Nvidia Aims to Occupy Every Chip in AI Data Centers

Nvidia has announced a new chip system, Vera Rubin, which will replace Grace Blackwell and form the core of its near future in the AI industry. The system includes 36 Vera CPUs paired with 72 Rubin GPUs in an NVL72 configuration, with the CPU also sold separately and potentially available for Chinese clients as early as August. Nvidia claims that the Vera Rubin NVL72 will process 10 times more tokens per watt than Grace Blackwell, while local memory subsystems will provide nearly three times the memory bandwidth. According to the company, the monolithic architecture of the Vera CPU outperforms the chiplet-based solutions of competitors AMD and Intel. The system is 100% liquid-cooled, described as "cable-free compute" and "hot-swappable," reducing the setup time for each rack from several hours to minutes. OpenAI is already using one Vera Rubin rack. The presentation took place amid competition with AMD, which recently revealed details about its Helios system. Nvidia emphasizes that Vera Rubin is on schedule, being sensitive to delays following overheating issues with the previous-generation Blackwell.