Why I’m Replacing a Homelab That Still Works

For several years my homelab has run on two 1U Supermicro servers — chimaera and basilisk — paired together as an XCP-ng pool. Each has two Xeons, 128 GB of ECC RAM, dual 10-gigabit networking, and local disk. Between them they offer 24 cores, 48 threads, and 256 GB of RAM. They work. Nothing is failing. That is the part that makes a setup hard to walk away from.

But “it works” hides four problems I had ignored for too long.

It’s over-provisioned for the actual workload

Twenty-four cores and a quarter-terabyte of RAM is a lot of capacity. Over its life the pool has run the lab’s core services — Active Directory, backups with Veeam, Mastodon, GitLab, a torrent box, Home Assistant, and monitoring — but never enough at once to justify two servers of that size. Retiring it is still in progress: the goal is to switch off both chimaera and basilisk and run everything on the two Proxmox mini PCs. basilisk is already powered down as an interim step; chimaera still carries what has not migrated yet.

It draws real power to do very little

Two dual-Xeon servers of this generation, idling for most of the day, draw on the order of 300 watts between them, continuously — roughly 2,300 kWh over a year if both run around the clock. In Maine, which has some of the highest residential electricity prices in the country, that is not a rounding error. They also dump heat into the room that then has to be removed, and 1U servers cool themselves with small, high-RPM fans that are audible from the next room. The full before-and-after numbers get their own post.

It depends on Windows licensing

Several core pieces are tied to Windows. Active Directory runs on two Windows Server 2019 Essentials domain controllers, vader and maul. Backups run on piett — a full Windows Server running Veeam Backup & Replication, not an appliance. Windows Server 2019 left mainstream support on 9 January 2024 and is now in security-only extended support through January 2029.¹ It still receives patches, but it is on the back half of its supported life, and each of those roles is a license to track and a machine that exists mainly to run one vendor’s product.

The platform is from another era

The Supermicro boards are from the early 2010s (Ivy Bridge-era Xeons), the RAM is DDR3-1333, and the final BIOS dates to 2015. XCP-ng 8.2.1 underneath is a long-term-support release, but an old one. None of it is broken. All of it is getting harder to justify keeping online.

The plan: consolidate, and remove the Windows dependencies

The replacement is small: two Minisforum mini PCs — proxmox-mini01 and proxmox-mini02 — running Proxmox as a cluster. They draw a fraction of the power, run quietly, and have ample capacity for a workload that already fits on a single old node.

Alongside the hardware swap, each Windows-dependent piece gets an open-source replacement:

• Active Directory moves from the Windows domain controllers (vader, maul) to Samba on a new DC, sidious.
• Veeam on piett gives way to UrBackup.
• XCP-ng gives way to the Proxmox cluster.

The rule for the rebuild is simple: if it isn’t captured as code, it didn’t happen. Hosts are provisioned with Terraform and configured with Ansible; real decisions are written down as architecture decision records; fiddly procedures become runbooks.

Working with Claude

I’m running the whole rebuild as a working collaboration with Claude Code. It drafts Terraform and Ansible, writes and cross-checks the decision records and runbooks, catches steps I would have skipped, and acts as a second set of eyes on every change before I run it. I make the decisions and do the physical work; the AI does much of the reading, drafting, and double-checking that would otherwise eat my evenings. That working relationship is a thread I’ll keep returning to, because it has changed both how much I can get done and how well I document it.

The payoff is already showing: quieter, cooler, lower-power, and reproducible from a git repository instead of from memory. Next: what the old servers actually cost to run, what the mini PCs cost instead, and — since part of the point was to be greener — how that compares to the energy the AI itself used.

AI-assistant disclaimer: This post was drafted by Claude Code (Claude Opus 4.8, claude-opus-4-8) from my homelab session notes, commit history, and architecture decision records, then reviewed and edited by me. It may contain inaccuracies; verify specifics before relying on them.