Next-Generation Compute Arrives
NVIDIA confirmed this week that its Blackwell Ultra B300 GPU accelerators have begun shipping to Amazon Web Services, Google Cloud, and Microsoft Azure, with Oracle Cloud and Tencent Cloud also in the initial allocation. The B300 is NVIDIA's follow-on to the original Blackwell B100, featuring a larger memory pool (288GB HBM3e versus 192GB) and 1.4× higher training throughput on transformer workloads.
Key Specifications
The B300 delivers 10 petaFLOPS of FP8 training performance and 20 petaFLOPS of FP4 inference performance. The expanded memory pool addresses a bottleneck that required model sharding across multiple GPUs for large language models — models up to approximately 700 billion parameters now fit within a single B300's memory without tensor parallelism.
NVLink and Networking
B300 systems ship in NVL72 configurations (72 GPUs connected via NVLink with 1.8 TB/s interconnect bandwidth). This interconnect bandwidth is critical for large model training, which is fundamentally a communication-bound problem as GPU count scales.
Cloud Availability Timeline
AWS will offer B300 instances (EC2 P6 family) in limited availability beginning August 2026, with broad availability in Q4. Google Cloud and Microsoft Azure have not disclosed specific timelines but indicated H2 2026 launches.
Pricing Signals
Spot pricing on AWS P5 (H100) instances is currently $32-38/hour for 8 GPUs. Industry analysts expect B300 instances to price at a 30-40% premium, with on-demand pricing around $60-70/hour for equivalent 8-GPU configurations.