Introduction: NVIDIA Expands the Jetson Thor Family
NVIDIA has introduced two new AI computing modules, the Jetson T3000 and Jetson T2000, designed to accelerate the adoption of robotics, autonomous machines, and edge AI applications. Built on the company’s advanced Thor architecture, these compact and power-efficient systems are aimed at bringing powerful artificial intelligence directly to devices operating outside traditional data centers. As industries increasingly adopt AI-powered robots for manufacturing, logistics, healthcare, retail, agriculture, and smart cities, NVIDIA’s latest Jetson platform provides developers with the computing performance needed to run sophisticated AI models locally while reducing latency, improving reliability, and lowering operational costs.
Meeting the Growing Demand for Edge AI
General-purpose robots and intelligent machines are rapidly moving beyond research laboratories into commercial and industrial environments. Modern robots must perform complex tasks such as understanding spoken language, recognizing objects, navigating dynamic surroundings, and making decisions in real time. These capabilities rely on large AI models that traditionally required cloud-based processing.
NVIDIA’s new Jetson Thor family addresses this challenge by delivering high-performance AI computing directly on the device. This allows robots to process information locally without depending on internet connectivity, enabling faster response times, enhanced privacy, and greater reliability in environments where cloud access may be limited.
The Jetson Thor platform is already gaining traction among robotics leaders, with companies including Amazon Robotics, Boston Dynamics, FANUC, Hitachi, Techman Robot, Agile Robots, and 1X building next-generation robotic systems on NVIDIA’s technology.
Jetson T3000: High Performance in a Compact Package
The flagship Jetson T3000 module is designed for advanced humanoid robots, autonomous machines, and industrial automation systems that require exceptional AI performance within a compact footprint.
Powered by NVIDIA’s latest Blackwell GPU architecture, the module delivers up to 865 FP4 teraflops of AI computing power while consuming significantly less power than previous high-end solutions. Despite being approximately half the size and power consumption of the Jetson T5000, the T3000 delivers comparable inference performance across demanding AI workloads.
The hardware combines:
- NVIDIA Blackwell GPU
- Eight-core Arm Neoverse CPU
- 32GB LPDDR5X memory
- 273GB/s memory bandwidth
- 25 Gigabit Ethernet connectivity
This configuration enables developers to deploy large AI models directly on robots without sacrificing performance.
The industrial-grade IGX T3000 version extends these capabilities by incorporating functional safety features essential for robots working alongside humans. It also supports NVIDIA’s Halos for Robotics safety platform, helping developers build machines that meet strict safety requirements for collaborative environments.
Running Advanced AI Models at the Edge
One of the biggest strengths of the T3000 is its ability to efficiently execute modern AI foundation models locally.
These include:
- Large Language Models (LLMs)
- Vision Language Models (VLMs)
- Vision Language Action Models (VLAMs)
- World Foundation Models
These AI models allow robots to understand both visual and language inputs simultaneously, making them capable of interpreting instructions, recognizing objects, understanding environments, and executing complex actions.
Because the T3000 achieves similar performance to the larger T5000 while using less memory, developers can reduce hardware costs during a period of rising memory prices without compromising AI capabilities.
Jetson T2000 Brings Thor Architecture to More Devices
Alongside the T3000, NVIDIA introduced the more affordable Jetson T2000, expanding access to Thor-based AI computing for a broader range of edge applications.
The T2000 delivers:
- 400 FP4 teraflops of AI performance
- 16GB memory
- Compact, energy-efficient design
The module is intended for developers building:
- Visual AI assistants
- Autonomous mobile robots
- Industrial robotic arms
- Smart cameras
- Intelligent transportation systems
- AI-powered inspection devices
By offering strong AI performance at a lower cost, the T2000 enables smaller businesses and developers to build advanced AI-powered machines without requiring enterprise-level hardware budgets.
With the launch of both modules, NVIDIA’s Jetson portfolio now spans AI performance ranging from 70 TOPS to 2,000 teraflops, giving developers scalable computing options for virtually every edge AI workload.
AI Agent Skills Simplify Development
Beyond introducing new hardware, NVIDIA announced new Jetson Agent Skills, software capabilities designed to automate memory optimization, system configuration, and deployment.
