NVIDIA Jetson AGX Thor: Edge AI That Thinks On-Site

NVIDIA has just unlocked a new level of on-device intelligence with Jetson AGX Thor, a system-on-module designed to serve as the “brain” for the next generation of robots. Available now for developers and manufacturers, the Thor kit brings server-class AI capabilities to autonomous machines that don’t want to rely on a steady cloud connection. In practice, this means smarter perception, faster decisions, and more reliable operation in environments where latency and bandwidth are at a premium.

What’s inside the Thor brick

• A unified architecture: Thor runs on NVIDIA’s latest Blackwell GPU, marking the first time the company unifies its server, PC, and edge computing lines under a single design philosophy. Paired with a 14-core Arm Neoverse-V3AE CPU and a generous 128GB of high-speed memory, Thor aims to slice through complex inference tasks that once required a data-center round trip.
• Raw AI power, on the edge: NVIDIA cites up to 2,070 FP4 teraflops of AI compute, a dramatic leap over the Jetson Orin and a clear signal that edge devices can handle sophisticated generative models closer to the source.
• Efficiency matters: The platform is engineered for energy efficiency, offering a configurable power envelope between 40 and 130 watts. That range matters for mobile robots where every watt matters for battery life and thermal management.

Edge AI for the era of physical AI

The press materials frame Thor as a cornerstone of a broader shift toward physical AI — robots that perceive, reason, and act in unstructured, human-centric environments without constantly pinging a cloud server. That shift isn’t just about speed; it’s about resilience. If a warehouse robot loses connectivity, a Thor-powered system can keep moving, still able to reason about where to pick a package or how to re-route a task in real time.

• Generative models at the edge: Thor is designed to run a wide range of models on-device, from vision-language models (VLMs) and large language models (LLMs) to vision-language-action (VLA) pipelines. NVIDIA even highlights its Isaac GR00T N1.5 as a reference point for humanoid-capable AI, framed to operate where humans and machines share the same workspace.
• MIG, for predictability: The Multi-Instance GPU feature lets a single Thor GPU be partitioned into isolated, predictable slices. That means critical perception or control tasks won’t be crowded out by background processes running on the same hardware.

Real-world impact and early adopters

Industry leaders are already eyeing Thor as a potential accelerant for next-gen automation. NVIDIA notes that more than 2 million developers are part of the Jetson ecosystem, with early pilots across robotics and automation programs. A few high-profile names are exploring or integrating Thor into their systems:

• Agility Robotics: Will outfit the sixth generation of its Digit humanoid with Thor to boost real-time perception and decision-making for logistics tasks.
• Boston Dynamics: Plans to equip Atlas with server-level compute on-device to enable richer on-board reasoning.
• Figure and Amazon Robotics: Exploring or using Thor to power more capable, edge-resident AI in their robotics work.

Beyond humanoids, the Thor platform is positioned to support surgical assistants, agricultural machinery, and industrial manipulators, all of which benefit from running increasingly capable AI models close to the point of action.

The larger ecosystem you’re buying into

NVIDIA isn’t selling Thor in isolation. The developer kit is part of a broader robotics software stack intended to streamline the path from lab to field:

• NVIDIA Isaac: Robotics simulation and development environment for testing and refining behaviors before you deploy.
• Isaac GR00T foundation models: A set of base models designed for humanoid and other robot applications.
• NVIDIA Metropolis: Vision AI tools for large-scale, camera-driven deployments.
• NVIDIA Holoscan: Real-time sensor processing to fuse data streams from multiple sensors.

This integrated stack is meant to smooth the journey from cloud-based training to edge deployment, helping teams iterate faster while keeping latency under control.

Pricing, availability, and what it means for the market

The Jetson AGX Thor developer kit is priced at $3,499. NVIDIA emphasizes broad accessibility to researchers, startups, and established players looking to accelerate edge AI robotics programs. With a robust software ecosystem backing it, Thor isn’t just a compute upgrade — it’s a pathway to practical, on-device intelligence that can transform how robots perceive, reason, and interact with the world.

Bottom line

Thor signals a new era where the power once reserved for the data center creeps closer to the robot’s “brain.” For teams building autonomous machines that must operate quickly, safely, and independently, this is the kind of hardware that makes the dream of truly capable edge AI feel tangible. It’s not just about faster chips; it’s about changing how robots think and act in real time, in the environments we share with them.