AWS launches EC2 P6 instances with Nvidia Blackwell GPUs in limited preview
Enzo Reno ·
Amazon Web Services opened limited preview of EC2 P6 instances on Tuesday, November 17, 2026, bringing Nvidia Blackwell GPUs to customers in US East N. Virginia, US West Oregon, and EU Ireland. AWS said P6 targets large language model training and high-throughput inference, aiming to reduce time to train and serve models for enterprise customers.
The launch matters because Blackwell is Nvidia’s first major architecture shift since Hopper. Nvidia said at GTC in March 2024 that Blackwell delivers up to 4 times training throughput and up to 25 times lower energy for LLM inference compared with H100, a baseline many enterprises still use.
AWS is offering two flavors at launch, single node P6 instances with B200 GPUs and GB200 backed UltraClusters for multi node training jobs. The clusters connect thousands of GPUs with Elastic Fabric Adapter and Nitro networking, and are accessible through EC2, Amazon SageMaker, Amazon EKS, and AWS Batch.
AWS said P6 integrates with SageMaker distributed training libraries, the Training Compiler, and managed checkpointing to Amazon FSx for Lustre. Customers can run Nvidia’s NIM inference microservices and NeMo framework through containers on EKS, which aligns with enterprise Kubernetes practices.
AWS did not publish on demand pricing today. The company said Savings Plans and Reserved Instances will be available at general availability later this quarter, which mirrors pricing models used for P5 H100. For context, AWS priced the P5.48xlarge at just under one hundred dollars per hour in 2024 in US regions, according to AWS pricing pages.
Nvidia’s GB200 pairs two B200 GPUs with a Grace CPU and NVLink Switch fabric. AWS said its UltraClusters expose this topology for higher memory bandwidth and interconnect scale, which is intended to support multi trillion parameter training and faster fine tuning of 70 billion parameter models common in production.
AWS framed the P6 rollout as capacity relief for customers juggling model training backlogs and latency constraints. The company pointed to internal benchmarks that showed multi node training jobs completing in a fraction of the time compared with P5, though it did not publish model names or dataset details today.
Competitively, Microsoft and Google have already previewed Blackwell based instances this quarter on Azure and Google Cloud. Synergy Research Group placed AWS at 32 percent of public cloud IaaS revenue in 2024, with Microsoft at 25 percent and Google at 11 percent, a spread that makes early Blackwell capacity a factor in workload placement decisions.
For enterprises, the takeaway is that Blackwell class GPUs are moving from roadmap to rentable capacity with managed tooling. That accelerates consolidation of training and inference on a single fleet, which can change vendor selection for AI platforms, MLOps pipelines, and long term capacity reservations in 2027.