GB300 NVL72 & HGX B300 NVL16: Hardware Implications

Question

Jensen announced at Computex 2026 a platform with two key configurations: the GB300 NVL72 rack-scale solution with 72 Blackwell Ultra GPUs and 36 Grace CPUs delivering 1.1 exaflops of FP4 computation, and the HGX B300 NVL16 system offering 11x faster inference on large language models compared to Hopper-based systems. From a hardware perspective, what are the implications? Is it sellable today? Can it be plugged into existing infrastructure? What's the impact on the ecosystem today? Focus on power/cooling requirements, rack compatibility, networking needs, supply chain readiness, and competitive positioning.