Overview
Jetson is NVIDIA's computing platform for edge and robotics work. The defining change in this generation is the move to the Blackwell GPU architecture, the same family used in data centre parts but configured for an embedded power and size envelope.
Its advantage was never the silicon alone. JetPack bundles CUDA, cuDNN and TensorRT together with board support packages, keeping the training, simulation and deployment path continuous, which is why robotics teams find it hard to move away.
Key Features
- Module family and compute ladder: The flagship T5000 delivers 2,070 FP4 sparse TFLOPS with 128GB of LPDDR5X across a 40W to 130W range; the T4000 delivers 1,200 TFLOPS with 64GB across 40W to 70W. Both share the same 87×100mm module format and pinout, so one carrier board can serve both.
- Developer kit to start: The Jetson AGX Thor Developer Kit shipped in August 2025 with a reference carrier board, active cooler and power adapter, suited to validating before moving to production modules.
- Mainstream and entry tiers coming: In July 2026 NVIDIA expanded the Thor family with the T3000 at 865 TFLOPS and 32GB, and the T2000 at 400 TFLOPS and 16GB, plus an industrial IGX T3000 with integrated functional safety. Both modules are scheduled for Q1 2027.
- JetPack as a unified stack: JetPack 7.2, released June 2026, is the first release to support the entire Orin family alongside the Thor developer kit, built on Jetson Linux 39.2.0 GA, Ubuntu 24.04, CUDA 13.2.1 and TensorRT 10.16.2. Orin and Thor now share one software branch, cutting the burden of maintaining two images.
- An edge runtime for large models: JetPack 7.1 introduced TensorRT Edge-LLM, an open-source C++ SDK for robots and real-time systems that need to run LLMs and VLMs inside strict latency and memory budgets, supporting FP8, NVFP4 and INT4 quantisation.
- Ecosystem scale: More than 3 million developers build on NVIDIA's robotics stack, and vendors including AAEON, ADLINK, Advantech and Seeed Studio supply carrier boards and complete systems to shorten the path from selection to production.
Use Cases
- On-device inference and control for humanoid and autonomous mobile robots
- Multimodal robots that need vision and language models on a single chip
- Edge AI nodes in industrial vision, smart infrastructure and automated production lines
- Real-time perception and path decision-making for drones and delivery robots
Pros
- GPU architecture shared with the data centre keeps model migration to the edge small
- JetPack covers the whole path from CUDA to TensorRT with a mature ecosystem
- Unified module format and pinout allow high and low configurations on one carrier board
- Mainstream and entry tiers will widen the range of choices
Pricing
Sold as modules and developer kits. The Jetson AGX Thor Developer Kit is priced at $3,499. T5000 and T4000 are production modules; T3000 and T2000 are scheduled for Q1 2027.
Summary
Jetson is currently the most complete way to run large models at the edge, and the maturity of the software stack is the real moat. The costs are just as clear: power draw and price sit high, and the T5000's 130W ceiling is more than many embedded scenarios can absorb. A practical approach is to fix the target model's parameter count and latency requirement first, then choose between T4000 and T5000. For pure vision work without large models, a smaller accelerator is enough.
Version History
- Thor family expanded (2026-07-15): T3000 at 865 TFLOPS and 32GB, T2000 at 400 TFLOPS and 16GB, both scheduled for Q1 2027, plus the industrial IGX T3000
- JetPack 7.2 (2026-06): First release to support the whole Orin family alongside Thor, moving to Jetson Linux 39.2.0 GA, Ubuntu 24.04 and CUDA 13.2.1
- TensorRT Edge-LLM (2026-03): JetPack 7.1 introduced an open-source C++ LLM inference SDK for real-time robotics
- Jetson AGX Thor Developer Kit (2025-08): Released with a Blackwell GPU at $3,499