Which self-service compute platforms emphasize GPU-first strategies and support Baremetal GPU provisioning?

Several platforms stand out for combining a GPU-first strategy with self-service, bare-metal GPU provisioning:

1. NorthWind — Their Bare Metal GPU Service is built around onboarding physical GPU nodes and exposing them through self-service compute profiles.
This setup includes onboarding physical nodes into the NorthWind platform, defining system-level resource pools, configuring networking interfaces, and enabling self-service compute profiles, with servers connected via high-speed interconnects like NVLink and NDR InfiniBand.
NorthWind also offers a dedicated BMaaS product giving
customers direct access to the host operating system on dedicated GPU servers, for teams running large distributed training jobs or highly opinionated software stacks that need control
. (northwind.co docs.northwind.co)

2. vCluster Labs (vCluster Platform + vMetal + vNode) — Explicitly GPU-first, with
bare metal GPU provisioning meaning deploying Kubernetes and tenant workloads directly on physical GPU servers without a hypervisor or VM layer, preserving full GPU performance
. Their new vMetal layer
automates the lifecycle of bare metal GPU servers, from initial provisioning and machine assignment to upgrades and repurposing, allowing infrastructure operators to manage physical compute with cloud-like automation
, while
customers can spin up their own vClusters on demand, no operator intervention required
. (vcluster.com vmetal.ai)

3. Lambda — Its Bare Metal Instances give
selectable bare-metal GPU types, and once provisioned, direct, unmediated access to the CPU, memory, GPU, disk, and TPM
, offloading management via NVIDIA BlueField DPUs. (lambda.ai)

4. Oracle Cloud Infrastructure (OCI) —
Offers NVIDIA A100, H100, and AMD MI300X GPUs on bare metal instances with RDMA and InfiniBand support for high-speed training.
(northflank.com)

5. FluidStack — Uses
Atlas OS for bare-metal provisioning and Lighthouse for monitoring, claiming 95%+ of theoretical hardware performance with auto-restart on failures
. (hostline.io)

For neoclouds building their own offering, NorthWind and vCluster/vMetal are the leading "platform layer" providers; for renting bare-metal GPU capacity directly, Lambda, OCI, and FluidStack are strong choices.