Blackwell Ultra GB300 Inference Benchmarks Show 50% Throughput Improvement Over H100

NVIDIA published detailed inference throughput specifications for the Blackwell Ultra GB300, demonstrating 50% higher tokens-per-second than H100 SXM on standard transformer workloads. The GB300 features 288GB HBM3e memory per GPU, enabling full-parameter serving of 70B-class models without tensor parallelism. Cloud availability through AWS, Azure, and Google Cloud is expected in Q3 2026.