# AWS outlines multi-team GPU sharing for SageMaker HyperPod with isolation controls

Digest AI · Enterprise & Industry · published 2026-10-08T16:20:04Z · updated 2026-10-13T15:17:09Z

Canonical: https://digestai.news/story/aws-outlines-multi-team-gpu-sharing-for-sagemaker-hyperpod-with-isolat

## Summary

AWS’s Machine Learning Blog detailed a reference architecture for **Amazon SageMaker HyperPod** to enable multiple teams within a company to share GPU clusters while enforcing isolation, fairness, and cost attribution. The setup uses **Amazon EKS** for cluster orchestration, **AWS IAM Identity Center** for centralized authentication, and **Kubernetes namespaces** to segment workloads by team. Each team gets its own **SageMaker AI domain**, **namespace**, and **compute quotas** via **HyperPod Task Governance**, ensuring they can’t overconsume resources or interfere with others. Storage is split into **POSIX-compliant file systems** (like FSx for Lustre) for shared datasets and **Amazon S3 buckets** with IAM-scoped access. The architecture also integrates **Amazon Managed Grafana** for team-specific observability dashboards, restricting each group to read-only access to their own metrics.

Key components include **EKS access entries** mapping IAM roles to Kubernetes permissions, **NetworkPolicies** to restrict pod-to-pod traffic, and **HyperPod Training/Inference Operators** for namespace-scoped jobs. The blog emphasizes this is designed for **trusted internal teams**—not for multi-cloud or adversarial isolation—highlighting that namespaces alone don’t prevent determined malicious actors from accessing shared nodes. The solution aims to balance operational simplicity with accountability, enabling teams to run distributed training, inference, and development workloads without administrative overhead.

## Key points

- AWS’s SageMaker HyperPod reference architecture lets multiple teams share GPU clusters with per-team isolation via Kubernetes namespaces and IAM roles
- Each team gets its own SageMaker AI domain, compute quotas, and storage (FSx/Lustre or S3) with IAM-scoped access to prevent overuse or interference
- Observability is restricted to team-specific Grafana dashboards, and workloads (training/inference) are namespace-scoped to enforce boundaries

## Why it matters

This architecture solves a critical pain point for AI teams sharing expensive GPU infrastructure, offering a scalable way to allocate resources fairly while maintaining security and cost tracking. It’s particularly relevant for enterprises running distributed training or inference workloads across multiple departments.

## Sources

1. [Share GPU clusters across teams with isolation and fairness using Amazon SageMaker HyperPod](https://aws.amazon.com/blogs/machine-learning/share-gpu-clusters-across-teams-with-isolation-and-fairness-using-amazon-sagemaker-hyperpod) (AWS Machine Learning Blog, 2026-10-08, primary source)
2. [Mixture-of-Experts kernel engineering on Amazon SageMaker AI](https://aws.amazon.com/blogs/machine-learning/mixture-of-experts-kernel-engineering-on-amazon-sagemaker-ai) (AWS Machine Learning Blog, 2026-10-13, primary source)

Part of the developing story: [AWS Expands SageMaker AI Agent Capabilities](https://digestai.news/thread/aws-shows-how-contextual-bandits-lift-conversions-in-acquisition-funnels) (4 stories)

## Cite

Digest AI, "AWS outlines multi-team GPU sharing for SageMaker HyperPod with isolation controls", 8 October 2026, https://digestai.news/story/aws-outlines-multi-team-gpu-sharing-for-sagemaker-hyperpod-with-isolat

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/aws-outlines-multi-team-gpu-sharing-for-sagemaker-hyperpod-with-isolat.json