Three servers from one provider each give you eight of something. One costs €69.99 a month, one €138.99, and the third — eight physical cores, 64 GB of RAM, unlimited traffic — €99.00 (Hetzner, November 2026). What separates the first two is how many virtual CPUs the hypervisor hands out per physical core, and no provider publishes that number. So dedicated server vs VPS is not really a question about isolation. It is a question about a ratio you cannot see — and this article covers what you can measure instead, and when the measurement justifies a move.
“8 vCPU” is three different purchases
On a shared-vCPU plan, a vCPU is a scheduling entitlement. DigitalOcean’s documentation is direct about it: “With shared CPU Droplets, the hyper-thread allocated to the Droplet may be shared between multiple other Droplets,” and when neighbours are busy, “a Droplet could receive fractions of hyper-threads instead of dedicated access to the underlying physical processors.”
On a dedicated-vCPU plan, the entitlement becomes exclusive. Hetzner: “With our dedicated plans, you get vCPUs that are assigned only to you. This means you have no ‘noisy neighbors.’” What you get exclusively is usually a hardware thread, not a core: DigitalOcean’s dedicated plans have “guaranteed access to the full hyper-thread at all times,” and AWS says each vCPU “is a thread of either an Intel Xeon core or an AMD EPYC core.” You own the thread — not the shared cache or memory controller of the chip it runs on.
On bare metal, the ratio is 1:1 by construction. A shared plan might be running at 2:1 or at 16:1, and the product page reads the same either way.
The ratio is a configuration default, not a specification
The default is published — just not by hosts. OpenStack Compute used an initial CPU allocation ratio of 16:1 until its 2023.1 release, when the default for new compute nodes dropped to 4:1 and the RAM default to 1:1 (OpenStack Nova release notes). The same note adds: “This will not affect any existing compute node resource providers.” A fleet enrolled before 2023 can still be running at 16:1 without anyone having decided anything since.
Operator guidance runs lower. VergeOS, for example, recommends 4:1 to 6:1 for desktops and file servers, 2:1 to 4:1 for a typical enterprise mix, 1:1 to 2:1 for databases and application servers, and “1:1 or less” for CPU-bound HPC. So the documented range behind a plan sold as “shared” runs from roughly 2:1 to 16:1 — an eightfold spread in what one vCPU means — and nothing stops an operator going higher.
Steal time is your only instrument — and the hypervisor supplies it
Linux has reported steal time since kernel 2.6.11: “Stolen time, which is the time spent in other operating systems when running in a virtualized environment” (proc_stat(5)). It is the st column in vmstat and top.
The caveat matters more than the metric: the number comes from the party you are auditing. On KVM the host fills in a shared memory area, and the kernel documentation states that “the hypervisor may update this structure at any time it sees fit,” and that “time during which the vcpu is idle, will not be reported as steal time.” A host that under-reports steal does not produce an error. It produces a clean graph.
What independent measurement found
Leitner and Cito measured CPU benchmark variability across Amazon EC2, Google Compute Engine, Microsoft Azure and IBM SoftLayer, reporting relative standard deviation per configuration (arXiv:1411.2429, 2014). The instance types are long retired; the pattern is what transfers.
| Configuration (2014 study) | Class in the paper | Across instances, EU | Across instances, NA | Within one instance |
|---|---|---|---|---|
| EC2 t1.micro | Bursting | 12.14% | 20.28% | 40.17% |
| EC2 m1.small | Non-bursting | 3.19% | 12.81% | 1.69% |
| EC2 m3.large | Non-bursting | 0.13% | 0.16% | 0.89% |
| GCE f1-micro | Bursting | 5.28% | 5.13% | 3.24% |
| GCE n1-standard-2 | Non-bursting | 1.71% | 1.16% | 0.86% |
| Azure Small | Non-bursting | 18.23% | 19.11% | 1.65% |
| SoftLayer 1 CPU / 2048 MB | Virtual, configurable | not tested | 0.11% | 0.22% |
Source: Leitner and Cito, “Patterns in the Chaos”, Tables 4 and 6. Within-instance figures are from European regions, except SoftLayer (North America only).
