At GTC Taipei, NVIDIA officially announced that its next-generation platform, Vera Rubin, has fully entered mass production. As a new infrastructure tailored for agentic AI, Vera Rubin leverages an open-source MGX rack-level architecture, deeply integrating Vera Rubin NVL72, Vera CPUs, Groq 3 LPX, BlueField-4 STX, and Spectrum-6 SPX to build a complete foundation for AI Factories. Compared to the previous generation Grace Blackwell, it delivers up to a 10x improvement in agent throughput at scale. Supply chain acceleration is underway simultaneously. Leading server manufacturers including Dell, HPE, Lenovo, and Supermicro have initiated large-scale manufacturing. In Taiwan alone, over 150 partners across more than 350 factories spanning 30 countries are collaborating on this effort. The network architecture has undergone significant upgrades. Vera Rubin debuts with the Spectrum-X Ethernet Photonics switch, featuring Co-Packaged Optics (CPO) technology and 200Gb/s SerDes interfaces. This reduces power consumption by fivefold while extending AI runtime duration by five times and accelerating deployment speed by 1.3 times, paving the way for million-GPU clusters. Paired with BlueField-4 DPUs supporting 800Gb/s speeds, the platform enables multi-tier tenant isolation and zero-trust networking policies. On security and operations fronts, Vera Rubin incorporates full-stack confidential computing, providing hardware-grade trusted execution environments and end-to-end encryption at the rack level. Supported by the DOCA software stack and DSX design-and-operate platform, it automates policy scheduling and energy optimization, helping enterprises deploy and operate AI Factories rapidly at minimal token costs—driving infrastructure evolution from "usable" toward "scalable autonomy."