AWS previews Blackwell-based P6 instances for AI training in two regions, pricing to be announced

Amazon Web Services on Tuesday opened preview access to EC2 P6 instances powered by Nvidia’s GB200 Grace Blackwell platform in Northern Virginia and Oregon, aiming to cut training times for large language and multimodal models. AWS said the new instances will scale in UltraClusters for enterprise workloads and model providers, with general availability and pricing to follow.

The move matters because the largest buyers of AI compute are weighing next-generation GPU options across AWS, Microsoft Azure, and Google Cloud. Nvidia announced Blackwell in March 2024 and said it would ship in 2025, and by mid 2026 the race to stand up production clusters has become a key differentiator for cloud providers courting AI startups and Fortune 500 customers.

AWS positioned P6 for customers that currently use P5 H100 instances and Trainium2 on Amazon SageMaker. Nvidia lists Blackwell B200 with 192 GB of HBM3e memory per GPU and higher NVLink bandwidth than H100, which provides 80 GB of HBM2e or HBM3 depending on SKU. AWS said P6 will be supported in managed services including SageMaker, EKS, and Batch, and will offer multi-petaflop scale per node when configured with Nvidia’s Grace CPU.

AWS did not release on-demand pricing. As a reference point, the p5.48xlarge H100 instance lists at 98.32 dollars per hour on demand on the AWS price pages. Customers that need predictable throughput can use provisioned capacity on SageMaker, a model AWS says reduces queueing for large training runs compared with on-demand spot capacity.

The company said early access customers will run benchmark suites on Llama and other open-weight models during the preview. Nvidia has previously published MLPerf Inference results that show gains for newer architectures over H100, and analysts expect Blackwell to improve both training time and energy use for long-context LLMs compared with Hopper-class systems, although independent results for cloud-scale clusters are still emerging.

Microsoft has previewed ND B200 v6 systems on Azure, and Google has expanded TPU and GPU options in Vertex AI, which underscores the competitive dynamics AWS faces for AI-heavy accounts. Synergy Research estimates AWS held 31 percent of global cloud infrastructure spend in Q2 2026, with Microsoft at 25 percent and Google at 13 percent, and AI workloads are a primary driver of new reservations.

For the broader technology landscape, Blackwell-class capacity entering mainstream cloud fleets signals that architectural choices and workload portability will sit at the center of enterprise AI strategy, from model selection and fine-tuning to multi-cloud deployment and cost governance.