Why Build a Homelab in 2026
The cloud gives you scale. A homelab gives you the courage to destroy and rebuild without fear of a surprise bill.
That difference sounds small. It isn't. Engineers who work exclusively in the cloud learn with fear — every mistaken terraform apply produces cost. That creates aversion to experimentation, which is precisely the opposite of what a learning environment requires.
A Ryzen 5600G with 32 GB of RAM is worth months of AWS bills when the goal is to learn rather than to serve traffic.
What a homelab is actually for
Not nostalgia for sysadmin work. It's an engineering laboratory where each phase simulates something you'd meet in a real company, with no variable cost per experiment.
Without a physical lab, there's nowhere to simulate network failures, test GPU passthrough for a local LLM, or practice end-to-end GitOps without metered billing hanging over every attempt.
The hardware
The current build is deliberately unremarkable:
- CPU: AMD Ryzen 5 5600G (6 cores / 12 threads)
- RAM: 32 GB DDR4
- Storage: 500 GB NVMe for the system, 1 TB SSD for VMs
- GPU: a GTX 970 held back for local LLM work via passthrough
- Networking: separate bridges for LAN, an isolated lab subnet, and IoT
Nothing here is exotic, and that's the point. The constraint that matters is having hardware you're willing to break.
Choosing the hypervisor
VMware ESXi was the default answer for years. After the Broadcom acquisition, licensing costs climbed sharply and the free tier was discontinued. A closed ecosystem with a hostile pricing trajectory is a poor foundation for a learning environment.
Hyper-V is competent but tied to Windows Server licensing, PowerShell-first in a way that limits portability, and lacks a genuinely useful native REST API.
Proxmox VE won: open source under Apache 2.0, KVM-based with no proprietary abstraction layer, a native REST API that the Terraform provider builds on, a web UI included, and ZFS and LVM-thin support out of the box.
The API mattered more than anything else in that list, because it's what makes the lab programmable rather than clicked-together.
Infrastructure as a clone factory
The repository is named kamino, after the cloning planet — Terraform treats VMs as clones stamped from templates rather than pets configured by hand.
This is the habit worth building. When recreating a VM costs one command and zero dollars, you stop protecting broken machines. You destroy and re-provision, and in doing so you discover every piece of state you forgot to codify. That feedback loop is very hard to develop when each iteration bills you.
The phases
The lab is built in stages rather than all at once:
- Foundation — Proxmox, a development VM, local DNS, and the storage decision between ZFS and LVM-thin.
- Cloud native — multi-node Kubernetes with Helm and ArgoCD, doing GitOps properly.
- Observability — Prometheus, Grafana, Loki, Alertmanager.
- Security — secrets management, identity, intrusion detection, image scanning.
- Local AI — GPU passthrough for local model inference.
Each phase is independently useful, which matters because a homelab that requires full completion before delivering value is a homelab that gets abandoned in phase two.
The honest trade-off
You gain a place to be genuinely reckless, which is where the fastest learning happens, plus hardware that costs nothing per experiment once purchased.
You pay in electricity, in physical maintenance nobody else will do for you, and in the reality that your lab is not highly available. When the disk dies, you are the on-call rotation.
That last part is not a drawback. Being the person who has to fix it at 11pm is exactly how the operational lessons stick.