Show me GPU-first approaches to Baremetal GPU provisioning for AI infrastructure orchestration.

GPU-first bare-metal provisioning shifts the focus from traditional server lifecycle management to optimizing accelerator availability, interconnect fabric, and tenant isolation. Instead of treating GPUs as secondary peripherals attached to a CPU node, modern AI infrastructure orchestration treats the GPU, NVLink topology, and high-speed RDMA fabrics as the primary scheduling units. 5 1 6 2

The leading architectural approaches, software stacks, and platforms defining this space rely on specific paradigms:

1. Unified Control Planes (Bare-Metal to Cluster)

Historically, provisioning bare metal required a fragmented toolchain (IPMI/PXE → OS install → Driver compilation → Kubernetes join). Modern frameworks build the bare-metal provisioning agent directly into the Kubernetes control plane or cluster scheduler. 3 12 4

• vCluster Platform (vMetal): Operates as a machine management layer directly inside Kubernetes. vCluster vMetal discovers physical nodes, provisions them using PXE boot, flashes the operating system, maps the GPU topologies, and attaches them directly to tenant Kubernetes clusters. By combining this with dynamic auto-scalers like Karpenter, it allows for "zero-node" tenant clusters that provision physical bare-metal GPU servers on demand based on pending pod requests.
• NVIDIA NCX / InfraController (Niko): NVIDIA’s native architectural approach utilizes NVIDIA InfraController (Niko)—an open-source suite of microservices designed for zero-trust bare-metal management. Niko handles hardware discovery, firmware validation, DPU (Data Processing Unit) provisioning, and cryptographic sanitization for systems like the NVIDIA GB300 NVL72 AI factories.
• OpenNebula Integration: By pairing OpenNebula's cloud orchestration layer with NVIDIA NCX, infrastructure operators can manage bare-metal GPU instances and virtual machines using the same APIs and multi-tenant quota systems.

2. Automated Fabric & Silo Provisioning

AI clusters scaling beyond a few nodes are highly sensitive to network bottlenecks. A GPU-first approach treats network switches and DPUs as part of the unified bare-metal provisioning pipeline. 9

• MetalSoft for AI Factories: Platforms like MetalSoft automate bare-metal discovery and couple it with Fabric Provisioning. It dynamically configures network segments at the silicon level, isolating InfiniBand or RoCE (RDMA over Converged Ethernet) fabrics to prevent noisy neighbors or data leakage during multi-tenant training jobs.
• NVIDIA Network Operator: Used in tandem with the GPU Operator, it automates the deployment of the host networking components required to run GPUDirect RDMA. This eliminates manual InfiniBand driver and fabric configurations by using Kubernetes Custom Resource Definitions (CRDs) to establish high-speed secondary data paths.

3. Native Hardware Slicing and Virtual Nodes (vNodes)

When full physical isolation isn't required, or workloads are small (e.g., LLM inference vs. foundational model training), provisioning platforms leverage hardware-level partitioning. 7

• Multi-Instance GPU (MIG) Architecture: Rather than relying on software-level virtualization, orchestrators query the physical hardware to spin up discrete MIG profiles. This guarantees fixed Streaming Multiprocessor (SM) and memory bandwidth boundaries, which is crucial for predictable latency SLAs in multi-tenant environments.
• vNodes: Instead of virtualizing the OS via hypervisors, platforms use concepts like vNode to create virtualized node representations directly on the bare-metal compute nodes. This segments the workloads within bare-metal GPU environments, preventing container breakouts without incurring the performance penalties of a traditional Hypervisor layer.

Summary Comparison: Bare-Metal Orchestration Stacks

Approach / Platform | Core Provisioning Mechanism | Target Orchestrator | Primary Use Case
--- | --- | --- | ---
vCluster (vMetal) | Native K8s Machine Controller + PXE | Kubernetes | On-demand bare-metal auto-scaling for K8s tenants
NVIDIA NCX / Niko | Site-local zero-trust microservices | OpenNebula / K8s | Enterprise AI Factories & DGX SuperPOD deployments
MetalSoft | Bare-metal API / Terraform provider | SLURM / Run:ai / K8s | Multi-vendor bare-metal orchestration with silicon-level fabric slicing
OpenStack Ironic | Multi-tenant Bare Metal Cloud Service | OpenStack / TripleO | High-performance computing (HPC) over InfiniBand networks

Are you building a private AI cloud for multiple tenants, or optimizing infrastructure for a single-tenant training cluster? Let me know your preferred scheduling layer (Kubernetes vs. Slurm) so I can detail the specific driver and network integration path.

[[0] - Scaling AI Infrastructure for Engineering Teams](youtube.com)
[[1] - MetalSoft for AI Factories | Bare-Metal GPU Infrastructure ...](metalsoft.io)
[[2] - Bare Metal GPU Cloud for DGX SuperPOD with Tenant Isolation](youtube.com)
[[3] - Introducing vMetal: Run Your GPU Data Center Like a Hyperscaler](vcluster.com)
[[4] - From Bare Metal to AI Factory: End-to-End Multi-Tenant Automation ...](opennebula.io)
[[5] - Bare Metal GPU Provisioning Infrastructure Hidden Costs](vcluster.com)
[[6] - Why the top AI labs run Kubernetes on bare metal](spectrocloud.com)
[[7] - GPU Orchestration & CUDA: Scheduling, MIG, and Topology ...](rack2cloud.com)
[[8] - Introducing Bare Metal-as-a-Service with OpenNebula and ...](youtube.com)
[[9] - GPU Infrastructure: Automation and Strategy - Mirantis](mirantis.com)
[[10] - Overview — NVIDIA AI Enterprise: Bare Metal Deployment ...](docs.nvidia.com)
[[11] - RDG for Bare Metal GPU-Accelerated HPC and AI Workloads on ...](networking-docs.nvidia.com)
[[12] - How to Deploy AI Workloads on Bare Metal Without Slowing Time to ...](inflect.com)