Last updated: 27 September 2026
Short answer: for most people learning cloud and DevOps skills, the best home lab is one used business mini PC with 32 GB of memory and an SSD, running the free Proxmox hypervisor. It idles at around 10 to 15 watts, costs about $25 a year in electricity in the US, and can run a small Kubernetes cluster, databases and practice servers. Add a GPU box with 16 GB of video memory only if you want to run AI models locally. Use cloud free tiers alongside it for the services you cannot run at home.
Why build a home lab at all?
Because certifications and courses teach concepts, but jobs need hands-on practice, including breaking things and fixing them. A home lab gives you:
- Unlimited practice with no meter running. A forgotten cloud server or database can quietly run up a bill; a lab box costs the same every month.
- Real infrastructure skills: virtual machines, networking, storage, backups, Linux administration and Kubernetes.
- A private place for AI experiments, with your data staying on your own hardware. I covered the software side in setting up a private AI assistant on your own PC.
Build 1: The starter lab (one used mini PC)
Hardware: a used business "tiny" PC such as a Lenovo ThinkCentre Tiny, Dell OptiPlex Micro or HP EliteDesk Mini, from a refurbisher or an auction site. Aim for:
- An Intel 8th-generation Core i5 or newer, for 6 cores
- 32 GB of memory, which you can often add cheaply yourself; memory, not processor, is what limits how many virtual machines you can run
- A 1 TB NVMe SSD
Power: ServeTheHome, which has measured dozens of these machines, found a Lenovo M920q idles at 12 to 15 watts. That is little more than an LED light bulb.
Good for: Linux administration, Docker, a three-node Kubernetes cluster in virtual machines, databases, a practice Active Directory, and learning infrastructure-as-code tools.
Alternatives: a new mini PC based on Intel's N100 processor, which has a 6-watt design power but supports only 16 GB of memory. That is fine for containers but tight for several virtual machines. A Raspberry Pi 5 is great for learning Linux and ARM, but memory-driven price rises in 2026 mean the 16 GB model now costs $305, more than a well-specified used mini PC.
Build 2: The cluster lab (three small machines)
Hardware: three of the same mini PC, each with 16 to 32 GB of memory, plus a small gigabit network switch.
Why three: real clusters are about what happens when a machine fails. With three physical nodes you can practise Proxmox clustering, Kubernetes with nodes that really go offline, distributed storage and rolling upgrades. This is the same pattern that businesses run at their branches, which I described in edge computing explained with three real examples.
Power: roughly 35 to 45 watts for all three at idle.
Good for: preparing for the Certified Kubernetes Administrator exam, learning high availability, and practising backup and restore properly.
Build 3: The local AI lab (GPU box)
Hardware: a desktop with a recent NVIDIA graphics card. The number that matters for AI is video memory (VRAM), because the whole model needs to fit in it to run quickly:
| Card | VRAM | Typical use |
|---|---|---|
| RTX 3060 (used) | 12 GB | Small models, about 7 to 8 billion parameters |
| RTX 4060 Ti 16 GB | 16 GB | Mid-sized models of about 12 to 14 billion parameters with compression |
| RTX 5060 Ti 16 GB | 16 GB | Same capacity, faster memory |
Several of these cards come in 8 GB versions too; for AI, buy the larger one. An Apple Mac mini is the quiet alternative: the new M6 model starts at $899 with 16 GB of memory shared between processor and graphics, configurable to 32 GB, and uses far less power than a GPU desktop.
Power: a GPU desktop can idle at 50 watts or more and draw a few hundred watts under load. Switch it off when you are not using it, or set it to sleep.
Good for: running local language models with the free, open-source Ollama, trying retrieval over your own documents, and learning how AI models are served. Whether this is cheaper than paying for cloud AI depends on how much you use it; I worked through the numbers in local vs cloud AI for small businesses.
What does it cost to run?
A machine running 24 hours a day uses its wattage × 8,760 hours per year. Using the US average residential price of 18.31 cents per kWh (EIA, July 2026) and the UK price cap of 26.32p per kWh (October to December 2026):
| Average draw | Energy per year | US cost | UK cost |
|---|---|---|---|
| 15 W (mini PC) | 131 kWh | about $24 | about £35 |
| 45 W (three mini PCs) | 394 kWh | about $72 | about £104 |
| 100 W (old tower or GPU box at idle) | 876 kWh | about $160 | about £231 |
This is why a free old tower server is rarely the bargain it looks like: over three years its electricity can cost more than a newer mini PC. A cheap plug-in power meter tells you what your own hardware really uses.
What software should you run?
All of this is free for home use:
- Proxmox VE: the hypervisor that turns one machine into many virtual machines and containers. It is open source, and every feature works without the optional paid subscription.
- Ubuntu Server or another Linux distribution for your virtual machines. The current long-term support release is 26.04.
- Docker for running applications in containers.
- K3s, a lightweight, fully compliant Kubernetes distribution that installs as a single small binary. It is ideal for a lab.
- TrueNAS Community Edition if you want to learn network storage.
- Ollama for local AI models.
A good first project: install Proxmox, create three Ubuntu virtual machines, install K3s across them, and deploy a small web application with a database. Then back it all up and practise restoring it.
How do you access the lab safely from outside?
Do not open ports on your home router. Anything exposed directly to the internet will be scanned within hours. Instead:
- Tailscale creates a private network between your devices. The free Personal plan covers up to 6 users for non-commercial use.
- Cloudflare Tunnel publishes a web service through an outbound-only connection, so no inbound ports need opening.
Also keep the lab on its own network or VLAN if your router supports it, update it regularly, and back up the configuration. My 30-minute backup setup applies to a lab as much as to documents. A small uninterruptible power supply (UPS) protects the storage from power cuts.
How does a home lab fit with cloud free tiers?
Use both. The lab is for things you want to run for weeks; the cloud is for learning the provider's own services, which you cannot install at home.
| Provider | New-account offer (September 2026) |
|---|---|
| AWS | $100 credit at sign-up plus up to $100 more for completing activities; the free plan lasts up to 6 months |
| Microsoft Azure | $200 credit for 30 days, plus selected services free for 12 months |
| Google Cloud | $300 credit for 90 days, plus one small always-free virtual machine |
| Oracle Cloud | Always Free resources including ARM compute equivalent to 2 cores and 12 GB of memory |
Set a billing alert on day one with every provider, and delete resources when you finish an exercise.
Which certifications pair well with a lab?
- AWS Certified Cloud Practitioner ($100) and Solutions Architect – Associate ($150): the lab teaches the networking and Linux fundamentals behind the cloud concepts.
- Microsoft Azure Fundamentals (AZ-900): an entry point for Microsoft-centred workplaces; the price depends on your country.
- Certified Kubernetes Administrator (CKA) ($445, including one retake): a hands-on exam where a three-node lab is excellent preparation.
FAQ
How much does a home lab cost to start?
A single used mini PC with upgraded memory is usually the cheapest useful start. Check refurbisher and auction prices in your country; memory prices rose in 2026, so compare the cost of a machine that already has 32 GB with upgrading one yourself.
Can I build a home lab on my laptop instead?
Yes, for a start. Virtualisation software on a laptop with 16 to 32 GB of memory is enough for a few virtual machines. A separate machine becomes worthwhile when you want things running all the time.
Do I need a GPU for a home lab?
Only for local AI. Cloud, Kubernetes and DevOps practice need memory and storage, not graphics.
Is a Raspberry Pi good for a home lab?
It is excellent for learning Linux, networking and low-power services. For several virtual machines or Kubernetes with room to grow, a used x86 mini PC now offers more memory for the money.
