A host's plan grid: vCPU, fair use and burst
Most hosts arrange their plans in a grid: several plan lines, each with resources growing in steps. The same "2 vCPU" means different things in different lines: in one it is a share of a core split with neighbours, in another the core is yours alone, and in a third it is the right to a short burst. Here is how to read the grid and what to check before you pay.
The short answer. First find the plan line in the grid: shared cores (shared vCPU), dedicated cores (dedicated vCPU) or "burstable" with a baseline. Then look at the memory-to-core ratio and how much traffic is included. A website, a bot or a control panel is usually fine on shared cores. Code builds, transcoding and calculations that keep the CPU busy for hours need dedicated ones.
How the grid is built
Usually each step in a plan line doubles the memory and cores of the previous one. The accsly.io catalog on 16 November 2026 lists 29,131 VPS plans at 2,176 hosts. The most common ratio is 2 GB of memory per core (36% of plans), followed by 1 GB (18.5%) and 4 GB (14.7%). Plans with 8 GB per core make up 4.4%. Large clouds split their grid into lines precisely by this ratio: at DigitalOcean, according to its documentation, General Purpose has 4 GB per vCPU, CPU-Optimized 2 GB and Memory-Optimized 8 GB.
Prices grow from step to step in different ways. Examples as recorded in the catalog, per month:
- Linode: 1 vCPU and 2 GB — $12, 2 vCPU and 4 GB — $24, 4 vCPU and 8 GB — $48. The price exactly doubles, traffic is 2, 4 and 5 TB;
- Hetzner: CPX12 (1 vCPU, 2 GB) — €11.99, CPX22 (2 vCPU, 4 GB) — €19.99, CPX32 (4 vCPU, 8 GB) — €35.99. Each next step is cheaper per core, and traffic in Germany and Finland is 20 TB.
That is why it is easier to compare plans by the price of a resource unit rather than the price of the plan: how to do that is covered in How to compare VPS prices: the price of a resource unit.
Shared and dedicated vCPU
With shared cores, several servers split the host's processor. DigitalOcean describes it like this: under load the hypervisor gives the server more hyper-threads, but if the neighbours are busy too, the server may get only fractions of a thread. Dedicated cores are assigned to a single customer, so performance does not depend on neighbours. The price difference for the same resources:
| Host | Shared | Dedicated |
|---|---|---|
| DigitalOcean, 2 vCPU and 4 GB | Basic — $24 | CPU-Optimized — $42 |
| Linode, 2 vCPU, 4 GB and 80 GB | Linode 4 GB — $24 | G6 Dedicated 4 GB — $36 |
| Hetzner, 2 vCPU | CPX22, 4 GB — €19.99 | CCX13, 8 GB — €43.49 |
Dedicated cores cost one and a half to two times more. They are worth paying for if the CPU is loaded constantly rather than in short spikes.
Burst and baseline performance
The third type is plans with a baseline. The server can always use a share of the CPU and, for a while, more than that, until it uses up its accumulated allowance. Amazon EC2 counts the allowance in credits: one credit is one core at 100% for one minute; t3.nano has two vCPU and a 5% baseline. When the credits run out, in standard mode the server falls back to the baseline.
Amazon Lightsail works the same way. The cheapest $5 plan has two vCPU, but the baseline is 5% per core; the 4 GB plan at $24 has 20%, the 8 GB plan at $44 has 30%. In Selectel's Shared Line you choose a guaranteed share of a core when ordering — 10, 20 or 50%. According to the documentation, performance never drops below that share and can temporarily reach 100%. In the catalog such a plan with a 10% share and 0.5 GB of memory costs 329 ₽ a month.
For a website with occasional peaks, burst is a good deal. For a task that loads the CPU for hours, it turns into a slow server: once the allowance is gone, you are left with 5–30% of a core.
Fair use: CPU and traffic
A fair use rule (fair use, AUP) is the clause in the terms where the host describes what load it considers normal. For the CPU it applies to shared cores: for example, a long stretch at 100% without a break may lead to throttling. Technically, hosts limit the CPU with scheduler tools. In the Linux kernel this is, for example, a CPU time quota per period: once it is used up, processes wait for the next period.
For traffic, fair use concerns "unlimited" plans. In the catalog, traffic terms are known for 21,849 VPS plans, and 9,482 of them (43.4%) are unlimited. Of these, the port speed is given for 6,776: 1 Gbit/s for 3,084, 100 Mbit/s for 1,301. What unlimited means on a particular port is covered in Bandwidth and traffic: what 1 Gbps and unlimited really mean, and the other pitfalls of a plan in How to read hosting plans: the hidden limits.
How to check the cores after buying
Inside a virtual machine you can see how long the CPU waited for the host. In top output this is the st field, in vmstat the st column; in /proc/stat the same value is described as steal — "involuntary wait". If steal stays above a few percent, or the result of the same benchmark varies noticeably with the time of day, the server is sharing its CPU with busy neighbours. Run the test several times during the first day — the procedure is described in How to test VPS ping and speed before you pay. Hourly billing is handy for this: you delete an unsuitable server and pay only for the hours used. Cloud hosts with this kind of billing are compared in the rating of cloud VPS with hourly billing.
What to choose
- Website, bot, control panel, office VPN — shared cores, 2 GB per core; how much a website needs is worked out in How much RAM and how many cores a website needs.
- Database and cache — a line with 4–8 GB per core.
- Builds, transcoding, calculations, game server — dedicated cores; for single-threaded work, high clock speed.
- Test environment and occasional tasks — burst or the cheapest plans; a server without IPv4 is cheaper still if the task allows it — see IPv6-only VPS: savings and limitations.