Quick Answer: Buy the Jetson AGX Thor Developer Kit if you need more than 64 GB of memory (70B-120B models, big vision-language-action models, multi-model robot stacks) or FP4 and the newest JetPack 7 / Isaac software. Buy the Jetson AGX Orin 64GB Developer Kit if your models fit in 64 GB and you want a mature, well-documented platform for $2,000 less. Don’t buy Thor for raw chatbot speed: NVIDIA’s “7.5× AI compute” headline shrinks to about 1.3-1.4× faster text generation, because LLM decoding is limited by memory bandwidth — 273 vs 204.8 GB/s.
Prices after the July 2026 Jetson hike
| Developer kit | Launch list | List since July 2026 |
|---|---|---|
| Jetson AGX Thor Developer Kit (T5000, 128 GB) | $3,499 (Aug 2025) | $5,499 |
| Jetson AGX Orin 64GB Developer Kit | $1,999 | $3,499 |
| Jetson T4000 module (1KU+) | $1,999 | $2,999 |
| Jetson AGX Orin 64GB module (1KU+) | $1,599 | $2,999 |
NVIDIA’s Jetson pricing FAQ as reported by CNX Software (July 22, 2026); another price tracker confirmed the $5,499 and $3,499 developer-kit prices on October 21, 2026. NVIDIA didn’t give a reason; observers point to memory costs. Several kits were intermittently out of stock in the US — check availability before planning around a ship date.
The ratio barely moved: Thor costs about 1.57× an AGX Orin kit, and it has 2× the memory. Per gigabyte of memory, Thor is now the cheaper kit (about $43/GB vs $55/GB).
Specs side by side
| Jetson AGX Thor (T5000) | Jetson AGX Orin 64GB | |
|---|---|---|
| GPU | Blackwell, 2,560 CUDA cores, 96 5th-gen Tensor cores | Ampere, 2,048 CUDA cores, 64 Tensor cores |
| AI compute (vendor) | Up to 2,070 FP4 TFLOPS (sparse) | 275 INT8 TOPS (sparse) |
| CPU | 14-core Arm Neoverse-V3AE | 12-core Arm Cortex-A78AE |
| Memory | 128 GB LPDDR5X, 273 GB/s | 64 GB LPDDR5, 204.8 GB/s |
| Power | 40-130 W | 15-60 W |
| Networking | QSFP28 (up to 4× 25 GbE) plus RJ45 | 10 GbE RJ45 |
| Software | JetPack 7 | JetPack 6 |
Read the compute row carefully: Thor’s 2,070 is a sparse FP4 figure; Orin’s 275 is sparse INT8. That’s where much of the “7.5×” headline comes from — Blackwell’s FP4 support is real and useful for optimized inference, but it is not a like-for-like number.
What the benchmarks actually show
| Test (Ollama, same settings) | Thor | AGX Orin 64GB | Thor advantage |
|---|---|---|---|
| gpt-oss-20b — prompt processing | 135.8 tok/s | 131.1 tok/s | 1.04× |
| gpt-oss-20b — generation (decode) | 36.6 tok/s | 25.8 tok/s | 1.42× |
| gpt-oss-120b — generation | 17.5 tok/s | aborted (likely out of memory) | Thor only |
Independent test by youtalk.jp (September 2025), using a stock/custom Ollama build rather than NVIDIA’s optimized TensorRT-LLM or vLLM stacks. Treat it as a lower bound on Thor’s advantage for compute-heavy work, and as a realistic picture for single-user chat.
Two takeaways. First, decode speed tracks bandwidth: a 1.33× bandwidth advantage became a 1.42× generation advantage, nowhere near 7.5×. Second, capacity is the real upgrade: the 120B model simply didn’t run on Orin. If your robot or edge server needs a large model resident in memory — or several models at once (perception, planning, a VLA policy, speech) — that 128 GB is what you’re paying for. Thor’s FP4 Tensor cores pull further ahead on prompt processing, batching and vision transformers once you move to NVIDIA’s optimized runtimes.
Jetson AGX Thor — for big models and next-gen robotics
NVIDIA Jetson AGX Thor Developer Kit
- 128 GB LPDDR5X — runs models the 64 GB Orin can't load (gpt-oss-120b at about 17.5 tok/s).
- Blackwell GPU with FP4 Tensor cores and Multi-Instance GPU; JetPack 7 and the latest Isaac stack.
- Sensor-heavy I/O with 25 GbE-class networking for camera and lidar fusion.
Thor is NVIDIA’s platform for “physical AI” — humanoids and autonomous machines running large foundation models on board. If that’s your roadmap, it’s the kit to prototype on, and the T4000 module ($2,999 at 1KU) gives a cheaper production path. Buying a kit for a lab or company? An Amazon Business account is free and adds business pricing and tax-exempt purchasing; for personal orders, a free 30-day Prime trial covers two-day delivery on the extras.
Jetson AGX Orin 64GB — the mature workhorse
NVIDIA Jetson AGX Orin 64GB Developer Kit
- 275 sparse INT8 TOPS, 64 GB at 204.8 GB/s; 15-60 W.
- Years of JetPack 6 documentation, community projects and production deployments.
- About 70% of Thor's LLM generation speed for about 64% of the price.
For multi-camera vision, classic robotics stacks, and LLMs or VLMs up to roughly 30B parameters at 4-bit, AGX Orin is still a capable and lower-power kit. It draws a maximum of 60 W versus Thor’s 130 W — relevant on a battery-powered robot. The 64 GB eMMC fills up fast; add an NVMe drive for models and datasets.
2 TB NVMe SSD for AGX Orin (M.2 2280)
- Model files and ROS bags fill the onboard 64 GB eMMC quickly.
- A mainstream PCIe Gen4 drive is plenty; the kit doesn't benefit from Gen5.
Which one should you buy?
| Your project | Best kit |
|---|---|
| Humanoid / VLA robotics, multi-model stacks | Jetson AGX Thor |
| 70B-120B models on the edge | Jetson AGX Thor |
| Multi-camera vision, AMRs, models up to about 30B | Jetson AGX Orin 64GB |
| Battery-powered robot with a tight power budget | Jetson AGX Orin 64GB |
| Learning, prototyping, models up to 8B | Jetson Orin Nano Super ($399) |
| Big local LLMs on a desk, no robot | A desktop GPU or unified-memory box — see DGX Spark vs Mac Studio |
The bottom line: Thor’s headline compute is real but mostly matters for optimized FP4 inference and heavy multi-model robotics; for plain text generation it’s about 1.4× an AGX Orin. Pay the extra $2,000 for the 128 GB and the Blackwell/JetPack 7 future, not for chatbot speed. If your models fit in 64 GB today, AGX Orin remains the better-value kit.