FAQ

Questions, answered.

The short version of how Raw Metal works. Still curious? The contact page goes straight to us.

What does “whole-server” actually mean?

You get the entire physical machine — bare metal, full root access, no hypervisor and no other tenants. Every GPU, NIC, and PCIe lane is yours.

How is it billed?

A flat price for the whole server, billed at 100% over the contract term. There's no per-instance metering, no utilization tracking, and no surprise overage — you know the number up front.

Which GPUs do you offer?

Current-generation NVIDIA Blackwell: B200 and B300 nodes, eight GPUs per node, with NVIDIA ConnectX-8 SuperNICs included on board.

Can I run multi-node training clusters?

Yes. Single nodes run on internal 5th-gen NVLink, and multi-node clusters are connected with non-blocking InfiniBand (copper or optical) so large jobs scale across racks as one fabric.

How long are the contracts?

Servers are reserved on 36-month terms, which gives training and inference fleets a stable, committed home instead of re-bidding for capacity every quarter.

What if I already own the hardware?

We also broker power-dense colocation — sourcing the cabinets, power, and InfiniBand-ready space your cluster needs, and matching you to the right facility.

Where are the servers hosted?

In power-dense colocation engineered for sustained GPU draw, with up to 80 kW per cabinet and 415 V 3-phase power. Nodes are racked, cabled, and burned in before handoff.

How do I get started?

Send your workload details through the contact page — GPU count, node type, topology, and timeline. We come back with hardware, a topology, and a committed quote.