NVIDIA released two new members of its Jetson Thor line, the T3000 and T2000, built to pack Blackwell-generation AI compute into machines small enough for a delivery robot arm or a warehouse forklift. The launch pushes NVIDIA further into the physical hardware that runs robots and vision agents on a factory floor, not just the training clusters that build the models behind them. Notably, the company frames the announcement around who is already buying the platform rather than around new benchmark numbers.
The T3000 module delivers 865 FP4 teraflops from a Blackwell GPU paired with an eight-core Arm Neoverse CPU, 32GB of LPDDR5X memory, 273GB/s of memory bandwidth, and 25 gigabit Ethernet connectivity. NVIDIA says it matches the multimodal inference performance of the existing T5000 module at roughly half the size and half the power draw. That comparison is NVIDIA’s own: the announcement does not point to an independent benchmark confirming the parity claim. A safety-certified variant, the IGX T3000, adds NVIDIA’s Halos for Robotics stack for machines built to work directly alongside people.
The T2000, an entry-level module with 400 FP4 teraflops and 16GB of memory, targets cheaper systems: mobile robots that navigate on their own, industrial manipulator arms, and camera-based visual AI agents. Combined with its existing lineup, NVIDIA now sells edge AI compute spanning 70 TOPS to 2,000 teraflops. That spread lets a robotics buyer pick a compute tier for each product line instead of over-provisioning one expensive chip across an entire fleet.
Current Jetson AGX Thor customers include Amazon Robotics, Boston Dynamics, FANUC, 1X, Agile Robots, Hitachi, and Techman Robot, according to NVIDIA. That customer list, disclosed ahead of the new modules even shipping, does more work in this announcement than the teraflop figures. It tells buyers the platform already has production references in warehouses and factory lines, not just developer kits sitting unused on a bench.
NVIDIA is also shipping new Jetson agent skills, an automated tool for cutting memory footprint across its device lineup. The company says UBTech and Agile Robots cut memory usage by up to 15GB, letting them move from a 64GB Jetson AGX Orin down to a 32GB module. NoTraffic, which builds intelligent-transportation systems, reported a 30 percent memory reduction on an older Jetson TX2 NX unit. Those figures come from NVIDIA’s own customer citations rather than a third-party audit.
The launch also introduces Cosmos 3 Edge, a 4 billion parameter robot foundation model that NVIDIA says developers can post-train for a new robot body in roughly a day. Software emulation for the T3000 arrives in late July 2026 with JetPack 7.2.1; emulation for the T2000 follows in a later release. Physical modules do not ship until the first quarter of 2027. Everything NVIDIA announced this week, in other words, is a software preview running on existing Thor hardware, not new silicon a customer can hold.
Robotics teams evaluating a Thor-based design should start on the JetPack 7.2.1 emulation kit now, since it lets code run against the new memory and compute profile more than six months before any T3000 module physically ships. Whether the size and power cuts hold up outside NVIDIA’s own comparisons will show up in fleet deployments at Amazon Robotics or FANUC, not in this week’s blog post.
According to NVIDIA’s corporate blog (blogs.nvidia.com), published July 15, 2026.
