On January 6, 2026, CES 2026, the world’s largest and most influential technology event, officially opened in Las Vegas. NVIDIA unveiled its next-generation artificial intelligence infrastructure platform—the Rubin platform. At the heart of this platform is the brand-new superchip, Vera Rubin, marking another significant leap for NVIDIA in the field of high-performance computing and once again pushing data center high-speed interconnect technology to new heights.
The NVIDIA Rubin Platform
The Rubin platform undergoes comprehensive upgrades centered around three core capabilities: «computing power, interconnect, and energy efficiency,» and is composed of six key components:
- Vera CPU: A new-generation processor optimized for AI workloads, providing the system with enhanced general-purpose computing and scheduling capabilities.
- Rubin GPU: Building upon the evolution of the Hopper and Blackwell architectures, designed for trillion-parameter models.
- Sixth-Generation NVLink Switch Chip: Enables higher bandwidth and lower latency interconnects between GPUs.
- ConnectX-9 Super NIC: Supports higher-speed Ethernet and InfiniBand connections.
- BlueField-4 DPU: Offloads network, storage, and security tasks to improve overall system efficiency.
- Spectrum-6 Ethernet Switch Chip: Signifies the further adoption of Co-Packaged Optics (CPO) in AI networks.
The Rubin platform introduces five key innovations, including the latest generation of NVIDIA NVLink interconnect technology, Transformer Engine, Confidential Computing and RAS engines, and the NVIDIA Vera CPU. These breakthroughs will accelerate agentic AI, advanced reasoning, and inference for large-scale Mixture-of-Experts (MoE) models, reducing the cost per token by 10x compared to the NVIDIA Blackwell platform. Compared to its predecessor, the NVIDIA Rubin platform requires 4x fewer GPUs to train MoE models, thereby accelerating the democratization of AI.
The Vera Rubin Superchip
At the core of the Rubin platform is the NVIDIA Vera Rubin superchip. This fundamental computing unit tightly integrates AI execution capability with high-bandwidth data movement and collaborative management. Each superchip fuses two Rubin GPUs with one Vera CPU into a cohesive whole through a memory-coherent NVLink-C2C interconnect, breaking down the traditional boundaries between CPU and GPU to create a unified rack-scale compute domain.
Within the Vera Rubin superchip, the CPU acts as a data engine closely coupled with GPU execution. This coupling enables low-latency coordination, shared memory access, and efficient orchestration across training, post-training, and inference workloads. The Vera CPU does not operate as an external host but participates directly in the execution process, handling data movement, scheduling, synchronization, and execution flow without introducing bottlenecks.
Industry-Wide Demand Growth from Chips to Systems
As the Rubin platform enters full production, products based on this platform will be launched to the market by partners in the second half of 2026. The focus of AI infrastructure development is also rapidly shifting from single-chip performance to system-level coordination and large-scale deployment capabilities.
Within the Rubin platform architecture, the ConnectX-9 Super NIC is a core component that embodies this system-level interconnect capability.
ConnectX-9 Super NIC Specifications：
- Protocols:Ethernet, InfiniBand
- Max Bandwidth: 800 Gb/s
- Port Configuration & Speeds: Single port up to 800 Gb/s, or dual ports up to 400 Gb/s per port
- Supported Speeds & Modulation: 200/100/50 Gb/s PAM4 and 25/10 Gb/s NRZ
- Optical Interface Form Factors: OSFP, QSFP112
With the unleashing of the Rubin platform’s bandwidth capabilities, the importance of high-speed optical modules and fiber connectivity solutions becomes even more pronounced. Whether for server-to-switch or switch-to-switch connections, stable, low-loss optical links are crucial for ensuring system performance and reliability.
