.png)
NVIDIA introduced the Jetson T3000 and T2000 modules for robotics and edge AI, extending the Jetson Thor family from the AGX Thor platform. The announcement is aimed at moving foundation-model-heavy robotics workloads out of research setups and into production machines with tighter power, size, and cost constraints.
T3000 is the higher-end part of the launch: a compact module built on the Thor architecture with 865 FP4 TFLOPS of AI compute, an NVIDIA Blackwell GPU, an 8-core Arm Neoverse CPU, 32GB of LPDDR5X, 273GB/s of memory bandwidth, and 25GbE connectivity. NVIDIA positions Jetson T3000 for mass-market robotics where multimodal inference needs to stay local, including large language models, vision-language models, vision-language-action models, and world foundation models.
The practical signal here is not just raw throughput. NVIDIA says T3000 delivers similar inference performance to the T5000 for these multimodal workloads while using about half the size and power. For teams shipping humanoids or autonomous machines, that is the kind of tradeoff that changes BOM math, thermal design, and enclosure constraints. If the stated performance holds in your workloads, T3000 looks like the new default target for deployments where T5000-class capacity is unnecessary or too expensive under current memory pricing.
The IGX T3000 variant keeps the same compute profile but adds integrated functional safety and runs the NVIDIA Halos for Robotics safety system. That makes it relevant for robots that operate alongside people or in environments where deterministic safety handling is part of the product requirement rather than an add-on.
T2000 brings the same Thor architecture down to a smaller entry point: 400 FP4 TFLOPS and 16GB of memory. NVIDIA is clearly aiming this part at a wider set of edge AI systems than humanoids alone, including visual AI agents, autonomous mobile robots, and industrial manipulators.
For engineers, T2000 looks like the part to evaluate when the workload is more constrained around perception and local decision-making than around heavyweight multimodal reasoning. The memory ceiling is the important limiter here. If your deployment needs large context, multiple concurrent models, or aggressive video batching, 16GB will force tighter model selection and more careful memory budgeting. If the system is primarily an on-device perception-and-control stack, T2000 may be enough and materially easier to deploy at scale.
The larger strategic move is that NVIDIA is turning Jetson Thor into a recognizable deployment tier for robotics. The company says AGX Thor is already seeing adoption from 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot.
That matters because robotics platforms often fail when every production team has to treat hardware as a one-off integration. A clearer hardware baseline makes software optimization, benchmarking, and fleet validation easier. If the ecosystem converges on Thor modules, teams can spend less time porting across bespoke edge boards and more time on model behavior, sensor fusion, and real-world evaluation.
The one detail I would watch closely is software parity across the family. The announcement is hardware-forward, and the real production question is how quickly the developer kit, safety stack, and deployment tooling map cleanly onto T3000 and T2000 across the usual robotics workflow: simulation, hardware-in-the-loop, field logging, rollback, and fleet updates. If NVIDIA keeps that path smooth, Thor becomes more than another Jetson generation; it becomes the reference platform for compact robotics inference at scale.
