U1 Foundations
Choosing homelab hardware without regret
The best first server is usually one you already own. This guide explains how to size a lab, which hardware families suit which jobs, and how to buy used equipment without inheriting someone else's problems.
At a glance
- Start with hardware you already have. The only hard requirement for a virtualization host is a CPU with virtualization extensions enabled in firmware.
- Memory runs out long before CPU in almost every homelab. Buy RAM first and cores second.
- Measure idle power, not peak power. A lab spends most of its life idle, and idle watts are what you pay for every hour of the year.
- Decide between tower and rack form factors before you buy, because changing your mind later means selling and buying again.
- Test every used machine with a memory test and a burn-in before you trust it with data.
Start with what you already own
A homelab is an environment you own and control, used to learn and to run services for yourself. It does not need to be big, new or loud. A retired desktop, a laptop with a cracked screen or a small office PC from a surplus sale will run a hypervisor, a handful of containers and a monitoring stack without complaint.
Old laptops deserve special mention. They are quiet, they sip power, and the built-in battery acts as a small uninterruptible power supply that rides through short outages. Put one on a shelf with the lid closed and you have a respectable first server.
Before you buy anything, check two things on the hardware you have:
- Virtualization support. Look in the firmware setup for the CPU virtualization options and the IOMMU option, and enable them. Without these, a hypervisor will either refuse to run virtual machines or run them very slowly.
- Memory ceiling. Find out how much RAM the board supports and how many slots it has. A machine that tops out at 8 GB will cap your lab quickly; one that accepts 32 or 64 GB can grow with you.
The four hardware families
Almost everything people run at home falls into one of four families. Each is a sensible choice for some jobs and a poor choice for others.
| Family | Typical idle draw | Noise | Good for | Watch out for |
|---|---|---|---|---|
| Mini PCs and small form factor desktops | 6 to 20 W | Silent to quiet | Hypervisor nodes, container hosts, firewalls, small clusters | Few drive bays, limited PCIe expansion, firmware support varies by vendor |
| Used enterprise servers | 70 to 200 W | Loud, especially 1U | Learning enterprise gear, many drive bays, lots of RAM slots, remote management | Power bills, heat, noise, licensing on management controllers |
| Single-board computers | 3 to 8 W | Silent with passive cooling | DNS filtering, monitoring probes, home automation, print servers, small always-on jobs | Software compatibility on non-x86 processors, weak storage I/O, unreliable SD cards |
| Workstations and self-built towers | 30 to 80 W | Quiet with large fans | An all-in-one learning lab, GPU work, storage builds with many drives | Desk space, consumer boards without ECC or remote management |
The ranges above are measured at the wall on typical configurations and will vary with drives, network cards and firmware settings. Use the power cost calculator to turn idle watts into a yearly figure before you decide.
Mini PCs
For most people starting today, two or three small form factor PCs are the best value in the hobby. They typically idle below 20 W, they are nearly silent, and three of them make a real cluster that you can take apart one node at a time. Business-class models from the off-lease market are especially good buys because their parts are standardized and their firmware gets updates for years. Before buying, check whether the memory is soldered or socketed and whether a second M.2 slot supports NVMe or only slower SATA drives; our Mini PC Hypervisor Build project lists what to check.
Fanless mini PCs with four to six network ports are popular as firewalls. They are excellent at that job, but many are sold under rotating model names by anonymous importers. Before you rely on one, find out whether firmware updates exist and where they come from. A firewall you cannot patch is a liability.
Used enterprise servers
Off-lease rack servers are cheap to buy because businesses retire them on a schedule, not because they are worn out. They bring dozens of memory slots, redundant power supplies, hot-swap drive bays and a built-in management controller. They also bring noise and power draw that surprise almost everyone. Small 1U chassis are the loudest because their small fans must spin fast to move enough air; 2U and larger chassis with bigger fans, or tower servers, are far easier to live with.
From the Bench
A 1U server that passed through the Bench in 2024 arrived with its previous owner's firmware profile still set to maximum performance and maximum fan speed. It idled at 131 W and could be heard two rooms away. Resetting the fan profile to automatic and the power profile to balanced brought it to 84 W and made it bearable in a basement. Check those settings before you judge any used server.
Single-board computers
Credit-card sized single-board computers are real tools, not toys, as long as you respect their limits. They shine at small always-on jobs: a DNS filter, an uptime monitor, a home automation controller, a print server that rescues an old USB printer, or a remote console for your other machines.
- Boot from or move busy directories to an SSD. SD cards are slow and are the most common point of failure.
- Buy the power supply the board was designed for. An underpowered adapter causes silent throttling and strange crashes.
- Newer, faster boards need active cooling. Budget for the cooler.
- Many appliance operating systems, including most firewall and NAS distributions, are built only for x86. Check before you plan around a non-x86 board.