Developing AI applications for embedded systems often involves manually optimizing memory usage, tuning software stacks, and balancing workloads across CPUs, GPUs, and accelerators. These tasks can take weeks and require deep technical expertise.
Jetson Agent Skills automate much of this work, allowing developers to optimize applications in days rather than weeks.
The software supports the complete Jetson lineup, including both the new Thor modules and existing Jetson Orin devices.
By reducing memory consumption, developers can run larger AI workloads on lower-memory hardware, lowering overall system costs while maintaining performance.
Real-World Success Stories
Several companies have already demonstrated significant improvements using NVIDIA’s optimization technologies.
Humanoid robotics companies UBTech and Agile Robots, along with industrial computing company Connect Tech, reduced memory usage by as much as 15GB. This optimization enabled them to migrate from the 64GB Jetson AGX Orin module to the less expensive 32GB version.
In the retail sector, SandStar lowered memory requirements by approximately 4GB, making it possible to deploy AI applications on the Jetson Orin NX 8GB module instead of the 16GB model.
Japanese robotics company GROOVE X, creator of the companion robot LOVOT, optimized workload distribution across Jetson’s heterogeneous AI accelerators, reducing memory usage while maintaining application performance.
Meanwhile, intelligent traffic management company NoTraffic achieved a 30% reduction in memory usage on the Jetson TX2 NX platform, creating additional capacity for new AI features without upgrading hardware.
These examples demonstrate how software optimization can significantly reduce deployment costs while improving scalability.
Cosmos 3 Edge Expands Robot Intelligence
NVIDIA also expanded its Cosmos 3 family of world foundation models with the introduction of Cosmos 3 Edge, a lightweight AI model specifically designed for Jetson Thor platforms.
Featuring 4 billion parameters, Cosmos 3 Edge enables robots to:
- Understand complex environments
- Reason about real-world situations
- Predict future actions
- Generate intelligent responses
- Perform real-time decision making
Unlike cloud-based AI models, Cosmos 3 Edge runs entirely on-device, reducing latency while enabling autonomous operation even without internet connectivity.
Using NVIDIA’s open Cosmos framework, developers can customize the model for specific robot designs, sensors, and environments within approximately one day, helping bridge the gap between simulation and real-world deployment.
Faster Development Through Emulation
To help developers begin building applications before the new hardware becomes widely available, NVIDIA introduced emulation mode for the Thor family.
Using the existing Jetson AGX Thor Developer Kit, developers can emulate the performance of both the T3000 and T2000 modules while continuing software development.
Because all Thor products share the same chip architecture and software ecosystem, applications developed today can be migrated seamlessly to future production hardware.
Developers can also leverage NVIDIA’s broader physical AI software stack, including:
- NVIDIA Isaac for robotics simulation
- NVIDIA Nemotron language models
- Cosmos 3 world foundation models
- Isaac GR00T foundation models for humanoid robotics
Together, these tools simplify the creation of intelligent robots capable of perception, reasoning, navigation, and manipulation.
Availability and Ecosystem
NVIDIA plans to release T3000 emulation support later this month through JetPack 7.2.1, while T2000 emulation will follow in a future software update.
Commercial availability of both Jetson T3000 and Jetson T2000 modules is scheduled for the first quarter of 2027.
A broad ecosystem of hardware partners—including ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK, and YUAN—is already developing Thor-based products. Software partners such as Antmicro, Neurealm, REBOTNIX, and RidgeRun are also preparing migration and emulation solutions to help customers transition smoothly to the new platform.
A New Foundation for Physical AI
As robotics and embodied AI continue moving into mainstream industries, NVIDIA’s Jetson T3000 and T2000 modules provide a scalable platform capable of supporting everything from entry-level intelligent devices to advanced humanoid robots. Combined with powerful AI software, optimized development tools, and a growing partner ecosystem, the new Jetson Thor computers are positioned to accelerate the next generation of autonomous machines, enabling developers to bring intelligent robots into factories, hospitals, warehouses, retail stores, and everyday environments around the world.
Source link: https://blogs.nvidia.com