NVIDIA Blackwell Ultra B300 Ships to Hyperscalers

Summary

NVIDIA's Blackwell Ultra B300 GPU has entered volume production and is shipping to Microsoft Azure, Google Cloud, and Amazon Web Services. The B300 features 288 GB of HBM3e memory and delivers approximately 1.5x inference throughput for large language models compared to its predecessor.

So What? — Market Impact

The B300 closes the memory gap that forced enterprises to use multi-GPU configurations for the largest models. The competitive pressure on AMD's MI400 intensifies.

Key Specifications

GPU Architecture: Blackwell Ultra

Memory: 288 GB HBM3e at 12 TB/s bandwidth

Inference: 1.5x throughput vs B200 on LLM workloads

TDP: 1200W (liquid-cooled)

Interconnect: NVLink 6, 1.8 TB/s GPU-to-GPU

Sources

NVIDIA GTC 2026 Keynote

Reuters: "NVIDIA ships next-gen AI chips to cloud giants"