Bare Metal, VPS, or Cloud VM: When a Dedicated Server Is Worth It

Once an application outgrows shared hosting, the next step is usually a VPS, and we covered that move in a separate article. The question that follows a year or two later is different: the VPS bill keeps growing, the database is slow at busy hours even though the plan looks big enough, and someone suggests renting a dedicated server instead. Sometimes that is exactly the right move. Sometimes it swaps a hosting bill for a hardware problem the team is not ready for. This article explains how bare metal differs from a VPS and a cloud VM, what it costs in 2026, and how to tell which one your workload actually needs.

Three Ways to Rent a Server

A VPS is a virtual machine running on a physical server that the provider shares among many customers. A hypervisor such as KVM divides the CPU, memory, and disks, and each customer sees what looks like a separate server. A cloud VM, such as an EC2 instance on AWS or a Compute Engine VM on Google Cloud, is technically the same thing, but it comes with much more around it: hourly or per-second billing, an API to create and delete servers, snapshots, network block storage, autoscaling, and managed databases and queues next to it.

A bare metal server, usually called a dedicated server, is a whole physical machine that only you use. There is no hypervisor and no other tenant. You get every core, all of the memory, and the disks directly. The traditional form is a monthly rental with a setup fee. A newer form, often called bare metal cloud, lets you order physical servers through a portal or API and pay by the hour or month, like cloud VMs.

What a vCPU Really Is

On most cloud and VPS plans, one vCPU is one hardware thread, which is half of a physical core. A cloud VM with 16 vCPUs therefore has roughly 8 physical cores. Cheap VPS plans may also sell the same core to several customers at once, a practice called overcommitting, so check whether a plan says dedicated or shared vCPU. On a bare metal server, 8 cores and 16 threads means exactly that, and all of it is yours.

There is a quick way to see whether neighbours are taking your CPU on a VPS. Run top and look at the st value, which stands for steal time. It shows the share of time your VM wanted the CPU but the hypervisor gave it to someone else. If it stays above a few percent during busy hours, the slowness is not in your code, and moving to a dedicated-vCPU plan or bare metal will help.

What It Costs: A Server with 8 Cores and 64 GB of RAM

The AWS price comes from AWS's own price list for the Jakarta region in November 2026, and the Hetzner price from its product page. The Jakarta figures are indicative ranges from providers' published plans, which vary a lot in hardware generation and what is included, so ask for a written quote. The pattern is clear, though. For a load that runs around the clock, a physical server can cost several times less per core than an on-demand cloud VM. What the cloud price includes is everything you would otherwise build yourself: snapshots, automatic recovery when hardware fails, fast resizing, and managed services next door.

Server prices have also moved a lot this year. Demand for memory and NVMe drives from AI data centres pushed component prices up, and providers passed it on. Hetzner raised prices in April and again in June 2026, and the AX42 above now costs more than double what it did at the start of the year. Most providers keep the old price for servers that are already running and apply new prices to new orders and resizes. If you rent servers today, check your provider's terms before you cancel and reorder a machine, because the replacement may cost much more.

When Bare Metal Makes Sense

• Large, busy databases. A PostgreSQL or MySQL database of hundreds of gigabytes with heavy reads and writes benefits most from local NVMe disks, which deliver far more IOPS than network block storage at the same price.
• Constant heavy CPU load. Video encoding, batch processing, CI runners, and game servers that run 24 hours a day waste money on hourly billing, because the flexibility you pay for is never used.
• GPUs that are used every day. For AI inference or a private language model that serves staff all day, a rented GPU server usually costs less over a year than hourly GPU instances. We covered running models in-house in our article on private LLMs with Ollama.
• Heavy outbound traffic. File downloads, video, and large API responses get expensive on cloud egress pricing. Many dedicated server offers include a large or unmetered monthly traffic allowance.
• Licences tied to physical cores. Some commercial software, such as Oracle Database, is licensed by physical cores or sockets. Counting cores is simpler and often cheaper on hardware you control.
• Single-tenant requirements, or running your own platform. Some contracts and regulators ask for hardware that is not shared. Bare metal is also the natural base if you want to run Proxmox, VMware, or your own Kubernetes cluster.

