Dedicated CPU VPS: Which Plans Are Actually Dedicated
There is exactly one word that tells you a VPS has dedicated CPU cores, and it’s “dedicated.” Not premium, not turbo, not high-frequency, not performance. Those describe clock speed or marketing intent, and a plan can carry all of them while still sharing its cores with strangers.
The distinction is expensive enough to be worth getting right. Here are two verified examples of the same provider’s shared and dedicated tiers:
| Provider | Shared plan | Dedicated plan | Premium |
|---|---|---|---|
| DigitalOcean | Basic, 4 GiB / 2 vCPU — $24 | CPU-Optimized, 4 GiB / 2 vCPU — $42 | 1.75× |
| Hetzner | CX43, 16GB / 8 shared — €15.99 | CCX23, 16GB / 4 dedicated — €85.99 | 5.4× |
Read the Hetzner row twice. At the same 16GB of RAM, the dedicated plan costs more than five times as much and gives you half the cores. That is the price of not sharing.
The names that don’t mean dedicated
The brief version: if the plan page doesn’t say the cores are dedicated, reserved or not oversubscribed, assume they’re shared. Specific traps worth naming:
- “High Frequency” — Vultr’s High Frequency line runs on higher-clocked CPUs with NVMe. It is faster. It is still shared. Vultr’s genuinely dedicated line is Optimized Cloud Compute, sold separately.
- “CPX” — Hetzner’s CPX plans sound like a premium tier and cost considerably more than the CX line, but they are shared AMD vCPUs. Only the CCX line is dedicated. Within Hetzner’s own catalogue, price does not indicate dedication: CPX42 at €69.49 is shared, while CX43 at €15.99 is also shared, and only CCX carries the guarantee.
- “Premium”, “Turbo”, “Boost”, “Performance” — descriptive marketing, no resource commitment.
- “Dedicated resources” in a shared-hosting context often means a RAM allocation, not CPU isolation.
DigitalOcean is the clearest of the major providers on this: its Basic Droplets are explicitly documented as sharing vCPUs with other Basic Droplets, while General Purpose, CPU-Optimized and Memory-Optimized are all dedicated. The two dedicated lines differ only in memory ratio — CPU-Optimized gives 2GB per vCPU, General Purpose 4GB per vCPU.
Measure before you pay
Before spending 2–5× on isolation, find out whether contention is actually your problem. The measurement is one number: CPU steal time.
Steal time is the proportion of time your VM was ready to run but the hypervisor didn’t schedule it — because another tenant was using the physical core. It is the definitive evidence of the noisy-neighbour effect, and it’s free to check:
vmstat 1 10
Look at the st column, the last one. Then get a per-core view:
top # then press 1 — the %st figure appears per CPU
mpstat -P ALL 1 5
How to read it:
| Steal time | Meaning |
|---|---|
| 0–2% | Normal. Contention isn’t your problem. |
| 2–5% | Mild contention, usually tolerable. |
| 5–10% | Real contention. Worth investigating. |
| Above 10% | You are materially losing CPU you paid for. |
The arithmetic is worth stating plainly: a plan with 8 vCores and 50% steal time delivers 4 effective cores. And note that steal time only catches one of the three ways a shared plan slows down — quota throttling and credit exhaustion both show zero steal. Our shared vs dedicated vCPU explainer covers all three mechanisms and which diagnostic catches which. A high core count on a contended host is a number on an invoice rather than capacity.
Two other signals point the same way — sustained CPU usage at 80–90% for extended periods, and rising p95 or p99 latency while average latency looks fine. Averages hide contention; percentiles expose it.
If your steal time is under 2% and your CPU sits at 30%, dedicated cores will change nothing and you’ll have doubled your bill to prove it. That is the single most common way money gets wasted on this keyword.
The providers with genuinely dedicated cores
1. Hetzner CCX — dedicated, and still the cheapest
Hetzner’s CCX line is explicitly dedicated vCPU with guaranteed performance and no noisy neighbours, and despite the eye-watering premium over its own shared tiers it remains the cheapest genuine dedicated-core hosting among mainstream providers.
