Building smart machines is getting cheaper and easier. Nvidia has expanded its robotics hardware lineup with 2 new entry points for its Thor architecture, aiming to push advanced AI out of research labs and straight into commercial hardware. These new modules, the T3000 and the T2000, are designed to run heavy AI models directly on edge devices without melting your power budget.
According to the official product announcement from Nvidia, the T3000 acts as the main hardware option for companies building next generation humanoids. The system combines an Nvidia Blackwell GPU with an 8 core Neoverse Arm CPU and 32GB of LPDDR5X memory. This hardware pushes out 865 FP4 teraflops of AI computing power. The best part is the footprint. It occupies about half the physical space and pulls half the power of the larger T5000 module while matching its performance on heavy visual workloads. For systems operating alongside humans, the IGX T3000 variant integrates the Halos for Robotics safety system to ensure secure operation.
For developers with tighter budgets, the T2000 module cuts down the entry barrier. It packs 400 FP4 teraflops of computing speed and 16GB of memory. This card targets smaller machines like visual AI agents and factory robotic arms. Major industry names are already on board. Companies like Boston Dynamics, Amazon Robotics, 1X, FANUC, and Hitachi are actively building on the Jetson AGX Thor platform to deploy autonomous hardware.
Memory prices are high right now. The catch with modern AI is memory cost. To fight this, Nvidia released new agent software tools that automate code optimization across their entire hardware portfolio. Instead of spending weeks manually configuring systems, developers can shrink memory footprints in just days. This software allows companies to run advanced workloads on cheaper hardware configurations with smaller memory capacities.
Real world testing shows the software works. Robotics developers at UBTech and Agile Robots managed to shave up to 15GB of memory usage. This allowed them to step down from a 64GB hardware configuration to a cheaper 32GB module without losing speed. Smart retail provider SandStar shaved 4GB of memory to run on an 8GB Orin module instead of a 16GB version. In the transportation sector, NoTraffic managed a 30% memory reduction on older Jetson TX2 NX hardware, giving their traffic platforms extra room for new features without upgrading the physical chips.
Alongside the hardware, Nvidia is expanding its open world foundation model family with Cosmos 3 Edge. This is a 4 billion parameter model designed to run on the Thor platform. It allows robots to perceive their surroundings, reason about situations, and generate actions locally. Developers can post train the model for specific hardware sensors in about 1 day, which helps close the gap between virtual simulations and real world physical deployments.
Developers can begin testing these systems immediately. Emulation mode for the T3000 arrives later this month with the JetPack 7.2.1 software update, allowing engineers to mimic the performance on existing developer kits. The emulation mode for the T2000 is scheduled for a future release. Physical T3000 and T2000 modules are slated to ship in Q1 2027 through partners such as Advantech, AAEON, and Seeed Studio.