When It Is the Wrong Choice

• The application is small. A typical business application runs comfortably on a VPS with 2 to 4 vCPU and 4 to 8 GB of RAM. A dedicated server would sit idle most of the day.
• Traffic comes in sharp peaks, for example during campaigns, ticket sales, or school registration periods. Cloud servers can be added in minutes and removed afterwards. Physical servers cannot.
• Nobody on the team handles servers. Bare metal is the most hands-on option. Without someone watching disks, updates, and backups, a cheap server becomes an expensive outage.
• You rely on managed services such as managed databases, queues, and object storage. Those live in the cloud, and splitting an application across a dedicated server and a distant cloud region adds latency and data transfer costs.
• The plan is one server for a critical system. A single physical machine with no second server and no tested backup is a single point of failure.

The Work That Comes with Bare Metal

• Disks will fail. Use at least two disks in RAID 1, monitor SMART values and the RAID status, and set alerts for a degraded array. RAID keeps the server running when one disk fails. It is not a backup.
• Out-of-band access. Learn how to use the provider's remote console, such as IPMI, iDRAC, iLO, or a rescue system, before you need it. It is the only way in when a firewall rule or a broken kernel locks you out of SSH. Never expose it to the public internet.
• Hardware replacement time. Ask how quickly a failed disk, power supply, or memory module is replaced, and whether that is written in the contract. Four hours and best effort are very different promises.
• No snapshots. There is no one-click snapshot of a VM to roll back to. You need proper backups of files and databases, stored off the server, and you need to test restoring them. Our article on backup and disaster recovery covers how to plan this.
• A second server, or a fallback in the cloud. For anything important, replicate the database to another machine or keep a recovery plan that can bring the service up elsewhere within an agreed time.
• Firmware as well as the operating system. BIOS, RAID controller, and network card firmware get security updates too, and they are easy to forget.
• Network protection. Small dedicated server providers often have limited DDoS protection. Putting Cloudflare or a similar service in front of web traffic helps, as long as the server only accepts connections from it.

Dedicated Server, Bare Metal Cloud, or Colocation

Whatever the model, keep server setup in code with tools such as Terraform or Ansible, so that moving to another provider is a project of days, not months. Providers do leave the market: Equinix shut down its Equinix Metal bare metal service on 30 June 2026, and its customers had to migrate everything elsewhere.

Where the Data Lives

Under PP 71/2019, private electronic system operators may generally store data outside Indonesia, but sector regulators such as OJK for finance can set stricter rules, and the Personal Data Protection Law (UU PDP) requires adequate safeguards for transfers abroad. A bare metal server or colocation in a Jakarta data centre gives you the clearest answer to where data is stored and who can physically reach it. It also keeps latency low for users in Indonesia, which is one reason a cheap server in Europe is rarely a good fit for an Indonesian application.

A Combination That Often Works

Few systems need to be all bare metal or all cloud. A common setup for a growing business application is to run the web and application servers on cloud VMs or VPS behind Cloudflare, so they can be scaled and replaced easily, and to put the database on a bare metal server in Jakarta with a replica on a second machine. Backups go to object storage in another location. The reverse also works: bare metal for the steady base load and cloud servers added only for peaks.

Put steady, predictable load that you understand well on bare metal. Put anything uncertain in the cloud. Most businesses end up with some of both.

Our cloud and DevOps team helps clients measure their current servers, estimate the cost of each option over one to three years, and plan migrations between VPS, cloud, and bare metal. If your server bill or database performance is becoming a problem, start by collecting a month of CPU, memory, disk, and traffic data. That usually makes the answer obvious.