price jumped; robotics-first
Clip
Same 128GB class as Spark, but this stack is for robotics — the tiny module goes in the robot; power/battery matter.
sentdex. 20:21–21:20.
What it is
The Jetson AGX Thor Developer Kit with the T5000 module: 2560-core Blackwell GPU, 96 fifth-gen Tensor Cores, Multi-Instance GPU, 14-core Arm Neoverse-V3AE CPU, 128 GB 256-bit LPDDR5X at 273 GB/s, 1 TB NVMe, QSFP28 for 4×25 GbE, HDMI/DP, USB, camera and sensor headers. Power envelope 40–130 W. Ships as a complete kit with carrier, cooling, and SSD. Production modules (T5000 / T4000) are sold separately for volume robotics builds.
Who it is for
Robotics and physical-AI teams who need CUDA, Isaac, Holoscan, and high-bandwidth sensor I/O in one box. Also useful for anyone who wants 128 GB of unified memory on a Jetson stack and will live with Arm + JetPack. Wrong if you want a silent desk cube, repairable consumer hardware, or the lowest price per token for pure language models.
What it actually runs
Comfortable: 20B–35B dense and MoE models at usable token rates; vision-language and VLA models for robotics.
Possible: 70B-class and larger sparse MoE (reports of GPT-OSS-120B and similar fitting in the 128 GB pool at ~10 tok/s range under tuned llama.cpp / SGLang / TensorRT setups). Capacity is the win; bandwidth still caps decode.
Not the win: cheapest tokens, Windows desktop convenience, or expecting discrete-GPU desktop bandwidth. Ignore pure marketing TFLOPS slides.
Standout details
Price snapshot early Sep 2026: launch list was $3,499; NVIDIA raised the developer kit to $5,499 in July 2026. Resellers and partners have tracked in the $3,499–$5,499 band depending on stock and region. Module pricing also rose.
128 GB LPDDR5X soldered, 273 GB/s. Same order of bandwidth as other high-end unified-memory boxes, higher capacity than the AGX Orin 64GB kit.
Native CUDA / JetPack / TensorRT / Isaac path. FP4 and Transformer Engine matter for quantized generative workloads.
Networking and camera I/O are the robotics differentiator: QSFP28 4×25 GbE, sensor bridge support, PVA.
Power and thermals are real: 40–130 W configurable. Not a silent living-room appliance.
Honest verdict
The high-end Jetson for physical AI. Buy it if you need Blackwell + 128 GB + sensor/network I/O under JetPack. Same memory class as DGX Spark / GX10 but Arm, edge form factor, and robotics software. The 64GB workhorse on the older Ampere stack is AGX Orin. Price has moved up; treat current street numbers as snapshots.
Prices and stock move. Affiliate links stay off until this directory has real traffic.