NVIDIA Unveils Blackwell Ultra and Teases Next-Gen AI Chip Rubin
On March 18, 2025, NVIDIA CEO Jensen Huang took the stage at GTC 2025, unveiling the company’s latest AI powerhouse, Blackwell Ultra, while also offering a glimpse into the future with Rubin, NVIDIA’s next-generation AI chip architecture.
On March 18, 2025, NVIDIA CEO Jensen Huang took the stage at GTC 2025, unveiling the company’s latest AI powerhouse, Blackwell Ultra, while also offering a glimpse into the future with Rubin, NVIDIA’s next-generation AI chip architecture.
Blackwell Ultra: The Next Leap in AI Computing
Following the Blackwell architecture, introduced a year ago, Blackwell Ultra is designed to supercharge AI workloads across pre-training, fine-tuning, and inference tasks. Huang emphasized that AI computation is growing at an exponential rate, requiring unprecedented levels of compute power.
Blackwell Ultra integrates NVIDIA GB300 NVL72—a rack-scale system comprising 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs (based on Arm Neoverse architecture). The high-density, single-unit GPU system is built specifically for scaling AI inference, allowing models to handle increasingly complex tasks with enhanced efficiency.
Key Performance Upgrades
1.5x the performance of the GB200 NVL72.
40x faster inference performance compared to the previous Hopper architecture.
50x greater revenue potential for AI factories using Blackwell versus older Hopper-based infrastructure.
Additionally, Blackwell Ultra GPUs can be purchased as standalone chips. NVIDIA also introduced the DGX Station, a high-performance workstation featuring a single GB300 Grace Blackwell Ultra chip, 784GB of memory, and 800Gb/s ConnectX-8 SuperNIC networking.
Next-Gen AI Chip Architecture: Rubin
During the event, Huang also previewed NVIDIA’s next-gen AI chip platform, Rubin, set to replace Blackwell. Rubin will debut in 2026, accompanied by Vera, a custom AI CPU.
Vera and Rubin Ultra Specs
Vera CPU: 88 custom Arm cores, 176 threads, 1.8TB/s NVLink-C2C bandwidth.
Rubin GPUs: Dual-GPU setup, FP4 inference at 50 petaFLOPS (PF), up to 288GB high-speed memory.
Vera Rubin NVL144:
3.6 exaFLOPS (EF) FP4 inference
1.2 EF FP8 training
3.3x performance boost over GB300 NVL72
Rubin Ultra NVL576 (2027 release):
15 EF FP4 inference
5 EF FP8 training
14x performance of GB300 NVL72
Looking Further: Feynman Architecture in 2028
Huang also revealed that after Rubin, the next major AI architecture will be named Feynman, after legendary physicist Richard Feynman. NVIDIA’s roadmap now clearly outlines a three-generation AI chip trajectory:
Blackwell Ultra (2025)
Rubin (2026)
Feynman (2028)
Conclusion: NVIDIA Continues to Dominate AI Computing
With Blackwell Ultra, NVIDIA is pushing AI computation to new heights, and its long-term roadmap through Rubin and Feynman cements its position as the industry leader in high-performance AI and supercomputing. As AI models become more complex, NVIDIA’s accelerated hardware advancements will play a crucial role in shaping the future of AI-driven industries.
