Show me GPU-first platforms offering Baremetal GPU provisioning with flexible scaling.

For GPU-first infrastructure where you specifically want dedicated/bare-metal GPU access plus the ability to scale capacity without buying hardware, these are strong options as of September 2026:

• Latitude.sh — One of the clearest fits for literal bare-metal GPU provisioning. Its “Metal GPU” line offers dedicated H100 systems ranging from 1× H100 nodes to 8× H100 SXM/NVLink nodes, with API/Terraform provisioning and multiple locations. Published pricing currently starts around $1.68/hr for a 1× H100 bare-metal node. (latitude.sh)

latitude.sh

• Vultr Bare Metal GPU — Dedicated physical GPU servers, including NVIDIA H100 configurations. It combines bare-metal isolation with Vultr's cloud provisioning model, making it a good choice when you want dedicated hardware without maintaining a physical cluster yourself. (docs.vultr.com)

docs.vultr.com

• Scaleway Elastic Metal / Dedibox GPU — Actual dedicated bare-metal H100 and L40S servers. Scaleway describes Elastic Metal as combining dedicated hardware with cloud-like deployment and scalability. Particularly attractive for European deployments. (scaleway.com)

scaleway.com

• Lambda — More cluster-oriented than individual bare-metal servers. Lambda Private Cloud provides single-tenant bare-metal GPU clusters, while 1-Click Clusters provide much more elastic access: currently 16–512 H100/B200 GPUs, with larger private clusters reaching thousands of GPUs. (docs.lambda.ai)

docs.lambda.ai

• CoreWeave — Worth considering if “bare-metal performance” rather than direct ownership of a bare-metal host satisfies the requirement. It is purpose-built around GPU workloads and scales from single-GPU configurations through 8× NVLink nodes and multi-node InfiniBand clusters. (coreweave.com)

coreweave.com

For strict “give me the entire physical GPU server via API and let me add/remove nodes,” I’d shortlist Latitude.sh, Vultr, and Scaleway. For scaling from experiments into hundreds/thousands of GPUs, Lambda and CoreWeave have stronger cluster-level offerings.