Hetzner Cloud or dedicated: which one fits the workload

Hetzner sells virtual cloud servers and physical dedicated servers, and the price gap between them per unit of raw resource is wide enough that the choice deserves real thought rather than defaulting to whichever you used last.

The deciding question is not how much compute you need. It is how quickly your requirements change, and whether you can keep a whole machine usefully busy.

The two products, honestly described

Cloud servers are virtual machines on shared physical hosts. They provision in seconds through the console or API, bill by the hour, resize between plans, and can be created and destroyed by automation. Snapshots, backups, private networks, floating IP addresses and load balancers all integrate cleanly.

Dedicated servers are physical machines allocated to you alone. Provisioning takes from minutes to a day depending on the configuration and whether it comes from the standard range or the auction market. Billing is monthly. You get the entire machine: every core, all the memory, the full disk throughput, and no noisy neighbours.

When elasticity is the deciding factor

If any of these are true, cloud servers are the answer regardless of price:

• Capacity needs to change within minutes rather than days.
• Servers are created and destroyed by automation, for CI jobs or ephemeral environments.
• You want to treat machines as disposable and rebuild rather than repair.
• The workload is small enough that a whole physical machine would sit largely idle.
• You need to be able to abandon the whole thing quickly without a monthly commitment.

Paying more per unit of compute for the ability to change your mind in seconds is usually a good trade at small scale.

When a whole machine is the answer

Dedicated servers become compelling when:

• Load is steady and predictable enough that you would keep the machine busy continuously.
• The workload is large enough to use a full machine’s cores and memory.
• Disk throughput and latency matter, because local NVMe on a dedicated machine outperforms most network-attached storage substantially.
• Performance consistency matters and you do not want to share cores.
• The workload is memory-hungry, where the price difference is at its most dramatic.

Large databases, storage-heavy applications, self-hosted platforms, build farms and long-running batch processing all fit this shape.

The hybrid that most people should actually run

The strongest arrangement is usually not one or the other. It is dedicated machines carrying the steady baseline, with cloud servers layered on for anything variable, joined by a private network so they behave as one environment.

A concrete example: the database and the primary application servers live on dedicated hardware, sized for steady load. Additional application servers are cloud instances that appear when traffic rises and disappear when it falls. CI runners are cloud instances that exist only while a build is running.

This gets the cost profile of dedicated hardware for the predictable majority of the workload and the flexibility of virtual machines for the part that genuinely varies.

The auction market, and its caveat

Hetzner runs a marketplace of previously-configured dedicated machines at reduced prices. The value can be excellent, particularly for storage-heavy or memory-heavy configurations.

Two caveats. Availability is whatever is listed at the moment, so you cannot rely on getting a specific configuration again later. And the hardware is older, which matters for power efficiency and for single-threaded performance, though rarely for throughput-oriented work.

Use the auction for workloads where the configuration is fungible and you can adapt to what is available. Avoid it where you need to provision an identical replacement quickly.

What people underestimate

Dedicated servers are not managed. Hardware failure means a replacement machine and a restore, not a live migration. Build for that: keep configuration in code, keep backups off the machine, and know how long a rebuild takes before you need to do one.

Provisioning time changes how you operate. With cloud servers you can respond to a capacity problem in minutes. With dedicated hardware you cannot, which means capacity planning has to happen before the pressure arrives rather than during it.

The private network is what makes the hybrid work. Connecting cloud and dedicated machines over a private network keeps traffic between them off the public internet and off your traffic allowance. Setting it up at the beginning is much easier than retrofitting it.

For the vendor’s own reference on the services involved here, see the Hetzner documentation.

Questions people ask