Amazon EC2 P6-B300 instances with NVIDIA Blackwell Ultra GPUs expand to new regions

AWS expanded availability of its EC2 P6-B300 instances to Asia Pacific (Hyderabad) and South America (Sao Paulo) on August 28, widening regional access to NVIDIA Blackwell Ultra-based compute for demanding AI training and inference. Each P6-B300 instance packs 8 NVIDIA Blackwell Ultra GPUs with 2.1 TB of high-bandwidth GPU memory, 6.4 Tbps of Elastic Fabric Adapter (EFA) networking, and 4 TB of system memory.

The generational gains are concrete: 2x networking bandwidth and 1.5x GPU memory versus the prior generation. For large-model training and high-throughput inference, memory capacity and interconnect bandwidth are frequently the binding constraints — bigger GPU memory reduces sharding overhead and lets larger context windows and batch sizes fit on-device, while faster EFA networking improves scaling efficiency across multi-node clusters.

Regional expansion into India and Brazil matters for data-residency and latency reasons. Customers in those markets can now train and serve frontier-class workloads locally rather than routing to US or European regions — relevant for sovereignty-conscious enterprises and governments, a theme echoed in Mistral's sovereign-AI deals this week.

The launch is part of the broader AWS-NVIDIA capacity surge (the 2 million-GPU expansion) and the industry-wide compute build-out. Competitively, it keeps AWS's GPU fleet current against Azure and Google Cloud, and against AWS's own Trainium silicon, which it is simultaneously pushing (OpenAI's ~2GW Trainium allocation). Watch pricing and availability constraints — Blackwell Ultra capacity remains supply-limited industry-wide — and whether more regions follow as AWS races to meet frontier-lab demand.