At CES 2026, NVIDIA announced the Rubin chip architecture platform that will replace the current Blackwell architecture. CEO Jensen Huang said the chips with this architecture will offer increased speed for artificial intelligence (AI) model training and inference while offering power savings.
Yesterday, NVIDIA announced that Vera Rubin has begun full-scale production, with it being deployed by CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud. The AI chip promises the highest performance per watt and lowest token cost for global AI operations.
The Rubin platform is named after astronomer Vera Rubin and combines six chips to create a complete AI ecosystem. The chips are the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX‑9 SuperNIC, BlueField‑4 DPU, and Spectrum‑6 Ethernet Switch.
AI model training performance is up to 3.5x faster than Blackwell, and up to 5x faster on inference tasks with processing power reaching 50 PFLOPS. In early testing, CoreWeave reported up to 10x the performance per megawatt compared to the Grace Blackwell NVL72.
In addition, the cooling system operates at a higher temperature of 45°C, allowing operation without the need for a chiller, saving millions of gallons of water per megawatt per year. At the same time, the design without cables, fans and hoses reduces the time to install the system tray from hours to minutes.
