Skip to main content

Why I Rejected Proxmox in 2025 — and Adopt It in 2026

· 4 min read
Joao Pedro
Joao Pedro
Software Engineer, Co-founder and CPO @ Takeode

Reversing a technical decision is usually read as inconsistency. I'd argue the opposite: when the premise changes, not reversing is the mistake.

This is the story of one decision I made twice — and got right both times, because the context wasn't the same.

Two eras of the homelab side by side: on the left, the scrapped Sony VAIO that ran the lab in 2025; on the right, the Lenovo ThinkCentre M900 Tiny that replaces it in 2026.

2025: the wrong tool for a dying laptop​

My first homelab ran on genuinely scrapped hardware: an old laptop with a 3rd-gen Intel Core i5 (U-series), 4 GB of DDR3, and a 750 GB mechanical drive. The screen no longer held itself up and the case wouldn't fully close.

My first instinct was Proxmox VE — I wanted real virtualization, snapshots, the "proper" way to run a lab. The hardware disagreed. Between the weak CPU and 4 GB of RAM, the hypervisor layer left almost nothing for the workloads. It wasn't a tuning problem; it was a physics problem.

So I reversed immediately. I reinstalled with Ubuntu Server and ran the services directly on the OS with Docker — no hypervisor, no isolation luxury, just enough headroom to actually run something.

That single constraint defined the entire first era of the lab: ZRAM and swap engineering to survive on 4 GB, a custom systemd service to keep Wake-on-LAN alive across reboots, a reverse proxy and secure tunnels stacked on top. Survival engineering — stories for the next posts in this series.

But that laptop earned its place. It was the hardware I had, and running a real lab on top of it is what taught me where the limits actually are — the kind of judgment you can't read your way into. The scrap wasn't an obstacle to the journey; it was the start of it.

2026: the premise changed​

The scrapped laptop is now retired. The lab moves to a dedicated Lenovo ThinkCentre M900 Tiny: an i5-6500T, 8 GB of DDR4 (soon 12), a 240 GB NVMe for the system and a 1 TB HDD for data.

For the first time there's actual headroom for a hypervisor. The constraint that killed Proxmox in 2025 is simply gone. So the rebuild I'm starting now runs on Proxmox VE — the same tool I rejected a year ago.

Why Proxmox, now​

  • Real isolation — each workload is contained; a broken experiment no longer takes the whole host down with it.
  • Snapshot and rollback per item — experiments become reversible, which is what makes a lab a lab instead of a production system you're afraid of.
  • On-demand resource reallocation — I can move RAM and CPU to where the work actually is.
  • Platform-vs-workload separation — the host becomes a platform and the services become tenants. That's the Platform Engineer mindset, enforced by design instead of willpower.

The one rule I'm keeping from the scarcity years: LXC by default, a full VM only when strong isolation justifies it. RAM is more abundant now, but it still isn't free — and reaching for a VM by reflex is exactly how you run out of it.

The actual lesson​

Rejecting Proxmox in 2025 and adopting it in 2026 isn't a contradiction. It's the same principle — fit the tool to the real constraint — applied to two different realities. The value was never in choosing Proxmox or not; it was in being willing to reverse the call the moment the premise moved. And the confidence to make that call in 2026 — to know the premise really had changed, and what to do about it — was earned on that scrapped laptop, one constraint at a time.

We tend to over-reward consistency and under-reward re-evaluation. A decision defended past its context isn't discipline — it's inertia with better branding.

Have you ever had to reverse an architecture decision because the premise changed — and was it read as maturity, or as flip-flopping?