NVIDIA Teases Next-Gen 'AI Factory' Infrastructure
NVIDIA's upcoming GTC 2026 conference is expected to showcase new infrastructure technologies—LPX, CPO, and Rubin—aimed at scaling 'AI factories.' The focus signals a continued push to build the underlying systems that power both digital and embodied AI.
The term 'AI Factory' is a concept championed by NVIDIA CEO Jensen Huang, who envisions future data centers as industrial plants that consume energy and data to produce intelligence in the form of tokens. He predicts that every industry will eventually operate two types of factories: one for producing their physical goods and another for generating the AI that enhances them. The upcoming Rubin platform is the successor to the current Blackwell architecture and is engineered to be the engine of these AI factories. It's not a single chip but a system of six co-designed components, including the Rubin GPU and the 'Vera' CPU, designed to reduce AI inference token costs by up to 10 times compared to Blackwell. Systems based on the Rubin platform are expected to begin shipping in the latter half of 2026. A key enabler for this scale is Co-Packaged Optics (CPO), a silicon photonics technology that integrates optical components directly with the networking switch ASIC. This shift from pluggable optical modules drastically shortens electrical pathways, which can reduce network power consumption by over 80% and increase bandwidth density, a critical factor for connecting tens of thousands of GPUs. This massive leap in computational and networking efficiency is aimed directly at enabling more sophisticated "agentic AI," which involves models that can perceive, reason, and act in the physical world. The infrastructure is being built to handle the massive long-context workflows and multi-step problem-solving required by advanced embodied AI. NVIDIA's strategy extends from the data center to the device, with platforms specifically for robotics. This includes Project GR00T, a general-purpose foundation model for humanoid robots, and the Jetson Thor computer designed to power them. These are complemented by simulation environments like Isaac Sim, which are used to train robots in a virtual world before real-world deployment. The company has also been heavily involved in robotics research, presenting papers on topics like 4D scene generation for autonomous vehicle training (DreamDrive) and unified neural controllers for humanoid robots (HOVER). This full-stack approach, from massive AI training clusters to on-robot computers and simulation tools, signals a major push to solve the challenges of embodied intelligence.
