Two servers with "2 vCPUs" can deliver very different performance. Providers sell CPU in three main ways: shared, burstable and dedicated. This guide explains shared vs burstable vs dedicated vCPU plans so you know what you are buying.
Before you start
- Background: CPU pinning and overcommit.
The three models
| Shared vCPU | Burstable | Dedicated vCPU | |
|---|---|---|---|
| Physical CPU | Shared with other VMs (overcommitted) | Baseline share plus credits to burst | Reserved cores or threads |
| Sustained 100% CPU | Possible, may be limited by fair use and neighbours | Only while credits last, then throttled to baseline | Yes, consistently |
| Price | Low | Lowest | Higher |
| Best for | Websites, small apps | Mostly idle workloads with occasional spikes | CI builds, databases, game servers, video encoding |
How burstable credits work
- The instance earns CPU credits while running below its baseline (for example 20% of a vCPU).
- Credits are spent when usage goes above the baseline.
- When credits run out, the CPU is throttled to the baseline, which can make the server feel very slow.
- Some providers offer an "unlimited" mode that charges for extra usage instead of throttling.
How to recognise contention
- Shared plans: rising CPU steal time at busy hours.
- Burstable plans: performance drops after a period of high use; check the provider's credit balance metrics.
Choosing
- Web servers with caching and modest traffic: shared vCPU is usually fine.
- Dev boxes, small bots, low-traffic apps: burstable can save money.
- Anything CPU-bound or latency-sensitive: dedicated vCPU.
- Unsure: start shared, monitor, and upgrade if steal or load shows contention.
Frequently asked questions
Is a dedicated vCPU a full physical core?
It depends on the provider: sometimes a full core, sometimes one hardware thread. Check the plan description.
Can I switch plan types later?
Often yes by resizing or migrating; check whether it needs a reboot or a new server.
Does shared hosting use vCPUs?
Shared hosting accounts get CPU limits per account rather than vCPUs; for example CloudLinux LVE limits on cPanel servers.