Most variance on non-bursting plans was placement, not contention. In North America, m1.small varied 12.81% across instances; grouped by the physical CPU model each instance landed on, the spread fell to 0.15–0.42% per model. Azure Small shows the same signature: 18.23% across instances, 1.65% within any one. The spread was mostly which silicon you got — and that has a different remedy. Bare metal has the same issue one layer down: two hosts’ “8-core, 16-thread” CPUs can differ six-fold in L3 cache, as covered in our guide to comparing dedicated server providers.
The bursting tier is a different product, not a worse one. The t1.micro’s within-instance swing came from its credit balance running down — a billing mechanism, not a neighbour.
Fixed virtual instances can come close to bare metal. A later study of three public clouds and one bare-metal server found variability ranging “from a coefficient of variation of 0.03% to more than 100%,” and that while “the bare-metal instance expectedly produces very stable results, however AWS is typically not substantially less stable” (Laaber, Scheuner and Leitner, Empirical Software Engineering, 2019).
One vendor, three tiers, eight cores
| CPX42 (shared vCPU) | CCX33 (dedicated vCPU) | AX42 (bare metal) | |
|---|---|---|---|
| Cores | 8 vCPU, shared | 8 vCPU, dedicated | 8 physical cores, AMD Ryzen 7 PRO 8700GE |
| RAM | 16 GB | 32 GB | 64 GB DDR5 ECC |
| Disk | 320 GB NVMe | 240 GB NVMe | 2 × 512 GB NVMe |
| Included traffic | 20 TB | 30 TB | Unlimited, 1 Gbit/s |
| Monthly | €69.99 | €138.99 | €99.00 |
| One-time | — | — | €49.00 setup |
| Billing and exit | Hourly, capped at the monthly price | Hourly, capped at the monthly price | No minimum term; cancellation immediately |
Hetzner product pages, November 2026, excl. VAT. Cloud prices are for Nuremberg; Helsinki is cheaper (CPX42 €59.49, CCX33 €117.99).
On this vendor’s own pages, the dedicated-vCPU tier is the most expensive way to buy eight cores. CCX33 costs about twice the shared plan for the same vCPU count, and 40% more per month than a bare-metal server with twice the RAM (19% more at the Helsinki price). Counting the setup fee, cumulative spend on bare metal falls below the dedicated-vCPU plan in the second month; over twelve months the gap is €430.88 at the Nuremberg price.
What the dedicated-vCPU premium buys is not more hardware but cloud mechanics: provisioning in minutes through an API, resizing, snapshots and no setup fee. If you need those, it is worth paying. If you do not, you are paying a premium for less hardware. The €49 is not the only bare-metal extra — see the line items a dedicated server quote leaves out.
The contention test
Run this before migrating. It takes one billing cycle.
Name the resource the workload saturates first — single-thread latency, aggregate throughput, memory bandwidth or IOPS. A test aimed at the wrong resource will pass on a machine that cannot run the workload.
Record %st per vCPU at one-minute resolution for seven days, not as an average. Contention is bursty: a single 20% hour on an otherwise clean day still averages under 1%.
Correlate steal against your own p99 latency, not against CPU utilisation. The question is whether contention moves the number your users feel.
Capture /proc/cpuinfo — model name, flags, reported clock. You will need it in step 5.
Destroy and recreate the instance three times, same plan, same region, and repeat steps 2–4 each time. If the spread tracks the CPU model, it is placement.
If it is placement, keep re-provisioning. On hourly billing that costs hours, and measurement suggests the spread within one CPU model is far narrower than across a fleet.
If steal survives re-provisioning and tracks your p99, you have a ratio problem. Price dedicated vCPU and bare metal from the same vendor on the same day, and apply the break-even above.