The catalogue is worth learning because the naming is unforgiving:
| Line | Type | Example at 16GB |
|---|---|---|
| CX | Shared Intel, cost-optimised | CX43 — €15.99 |
| CAX | Shared Ampere ARM | CAX31 — €20.99 |
| CPX | Shared AMD | CPX42 — €69.49 |
| CCX | Dedicated | CCX23 — €85.99 |
If you buy CPX expecting isolation because it costs four times CX, you’ve bought an expensive shared server. Only CCX is dedicated.
Check Hetzner Cloud pricing → · Read our Hetzner review
2. DigitalOcean — the clearest documentation
Three dedicated families — General Purpose, CPU-Optimized and Memory-Optimized — all with dedicated vCPUs, against Basic Droplets which explicitly share.
The reason it ranks this high isn’t price; it’s that DigitalOcean tells you plainly which is which, in documentation rather than sales copy. Its “Choosing the Right CPU Droplet Plan” guide carries a section headed “Shared CPU vs. Dedicated CPU” — the comparison stated as a concept to understand rather than buried in a footnote. In a category built on ambiguous adjectives, that’s worth something.
This is what the rest of the category should look like. The distinction gets its own heading.
General Purpose starts around $63/month and CPU-Optimized around $84, against $4 for entry Basic.
Pick CPU-Optimized for compute-bound work at 2GB per vCPU, General Purpose when you need more memory alongside the cores at 4GB per vCPU. Our 16GB tier comparison covers how these ratios play out at a fixed memory size.
Check DigitalOcean pricing → · Read our DigitalOcean review
3. Liquid Web — single-tenant, with support attached
Liquid Web’s dedicated offerings are genuinely single-tenant rather than a reserved slice of a shared host — its own marketing language for the GPU line is “single-tenant… fully dedicated servers,” and the same philosophy runs through its higher VPS and dedicated tiers.
What you’re buying beyond isolation is the managed layer covered in our managed VPS comparison: OS patching, panel management, monitoring with response, and 24/7 US-based support. For a production workload where contention is unacceptable and you don’t want to administer the box, that combination is the argument.
Priced well above the European hosts, as always.
Check Liquid Web pricing → · Read our Liquid Web review
4. Vultr — read the product name carefully
Vultr sells Optimized Cloud Compute with dedicated vCPUs, and separately sells High Frequency, which is faster-clocked but shared. Both sound premium. Only one is dedicated.
This is the single most common misunderstanding in this category, and it’s not Vultr’s fault — the naming is accurate if you read it literally, since “high frequency” describes clock speed and never claimed isolation. But a buyer scanning for the most impressive-sounding tier will pick wrong.
Vultr’s genuine advantage remains its 30-plus regions. We don’t have an affiliate relationship with Vultr, so there’s no tracked link here.
5. Nexcess — dedicated resources, application-tuned
Nexcess provisions with dedicated resources on plans tuned per application — WordPress, WooCommerce, Magento — rather than as a generic compute product.
Worth knowing that Nexcess and Liquid Web are the same group: Liquid Web’s site brands itself “Liquid Web by Nexcess” and carries a switcher between the two. If you’re gathering competing quotes, you’re talking to one company.
Check Nexcess pricing → · Read our Nexcess review
6. Cloudways — inherits from whatever you pick
Cloudways is a managed layer over DigitalOcean, AWS, Google Cloud, Vultr or Linode, so your CPU isolation is determined by the underlying instance type you choose rather than by Cloudways itself.
That’s a genuine feature here: you can select a dedicated-CPU instance and keep the managed platform on top. Just be aware the choice is yours to make and the cheaper default options are shared.
Check Cloudways pricing → · Read our Cloudways review
7. ScalaHosting — managed cloud with resource guarantees
ScalaHosting’s managed cloud VPS allocates resources per account with one-click vertical scaling, and its managed scope includes performance tuning — which several competitors explicitly exclude.
