At the GPU Technology Conference 2025, Nvidia CEO Jensen Huang announced a lineup of advanced AI accelerator GPUs, including the Blackwell Ultra B300, Vera Rubin and Rubin Ultra. These GPUs are designed to significantly improve AI performance, especially for training and inference workloads. The standout is the Nvidia Blackwell Ultra B300, which promises to reshape the AI industry with more processing power and a substantial jump in memory capacity.
Nvidia Blackwell Ultra B300: A Major Leap Forward for AI
Expected in the second half of 2025, the Blackwell Ultra B300 raises memory capacity from 192GB to 288GB of HBM3e and delivers 50% more FP4 tensor compute performance than the Blackwell GB200. That supports larger AI models and improves inference on frameworks such as DeepSeek R1.
In a full NVL72 rack configuration, the Blackwell Ultra B300 can reach 1.1 exaflops of FP4 inference compute, far surpassing the previous generation. Beyond the standalone GPU, Nvidia also introduced B300 NVL16 server solutions, the GB300 DGX Station and the GB300 NV72L full rack, helping enterprises deploy more powerful AI supercomputers.
Combining eight NV72L racks forms the Blackwell Ultra DGX SuperPOD, which includes 288 Grace CPUs, 576 Blackwell Ultra GPUs and 300TB of HBM3e memory, for a total of 11.5 ExaFLOPS of FP4 processing power. These systems can be linked together to build large-scale AI supercomputers, which Nvidia calls "AI factories."
Vera Rubin and the Roadmap for Next-Generation GPUs
First shown at Computex 2024, the Vera Rubin GPU generation is expected in the second half of 2026, bringing major advances in AI training and inference. The GPU offers tens of terabytes of memory and pairs with the Vera CPU, which Nvidia designed in-house with 88 custom Arm cores and 176 threads.
With an architecture of two dies on a single package, Vera Rubin delivers 50 petaflops of FP4 inference performance per chip. Configured in an NVL144 rack, the system can reach 3.6 exaflops of FP4, speeding up the training of complex AI models.
Building on the Vera Rubin architecture, the Rubin Ultra version is expected in the second half of 2027. This GPU will use an NVL576 rack configuration, with four reticle-sized dies per GPU, providing 100 petaflops of FP4 per chip. The total output of the Rubin Ultra system reaches 15 exaflops of FP4 inference and 5 exaflops of FP8 training, well ahead of Vera Rubin. Notably, each Rubin Ultra GPU will have 1TB of HBM4e memory, bringing the total memory of the entire rack to 365TB.
Nvidia's Direction and the Future of AI
Beyond Blackwell Ultra B300 and Vera Rubin, Nvidia also announced a new-generation GPU architecture named Feynman, expected in 2028 alongside the Vera CPU. Although few details are available, Feynman is expected to push AI computing to a new level.
In his keynote, CEO Jensen Huang stressed the importance of the "AI factory," where data centers act as production lines that turn out data for AI models. He also pointed to the potential of physical AI, which could help develop smarter humanoid robots by training AI models in simulated environments before applying them in the real world.
Wrapping Up: Nvidia's Outlook in the AI Race
https://www.youtube.com/watch?v=_waPvOwL9Z8
The launch of the Nvidia Blackwell Ultra B300 and plans for new GPU generations such as Vera Rubin and Rubin Ultra underline Nvidia's ambition to lead the AI industry. With major gains in memory, performance and scalability, these GPUs not only meet the growing demands of AI but also open up a new future for AI supercomputers. Nvidia's roadmap shows a strong focus on building an advanced AI platform, helping the company keep its leading position in high-performance computing.
