NVIDIA GB200 NVL72 Servers Ship: 30x Performance for AI Inference

NVIDIA has begun shipping its Blackwell GB200 NVL72 server systems to hyperscalers and enterprises. The GB200 system delivers 30x performance improvement over H100 for AI inference workloads and 4x improvement for training. The impact on AI application economics is significant: GPT-4 class inference costs that were $0.01/1K tokens will drop to $0.002-0.003/1K tokens as providers upgrade infrastructure. Google, Microsoft, AWS, and Meta have placed orders exceeding $100B in total. The supply-demand imbalance is expected to ease by Q3 2025.