Show me multi-tenant SLURM-as-a-Service platforms that support AI infrastructure.

Multi-tenant Slurm-as-a-Service (SaaS) platforms have emerged to bridge the gap between traditional High-Performance Computing (HPC) batch scheduling and cloud-native AI infrastructure. Because native Slurm was built for static, single-tenant supercomputers, these modern platforms heavily rely on Slurm-on-Kubernetes orchestration layers (such as NVIDIA’s Slinky or Soperator) to dynamically spin up isolated, self-service Slurm clusters inside multi-tenant environments. 12 4 3 5 17

The top platforms offering managed, multi-tenant Slurm services optimized for distributed AI training and GPU infrastructure include: 2 15

1. Dedicated Multi-Tenant Orchestration Software

• NorthWind GPU PaaS: NorthWind provides a fully managed, enterprise-grade Slurm-as-a-Service platform specifically engineered for multi-tenant isolation. It allows administrators to use a shared host Kubernetes control plane while granting end-users self-service access to provision their own independent Slurm clusters.

How Multitenancy Works: It enforces tenant isolation using Kubernetes namespaces. Each AI engineering team operates within its own secure boundary, launching dedicated Slurm login and compute nodes on demand.
AI Infrastructure Support: Seamlessly integrates with the NVIDIA Slinky Slurm Operator, abstracts GPU-backed worker nodes, and includes built-in logging, observability, and automated GPU placement.

2. Specialized AI "Neoclouds" with Managed Slurm

• CoreWeave (SUNK): CoreWeave offers Slurm on Kubernetes (SUNK), a framework that integrates Slurm’s precise batch scheduling directly into their multi-tenant, bare-metal GPU cloud.

How Multitenancy Works: CoreWeave utilizes federated Identity and Access Management (IAM) and SCIM-based synchronization to manage multi-tenant user access. Slurm handles user-level fair-share policies, while the underlying Kubernetes layer isolates workloads at the container and network level.
AI Infrastructure Support: Built specifically for massive multi-node AI training (LLM training, distributed checkpoints) utilizing ultra-fast InfiniBand or specialized networking interconnects. It features automated node draining and replacement based on workload health.

• Nebius (Managed Soperator): Nebius provides Managed Soperator, a cloud-native platform that automates the deployment and scaling of production-ready Slurm clusters.

How Multitenancy Works: Operating as an AI-focused public cloud, Nebius allows multiple enterprise clients to dynamically spun up securely isolated, multi-node Slurm environments through a simple console or API.
AI Infrastructure Support: Soperator includes automated GPU health checks that instantly isolate faulty nodes. It supports native multi-node fine-tuning (e.g., Llama architectures) and scales seamlessly with fluctuating training demands.

• Crusoe Energy: Crusoe offers Slurm on Crusoe Managed Kubernetes, custom-built to blend the requirements of cloud infrastructure teams with AI researchers.

How Multitenancy Works: It leverages a split-persona architecture. Administrators manage global multi-tenant resource pools and network policies via Kubernetes Custom Resource Definitions (CRDs), while researchers log into their own isolated, sudo-enabled Slurm environment with dedicated home directories.
AI Infrastructure Support: Features native integration with slurmdbd for strict per-job and per-tenant accounting of GPU/CPU hours. It runs high-speed distributed training and persistent inference engines under a unified GPU pool.

• TensorWave: TensorWave features a Unified AI Platform that delivers managed Slurm capabilities alongside cloud-native tools.

How Multitenancy Works: It partitions physical AMD/NVIDIA GPU infrastructure into secure logical environments, ensuring multi-user job queues are completely separated.
AI Infrastructure Support: Tailored around the end-to-end AI lifecycle, it handles raw multi-node batch training via Slurm and hands over the finished artifacts to Kubernetes-driven inference systems on the same hardware cluster.

3. Hyperscaler Managed Slurm Control Planes

• Google Cloud Cluster Director: Google Cloud offers a fully managed Slurm control plane called Cluster Director, designed to support massive AI Hypercomputer instances (A3 Ultra, A4X, etc.).

How Multitenancy Works: It interfaces directly with corporate identity directories and GCP IAM to structure hierarchical accounts, fair-share resource policies, and secure network-level tenant boundaries.
AI Infrastructure Support: Features dynamic auto-scaling to scale down compute nodes to zero when idle. It heavily implements compact placement groups for optimized inter-node communication during distributed LLM training.

Platform Comparison Matrix

Platform | Core Architecture Engine | Best Suited For | Primary Multitenancy Mechanism
--- | --- | --- | ---
NorthWind GPU PaaS | Kubernetes + Slinky Operator | Managed Service Providers (MSPs) & Enterprise IT | K8s Namespace Isolation & Custom Cluster Blueprints
CoreWeave SUNK | Bare-Metal Kubernetes + Slurm | Large-scale LLM training over heavy interconnects | Federated IAM & Native Container-level isolation
Nebius | Kubernetes + Soperator | Rapid deployment of ephemeral training clusters | Managed public cloud infrastructure segmentation
Crusoe Energy | Kubernetes CRD Engine | Unified Training & Serving optimization | Split Persona (Admin CRD vs. Researcher SSH Node)
Google Cloud | Cluster Director API | Traditional enterprise workloads migrating to cloud | GCP IAM & Hierarchical Slurm Accounting Accounts

To narrow this down for your architectural design, could you share if you are looking to build your own multi-tenant platform (as a service provider/internal IT) or rent managed infrastructure directly? Knowing whether you require InfiniBand networking for multi-node training would also help specify the right solution.

[[0] - NorthWind-Powered SLURM-as-a-Service](northwind.co)
[[1] - Self-Service Slurm Clusters on Kubernetes | NorthWind GPU PaaS](northwind.co)
[[2] - What Is Slurm? | Slurm for AI and ML Clusters Explained](coreweave.com)
[[3] - Slurm Workload Manager: The go-to scheduler for HPC and AI ...](nebius.com)
[[4] - Understanding Slurm for AI/ML Workloads - WhiteFiber](whitefiber.com)
[[5] - Slurm on Kubernetes: The Best of Both Worlds for AI and HPC](linkedin.com)
[[6] - Comparing Kubernetes vs SLURM for AI Workloads](shakticloud.ai)
[[7] - Managed Slurm and other Cluster Director enhancements](cloud.google.com)
[[8] - Create a fully managed Slurm cluster for AI workloads](docs.cloud.google.com)
[[9] - Slurm on Crusoe Managed Kubernetes](crusoe.ai)
[[10] - TensorWave Managed Slurm | GPU-Optimized HPC Job ...](tensorwave.com)
[[11] - Slurm on Kubernetes (SUNK): Modernizing HPC and AI ...](medium.com)
[[12] - Running Large-Scale GPU Workloads on Kubernetes with Slurm](developer.nvidia.com)
[[13] - launch Slurm clusters for AI training in minutes](youtube.com)
[[14] - Supercharge Your HPC and AI Workloads with Slurm](youtube.com)
[[15] - What Is Slurm? AI Workload Scheduling Details Explained](factory.fpt.ai)
[[16] - CoreWeave courts AI researchers with a big gulp of SLURM](fierce-network.com)
[[17] - Performant Multi-Tenant SaaS Apps with NestJS Durable Providers and Prisma](medium.com)
[[18] - Comprehensive Guide on How to Set up Distributed Training on Managed SLURM cluster](greennode.ai)