Jetson AGX Orin 64GB

What it is

The official NVIDIA Jetson AGX Orin Developer Kit with the 64GB module — not Orin Nano, not Orin NX, not AGX Thor. 2048-core Ampere GPU with 64 Tensor Cores, 12-core Arm Cortex-A78AE, 64GB 256-bit LPDDR5 at 204.8 GB/s, 64GB eMMC 5.1 plus NVMe via M.2. About a 110 × 110 × 72 mm cube with carrier and thermal. Power 15W–60W. The kit emulates Orin modules and runs the JetPack / CUDA / Isaac / ROS stack.

Who it is for

Robotics and edge people who already live in JetPack and need more than the Nano Super’s 8GB without stepping up to Thor’s 128GB. This is the workhorse a lot of local-AI camera and robot work actually sits on. Wrong if you want a Windows mini PC, a silent desk cube, or 128GB class capacity for 70B weights.

What it actually runs

• Comfortable: mid-size LLMs and VLMs that fit in 64GB with headroom for JetPack and cameras. Order of magnitude: well above Nano’s 8GB; below 128GB Thor / Spark capacity. We did not bench token rates.
• Possible: larger dense models at hard quantization if KV cache and OS still fit. Capacity is 64GB. Bandwidth is about 205 GB/s.
• The real job: robotics, vision, Isaac / ROS, generative AI at the edge under JetPack. Language models are secondary to physical AI.

Standout details

• Price snapshot Sep 2026: about $1,999–$3,499 USD. NVIDIA Marketplace and partner carts often list around $1,999; NVIDIA’s Developer FAQ lists AGX Orin Developer Kit MSRP at $3,499. Snapshot only. Prices move.
• 64GB is the ceiling versus Thor and Spark at 128GB. Bandwidth 204.8 GB/s.
• Up to 275 TOPS is an NVIDIA INT8 sparse claim, not a Local Boxes bench.
• Arm + Jetson software world. Not a drop-in Windows mini PC.
• Cheaper door: Orin Nano Super. Step up for 128GB: AGX Thor or DGX Spark.

Honest verdict

The 64GB Jetson workhorse. Buy it for JetPack, cameras, and robotics with real unified memory. Do not buy it as a 70B desk cube. Nano Super is the cheap door. Thor or Spark if you need 128GB.