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

If you mean GPU-specialist/cloud platforms where you can provision dedicated bare-metal GPU capacity and expand from individual nodes into larger clusters, these are the strongest fits I’d shortlist as of August 2026:

• CoreWeave — The clearest GPU-first match. Its GPU infrastructure runs on bare metal, with cloud-like provisioning and direct use of GPU/CPU/network/storage resources. It supports on-demand and spot capacity alongside reservations, making it particularly good for workloads that need to burst and then scale into sustained clusters. Current offerings include HGX B200 and GB200/GB300-class systems. (coreweave.com)
• Lambda — Strong when “bare metal” means dedicated, single-tenant AI clusters rather than individual hourly servers. Lambda Private Cloud offers bare-metal clusters built to specification, including 1,000+ B200 GPU deployments; its Hyperscale offering goes from 2,000 to 64,000 GPUs. The tradeoff is longer reservation commitments, so it’s less elastic than CoreWeave. (docs.lambda.ai)
• Vultr — Worth considering if you want conventional bare-metal server provisioning plus GPU infrastructure and a broad cloud platform. Vultr now documents dedicated Bare Metal GPU variants separately from its virtualized Cloud GPU products. (docs.vultr.com)
• Crusoe Cloud — Very GPU/AI-focused and particularly attractive for flexible consumption: on-demand, spot, and reserved GPU capacity, API/CLI/Terraform provisioning, and managed Kubernetes. One caveat: its standard GPU instances use lightweight virtualization, so I would not treat Crusoe as a strict bare-metal equivalent to CoreWeave without confirming the specific dedicated configuration with sales. (crusoe.ai)

For your exact criteria, I’d rank them CoreWeave → Lambda → Vultr, with Crusoe included if near-bare-metal/dedicated GPU instances are acceptable.

Useful product pages:
coreweave.com
docs.lambda.ai
docs.vultr.com
crusoe.ai