If contention is your problem and you’d rather someone else diagnose it, a managed provider that will actually investigate performance is worth more than raw isolation you have to troubleshoot yourself.
Check ScalaHosting pricing → · Read our ScalaHosting review
Who genuinely needs dedicated cores
Yes, pay the premium for:
- CI/CD runners and build servers. Builds are sustained, parallel and CPU-bound — precisely the profile shared plans throttle.
- Video encoding and media processing. Long, pegged-at-100% jobs.
- Game servers. Tick rate is latency-sensitive, and a stutter caused by a neighbour is visible to every player.
- Trading and anything with latency guarantees. Where a p99 spike has a financial cost.
- Databases under constant load, particularly with strict query-time SLOs.
- Agent loops and inference workers that run continuously rather than in bursts.
No, don’t:
- A WordPress site at 500 visits a day. Its CPU is idle almost all the time. This is the classic waste.
- A personal blog, portfolio or brochure site. Not close.
- A Discord bot or scheduled jobs. Bursty and tiny.
- A development or staging box. Occasional contention costs you nothing that matters.
- Anything where you haven’t measured steal time. Measure first.
The pattern: dedicated cores are for sustained load, not peak load. Shared vCPUs handle bursts perfectly well — that’s what the model is designed for. What they handle badly is running hot continuously, which is when the hypervisor starts scheduling you behind someone else.
How we picked
We applied one qualification rule: the provider must state that cores are dedicated, reserved or not oversubscribed. Marketing adjectives — premium, turbo, high-frequency — do not qualify, and we’ve named the specific plans that carry them while remaining shared. Prices and plan structures come from provider pricing pages and documentation read in August 2026.
We have not benchmarked steal time across these providers, and we’d distrust anyone who claims to have: steal time depends on which physical host you land on and what your neighbours are doing that week, so a single measurement isn’t a property of the provider. That’s why this article gives you the commands to measure your own instance rather than a table of someone else’s results.
Most links here are affiliate links. Note that the honest recommendation in the section above — measure first, and most readers will find they don’t need this — points away from the more expensive plans we’d earn more from.
FAQ
What is CPU steal time?
The percentage of time your virtual machine was ready to run but the hypervisor scheduled another tenant instead. Check the st column in vmstat 1 10. Consistently above 5–10% means you’re materially losing CPU you’re paying for; under 2% means contention isn’t your problem.
Is a dedicated CPU VPS worth the extra cost?
Only for sustained CPU load. Build servers, encoding, game servers, busy databases and always-on inference genuinely benefit. A typical website, bot or staging box does not — its CPU is idle most of the time, and paying 2–5× for isolation buys nothing measurable.
How much more does dedicated CPU cost?
Roughly two to five times, depending on provider. DigitalOcean’s own pricing page lists a CPU-Optimized 4 GiB / 2 vCPU Droplet at $42 against $24 for the equivalent Basic — about 1.75×. Hetzner’s dedicated CCX23 at 16GB costs €85.99 against €15.99 for the shared CX43 at the same memory, which is 5.4× for half the cores.
Does “high frequency” mean dedicated?
No, and this is the most common mistake in the category. High frequency describes clock speed — the cores are faster but still shared. Vultr sells High Frequency and Optimized Cloud Compute as separate products; only the latter is dedicated.
Can I tell if my current VPS has dedicated cores?
Check your plan documentation first, then measure. Sustained steal time near zero under load is consistent with dedicated cores; steal time that fluctuates with time of day strongly implies sharing. Note that a lightly-loaded shared host can also show near-zero steal, so a single quiet reading doesn’t prove isolation.
Are more shared cores better than fewer dedicated ones?
It depends entirely on duty cycle. For bursty work, eight shared cores usually beat four dedicated ones and cost far less. For continuous load, four dedicated cores can outperform eight contended ones — and if steal time hits 50%, those eight shared cores are delivering four anyway, without the guarantee.
Related Products
