A bare-metal node is a Hetzner Robot dedicated server that Syself Autopilot provisions and runs the same way it runs a cloud node. A dedicated server gives you the whole physical machine, with no hypervisor and no other tenants sharing its CPU, disk, or network.

This page is the entry point for the bare metal section. It mentions when bare metal fits, what its limits are, and then points you to the pages that set it up, operate it, and explain the detail.

When to use bare metal 

Reach for bare metal when the work keeps machines busy or needs real hardware:

• Steady base load that runs around the clock.
• Databases and other I/O-heavy workloads on local NVMe .
• GPU workloads , which need physical accelerator cards.
• The secure (Kata) runtime , which needs hardware virtualization (KVM).
• Dedicated control planes that stay independent of cloud capacity.
• Single-tenant physical hardware for compliance .
• Lower cost once the cluster grows, see the economics of owning versus renting for the detail.

The limits 

A bare-metal pool is the set of servers you have already ordered and registered, so you cannot spin up a node the moment you need one the way you can on cloud. Adding capacity means ordering hardware ahead of time, so size pools for your peak or pair them with a cloud pool for bursts.

Bare-metal servers keep their data on local disks that, unlike most cloud storage, are not replicated for you. That data lives on a single machine, so a workload rescheduled onto another node loses access to it. Replicate it at the application level or back it up if you cannot afford to lose it.

Pending csi-controller pods without cloud workers 

On a cluster with no cloud workers, the csi-controller pods in kube-system stay Pending. This is expected: the controller runs only on cloud workers, so it has nowhere to schedule. The Pending pods are harmless and do not affect the cluster or your workloads.

Where to go next 

The pages below are organized by tasks. Feel free to checkout one that matches your goal.

Set up

• Order and prepare a Robot server : buy a Hetzner dedicated server, clear its disks, and find the disk WWN.
• Add bare-metal servers : register servers as HetznerBareMetalHost resources and attach them to a cluster.

Mix with cloud

• Mix cloud and bare-metal pools : run a bare-metal base with cloud workers in one cluster.

Operate and maintain

• Wipe a disk before provisioning : clear RAID or a foreign OS from the rescue system.
• Machine health checks and remediation : how Syself Autopilot replaces unhealthy machines, and how to pause it for maintenance.

Access and troubleshoot

• Serial console and rescue system : get a shell on a server that will not boot.
• A server that will not provision : unstick a bare-metal host that is stuck provisioning.

Plan and cost

• Size a cluster for cost : where bare metal beats cloud and how to right-size the pools.
• Dedicate node pools per tenant : isolate tenants on their own bare-metal pools.

Related

• GPU nodes : GPU servers are bare metal and have their own section.
• For the deeper detail, see the node lifecycle .