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

Here are several GPU-first platforms offering true bare-metal GPU provisioning with flexible scaling:

Lambda —
Deploy NVIDIA B200, H100, A100, or GH200 instances in minutes with self-serve, first-come access, with production-ready clusters from 16 to 2,000+ NVIDIA B200 or H100 GPUs.
Offers on-demand hourly and reserved pricing. (lambda.ai)

CoreWeave —
CoreWeave's infrastructure is optimized for compute-heavy tasks like AI training, visual effects rendering, and scientific simulations, and it offers bare-metal performance with minimal virtualization overhead.

It supports large multi-GPU clusters connected by high-bandwidth, low-latency interconnects (InfiniBand and NVIDIA GPUDirect RDMA) for distributed training, allowing tightly coupled workloads to scale efficiently across nodes.
(via runpod.io)

Atlas Cloud —
Offers dedicated bare metal H100, H200, GB200, and A100 GPU servers across EU, US, Singapore, and Japan with no virtualization overhead, InfiniBand, and flexible billing.

Choose from on-demand, hourly billing, or monthly custom plans, with dedicated bare-metal clusters tailored to your project needs supporting multi-region operations.
(atlascloud.ai)

Vultr Bare Metal / Vultr Clusters —
Vultr Clusters delivers on-demand, self-service Vultr Bare Metal server clusters, provisioned through Vultr's console or API, providing scalable and customizable CPU and GPU compute solutions.

You can create configurations to satisfy exacting workload needs or scale infrastructure to meet varying demand with ease, with no reservations or special requests required.
(vultr.com)

Yotta Shakti Cloud —
Provides direct access to NVIDIA HGX H100 (8x SXM5 with NVLink + NVSwitch) and L40S GPUs with no virtualisation overhead, plus 900 GB/s GPU-to-GPU bandwidth for seamless multi-node scaling.

You can start with a single node and scale to clusters.
(shakticloud.ai)

Spheron —
Offers bare-metal access to both A100 and H100 instances with per-minute billing and no contract requirements, available on-demand and as spot instances with NVLink SXM configurations.
(spheron.network)

For evaluation, compare on GPU generation availability (H100/H200/B200), network fabric (InfiniBand vs Ethernet), billing granularity, and cluster size flexibility to match your workload's scaling needs.