Are there multi-tenant Baremetal GPU provisioning platforms?

Yes. There are platforms for multi-tenant bare-metal GPU provisioning, though the market splits into two somewhat different categories: provisioning entire GPU servers to tenants versus securely sharing GPUs/nodes among tenants.

The most relevant products/projects are:

• NVIDIA Infra Controller (NICo) — probably the closest match to a purpose-built “multi-tenant Bare Metal GPU cloud” control plane. It provides hardware discovery, firmware management, DPU/network isolation, bare-metal lifecycle management, and tenant sanitization. NVIDIA explicitly positions it for cloud operators allocating bare-metal AI infrastructure directly to tenants. (docs.nvidia.com)

docs.nvidia.com

• NVIDIA Base Command Manager — provisions and manages GPU-accelerated bare-metal clusters, from individual nodes to very large clusters. It can be combined with NVIDIA Run:ai, which supplies the higher-level tenant/project/resource-quota model. Run:ai's documentation specifically describes provisioning dedicated tenant Kubernetes clusters using BCM or OpenStack. (nvidia.com)

nvidia.com

• Canonical MAAS — general-purpose Bare-Metal-as-a-Service rather than GPU-specific. It handles discovery, PXE deployment, machine allocation/release and APIs, and can serve as the provisioning layer underneath a multi-tenant GPU cloud. Canonical explicitly describes the pattern where MAAS provisions customer servers while a higher layer provides multi-tenancy, UI and billing. (maas.io)

maas.io

• CoreWeave Kubernetes Service — an example of a commercial GPU cloud implementing this architecture rather than software you'd normally deploy yourself. CKS runs Kubernetes workloads directly on bare-metal GPU nodes and uses NVIDIA BlueField DPUs. (coreweave.com)

coreweave.com

There's an important distinction: "multi-tenant bare metal" usually means Tenant A gets physical GPU nodes 1–8 and Tenant B gets nodes 9–16, with network/storage/BMC isolation. The physical servers remain single-tenant during an allocation. Actually putting multiple mutually untrusted tenants on one physical GPU requires another isolation layer such as NVIDIA MIG, which supports hardware-partitioned GPU instances on supported GPUs. (docs.nvidia.com)

If you're looking to build something like a private CoreWeave/Lambda-style GPU cloud on your own hardware, I'd specifically evaluate NICo + BCM/Run:ai versus MAAS + Kubernetes/OpenStack.