Amazon just launched EC2 P6e-GB200 UltraServers powered by NVIDIA GB200 NVL72. These are the top-tier GPUs for AI training and inference on AWS.
The UltraServers link multiple EC2 instances with a high-bandwidth, low-latency interconnect. The NVIDIA Grace Blackwell Superchips combine two Blackwell GPUs and an Arm-based Grace CPU via NVLink-C2C. Each superchip pumps out 10 petaflops of FP8 compute and up to 372 GB of HBM3e memory.
You get up to 72 NVIDIA Blackwell GPUs in a single NVLink domain—totaling 360 petaflops of compute and 13.4 TB of ultra-fast GPU memory. The AWS Nitro System lets these UltraServers scale to tens of thousands of GPUs. 
Network speeds hit 28.8 Tbps with Elastic Fabric Adapter (EFAv4). EFA pairs with NVIDIA GPUDirect RDMA for GPU-to-GPU comms with OS bypass.
Specs break down into:
| UltraServer type | GPUs | GPU memory (GB) | vCPUs | Instance memory (GiB) | Storage (TB) | EFA Bandwidth (Gbps) | EBS Bandwidth (Gbps) |
|---|---|---|---|---|---|---|---|
| u-p6e-gb200x36 | 36 | 6660 | 1296 | 8640 | 202.5 | 14400 | 540 |
| u-p6e-gb200x72 | 72 | 13320 | 2592 | 17280 | 405 | 28800 | 1080 |
These servers target massive AI workloads: trillion-parameter models, mixture of experts, generative AI, reasoning, speech, video, code.
You can spin these up now in the Dallas Local Zone using EC2 Capacity Blocks for ML via the EC2 console. Pay upfront, run instances with AWS CLI, Management Console, or SDKs. Start with AWS Deep Learning AMIs preloaded with PyTorch, JAX, and more.
Integration works with:
- Amazon SageMaker Hyperpod for managed fault tolerance and ultra-performance.
- Amazon EKS for Kubernetes with multi-UltraServer node groups and topology-aware routing.
- Amazon FSx for Lustre and local NVMe storage (up to 405 TB) for high-throughput AI data.
Amazon EC2 P6e-GB200 UltraServers are available now at Dallas Local Zone (us-east-1-dfw-2a). Check the Amazon EC2 pricing page for details.
You can try them out in the Amazon EC2 console or learn more on the Amazon EC2 P6e instances page.
Channy from AWS shared this update.
