NVIDIA Blackwell Ultra GPUs (GB300)

The NVIDIA Blackwell Ultra GB300 delivers 1.5 petaFLOPs of FP4 AI compute per chip, a 50% leap over the B100, backed by 288GB of HBM3e memory for massive model training. Combined with NVLink 5 interconnects, the NVL72 server rack becomes the standard for frontier AI workloads at $50-100 billion compute scale. This power outperforms #1 by 3x in raw inference throughput, and at 1.5 petaFLOPs, it's 40% faster than the typical rival AI accelerator. For data centers, this means cutting training times by a third.