We Turned Three $20 Thrift-Store Desktops Into a Proxmox Cluster and Ran It for 90 Days

The pitch was simple enough to be dangerous: if a $6 VPS can run a small stack, three $20 desktops from Goodwill ought to run it better. More cores, more RAM, no monthly bill. We gave the idea 90 days and a Kill-A-Watt meter, and we wrote down everything.

What sixty dollars actually buys in 2026

Thrift-store PCs are a buyers market right now. Offices refreshed en masse during the Windows 10 sunset, and the machines that fell out of warranty landed on donation shelves across the country. On our first Saturday sweep we found three Dell OptiPlex 7010 towers at $20 apiece: Core i5-3470, 8GB DDR3, no drives, no OS, dust included at no extra charge.

We also grabbed a 5-port gigabit switch at a flea market for $18, three refurbished 240GB SATA SSDs at $21 each, and a 16GB DDR3 kit for $24 because one of the boxes only had a single 4GB stick.

Total outlay: $183. That is roughly 30 months of the $6 VPS we were benchmarking against.

One thing to know before you go hunting: DDR3-era machines are cheap but they are not free to run. Buy the tower only if you already have drives and a switch, or the math slides fast.

Building the cluster, and the first wall

Proxmox VE 8.x installed on all three without a hiccup. The installer has gotten genuinely friendly, and the ZFS mirror option on install is a single click. Within an hour we had three nodes, a quorum, and the smug feeling of someone who has clearly not been humbled yet.

Then the corosync networking ate our afternoon. A Proxmox cluster wants a stable, low-latency link between nodes, and the onboard Realtek NICs on these boxes are not it. We got flapping, phantom node drops, and one memorable evening where the cluster decided node two did not exist while node two was actively serving VMs. Moving cluster traffic to a dedicated VLAN on the switch helped. Adding a cheap Intel PCIe NIC to the node that complained most fixed the rest.

Lesson one: budget for a NIC before you budget for RAM.

Power draw: the number nobody puts on the Goodwill tag

Here is the part that changed how we think about e-waste. Measured at the wall with our meter:

• Idle cluster, all three nodes up, no VMs: 114W combined
• Typical load (a dozen containers and two small VMs): 148W
• Simulated patch window, all nodes working: 171W

At the US average residential rate of about $0.17 per kWh, 114W idle runs roughly $14 a month. Our 90-day power bill for the experiment came to just over $42.

The $6 VPS we compared against would have cost $18 over the same window.

That is the whole argument in two numbers. The hardware was free-ish. The electricity was not, and it quietly overtook the thing we were trying to beat.

Day 51: the drive that taught us about ZFS

One of the refurbished SSDs dropped out of the mirror at 3:12 in the morning on day 51. No warning, no smartctl story worth telling, just a pool in a degraded state and a node running on one leg.

This is where the lab earned its keep. ZFS mirrored the boot drive on install, so the node kept running. We swapped in a spare drive, ran a replace, and were whole again inside 40 minutes. Total cost of the failure: $21 for the replacement SSD and one hour of attention.

On a single-drive VPS you would never see this. You would also never have to think about it. Both of those are true at the same time.

Refurbished SSDs are a coin flip. Buy one more than you need and treat the spare as part of the build cost.

Cluster quirks that will eat your weekends

Three nodes is the smallest quorum-safe cluster, which means you have exactly zero redundancy in your redundancy. Lose one node and everything is fine. Lose two and the survivor fences itself and stops serving anything. That is correct behavior and it still feels like a betrayal at 11pm.

We also tried Proxmox's HA features for a couple of weeks and turned them off. HA assumes your fencing and your storage are solid. Ours were cheap. We got a VM migrated mid-write and a lock that took a manual clear to resolve. For a home lab, plain manual migration is boring and correct.

Local ZFS storage per node, no shared storage, no live migration. If you want the fancy stuff you are buying a NAS, and the NAS is the expensive part.

What it actually ran

For 90 days the cluster hosted:

• Pi-hole on one node, because DNS is the one thing that must never be down
• Home Assistant with a USB Zigbee stick passed through
• A Nextcloud instance for family file sync
• A Postgres container for a side project
• A backup target receiving nightly restic jobs from a laptop

Everything worked. Uptime across the cluster was 99.1 percent, and nearly all of the downtime was us, not the hardware. CPU load never crossed 40 percent on any node.

We were running a workload that a $6 VPS and a $5 storage add-on would have handled without complaint.

Where free hardware stops being worth it

Around day 60 we started noticing the tax. Not money, time. Twenty minutes here to chase a SMART warning. An hour there to debug why the switch dropped a link. A Saturday morning rebuilding the corosync config after the house lost power and the nodes came back in the wrong order.

Add it up across 90 days and we logged about 19 hours of troubleshooting and upkeep. At even a modest $25 an hour of your own time, that is $475 of labor attached to a $183 build.

That is the honest break-even point. Cheap hardware beats a small VPS the day your time is worth nothing, or the day you genuinely want the tinkering. If you are doing it to learn Proxmox clustering, this is the cheapest classroom you will ever find. If you are doing it to save $6 a month, you are paying for a hobby and calling it infrastructure.

The verdict

We are keeping the cluster. We are also keeping the VPS for anything that has to be up when we are not home.

E-waste is a fantastic lab. It is a mediocre production environment and a worse cost-saving strategy than the internet suggests. Buy the $20 desktops, learn the cluster, mirror your drives, and put a meter on the wall outlet before you tell anyone you saved money. The meter will keep you honest in a way that no spreadsheet does.