Dedicated servers explained
One physical machine, rented whole. No hypervisor taking a cut, no neighbours competing for disk, and no ambiguity about what you are paying for. Also no abstraction layer to save you when the hardware fails.
- You get
- The entire machine: all cores, all RAM, all disks, and usually an out-of-band console
- Performance
- Predictable — nothing else is scheduled on your CPUs
- Provisioning
- Hours to days, not seconds
- Failure mode
- Hardware is yours to wait on; redundancy means a second machine
What you actually gain over a large VPS
Three things, and they are narrower than the marketing suggests.
- Predictability. No oversubscription, no steal time, no neighbour's backup job saturating the disk queue. Latency percentiles get tighter, which matters far more than peak throughput for databases and real-time workloads.
- Hardware you specify. Specific CPU generations, NVMe in a particular RAID arrangement, ECC memory, a GPU, more disks than any VPS plan offers.
- No hypervisor tax. Modest for most work, significant for I/O-heavy and virtualisation-nested workloads.
What you do not gain is uptime. A single dedicated server is a single point of failure with a replacement time measured in hours. Two VPS instances in different failure domains are more available than one excellent machine.
Managed, unmanaged, and what sits between
| Unmanaged | Managed | |
|---|---|---|
| Hardware replacement | Provider | Provider |
| Network and power | Provider | Provider |
| OS install | Provider, once | Provider, ongoing |
| Patching, firewall, backups | You | Provider |
| Application stack | You | Varies — ask explicitly |
| Out-of-hours response | You | Per the SLA, if there is one |
Most dedicated servers are sold unmanaged with a remote console (IPMI, iDRAC, iLO) and a rescue mode. That console is the thing to verify before you need it: if the only way back into a misconfigured machine is a support ticket, your recovery time is the provider's queue length.
Reading a dedicated spec sheet
- Disks and RAID. Two disks mirrored is not a backup, but it is the difference between a failed disk and a failed weekend. Ask whether RAID is hardware or software, and whether the controller has a battery-backed cache.
- ECC memory. Standard on server hardware, absent on the cheapest consumer-grade offers. For anything holding data, insist on it.
- Bandwidth. Port speed is not the same as included transfer. A 1 Gbps port with a monthly cap is a common and reasonable arrangement; unmetered on a shared uplink is sometimes neither.
- IPMI access. Whether you get it directly or must ask.
- Setup fee and contract length. Cheap monthly rates often sit behind a one-off provisioning charge or a year's commitment.
When it is the right answer
Sustained high load rather than spiky load; databases that are sensitive to I/O latency; licensing that is priced per physical socket; compliance requirements that forbid shared tenancy; and the point at which the metered bill for equivalent cloud capacity exceeds the rental several times over — which, for steady workloads, arrives sooner than most teams expect.
It is the wrong answer for anything bursty. Paying for peak capacity around the clock is exactly what elastic platforms exist to avoid — see cloud hosting.