NVIDIA introduced the Thor-based T3000 and T2000 chips today, targeting mass-market robotics and edge AI deployment.
For robotics companies weighing custom silicon against off-the-shelf platforms, the calculation now includes a set of Thor-architecture modules built for the cost and power constraints that come with running foundation models outside a data centre.
NVIDIA’s Jetson AGX Thor family already sits inside humanoid and mobile robot programmes at 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, and Techman Robot, according to the company. The T3000 and T2000 extend that lineup downward, offering smaller footprints and lower memory ceilings aimed at the point where robotics programmes transition from pilot fleets to purchasing decisions made at volume.
Compute density against a shrinking power budget
The T3000 delivers 865 FP4 teraflops of AI compute in a package NVIDIA describes as roughly half the size and power draw of its existing T5000 module. The board pairs an NVIDIA Blackwell GPU with an eight-core Neoverse Arm CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth and 25 GbE connectivity.
A safety-oriented variant, IGX T3000, matches that compute figure while adding integrated functional safety and running NVIDIA’s Halos for Robotics stack, intended for machines operating in proximity to people.
NVIDIA states that despite the smaller footprint, T3000 matches T5000 inference performance across multimodal workloads including large language models, vision language models, vision language action models and world foundation models. The company frames the module as a cost lever against elevated memory pricing, though the safety stack itself doesn’t remove the compliance work that follows
Running Halos for Robotics gives an integrator a framework for human-proximate operation; certifying that a specific robot in a specific facility meets local safety regulation, and carrying the liability if it doesn’t, remains work for the manufacturer and the buyer, not NVIDIA.
The T2000 sits below the T3000 in NVIDIA’s stack, offering 400 FP4 teraflops and 16GB of memory. NVIDIA positions it as an entry point for developers building visual AI agents, autonomous mobile robots and industrial manipulators where the full T3000 specification isn’t necessary or affordable.
Combined with the rest of the Jetson range, the company now claims a platform spanning 70 TOPS up to 2,000 teraflops, which it says lets developers address most edge AI workloads from a single software base rather than re-architecting for each device class.
Memory optimisation claims arrive alongside the new silicon
Alongside the hardware, NVIDIA released what it calls Jetson agent skills: automated tooling meant to optimise memory configuration and deployment across its Jetson portfolio, including both Thor and the older Orin line. The company reports that the tooling lets developers achieve memory savings in days rather than weeks, and several named customers back that claim with figures of their own.
UBTech and Agile Robots – working alongside industrial solutions provider Connect Tech – reduced memory usage by up to 15GB, according to NVIDIA, moving from the Jetson AGX Orin 64GB module down to the 32GB configuration.
SandStar reports a 4GB reduction in a smart retail deployment, enough to run on an 8GB Orin NX module instead of 16GB. GROOVE X, maker of the LOVOT companion robot, says it used Jetson’s heterogeneous AI accelerators to redistribute workload and land on a lower-memory configuration without naming a specific figure. NoTraffic, meanwhile, reports a 30 percent memory reduction on Jetson TX2 NX within its smart traffic platform, freeing capacity for additional AI features rather than a smaller board.
These are vendor-reported figures from named partners, which puts them a step above unattributed marketing claims. However, they’re still figures generated in controlled optimisation runs rather than independently audited numbers pulled from months of field operation.
Memory headroom measured against a known workload in a lab differs from memory headroom under a live deployment dealing with intermittent connectivity, firmware drift across a fleet, or sensor data that arrives late or incomplete. Enterprises evaluating a SKU downgrade on the strength of these figures should expect their own validation cycle before committing procurement budgets to a smaller memory tier.
Cosmos 3 Edge puts a foundation model directly on the module
NVIDIA also expanded its Cosmos 3 open world foundation model family with an edge-specific version compatible with the Thor platforms. Cosmos 3 Edge runs at four billion parameters and is built to let embodied systems interpret their surroundings, reason over that input in real-time, and generate or predict actions through on-device inference rather than a round trip to the cloud.
Using the open Cosmos framework, NVIDIA says developers can post-train the model for a specific robot body and sensor set in about a day, a figure the company frames as closing the gap between simulation training and real-world performance. That timeline describes the post-training step itself; validating the resulting policy against a robot’s actual operating environment is a separate exercise, and one that determines whether the model performs outside the conditions it was tuned against.
Because the T3000 and T2000 share chip architecture and software stack with the rest of the Thor family, NVIDIA is opening a development path ahead of physical availability. Developers can start building now on the existing Jetson AGX Thor developer kit, sold through channel partners, and emulate T3000 and T2000 performance in software.
T3000 emulation mode arrives this month with JetPack 7.2.1; T2000 emulation follows in a later release NVIDIA hasn’t dated. The modules themselves are scheduled to ship in the first quarter of 2027, which leaves over a year between the announcement and physical hardware reaching customers.
Hardware partners including ADLINK, Advantech, AAEON, Aetina, Auvidea, AVerMedia, Connect Tech, ForeCR, JWIPC, NEXCOM Robotic Solutions, Realtimes, Seeed Studio, Twowin, TZTEK, and YUAN already build Thor-based systems and will presumably add T3000 and T2000 boards to their catalogues.
Software partners Antmicro, Neurealm, REBOTNIX, and RidgeRun are named as providing emulation and migration support for customers moving existing code onto the new modules. Whether that migration proves as straightforward as the shared-architecture pitch suggests will depend on how much of a given deployment’s software stack was tuned against T5000 memory bandwidth and Orin-generation constraints in the first place.
Companies planning around the Q1 2027 ship date have roughly a year to run that migration work through emulation before real silicon arrives.
See also: RoboLab expands robot policy evaluation beyond success rates
Want to learn more about the IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including AI & Big Data Expo and the Cyber Security Expo. Click here for more information.
IoT News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
