NVIDIA Jetson: The Complete Guide

Unlike NVIDIA’s data-center GPUs, Jetson modules and developer kits carry NVIDIA-published retail prices, sold directly through NVIDIA’s own marketplace store 234.

Both families add specialised hardware beyond the GPU. A DLA (Deep Learning Accelerator) is a fixed-function engine that offloads convolutional-network inference from the GPU; AGX Orin carries two second-generation DLAs, and the Xavier generation carried two first-generation DLAs 10. A PVA (Programmable Vision Accelerator) handles classical computer-vision tasks like stereo depth and optical flow alongside the GPU and DLA; AGX Orin carries one second-gen PVA, and AGX Thor carries a third-generation PVA 1013. Thor additionally adds fifth-generation Tensor Cores and Multi-Instance GPU (MIG) support, both firsts for the Jetson line 13.

Orin figures per NVIDIA’s Jetson Orin page 10; Thor figures per NVIDIA’s Jetson Thor page and technical blog 1413; prices per NVIDIA’s own marketplace store, checked September 2026 234. An entry-level Thor T4000 module also ships (1,200 FP4 TFLOPS, 64GB memory); NVIDIA’s marketplace did not list a standalone T4000 dev-kit price at the time of this check.

NVIDIA’s Product Change Notice PCN206280, dated 2 March 2020, is titled “Jetson Family: Foxconn Taoyuan Facility as Second-Source Contract Manufacturer” and is indexed on NVIDIA’s own developer site 15. NVIDIA does not publish a single factory address for current Orin or Thor modules beyond this notice.

Earlier generations had their own named adopters: Lockheed Martin and Musashi Seimitsu for AGX Xavier and Xavier NX 9. For Thor, NVIDIA additionally names 1X, John Deere, OpenAI and Physical Intelligence as evaluating the platform, and Stanford, Carnegie Mellon (AirLab) and the University of Zurich as research adopters 16. NVIDIA states “over 2 million developers” use its robotics stack in one release 12 and “over 3 million” in another 17 — both official, but inconsistent between releases, so this page reports both rather than picking one.

Both flagship kits have risen sharply from their launch prices: Orin Nano Super launched at $249 in December 2024 11 and now lists at $399, a 60% increase on the same hardware; AGX Thor launched at $3,499 in August 2025 12 and now lists at $5,499, a 57% increase. NVIDIA’s own storefront shows both flagship kits as “out of stock” even at the higher prices as of this check. No NVIDIA statement was found explaining the increases; any explanation (tariffs, component costs) reported elsewhere is trade-press analysis, not an NVIDIA claim, and is not repeated here.

NVIDIA’s Jetson benchmarks page lists MLPerf-style results for AGX Orin (ResNet: 6,424 samples/sec; BERT: 554 samples/sec) and Orin NX (ResNet: 2,641 samples/sec), plus generative-AI throughput for AGX Thor (Qwen3 30B-A3B: 61 tok/s single-stream, 226 tok/s at 8 concurrent streams; DeepSeek R1 7B: 41.3 tok/s single-stream, 305 tok/s concurrent) 18. These are NVIDIA’s own published figures, not an independently audited MLPerf submission.

NVIDIA’s own blog confirms two mid-range Thor modules, T3000 (865 FP4 TFLOPS, 32GB memory) and T2000 (400 FP4 TFLOPS, 16GB memory), targeted for Q1 2027, with T3000 software emulation support arriving in July 2026 19. Separately, NVIDIA has announced Jetson Orin Nano 2, targeted for the first half of 2027, claiming 78 TOPS and roughly 2x the inference performance of Orin Nano Super at 40% lower power, with early named partners Cognex, Doosan Bobcat, Matic Robots and Wing 17.

NVIDIA’s own December 2024 announcement states the Jetson Orin Nano Super Developer Kit uses the identical Orin Nano 8GB module already in customers’ hands, unlocked from 20 dense/40 sparse TOPS to 33 dense/67 sparse TOPS, with memory bandwidth raised from 65 GB/s to 102 GB/s and CPU clocks from 1.5GHz to 1.7GHz — entirely through a new power mode and a firmware update, while simultaneously cutting the price from $499 to $249 11. That same kit, per NVIDIA’s own current storefront, now lists at $399 — a 60% price rise since launch on hardware that has not changed at all 2.