NVIDIA GTC Focuses on Blackwell Ultra and Rubin

At GTC 2025, NVIDIA detailed Blackwell Ultra systems and its roadmap toward the next-generation Rubin architecture. The announcements emphasized rack-scale AI infrastructure rather than single accelerator cards.

Blackwell Ultra increased memory capacity and very-low-precision AI throughput, including FP4 operations aimed at large reasoning models. NVIDIA also demonstrated large inference systems where dozens of GPUs communicate through NVLink and high-speed networking as one coordinated platform.

The roadmap showed how rapidly AI hardware cycles were accelerating. Memory bandwidth, interconnect topology, cooling, and power delivery had become first-class design constraints because frontier models could not fit or run efficiently on one chip.