// Platform

Jetson AGX Thor T5000 for Edge AI — Specs, Sizing & Deployment Fit

Last updated: August 2026

NVIDIA’s generative-AI edge flagship: Blackwell-generation FP4 compute, 128 GB unified memory, and enough throughput for 70B-class on-device LLMs, VLA robotics models, and very high stream densities.

2,070 TOPS (FP4)
40–130 W
~32 streams (est.)
~$4,999 module

Planning Takeaway

NVIDIA’s generative-AI edge flagship: Blackwell-generation FP4 compute, 128 GB unified memory, and enough throughput for 70B-class on-device LLMs, VLA robotics models, and very high stream densities. Justify the ~$4,999 module (and 130 W envelope) with those workloads — for classical vision, AGX Orin remains the better spend.

Specifications

SpecValue
Compute2,070 TOPS-class FP4 (Blackwell GPU); NVFP4 + FP8 Transformer Engine
Memory128 GB unified, ~273 GB/s
Power envelope40–130 W
Video decode (modeled planning ceiling)~82× 1080p / ~10× 4K H.264 concurrent
Est. planning capacity~32 concurrent 1080p detection streams (planning cap)
Multi-instance GPU (MIG)Tech preview on T5000
CoolingActive required
Form factorJetson AGX module (100×87 mm)
SoftwareJetPack 7.2 (Ubuntu 24.04, CUDA 13, June 2026) — current; JetPack 6.x (Ubuntu 22.04) — legacy track; production JetPack 7.1 shipped Jan 2026 with the TensorRT Edge-LLM SDK
Indicative price~$4,999 module / ~$5,499 dev kit (as of mid-2026)

Sizing Fit

Built for on-device generative AI: 30B–70B LLMs, vision-language-action robotics, multi-model concurrency with hard isolation, and stream densities beyond AGX Orin.

Decision Framework

Choose the Thor T5000 if:

  • You run 30B+ parameter LLMs or VLA models on-device — NVFP4 and 128 GB are the enablers.
  • You need multiple concurrent models with real-time isolation (MIG is a tech preview on T5000).
  • Stream density beyond ~20 concurrent cameras on a single node matters more than power draw.

Choose a sibling instead if:

  • Jetson T4000 — you want the Thor generation at ~$2,999 with a lower power and thermal envelope.
  • Jetson AGX Orin 64GB — classical vision and ≤13B LLMs — it remains the production workhorse at ~$2,999.

Frequently Asked Questions

How much does Jetson Thor cost?

As of the July 2026 NVIDIA repricing, the Thor T5000 module lists at ~$4,999 (up from $2,999) and the AGX Thor Developer Kit at ~$5,499 (launched August 2025 at $3,499). The T4000 module lists at ~$2,999.

Is Thor worth it over the AGX Orin 64GB?

Only if you are running 30B+ parameter LLMs, Vision-Language-Action models, or many concurrent models on one device. For multi-stream vision and LLMs up to ~13B parameters, the AGX Orin 64GB at ~$3,499 (dev kit) remains the better choice — Thor’s price, 130 W draw, and cooling needs rule it out for constrained deployments.

What is special about Thor’s FP4 support?

NVFP4 is NVIDIA’s block-scaled 4-bit floating-point format: FP4 weights with FP8 activations, switched at runtime by the Transformer Engine. It is what lets a 70B-parameter model fit in 128 GB of unified memory with room for KV cache — AGX Orin (Ampere) has no FP4 hardware.

What JetPack does Thor run?

JetPack 7.2 (June 2026) is current — Ubuntu 24.04, CUDA 13 — across Thor and the Orin family. Thor’s production software line began with JetPack 7.1 (January 2026), which introduced the TensorRT Edge-LLM SDK. MIG on the T5000 is a tech preview.