What breaks in production
Credit exhaustion read as a neighbour problem. AWS T3 instances launch in unlimited mode by default: a depleted instance keeps bursting, and the surplus is billed per vCPU-hour. In standard mode, depletion becomes a throttle back to the baseline — 5% per vCPU on a t3.nano. One is a bill, the other a ceiling, and neither is a neighbour.
Benchmarking once. Microbenchmarking research found that even with twenty instances, a slowdown below 10% was reliably detectable in only 77–83% of cases. A single sysbench run on a fresh VPS tells you about that VM, that hour, that host.
Sizing dedicated vCPU against bare-metal cores one-for-one. Eight dedicated vCPUs are typically eight hardware threads, not eight cores. For single-threaded, latency-bound work, see single-thread workloads on dedicated hardware.
Forgetting the port while comparing the cores. Hetzner says of its cloud servers: “We do not offer bandwidth guarantees for our Cloud servers, but you can expect about 300-500 Mbits.” A gigabit port running flat out for a month moves about 324 TB, 10.8 times the dedicated plan’s 30 TB allowance — though few workloads use it that way; see how bandwidth is actually metered.
Treating the move as one-way. Running both tiers for one cycle against a slice of production traffic costs a month and replaces plan pages with your own numbers — the same logic as in hosting for high-traffic sites.
Dedicated server vs VPS: two questions
Does your p99 move with steal time after you have re-provisioned three times? If not, keep the shared plan and spend the difference elsewhere — you are being offered isolation whose absence you cannot measure. If it does, the ratio is your problem, and you have the evidence to price the fix.
Do you need API provisioning in minutes, resizing or snapshots? If yes, dedicated vCPU is the right product and the premium buys elasticity, not cores. If not, price bare metal first: on the published rates above it is the cheaper exclusive-hardware option, and the ratio question disappears rather than being answered.
Dedicated servers on INXY.hosting
INXY.hosting dedicated servers come with no setup fee and no minimum term, in the EU and the US — which removes the two costs that usually make bare metal the slower, more committed choice. Because we source across multiple providers, the same workload can be priced on several vendors’ hardware on the same day. If you have steal-time data and no decision yet, send it to us with your p99 target and we will price the options against your traffic — see dedicated and bare-metal hosting.
FAQ
What is a good steal time percentage on a VPS?
There is no published threshold, and a single average is the wrong statistic. What matters is whether steal correlates with your own p95 or p99 latency at peak. Sustained double-digit %st on a plan sold as dedicated is grounds for a support ticket; intermittent steal on a shared plan is the product working as described.
Is a dedicated server always faster than a VPS?
No. Measurement studies find fixed-performance virtual instances with CPU variability under 1%, and one study found AWS “typically not substantially less stable” than the bare-metal server it tested. Bare metal wins on memory bandwidth, cache and the absence of hypervisor overhead — advantages that show up only when the workload saturates those resources.
Can I find out my provider’s oversubscription ratio?
Usually not: the provider documentation we checked does not publish a vCPU-to-physical-core ratio for shared plans. Platform defaults and operator guidance span roughly 2:1 to 16:1. The measurable proxy is steal time, which the hypervisor supplies — treat the ratio as unknowable and the test as the substitute.
Does dedicated vCPU eliminate noisy neighbours?
It eliminates CPU-scheduling contention, which is what providers mean by the phrase. It does not give you exclusive cache, memory bandwidth or storage back-end — those remain shared on a multi-tenant host, and that interference does not show up in steal time.
Is dedicated vCPU always more expensive than bare metal?
Not always — it varies by vendor and location, so check on the day. On Hetzner’s published rates above, the dedicated-vCPU plan costs more per month than a bare-metal server with twice the RAM. Compare on the same day, include setup fees, and find the month at which cumulative cost crosses over.
When is a VPS the right answer for production?
When the workload runs at low to moderate utilisation, tolerates variable processor performance, and benefits from fast provisioning and resizing. Hetzner describes its shared tier as “suited for workloads that can handle variable processor performance” — plenty of production services meet that description.
