During his keynote address at CES 2026, Jensen Huang, CEO of NVIDIA, unveiled the company’s latest 'Rubin' computing architecture. He hailed it as the "most advanced technology" in today’s AI hardware landscape, noting that full-scale mass production is already in progress. The architecture, named in honor of the renowned astronomer Vera Rubin, integrates six independent chips that operate harmoniously. At its heart lies the Rubin GPU, supported by the newly developed Vera CPU. NVIDIA has made strategic enhancements to the Bluefield and NVLink systems. Additionally, the Rubin architecture introduces a cutting-edge external connected storage layer designed to streamline workflows.

In terms of performance metrics, the Rubin architecture significantly outpaces its predecessor, the Blackwell architecture. It offers 3.5 times faster AI model training speeds and 5 times quicker inference speeds. The architecture boasts a peak computing capability of 50 Petaflops and achieves an 8-fold increase in inference performance per watt. Leading cloud service providers are set to adopt Rubin chips, and these chips will also be deployed in supercomputers at Hewlett Packard Enterprise and the Lawrence Berkeley National Laboratory.

Previously, Jensen Huang forecasted that global investment in AI infrastructure would surge to between $3 trillion and $4 trillion over the next five years.