NVIDIA Accelerated Roadmap Transitions from Rubin Architecture in 2026 to Feynman Platform in 2028 with HBM4 Memory Vera CPUs and Advanced AI Factory Scaling Systems
NVIDIA Accelerated Roadmap Scaling from Rubin to Feynman Architecture
NVIDIA established an annual product release schedule for its data center and AI infrastructure systems starting in 2026. The organization has established its Extreme Co Design method as an aggressive performance improvement protocol which allows for continuous performance enhancements from single chips to complete AI production systems. The 2026 phase currently employs Rubin architecture while the 2028 Feynman platform base has already been constructed.
The Rubin platform represents a major shift in memory and processing efficiency. Rubin moves beyond Blackwell Ultra by establishing HBM4 memory as the new benchmark for AI workloads which require high performance capabilities. The 2026 system contains the following core components
The Rubin family of GPUs utilizes HBM4 memory for its standard model and HBM4e memory for its advanced model which delivers multiple times greater performance capacity than older GPU series. The Vera CPU serves as the Grace CPU replacement which enables better system performance through its direct NVLink connection to Rubin GPUs. The BlueField 4 DPU functions as advanced data processing units that handle all networking tasks needed to support Rubin based cluster systems.
The organization has implemented NVLink 6 and 7 switches which offer networking capabilities of 3600 GB/s through two separate systems CX9 1600G and Spectrum6 102T CPO.
NVIDIA has already outlined the technical specifications for the 2028 Feynman architecture. The current generation will develop custom die stacking technology together with integrated optics as a solution to solve existing physical barriers that restrict data transmission. The 2028 roadmap includes multiple projects which will be implemented throughout the year.
Feynman GPUs The new GPU technology combines custom HBM with advanced die stacking methods which enable unmatched memory density for future applications. The Rosa CPU is the advanced processor that will replace the Vera CPU as the next processor generation. The Feynman period will revolution networking through its introduction of NVLink 8 CPO and Spectrum7 204T CPO which double Rubin generation network capacity. The ConnectX 10 networking interface serves as the primary interconnect for all Feynman based AI factories.
NVIDIA's strategy emphasizes the Rack as the new unit of compute. The 2026 computing system uses three new products which demonstrate better memory performance than existing systems. The Oberon NVL576 and Kyber NVL144 systems will deliver all necessary system performance requirements needed to conduct successful massive language model training operations in 2026. The organization will establish the Kyber NVL1152 system in 2028 which will serve as a core component for all future industrial scale generative AI cluster systems.
Hardware refreshes have become synchronized with networking and software updates after the organization shifted to an annual cycle. This roadmap confirms NVIDIA's intent to dominate the compute stack by controlling the interconnects, the memory architecture, and the software ecosystem simultaneously.
