AWS Expands Blackwell GPU Instances Across More Global Regions

Amazon Web Services has widened the geographic footprint of its most powerful GPU instances, pushing both the P6-B300 and P6-B200 families into new regions as demand for large-scale AI training and inference capacity keeps climbing.

The P6-B300, AWS's top-tier offering, is now available as the p6-b300.48xlarge size in US West (Oregon), US East (N. Virginia), AWS GovCloud (US-East), Asia Pacific (Hyderabad, Seoul), and South America (Sao Paulo). Each instance packs eight NVIDIA Blackwell Ultra GPUs paired with 2.1 TB of high-bandwidth GPU memory.

Networking is where the B300 pulls ahead most sharply: it delivers 6.4 Tbps of Elastic Fabric Adapter bandwidth alongside 300 Gbps of dedicated ENA throughput. Compared to the earlier P6-B200, AWS says the B300 offers double the networking bandwidth, 1.5 times the GPU memory, and 1.5 times the FP4 GPU compute (without sparsity).

One detail in AWS's own materials doesn't fully line up. The official announcement's title and URL reference Jakarta as a newly available region for the P6-B300, and the page body even lists "Hyderabad, Jakarta, Seoul" among supported locations. Yet the regions independently confirmed as live for the B300 are Hyderabad, Seoul, and Sao Paulo, with no corroborating evidence that Jakarta is actually serving these instances yet. The discrepancy appears to be a labeling or publishing error on AWS's part rather than a confirmed rollout.

P6-B200 rollout continues in parallel

Separately, AWS also extended the P6-B200 line, accelerated by standard NVIDIA Blackwell GPUs, into the Asia Pacific (Hyderabad) region. That instance type is now available in p6-b200.48xlarge size across US West (Oregon), US East (N. Virginia, Ohio), both AWS GovCloud regions, and Asia Pacific (Hyderabad, Mumbai).

Each P6-B200 instance runs on the AWS Nitro System and comes equipped with 1,440 GB of high-bandwidth GPU memory, plus up to 3.2 Tbps of EFAv4 networking. AWS claims the B200 delivers up to twice the training and inference performance of the previous-generation P5en instances, along with a 60% jump in GPU memory bandwidth.

For enterprises running frontier-scale model training, the expanded regional availability matters as much as the raw specs. Teams with data residency requirements in India, South Korea, Brazil, or U.S. government cloud environments now have direct access to Blackwell-class hardware without routing workloads through a single hub region, reducing latency and easing compliance constraints for AI workloads that previously had limited regional options.