NVIDIA introduced two new AI processors built on its Thor architecture, designed to power the next generation of humanoid robots and autonomous machines by delivering compact, power-efficient computing that can run advanced AI models at the edge. The Jetson T3000 and T2000 modules, announced on July 16, 2026, represent a significant step toward making general-purpose robots practical for mass-market deployment, with leading robotics companies including 1X, Boston Dynamics, Amazon Robotics, and FANUC already building systems on the platform.

What Makes These New Chips Different From Previous Robotics Hardware?

The Jetson T3000 is roughly half the size and power consumption of NVIDIA's previous T5000 module, yet delivers comparable performance for the types of AI tasks humanoid robots need to perform. The T3000 packs 865 FP4 teraflops of computing power, which translates to enough raw processing capacity to run large language models, vision language models, and world foundation models directly on the robot itself, without needing to send data back to a cloud server.

For roboticists, this matters enormously. Running AI models on the robot itself means faster response times, better privacy, and the ability to operate in environments without reliable internet connectivity. The T3000 combines an NVIDIA Blackwell GPU, an eight-core Neoverse Arm processor, 32 gigabytes of memory, and 273 gigabytes per second of memory bandwidth, along with 25-gigabit Ethernet connectivity.

The smaller T2000 module brings Thor architecture to a broader range of applications. With 400 FP4 teraflops of compute and 16 gigabytes of memory, it serves as an entry point for developers building visual AI agents, autonomous mobile robots, and industrial manipulators. Together, the two new modules give developers a scalable platform spanning performance from 70 TOPS to 2,000 teraflops, enabling them to address virtually any edge AI workload.

How Are Robotics Companies Already Using These Chips?

Several humanoid and robotics leaders have begun optimizing their systems for the new Thor architecture. Companies including UBTech and Agile Robots have used NVIDIA's newly released Jetson agent skills, which automate memory optimization and system configuration, to reduce memory usage by up to 15 gigabytes. This optimization allowed them to move from higher-memory configurations to the 32-gigabyte module without sacrificing performance.

The agent skills represent a significant productivity boost for developers. Rather than manually tuning software to fit hardware constraints, developers can now use AI agents to optimize the entire software stack and achieve substantial memory savings in days instead of weeks. This approach supports the entire Jetson portfolio, including the older Jetson Orin, enabling developers to run more capable workloads on lower-memory configurations.

Steps to Get Started With the New Thor Processors

• Access the Developer Kit: Developers can begin building today using the Jetson AGX Thor developer kit available through channel partners, which shares the same chip architecture and software stack as the new T3000 and T2000 modules.
• Use Emulation Mode: Developers can emulate the performance of the T3000 and T2000 modules before the hardware becomes available, with T3000 emulation mode arriving later in July 2026 through JetPack 7.2.1, and T2000 emulation support following in a future release.
• Leverage the Physical AI Software Stack: Developers can use NVIDIA's full physical AI software stack, including NVIDIA Isaac for robotics simulation and perception, alongside open models such as NVIDIA Nemotron, Cosmos 3, and Isaac GR00T to accelerate development of next-generation robots and autonomous machines.

When Will These Chips Actually Be Available?

The Jetson T3000 and T2000 modules are scheduled to become available in the first quarter of 2027. A broad ecosystem of hardware partners is already preparing Thor-based solutions, including ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK, and YUAN. Software partners such as Antmicro, Neurealm, REBOTNIX, and RidgeRun are developing emulation and migration solutions to help customers transition to the new modules.

NVIDIA also expanded its Cosmos 3 frontier open world foundation model family with a lightweight version compatible with Thor platforms. Cosmos 3 Edge is a 4-billion-parameter model that helps embodied systems see the world, reason over it in real time, and predict and generate actions through on-device inference. Using the open Cosmos framework, developers can post-train Cosmos 3 Edge for specific robot embodiments and sensors in about a day, then deploy it on Jetson Thor for real-time vision analysis and on-device robot policy.

The timing is significant for companies like 1X, which has been working to move its Neo humanoid robot from pilot deployments toward broader commercial use. By providing compact, power-efficient processors specifically designed for humanoid robotics, NVIDIA is removing one of the major technical bottlenecks that has slowed the industry's transition from research labs to real-world mass-market deployment. As physical AI and embodied AI move toward mainstream adoption, these new processors give developers a scalable foundation for bringing intelligent humanoids and autonomous machines into practical, everyday use.