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

For GPU-first bare-metal provisioning in AI infrastructure orchestration, the most relevant approaches in your results are NVIDIA GPU Operator, vMetal, and Kubernetes-native automation layers that treat GPU nodes as first-class managed hardware.[1][2][14]

• NVIDIA GPU Operator: Best if you want Kubernetes-native GPU enablement on standard OS images. NVIDIA says the operator lets admins use the same base OS for CPU and GPU nodes, then installs and manages the GPU software stack via Helm, including GPU Feature Discovery for node labeling and scheduling.[1]
• vMetal: Best if you want a bare-metal machine management layer for AI factories or neoclouds. vMetal automates machine discovery, PXE boot provisioning, OS install, lifecycle management, and assignment of GPU servers into Kubernetes clusters or infrastructure pools.[2][3][7]
• Mirantis-style Kubernetes GPU orchestration: Best if you already run Kubernetes and want to combine device plugins, Node Feature Discovery, labels, and scheduling to place workloads on the right GPU types and sizes.[14]
• OpenStack Ironic / MAAS / Cobbler + Ansible/SaltStack: Best if you need a more traditional bare-metal provisioning stack. These tools are cited as automated deployment frameworks using PXE, IPMI, and Redfish, with configuration management for OS and driver setup.[5]
• Provider-managed bare metal GPU platforms: If you want infrastructure rather than software, Oracle OCI bare metal GPU offerings and DigitalOcean Bare Metal GPUs are examples of dedicated hardware options, though their provisioning model is more vendor-driven than orchestration-first.[10][11]

A practical GPU-first stack usually looks like this:

• 1. Discover and provision hardware with PXE/IPMI/Redfish or a machine-management layer like vMetal.[2][5]
• 2. Install OS and GPU drivers on a standard image.[1][5]
• 3. Register nodes into Kubernetes and apply GPU-aware labels/feature discovery.[1][14]
• 4. Schedule workloads using GPU labels, device plugins, and cluster policies.[1][14]

If you want, I can turn this into a vendor shortlist or a reference architecture for AI clusters.