Nvidia Announces Roadmap with AI GPUs, Rubin Architecture with HBM4 Set to Release in 2026
Nvidia CEO Jensen Huang revealed the company’s new roadmap for data center products during Computex. The company plans to launch a new architecture for AI GPUs in 2026, named Rubin, to succeed the recently announced Blackwell GPUs. During his Computex 2024 keynote, Huang showcased Nvidia’s data center lineup for the upcoming years. He initially announced a Blackwell refresh for next year, which will support twelve HBM3e stacks compared to the current eight. The upcoming Blackwell Ultra refresh is expected to support larger memory capacities than the initial Blackwell B200 GPU, set to hit the market later this year with 192GB of HBM3e memory.
Furthermore, Huang outlined his plans for 2026 and 2027, confirming that the next generation of Nvidia data center GPUs will be codenamed Rubin. The architecture is named after astronomer Vera Rubin. While details about the new Rubin GPUs are scarce as the current Blackwell cards are yet to be released, Nvidia has confirmed that the Rubin GPU will support faster HBM4 memory with eight HBM4 stacks. In 2027, Rubin Ultra will follow with twelve HBM4 memory stacks for increased total capacity.
Huang also confirmed the arrival of a new processor for data centers to succeed the current Grace-Arm CPUs, codenamed Vera. Additionally, there will be a Vera Rubin superchip, featuring a PCB combining two Rubin GPUs with a Vera processor, similar to the previous Grace Hopper superchip. This platform will also support new NVLink 6 switches, offering bandwidths of 3.6TB/s.
While Nvidia has not provided specific release dates for the new products, rumors have suggested the production of Rubin GPUs as successors to Blackwell by the end of 2025. Analyst Ming-Chi Kuo mentioned that the GPUs will be produced on TSMC’s N3 node, although Huang did not confirm this during his keynote presentation.
