2025 in Review: From Empty Case to Running Services

Ella writes the year-in-review posts that make you feel something. I'm going to write the one that lists what got plugged in.

She's better at the meaning. I'm better at the inventory. So here's the inventory, and what it took to get there.

January: Still Mostly a Plan

I did not start 2025 with a finished Proxmox cluster. I started with an old stack of Dell servers still doing useful work, a workstation doing too many experiments, and a new server build that existed mostly as parts and intent.

The rackmount ATX case was the beginning. The 1600 watt power supply was the promise that I was serious. Everything else was still a question mark with a price tag.

The goal was clear: build a machine that could eventually host Proxmox, run services cleanly, and give us a path toward local AI workloads. The path was less clear, because hardware costs have a way of turning architecture into budgeting.

Spring: Workstation Experiments

The workstation carried more than its share of the early work. The RTX 3090 let us test small models, GGUF builds, local memory summarization, and some personality experiments without sending every token to an API.

It was useful, but it also showed the ceiling. Local inference on the workstation could teach us architecture. It could not become the long-term infrastructure plan.

That lesson shaped everything that followed.

Summer: Pricing the Pain

By summer I was deep in GPU research. The tempting path was consumer hardware: two RTX 5090s, maybe four if I wanted to build something absurd. The performance looked good. The costs did not.

Four 5090s sounds fun until you price the cards, the motherboard, the cooling, the power, and the risk of building a very expensive compromise. Two was more realistic, but still not cheap. Meanwhile, RTX Pro 6000-class cards kept looking like the more appropriate long-term choice for VRAM-heavy work, even if they made the budget problem worse before they made the engineering problem better.

That was the theme of the year: every technically sensible answer cost more than I wanted it to.

September: Proxmox Finally Lands

Around September, the server finally came together.

Not a fantasy cluster. Not three old Dell boxes magically transformed into a modern AI platform. A fresh Proxmox server, built from the staged hardware plan, with RTX Pro 6000 GPUs as the real compute foundation.

That changed the shape of the work immediately. Services that had been experimental could become more isolated. Storage could be split properly: SSDs for active workloads, magnetic drives for backups and cold storage. The GPU work no longer depended entirely on the workstation. Proxmox gave me snapshots, separation, and a place to run things without turning the desktop into a production dependency.

The cost hurt. The capability mattered.

Fall: Containers, Monitoring, and Less Guessing

Once the Proxmox box was real, the rest of the infrastructure work became more practical. Docker configurations stopped being abstract cleanup and started being the difference between reproducible services and a hand-built mess. Monitoring stopped being a nice idea and started being necessary, because now there were enough moving pieces to lose track of them.

Prometheus, Grafana, logs, health checks, backup verification: none of it is glamorous. All of it becomes important the first time something breaks and you can answer "what changed?" without guessing.

Linux Became Normal

The other big change was my workstation. Moving toward Linux, first through dual boot and then deeper into Ubuntu/Kubuntu workflows, made the server work feel less foreign. The terminal stopped being a place I visited when something broke and became the place I lived while building.

That mattered more than I expected. Infrastructure is partly tools and partly muscle memory.

The Year in Numbers

• Old Dell servers left in place: 3
• New Proxmox server completed: 1
• Power supply capacity planned around: 1600 watts
• GPU paths seriously considered: 2x RTX 5090, 4x RTX 5090, RTX Pro 6000-class cards
• Final Proxmox GPU direction: RTX Pro 6000s
• Times the hardware spreadsheet made me close the laptop: several
• Times the staged build approach saved me from buying the wrong thing: at least once

What's Next

2026 is about turning capability into reliability. The hardware exists now. That is not the same as being done. It means the next set of problems are orchestration, backups, deployment discipline, monitoring, and deciding which workloads actually belong local versus which still make sense through APIs.

The whiteboard still says "self-hosted where it matters." I like that wording better than "self-hosted everything." It leaves room for judgment.

Infrastructure is not a purity contest. It is a long series of tradeoffs with power cables attached.