AWS expands NVIDIA Blackwell EC2 capacity with P6-B200 and P6-B300 in new Asia-Pacific regions

AWS expanded its NVIDIA Blackwell footprint across two new Asia-Pacific regions simultaneously. The P6-B200 instances arriving in Hyderabad pack 8 Blackwell GPUs with 1,440 GB of aggregate GPU memory and 60% higher memory bandwidth, delivering up to 2x the AI training and inference performance of the prior P5en generation. The P6-B300 Blackwell Ultra instances in Jakarta go further with 8 GPUs, 2.1 TB of GPU memory and 6.4 Tbps of EFA networking — 2x the networking bandwidth and 1.5x the GPU memory of the previous generation.

The regional angle is the story: bringing frontier training and inference capacity to Hyderabad and Jakarta reduces latency and addresses data-residency needs for the fast-growing Indian and Southeast Asian AI markets, rather than forcing customers to route to US or European regions. The higher memory ceilings (2.1 TB on B300) matter specifically for serving very large models and long-context inference without sharding across nodes.

Competitively, this is AWS racing to deploy Blackwell capacity as demand for training compute stays acute and NVIDIA aggressively finances buildout across the industry. It also reinforces the AWS-NVIDIA relationship even as AWS pushes its own Trainium silicon. For customers, more regional Blackwell availability means more choice on where to run workloads under local compliance regimes, and the memory/networking upticks directly enable larger agentic and multimodal deployments. The practical watch item is capacity availability — new-region GPU instances frequently launch constrained, and whether supply meets the region's surging demand will determine real-world impact.