NVIDIA Blackwell B200 Ships in Volume: 2.5x H100 Performance Transforms AI Training Economics

NVIDIA's Blackwell B200 GPU has begun shipping in volume to major cloud providers and hyperscalers, delivering 2.5x the training throughput of the H100 for transformer models. The B200 features 192GB HBM3e memory and 10 PB/s NVLink bandwidth in an 8-GPU NVL72 configuration. AWS, Google Cloud, Microsoft Azure, and Oracle Cloud have announced Blackwell-based instances. First customer benchmarks show GPT-4 class model training 40% faster than H100 DGX at comparable cost, significantly changing the economics of large-scale AI development.