What's New

AWS has expanded the availability of Amazon EC2 P5.48xlarge instances on SageMaker Studio notebooks to seven additional regions across North America, Asia Pacific, and Europe. These instances are powered by NVIDIA H100 Tensor Core GPUs and are designed to accelerate deep learning, HPC, and generative AI workloads. The expansion enables more global customers to leverage high-performance GPU compute directly within SageMaker Studio's JupyterLab and CodeEditor environments.

How It Works

• P5.48xlarge instances are equipped with NVIDIA H100 Tensor Core GPUs, which provide high-bandwidth memory (HBM2e) and NVLink interconnects for large-scale parallel computation.
• Users can select P5.48xlarge as the kernel instance type when launching JupyterLab or CodeEditor applications within SageMaker Studio, with no additional infrastructure setup required.
• The instances integrate natively with SageMaker Studio's managed notebook environment, allowing direct access to AWS storage, datasets, and SageMaker training/inference APIs.
• Workloads such as LLM fine-tuning, diffusion model training, and HPC simulations run directly on the H100 hardware, benefiting from NVIDIA's Transformer Engine and FP8 precision support.
• Pricing follows the standard SageMaker Studio notebook instance model, billed per second of instance usage, with rates available on the AWS pricing page.

Why It's Important

• Expanding to seven new regions reduces data residency and latency concerns for customers in APAC and Europe who previously had to route workloads to distant regions to access P5 instances.
• The H100-powered instances deliver up to 4x faster time-to-solution compared to previous-generation GPU instances (e.g., P4), directly compressing model iteration cycles for research and production teams.
• A claimed 40% reduction in ML training costs makes large-scale generative AI experimentation more economically viable, lowering the barrier for enterprises and startups alike.
• Availability within SageMaker Studio notebooks means data scientists can prototype, train, and evaluate LLMs and diffusion models in a single managed environment without provisioning separate EC2 clusters.
• Broader regional availability supports compliance with local data sovereignty regulations, particularly relevant for customers in the EU (London, Stockholm) and regulated APAC markets (Mumbai, Jakarta).

How It's Different

• Compared to P4d/P4de instances (A100 GPUs), P5.48xlarge offers H100 GPUs with FP8 Transformer Engine support, delivering significantly higher throughput for transformer-based model training.
• Unlike using P5 instances directly on EC2, the SageMaker Studio integration provides managed Jupyter/CodeEditor environments with built-in IAM, VPC, and lifecycle management, reducing operational overhead.
• P5.48xlarge provides 8x NVIDIA H100 GPUs with 640 GB of HBM2e memory per instance, enabling training of models that exceed the memory capacity of prior-generation GPU instances.
• The 3200 Gbps of EFA (Elastic Fabric Adapter) networking on P5 instances enables near-bare-metal MPI and NCCL communication performance for distributed training, which is not available on standard GPU instance families.
• Compared to cloud competitors offering H100-based notebook environments, the SageMaker Studio integration provides tighter coupling with AWS data services (S3, FSx for Lustre, SageMaker Pipelines) for end-to-end MLOps workflows.

When to Prefer It

• Choose P5.48xlarge on SageMaker Studio when fine-tuning or pre-training large language models (e.g., 70B+ parameter models) that require high GPU memory and fast inter-GPU communication.
• Prefer this option when your team needs an interactive notebook environment for rapid experimentation with diffusion models for image, video, or audio generation without managing underlying infrastructure.
• Use P5.48xlarge when your workload is latency-sensitive to regional proximity and your team is located in or serving customers in APAC (Tokyo, Mumbai, Sydney, Jakarta) or Europe (London, Stockholm).
• Select this instance type for HPC workloads such as molecular dynamics, climate modeling, or computational fluid dynamics that benefit from H100's double-precision FP64 performance.
• Opt for P5 on SageMaker Studio when you need to combine interactive prototyping with seamless handoff to SageMaker Training Jobs or Pipelines for production-scale runs.
• Consider this when cost efficiency is critical for large training runs, as the 40% cost reduction over previous-generation instances can yield significant savings at scale.

Availability

• Status: Generally Available (GA) as of May 12, 2026.
• New Regions: US West (San Francisco), Asia Pacific (Tokyo), Asia Pacific (Mumbai), Asia Pacific (Sydney), Asia Pacific (Jakarta), Europe (London), Europe (Stockholm).
• Supported Environments: Available as notebook instance types within SageMaker Studio's JupyterLab and CodeEditor applications.
• Pricing Model: Billed per second of instance usage under the SageMaker Studio notebook pricing model; specific rates available on the AWS SageMaker pricing page.
• Limitations: P5.48xlarge instances are subject to regional EC2 capacity availability and account-level service quotas, which may require a quota increase request via the AWS Service Quotas console.
• Prerequisites: Users must configure SageMaker Studio with appropriate IAM roles and, if required, VPC settings to launch P5 instances; refer to the SageMaker developer guide for setup instructions.