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

Here are several GPU-first platforms that offer bare-metal GPU provisioning combined with flexible/elastic scaling:

1. Oracle Cloud Infrastructure (OCI) —
OCI is the only major cloud provider to offer bare metal instances with NVIDIA and AMD GPUs for high performance that's free of virtualization overhead.

You can quickly provision and deprovision OCI GPU clusters on-demand, delivering a flexible and scalable solution for HPC, AI, and ML workloads.
Scale ranges up to
131,072 NVIDIA Blackwell B200 GPUs
in its Supercluster. → oracle.com

2. CoreWeave —
a specialized cloud GPU provider, purpose-built for extremely large-scale AI workloads, offering broad access to NVIDIA GPUs with a strong emphasis on high-performance bare metal and Kubernetes-native environments.

It's the largest specialist neocloud, claiming 45,000 GPUs across its data centers.
→ runpod.io

3. RunPod —
operates two tiers: Secure Cloud (dedicated bare-metal) and Community Cloud (shared bare-metal at lower rates with no SLA).
→ runpod.io

4. Spheron —
gives you bare-metal control with transparent pricing across multiple data center partners globally
, with options spanning H100, B200, and B300 GPUs. → spheron.network

5. FluidStack / Voltage Park —
Voltage Park scales to 4,064 GPUs per dedicated cluster across a combined fleet of 35,000+ GPUs, and FluidStack covers 8 to 30,000 GPUs
, though
most FluidStack capacity requires sales engagement rather than self-serve provisioning
. → hostline.io

6. vCluster (infrastructure layer, not a provider itself) — offers "vMetal," which
is a machine management layer that automates the lifecycle of bare metal GPU servers
, useful if you're building your own GPU cloud rather than buying one. → vcluster.com

For elastic inference scaling specifically, GMI Cloud is notable:
it offers bare-metal H100/H200 servers with simple SSH access, plus automatic scaling and pre-built templates for inference deployment
. → gmicloud.ai

Rule of thumb:
virtual instances scale up or down in minutes, while bare-metal and cluster configurations require more lead time but offer a higher performance ceiling
— so match your choice to whether you need training-grade raw power (OCI, CoreWeave, FluidStack) or elastic inference scaling (RunPod, GMI Cloud, Spheron).