- Pull container images from maintainers who publish builds for your board's processor architecture, rather than trying random images until one starts.
Workstations and self-built towers
One capable desktop with 12 cores and 32 to 64 GB of memory, running a hypervisor, is a complete lab. It is also the easiest way to add a GPU or many drives. Choose a motherboard with headroom, buy the minimum memory you need today and fill the remaining slots later.
Sizing CPU, memory and storage
Homelab workloads are mostly idle. A file server, a password vault, a photo library and a dozen small containers will barely register on a modern CPU. A firewall handling a home internet connection typically sits at a few percent CPU even on a mobile processor that is several years old. What runs out first is memory, then storage performance.
| Lab size | Memory | CPU | Storage |
|---|---|---|---|
| Learning lab, a few VMs and containers | 16 GB | 4 cores | One 500 GB to 1 TB SSD |
| Household services plus experiments | 32 GB | 6 to 8 cores | Mirrored SSDs for VMs, separate NAS for bulk data |
| Cluster, heavy experiments, full desktop VMs | 64 GB or more per node | 8 cores or more per node | Fast SSDs per node, 10 GbE to shared storage |
Error-correcting (ECC) memory is worth having on machines that hold your only copy of important data, and it is standard on used servers. It is not a requirement for a homelab, and its absence is not a reason to give up on a filesystem with checksums. More on that in ZFS and Filesystems.
Buying used without inheriting problems
Most of the hardware in our own lab was bought used, and failures have been rare. The reason is not luck. It is a routine that catches problems before data lands on the machine.
- Inspect before powering on. Look for bulging capacitors, corrosion, bent pins and missing heat sinks. Clean out dust.
- Reset firmware to defaults. Previous owners leave fan, power and boot settings in strange states. Update the firmware if updates are freely available.
- Run a memory test. A short pass catches obvious faults in minutes. Let a full pass run overnight before you install anything permanent.
- Burn in the drives and the CPU. Run sustained load for 24 hours while watching temperatures. Failures cluster early in a component's life.
- Read the hardware event log. Server management controllers keep a log of memory errors, fan failures and power events. It tells you how the machine was treated.
- Replace the consumables. Buy power supplies and SSDs new. Replace the small firmware battery on older boards; it costs almost nothing.
Caution
Some server vendors gate firmware updates, remote console features or virtual media behind paid licenses. Before buying a used server, find out which management features are licensed, whether the license transfers with the hardware, and whether firmware updates are downloadable without a support contract.
Older hardware and instruction-set cliffs
Memory and drive prices rose sharply through 2025 and 2026, which pushed many people toward older platforms. That can be a smart move. Six-core server CPUs from the DDR3 era cost very little, and 32 GB of DDR3 is inexpensive compared with current memory. These machines still run hypervisors, containers and file services well.
There are hard floors, though. Very old CPUs lack newer instruction set extensions such as AVX2 or SSE4.2, and older boards lack a TPM 2.0 module. Some current software will not start at all on such hardware, no matter which operating system you install. Check the requirements of the software you plan to run before you buy the platform. And remember that low power is not the same as efficient: a slow netbook-class chip that idles at 10 W may do so little work per watt that a newer mini PC beats it on both speed and energy.
We go deeper on this trade-off in A Second Life for Old Hardware.
Tower or rack: decide early
Rack-mount and tower equipment do not mix well. Once you own a rack, towers waste shelf space; once you own towers, a rack is an expensive shelf. Make the decision before your second purchase.
Choose towers and mini PCs if
- The lab lives in a living space or an office.
- Noise matters to anyone in the house.
- You plan to run three to five machines.
Choose a rack if
- You have a basement, garage or utility room with airflow.
- You want to learn on enterprise servers and switches.
- You expect the lab to grow past a single shelf.
Racks, shelves, depth and cooling are covered in Racks, Power and Cooling.
Do not build one box to do everything
It is tempting to buy one large server and run the firewall, the NAS, the hypervisor and every service on it. It works until the day you need to reboot it. When the hypervisor goes down for maintenance, the house loses internet, and you lose the ability to download the fix or read the documentation you need.
Rule of thumb
Keep the firewall on its own small, dedicated device. Keep bulk storage on its own machine if you can. Run experiments on everything else. Virtualizing a firewall or NAS is a fine way to learn, but it should not be what the household depends on.
A sensible first shopping list
If you are starting from nothing and want a lab that will grow without being rebuilt, this is the order we recommend:
- A small dedicated firewall appliance with at least two network ports.
- A managed switch that supports VLANs.
- One mini PC or small desktop with 32 GB of memory as your first hypervisor host.
- An external drive or small NAS for backups, before any service holds data you care about.
- A second and third host only when the first one is genuinely full.
Our Reference Builds show three complete labs built this way, with parts by category, measured power draw and what each one runs. For current price ranges and running costs, see Budgeting a Homelab.