Skip to content

Independent marketplace for verified cloud accounts

Bare Metal vs Virtualized Compute: When Dedicated Hardware Wins

Virtualization rarely makes you slower on average. It makes you less predictable. The specific workloads where owning the whole machine is faster and cheaper.

Virtualization won so completely that dedicated hardware now has to justify itself. That is the right way round, but it has left a lot of teams assuming the answer is always virtual, and paying for it in jitter, in licensing, and occasionally in a performance problem that no amount of instance resizing will fix.

Bare metal is not a nostalgia purchase. There are specific, identifiable workloads where owning the whole machine is both faster and cheaper, and the way to find out whether yours is one of them is to understand precisely what the hypervisor takes from you.

What the hypervisor actually takes

Modern virtualization is efficient. Hardware assisted CPU virtualization means instructions mostly run natively, and device passthrough means network and storage adapters can be addressed almost directly. The raw overhead on compute bound work is small enough that for most applications it is not the deciding factor.

The costs that remain are less about average throughput and more about consistency and control. Interrupt handling adds latency. Memory access patterns change because your guest’s view of locality does not match the physical layout, which matters on multi socket machines where crossing between processor sockets is slower than staying put. Some processor features are hidden from guests or exposed inconsistently. And the scheduler on the host can pause your virtual CPU while it attends to someone else, producing a stall that your application cannot see, explain or prevent.

That last one is the heart of the matter. Virtualization does not usually make you slower on average. It makes you less predictable.

Noisy neighbours and the shape of the variance

On a shared host your performance depends on what other tenants are doing. Providers mitigate this with CPU allocation guarantees, dedicated cores on larger instance types, and storage quality of service. The mitigations are real and they are not complete, because cache and memory bandwidth are physically shared in ways that are hard to partition.

The symptom is a distribution with a long tail. Average response time looks fine, the median looks fine, and then a few percent of requests take many times longer for no reason visible in your own metrics. On a system where one slow operation holds a lock or occupies a connection from a limited pool, a small tail can produce a large user facing problem.

If you have ever chased a latency spike through your application, your database and your network and found nothing, the host is a legitimate suspect. Smaller shared instance types on any provider are the most affected, because they pack the most tenants per machine. Moving the same workload to a dedicated machine and watching the tail flatten is a diagnosis as much as a fix.

Workloads where dedicated hardware pays

A handful of categories come up again and again.

  • Databases with strict latency targets, where the tail matters more than the mean and the working set is large enough that cache behaviour dominates.
  • Sustained high throughput work that runs continuously rather than in bursts, because burstable allowances never apply and you are paying a premium for elasticity you do not use.
  • Anything licensed per physical core or per socket, where the licence cost dwarfs the hardware cost and you want to control exactly how many cores you are declaring.
  • Workloads needing specific hardware features: particular processor instruction sets, local non volatile storage, or direct device access that a hypervisor abstracts away.
  • Nested virtualization, such as running your own hypervisor or certain container runtimes and emulators, which many virtual instance types do not support at all.
  • Deployments with a regulatory requirement for physical isolation of tenancy, where a contractual single tenancy statement is easier to produce for dedicated hardware.

Notice what is not on that list: general web applications, API backends, queue workers, and almost anything that scales horizontally. Those belong on virtual instances, and should be designed to be replaced rather than preserved.

Provisioning time changes how you architect

A virtual instance appears in under a minute. Bare metal can take anywhere from a few minutes at providers with pre racked automated pools to days where the machine is built to order. That difference is not merely an inconvenience, it changes the architecture you are allowed to use.

Autoscaling in response to traffic is impractical if capacity arrives tomorrow. Replacing a failed node automatically requires a warm spare rather than a fresh provision. Treating servers as disposable, which is the central habit of good cloud operations, becomes much harder when each one takes real time and often a minimum billing term to acquire.

There is a second consequence that is easy to miss. Slow provisioning pushes you toward larger machines, because if acquiring capacity is painful you buy headroom in advance. Larger machines then concentrate more of your workload behind a single failure domain, so the loss of one box hurts more. You end up with fewer, bigger, more precious servers, which is precisely the operating model cloud computing was meant to replace. That is acceptable if you have chosen it knowingly and have the standby capacity to match. It is a problem if it happened by drift.

