Buying guides
Rack server buying guide
Matching compute, memory and storage to the workload, with enough headroom to grow and enough redundancy to survive a failure.
Published
July 9, 2026
Topic
Buying guides
Talk it through
Bring us the requirement and we will price it.
Servers are bought less often than laptops and hold more of the business, which makes the specification decision unusually consequential. Buying too little means an early replacement; buying too much means paying for capacity that will never be used. This guide covers how to size a host around the workload, and the redundancy decisions that determine whether a single failure is an inconvenience or an outage.
Sizing around the workload
Start from measurements of the systems you are replacing, taken over a representative period. Peak anecdotes are a poor guide, and a specification built on one bad afternoon leads to an expensive machine doing very little.
- Processor: core count against sustained utilisation, with headroom for growth
- Memory: the most common bottleneck, and the most expensive thing to add later in some platforms
- Storage: separate the capacity requirement from the throughput requirement, because they are different decisions
- Network: confirm whether the workload needs 10Gb or whether 1Gb with teaming is genuinely enough
Redundancy that earns its cost
- Dual power supplies on separate circuits, so a single supply or circuit does not stop the server
- RAID with a hot spare, and a rebuild tested rather than assumed
- Redundant networking so one failed switch port does not isolate the host
- A second host or a documented recovery path if the workload cannot tolerate downtime
A server is only as available as the weakest thing in front of it. Check the UPS, the switch and the cooling as part of the same decision.
Virtualisation and service life
Most servers now run virtual machines rather than a single operating system, which changes the sizing question. A host needs enough memory and storage for every workload it will carry, plus room to absorb a workload moved from a failed machine.
- Size hosts so the estate survives losing one of them
- Confirm the virtualisation platform licensing and the guest operating system entitlements
- Plan the storage for growth, since storage is usually extended long before processors are
- Match the warranty to the intended service life and keep it continuous