Show me self-service Baremetal GPU provisioning options.

If by “self-service Bare Metal GPU provisioning” you mean dedicated physical GPU servers that developers can provision/deprovision through a cloud console/API rather than sales-managed clusters, these are the strongest options I’d evaluate as of August 2026:

• Lambda Bare Metal Instances — Probably the closest fit. Lambda recently introduced API-driven bare-metal GPU instances, with self-service APIs covering instance lifecycle, VPCs, storage volumes, and governance. This is distinct from Lambda Private Cloud, which remains sales/reservation oriented. (lambda.ai)
• Vultr Bare Metal GPU — Dedicated physical GPU servers with H100, A100 PCIe/SXM, L40S, GH200 and AMD MI300X options. Vultr also offers self-service cluster provisioning and automates high-speed RoCEv2/InfiniBand networking where applicable. (docs.vultr.com)
• Oracle Cloud Infrastructure (OCI) Bare Metal GPU — Mature cloud-style provisioning with shapes such as BM.GPU.H100.8 (8× H100), alongside A100, A10, MI300X and other GPU families. Strong choice if you need VCN/IAM/API/Terraform-style infrastructure around bare metal and multi-node RDMA clusters. (docs.oracle.com)
• CoreWeave Bare Metal — NVIDIA-focused infrastructure including H100/H200, B200 and newer accelerator configurations. It provides cloud-like provisioning of dedicated hardware, although CoreWeave's platform is more Kubernetes/AI-infrastructure oriented than a conventional “rent a server and SSH into it” bare-metal cloud. (coreweave.com)

One former option to exclude from a new evaluation is Equinix Metal: it was shut down on June 30, 2026. (docs.equinix.com)

Useful starting points:
lambda.ai
docs.vultr.com
oracle.com
coreweave.com

For “credit card → API → physical 8-GPU node in minutes,” I’d shortlist Lambda, Vultr, and OCI first.