The usual resolution is a hybrid: dedicated machines for the stable, performance sensitive core, and virtual instances for everything that flexes. That pattern is well supported at providers who sell both from one account, and it is a good reason to look at operators with a dedicated hardware line rather than assuming you need a second vendor. Accounts such as Buy Hetzner Account and Buy OVH Account are common choices for exactly this mix, and the Hetzner and OVHcloud comparison covers how their dedicated offerings differ in practice.

Licensing, nested virtualization and compliance

Software licensing is frequently the whole argument. Products priced per physical core, per socket or per host can make the infrastructure decision for you, because on a shared virtual host you may be required to licence capacity you cannot see. Dedicated hardware, or a dedicated host product that gives you a whole machine while keeping the virtualization layer, lets you count and control what you are licensing.

Check the licence text rather than assuming. Some vendors allow virtual core counting, some require physical, and some have specific provisions for public cloud that differ from on premises terms. Getting this wrong is expensive in a way that no engineering optimisation will recover.

On the compliance side, the honest position is that physical isolation is sometimes genuinely required and more often requested out of habit. Ask whether the requirement is written down or assumed. If it is written down, dedicated hardware is the straightforward answer. Our guide to cloud compliance obligations is a reasonable starting point for working out which category you are in.

The cost comparison people get wrong

The common error is comparing an hourly virtual instance rate against an hourly bare metal rate and concluding that bare metal is expensive. The comparison is wrong in both directions.

On one side, a bare metal machine typically offers substantially more capacity than the virtual instance you were pricing, so the correct comparison is against however many virtual instances it replaces, including their storage and the network traffic between them. On the other side, bare metal usually carries a minimum term or a monthly rate, so if your workload runs eight hours a day the virtual option wins easily on the same hardware.

Build the comparison on monthly cost at your real duty cycle, and include the things people forget: the licence difference, the inter instance traffic you no longer pay for because the components now share a machine, and the engineering time of operating a fleet versus a single box. Our guide to right sizing instances covers the measurement side of that, and the figures it produces are the ones that should drive this decision.

A practical way to decide

Run the workload on a virtual instance first. It is faster to provision, cheaper to abandon, and it generates the measurements you need. Then look for three signals: a latency distribution with an unexplained tail, utilisation that sits high and flat for most of the day, and a licence or compliance constraint tied to physical cores.

One signal is a reason to investigate. Two is a reason to run a trial on dedicated hardware for a month and compare the same metrics. Three means you probably already know the answer. Absent all three, stay virtual and spend the attention somewhere it will pay back, such as the choice of compute model itself, which for many teams turns out to matter more than the hardware underneath it. If you are at the stage of pricing options, the account catalogue shows which providers offer both tiers from a single login.

Related comparisons

Hetzner vs Digital Ocean
vs

Hetzner vs Digital Ocean

Hetzner sells hardware value from a mostly European footprint. DigitalOcean sells a developer platform with managed add-ons and deep documentation.…

Google Cloud vs Hetzner
vs

Google Cloud vs Hetzner

Hetzner sells compute by the unit at a low price. Google Cloud sells a platform. The comparison is really about…

AWS vs Hetzner
vs

AWS vs Hetzner

For steady always-on compute the price gap between Hetzner and AWS is wide. This comparison works out what that gap…

Hetzner vs Contabo
vs

Hetzner vs Contabo

Both are German hosts competing on low prices, and Contabo is usually the cheaper number on the page. The useful…

Accounts discussed in this guide

Buy Hetzner Account
Buy Hetzner Account
Hetzner

Buy Hetzner Account

Hetzner has earned a reputation among developers as one of the most powerful and affordable cloud platforms in the world.

From$50 3 options
Buy OVH Account
Buy OVH Account
OVHcloud

Buy OVH Account

OVHcloud has built a formidable reputation as one of Europe's largest and most independent cloud providers, offering a compelling alternative to the hyperscalers for…

From$30 2 options

Keep reading

Reserved, Spot and On-Demand Pricing Compared
Cost Management 7 min read

Reserved, Spot and On-Demand Pricing Compared

The same machine sells three ways at three prices. What changes is the commitment you make and the interruption risk you accept.…

Reading Cloud Storage Performance Specifications
Infrastructure 7 min read

Reading Cloud Storage Performance Specifications

IOPS, throughput and latency are three promises, not one. How to read a storage spec sheet so the performance you get matches…

Need the account, not just the theory?

Browse documented configurations across every provider in the catalog, or ask us which one fits your workload.

Telegram