NVIDIA Jetson Thor Computers Bring Scalable AI Power to Mainstream Robotics

Robots are leaving controlled demos behind and entering warehouses, factories, stores and shared human spaces. NVIDIA’s new Jetson Thor computers are designed to give those machines the compact, power-efficient AI compute needed to perceive, reason and act directly at the edge.

The newly introduced T3000 and T2000 modules expand the NVIDIA Thor lineup for developers building humanoids, autonomous mobile robots, industrial systems and visual AI agents at scale.

Why Jetson Thor computers matter for physical AI

General-purpose robots need much more than fast vision. They increasingly run multimodal models that combine language, images, sensor data and action planning, often in situations where sending every decision to the cloud is impractical.

NVIDIA positions Jetson AGX Thor as the computing foundation for this next phase of embodied AI, with companies including 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi and Techman Robot building on the platform.

• On-device intelligence: Edge inference allows machines to respond where they operate.
• Broad model support: The Thor architecture targets workloads such as large language models, vision-language models, vision-language-action models and world foundation models.
• Scalable deployment: NVIDIA’s Jetson range now spans from 70 TOPS to 2,000 teraflops, giving developers options across a wide range of edge AI requirements.

The T3000 targets high-performance robotics in a smaller package

The Jetson T3000 is the key new option for robotics teams that need substantial AI throughput without the footprint and power profile of the higher-end T5000. NVIDIA says it delivers similar multimodal inference performance to the T5000 while coming in at roughly half the size and power.

• AI performance: 865 FP4 teraflops.
• Core architecture: An NVIDIA Blackwell GPU paired with an eight-core Neoverse Arm CPU.
• Memory: 32GB LPDDR5X with 273GB/s of memory bandwidth.
• Networking: 25 GbE connectivity.
• Safety option: IGX T3000 provides the same performance with integrated functional safety and support for the NVIDIA Halos for Robotics full-stack safety system.

That combination matters especially for robots working near people. Functional safety is not simply an add-on in these settings; it is part of the platform requirement for machines expected to operate reliably in human environments.

Jetson T2000 makes Thor architecture more accessible

The Jetson T2000 brings the Thor architecture to a lower-memory, lower-compute tier without abandoning the kinds of applications that make edge AI useful. It is positioned as an entry point for developers creating capable machines that do not require the T3000’s full performance envelope.

• AI performance: 400 FP4 teraflops.
• Memory capacity: 16GB.
• Intended applications: Visual AI agents, autonomous mobile robots, industrial manipulators and other intelligent machines.
• Platform continuity: It shares the Thor family’s chip architecture and software stack, helping teams move across modules without starting their software work from scratch.

For product builders, this is a practical expansion rather than a purely theoretical one. A broader performance range lets teams align hardware cost and capability more closely with the machine they are actually shipping.

New agent skills focus on the memory bottleneck

Compute is only one part of deployment. Memory limits can determine whether a promising model fits on a device at all, which is why NVIDIA’s newly released Jetson agent skills focus on automating memory optimization, system configuration and deployment work.

These skills support the entire Jetson portfolio, including Jetson Thor and Jetson Orin. NVIDIA says they can help developers optimize the software stack in days rather than weeks and potentially move down one memory SKU in the same product tier without sacrificing performance.

• UBTech, Agile Robots and Connect Tech reduced memory use by up to 15GB, enabling moves from Jetson AGX Orin 64GB to the 32GB module.
• SandStar reduced memory use by up to 4GB, allowing deployment on Jetson Orin NX 8GB rather than the 16GB configuration.
• GROOVE X, creator of the LOVOT robot, uses Jetson’s heterogeneous AI accelerators to distribute workloads and reduce memory use.
• NoTraffic reduced memory use by 30% on Jetson TX2 NX, creating room for additional smart-traffic AI capabilities without increasing hardware requirements.

Alongside NVIDIA NemoClaw blueprints for orchestrating intelligent agents, these tools aim to make Jetson a more agentic-ready platform for physical AI development.

Cosmos 3 Edge brings a robot foundation model onto Thor

NVIDIA also expanded its Cosmos 3 frontier open world foundation model family with Cosmos 3 Edge, a lighter model compatible with Thor platforms. Built for embodied systems, it is intended to help a robot understand a scene, reason over it in real time and predict or generate actions through on-device inference.

• Model size: 4 billion parameters.
• Adaptation path: Developers can post-train Cosmos 3 Edge for particular embodiments and sensors in about one day using the open Cosmos framework.
• Deployment role: Jetson Thor can run real-time vision analysis and an on-device robot policy.
• Development objective: The workflow is designed to help close the sim-to-real gap between simulated training and physical deployment.

This is an important distinction: the hardware is not being presented only as a generic accelerator. NVIDIA is pairing it with models and robotics software intended for the full loop of sensing, reasoning and action.

Developers can start before the modules ship

The new modules are scheduled to become available in Q1 2027, but developers do not have to wait to begin software work. The available Jetson AGX Thor developer kit can emulate the performance of the upcoming T3000 and T2000 modules because the Thor family shares a common architecture and software stack.

• T3000 emulation: Expected later this month through JetPack 7.2.1.
• T2000 emulation: Planned for a future software release.
• Core software: NVIDIA Isaac for robotics simulation and perception, plus open models including NVIDIA Nemotron, Cosmos 3 and Isaac GR00T.
• Hardware ecosystem: Partners include ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK and YUAN.

Software partners Antmicro, Neurealm, REBOTNIX and RidgeRun are also set to provide emulation and migration solutions for customers moving to the new modules.

Key takeaways

• T3000 offers 865 FP4 teraflops for demanding multimodal robotics workloads in a smaller, lower-power form factor than T5000.
• T2000 offers 400 FP4 teraflops and 16GB of memory for a wider set of edge AI systems.
• Jetson agent skills target a real deployment constraint: reducing memory use so capable workloads can run on less expensive configurations.
• Cosmos 3 Edge adds a 4-billion-parameter robot foundation model designed for real-time on-device inference.
• Development can begin with the Jetson AGX Thor developer kit before the T3000 and T2000 reach the market in Q1 2027.

What comes next for embodied AI

The significance of the T3000 and T2000 is not just that they add more product choices. They give robotics teams a common path from development kit to scalable deployment, with hardware, optimization tools, simulation software and robot-focused foundation models designed to work together.

As physical AI moves into more practical, cost-sensitive deployments, that continuity could be as valuable as raw compute: it gives developers a way to build once, tune carefully and select the Thor configuration that fits the machine in front of them.