NVIDIA Jetson Thor Powers Edge AI Revolution as NVDA Stock Holds $178
Rongchai Wang Mar 10, 2026 17:41
NVIDIA's Jetson platform enables enterprise-grade AI to run locally on industrial equipment, from Caterpillar excavators to dual-arm robots, without cloud dependency.
NVIDIA's push to move AI processing from data centers to physical machines is gaining serious traction. The company's Jetson platform now runs generative AI models locally on everything from eight-ton excavators to dual-arm robots, eliminating cloud latency and ongoing compute costs that have plagued industrial AI deployments.
At CES earlier this year, Caterpillar demonstrated its Cat AI Assistant running on Jetson Thor inside a 306 CR mini-excavator—a machine small enough to fit in a shipping container but complex enough to require extensive operator training. The system uses Qwen3 4B for natural language processing and NVIDIA Nemotron speech models, all executing locally with no internet connection required.
Why Edge Matters for Industrial AI
The shift addresses a fundamental tension in industrial AI. Cloud deployments work fine for chatbots, but physical systems need something different: sub-millisecond response times, consistent behavior regardless of network conditions, and the ability to operate in environments where connectivity isn't guaranteed.
Memory shortages across the semiconductor industry have complicated matters further, driving up costs for discrete component approaches. Jetson's system-on-module design bundles compute and memory together, simplifying hardware sourcing for manufacturers.
NVIDIA stock traded at $178.03 on March 10, down 1.7% on the day, with the company's market cap holding at $4.57 trillion. The Jetson business represents a smaller but strategically important piece of NVIDIA's broader AI infrastructure play.
Real-World Deployments Accelerating
The developer ecosystem around Jetson has expanded rapidly. Franka Robotics ran the NVIDIA GR00T N1.6 vision-language-action model entirely onboard its FR3 Duo dual-arm system at CES—perception to motion, no task scripting required.
NVIDIA's own GEAR Lab trained a humanoid controller on 100 million frames of motion-capture data, then deployed it on a physical robot where the kinematic planner runs on Jetson Orin at roughly 12 milliseconds per pass. The policy loop executes at 50 Hz, all onboard.
A UIUC robotics team built a matcha-making robot on Jetson Thor that won first place at an NVIDIA embodied AI hackathon. NYU's Center for Robotics recently ran its YOR robot on the platform, showing improved generalization on pick-and-place tasks.
Model Performance Numbers
Jetson Thor delivers 52 tokens per second for Mistral 3 models at single concurrency, scaling to 273 tokens per second with eight concurrent requests. The Qwen 3.5-35B-A3B model reasons at 35 tokens per second. Physical Intelligence's PI 0.5 model generates 120 action tokens per second for robotics applications.
ABB Robotics announced a partnership with NVIDIA on March 9 focused on industrial-grade physical AI deployment. Texas Instruments followed on March 5 with its own collaboration targeting next-generation physical AI systems.
NVIDIA plans to showcase these capabilities at GTC 2026 next month, including a panel on industrial autonomy. For developers already building on the platform, the message is clear: the models are ready, the hardware exists, and the question has shifted from whether edge AI works to how fast it can scale.
Image source: Shutterstock.jpg)