On 15 July 2026, NVIDIA announced two new edge computing modules built on its Thor architecture: the T3000, which pairs a Blackwell GPU with 32GB of memory to deliver 865 FP4 teraflops, and the entry-level T2000 at 400 teraflops. The modules are aimed at putting high-performance perception directly inside robots, industrial machines, and visual edge systems, rather than relying on the cloud. Alongside the hardware, NVIDIA presented Cosmos 3 Edge, a 4-billion-parameter foundation model that lets embodied systems see their surroundings, reason about them in real time, and generate actions through on-device inference. The platform runs multimodal workloads including vision-language and vision-language-action models, and developers can post-train Cosmos 3 Edge for a specific machine in roughly a day. Emulation support begins in late July 2026, with hardware modules expected in the first quarter of 2027.
The move underlines how perception, reasoning, and control are converging onto small, low-power devices at the edge. For Ergonet, whose work spans machine vision and robot-agnostic warehouse automation, advances that make real-time on-device vision cheaper and more capable point in the same direction as its own Physical AI focus.
Source: NVIDIA