NVIDIA Vera Rubin & Rubin Ultra: Next-Gen AI Infrastructure

On March 16, during his keynote at GTC 2026, NVIDIA CEO Jensen Huang unveiled the next-generation AI infrastructure blueprint centered on Vera Rubin and showcased the upgraded AI supercomputer system Rubin Ultra. He also announced the next-generation architecture Feynman, expected to launch in 2028. Taiwanese ODMs including Foxconn (2317), Quanta (2382), and Wistron (3231) immediately confirmed they would release products based on NVIDIA’s new platform to capture market opportunities.

The market is most focused on Huang’s emphasis that AI computing demand continues to rise. Future data center interconnects will require more copper cabling, optical communications, and CPO (co-packaged optics) capacity, signaling a “hybrid optical-copper” strategy for AI data center interconnects.

This differs from earlier expectations that the Vera Rubin generation would favor “optical over copper,” causing optical communication stocks such as BrightWin (3163) and Wah Hong Optoelectronics (4979) to drop more than 7% in Taipei trading on March 17.

With NVIDIA unveiling the new Vera Rubin platform, Taiwanese ODMs quickly responded. Companies including Foxconn, Quanta, Wistron, Wiwynn (6669), Compal (2324), ASUS (2357), and Pegatron (4938) announced plans to release AI servers compatible with the Vera Rubin platform. Analysts expect ODMs to begin delivering Vera Rubin servers in Q3 and reach peak shipments in Q4.

The Vera Rubin platform is NVIDIA’s latest AI computing infrastructure, integrating GPU, CPU, and high-speed interconnects, optimized for high-throughput computation and large-scale inference, serving as the backbone of current AI data centers.

Huang noted that AI has evolved from generative models to agent-based AI with reasoning and action capabilities. Each inference requires massive token computation, transforming data centers from traditional storage hubs into “AI factories” for token production. Vera Rubin is the critical platform enabling this transformation.

NVIDIA also introduced the Rubin Ultra system, enhancing large-scale compute integration. Using the new NVLink interconnect technology, Rubin Ultra can connect up to 144 GPUs in a single system, forming a unified compute domain and significantly improving overall throughput and inference efficiency.

Industry experts believe this “scale-up” architecture breaks the limitations of traditional server nodes, enabling AI computation at full-rack and system-level deployment. It is particularly suited for long-context, large-model, and complex inference workloads.

NVIDIA also publicly revealed the advanced Feynman architecture, expected in 2028. While CPUs will continue to use the Vera design, Feynman will include next-generation NVLink technology and upgrade switch specifications to 204G CPO, supporting continued growth in AI computing capabilities.