Are there multi-tenant Baremetal GPU provisioning platforms?

Yes, multi-tenant bare-metal GPU provisioning platforms exist and have become crucial for AI cloud providers, "neoclouds," and enterprises building internal AI factories. These platforms automate physical server life cycles while ensuring that tenants remain securely isolated from one another without sacrificing raw hardware performance. 0 5 7 11 12

The market generally splits into native bare-metal orchestration platforms, Kubernetes-driven control planes, and traditional infrastructure-as-a-service (IaaS) stacks. 13

1. Dedicated Bare-Metal GPU Orchestration Platforms

These tools handle the entire pipeline from discovering unconfigured physical servers in a rack to carving out secure, isolated hardware environments for specific tenants. 14 15

• vCluster / vMetal: This platform combined the vMetal automation layer with virtual control planes. vMetal automates the zero-touch PXE boot, OS installation, and NVIDIA driver setup. Simultaneously, it isolates tenants structurally by deploying standalone virtual Kubernetes control planes straight onto the bare metal, avoiding standard shared-cluster boundary vulnerabilities.
• NorthWind Bare Metal GPU-as-a-Service (BMaaS): NorthWind matches incoming multi-tenant requests to an automated hardware inventory. It dynamically boots physical GPU servers (such as NVIDIA DGX systems) and mounts isolated per-tenant storage. Multi-tenancy is enforced at the network layer using NVIDIA UFM (Unified Fabric Manager), EVPN, and Pkey isolation so that one tenant's east-west InfiniBand traffic cannot leak to another.
• Rapt.AI: This software acts as an intelligent layer on top of bare-metal servers to help cloud providers maximize GPU usage. Rather than manually partitioning hardware configurations, Rapt.AI optimizes dynamic multi-tenancy with model-aware slicing and usage-based billing features.

2. Multi-Tenant Bare-Metal Infrastructure Stacks

These open systems allow operators to combine traditional compute management tools with specific GPU slicing capabilities.

• OpenNebula + NVIDIA Niko: OpenNebula integrates bare-metal instances directly into standard management workflows. By combining it with hardware-level scheduling automation, operators can assign bare-metal physical GPU servers to distinct tenants using the same multi-tenant guardrails historically reserved for virtual machines.
• OpenStack (IaaS): Often utilized by operators like OpenMetal, a bare-metal architecture can be paired with OpenStack ironic (for physical machine provisioning). Multi-tenancy is maintained inside strict OpenStack project limits, while the raw GPUs are segmented using NVIDIA's Multi-Instance GPU (MIG) mode to guarantee hardware-level isolation down to the physical compute and memory blocks.
• Mirantis k0rdent AI + NVIDIA BlueField-3: This platform separates host operating system orchestration from administrative controls by leaning heavily on Data Processing Units (DPUs). By provisioning and configuring clusters entirely via the DPU, it provides absolute multi-tenancy for north-south traffic and prevents malicious tenants from breaking out into host storage networks.

Summary of Isolation Approaches

Depending on the specific platform chosen, tenant isolation is maintained through distinct methods:

Approach | Technology Used | Best For
--- | --- | ---
Network-Level | EVPN, VXLAN, InfiniBand Pkeys | Distributed multi-node training clusters.
API Control Plane | Virtual K8s clusters (vCluster) | Teams needing individual root/admin Kubernetes access.
Hardware-Level | NVIDIA MIG (Multi-Instance GPU) | Splitting massive bare-metal GPUs securely among multiple small tasks.
DPU-Level | SmartNIC / BlueField-3 Offloading | High-security environments where the host network must be completely isolated.

If you are looking to narrow down your options, tell me:

• What specific GPUs are you provisioning (e.g., NVIDIA H100Go to product viewer dialog for this item., A100Go to product viewer dialog for this item.)?
• Do your tenants need individual physical servers or sliced fractions of a single GPU?
• Are you looking for a Kubernetes-native interface or a traditional cloud API / dashboard?

This will help tailor the exact stack for your deployment architecture. 20 2

[[0] - Bare Metal Kubernetes GPU Tenant Isolation with vCluster](vcluster.com)
[[1] - Bare Metal GPUs-as-a-Service (BMaaS) | NorthWind Platform](northwind.co)
[[2] - Mirantis Delivers Secure Multi-Tenant Networking with NVIDIA ...](mirantis.com)
[[3] - Bare Metal Kubernetes Distributions for GPU Workloads - vCluster](vcluster.com)
[[4] - Introducing vMetal: Run Your GPU Data Center Like a Hyperscaler](vcluster.com)
[[5] - Multi-Tenant GPU Infrastructure Platform for Cloud Providers](rapt.ai)
[[6] - Building an On-Demand GPU Cloud: A Guide for ... - OpenMetal](openmetal.io)
[[7] - Top Bare Metal GPU Providers for AI Workloads - vCluster](vcluster.com)
[[8] - Provision & Use Bare Metal GPU Servers with NorthWind ...](youtube.com)
[[9] - Building a Managed Kubernetes Service for Your GPU ...](youtube.com)
[[10] - Introducing Bare Metal-as-a-Service with OpenNebula and ...](youtube.com)
[[11] - What is bare-metal provisioning? – TechTarget Definition](techtarget.com)
[[12] - How to Choose a Bare Metal Hosting Provider for Your AI Project](ipxo.com)
[[13] - Inside the AI Infrastructure Stack: From GPUs to Production LLM Applications](medium.com)
[[14] - Metal Cloud | AI Factory Documentation](ai-docs.fptcloud.com)
[[15] - Bare Metal GPU Provisioning Infrastructure Hidden Costs - vCluster](vcluster.com)
[[16] - Top 9 bare-metal cloud providers of 2025](techtarget.com)
[[17] - AI and OpenStack Ironic: Bare Metal Provisioning and Cloud-Native Infrastructure](linkedin.com)
[[18] - Fixed Capacity Spatial Partition, FCSP : GPU Resource Isolation Framework for Multi-Tenant ML Workloads](budecosystem.com)
[[19] - What Is GPU Sharing in Kubernetes? Strategies for AI Efficiency](vcluster.com)
[[20] - Why Enterprises Are Moving Generative AI On-Premises | Benefits, ROI & Deployment Guide](pryon.com)