At CES 2026, NVIDIA unveiled its next-gen Rubin platform, marking a significant leap in AI compute. Designed for demanding large language models, the Rubin GPU delivers up to 5x greater inference and 3.5x greater training performance than its Blackwell predecessor.
The Vera Rubin NVL72 rack integrates six new silicon types, including the Vera CPU and Rubin GPU, linked by a high-bandwidth NVLink 6 fabric providing 260 TB/s of intra-rack bandwidth. Key innovations include an AI-optimized memory layer via BlueField-4 DPUs and rack-scale security. This architecture offers substantial efficiency gains, requiring 4x fewer GPUs for training some MoE models and reducing per-token inference costs by up to 10x.
Rubin is now in production, with cloud instances from AWS, Google, Microsoft, and others expected in late 2026. NVIDIA also announced the Mercedes-Benz CLA will debut its DRIVE AV software in the US and introduced Alpamayo, an open-source model suite for autonomous vehicle development.
