Building the Proxmox Server: Virtualization for AI Workloads

It's 8:15 PM, and tonight's infrastructure work is not glamorous. No cluster is humming. No migration demo is running. There is no triumphant screenshot of a Proxmox dashboard with three green nodes.

There is a case. There is a power supply. There is a spreadsheet with numbers that are, frankly, rude.

That is where the server build actually starts.

The Old Stack Stays Put

I already have an old server stack: three Dell servers that have been doing useful work and will keep doing useful work. They are not the story here. I am not tearing them down to pretend I discovered virtualization in a blaze of enterprise recycling.

The Dell boxes are staying in place.

The new idea is different: build a fresh machine specifically for the next phase of Sorren.ai infrastructure. Something I can design around AI workloads, storage separation, virtualization, and future expansion instead of inheriting whatever constraints came with the old hardware.

That sounds simple until you start pricing parts.

The First Pieces

The first two purchases are not exciting, but they are foundational:

• a rackmountable ATX server case
• a 1600 watt power supply

The case matters because I want this machine to live with the rest of the infrastructure eventually. It is an ATX-style build, not a proprietary server chassis, but it can go in a rack. That gives me room to use commodity parts while still keeping the setup organized.

The power supply matters because AI workloads have a way of making "I'll just upgrade later" sound naive. GPUs are not subtle. They want power, cooling, airflow, and enough headroom that the machine does not become a space heater with aspirations.

A 1600 watt supply is overkill for the machine as it exists today, because the machine as it exists today is mostly an idea. But it is not overkill for the machine I am trying to build toward.

Buy once, cry once, as the hardware people say. The problem is that the crying starts immediately.

The Budget Reality

The costs are staggering.

That is not a complaint exactly. It is just the reality of trying to build infrastructure for AI work without pretending cloud bills are free or that consumer hardware magically becomes a datacenter because I used the word "server."

Right now I am low enough on funds that the staged approach is the only sane approach. Case and power supply first. Then figure out what comes next: motherboard, CPU, RAM, drives, GPU plans, cooling, rails, cabling, and all the little parts that somehow add up to another computer hiding inside the first computer.

This is one of the less romantic parts of building a company. Sometimes the strategic roadmap is: buy the box, stare at the box, decide what can go in the box next month.

Why Proxmox

The target is still Proxmox.

I want virtualization because I do not want every service living directly on bare metal. Databases, development environments, background workers, inference experiments, monitoring, and storage services all need isolation. Snapshots are useful. Clean rebuilds are useful. Being able to separate "this experiment broke" from "the whole machine is cursed" is useful.

Proxmox is the right fit for the kind of infrastructure I want: open, Linux-based, practical, and not allergic to homelab reality. It gives me VMs and containers without requiring me to turn this into a VMware licensing exercise. It also gives me a path to GPU passthrough later, which matters if this box becomes part of the AI workload stack.

But I am not building the cluster tonight. I am building toward the machine that will eventually host it.

Storage Plan: Fast Where It Matters, Cheap Where It Doesn't

The storage plan is simple in principle:

• SSDs for anything actively running
• old magnetic drives for backups and cold storage

The active workloads need fast storage. VMs, containers, databases, indexes, and anything touching model or memory pipelines should live on SSDs. Waiting on disk because I saved a few dollars is the kind of optimization that punishes you every day afterward.

But not everything needs to be fast. Backups do not need NVMe. Cold file storage does not need to win a benchmark. Old magnetic drives are still useful if you ask them to do the right job: hold data, hold backups, hold things that need to exist but not necessarily respond in milliseconds.

That split feels right for Sorren.ai at this stage. Spend money where latency matters. Use boring storage where boring storage is enough.

Learning Linux the Practical Way

The other half of this work is not in the server case at all. It is on my workstation.

I have been switching my workstation toward Linux in a dual-boot configuration so I can get more comfortable with Ubuntu and more comfortable living in the terminal. That sounds like a small personal preference, but it is really part of the infrastructure plan.

If the servers are going to run Linux, and the tooling is going to be Linux-first, and the deployment path is going to involve SSH, systemd, Docker, Proxmox, logs, mounts, permissions, and the usual pile of command-line reality, then I need those motions to become normal.

You can administer Linux from Windows. I have done it. But there is a difference between using the terminal when forced and thinking in it naturally. Dual boot is my way of closing that gap without pretending I am already where I want to be.

What This Build Is Really For

This server is not just about compute. It is about having a place where the architecture can grow without turning every experiment into a production risk.

Companion needs databases, memory pipelines, background summarization, vector search, and eventually model workloads. Tutor will need its own services. The company site, internal tools, testing environments, and automation loops all need places to live. Some of that can run on existing hardware for now. Some of it should not.

A proper Proxmox host gives me room to separate those concerns.

Maybe later this becomes a cluster. Maybe the old Dell servers get folded into a wider virtualization setup. Maybe this new machine becomes the main node and the old stack keeps doing backup duty. I do not know yet. That is part of why I am building carefully instead of pretending the final architecture is obvious.

The Honest Status

So here is the honest status: the build has started, but barely.

The old Dell servers are still in place. The new rackmount ATX case and 1600 watt power supply are the first step toward a dedicated Proxmox box. The storage plan is taking shape: SSDs for active workloads, magnetic drives for backups and cold storage. The workstation is moving toward dual-boot Linux because I need Ubuntu and the terminal to feel like home, not like a place I visit when something breaks.

It is not a cluster yet.

It is a foundation.

Sometimes that is what infrastructure looks like: not a rack full of blinking lights, but a big empty case, an oversized power supply, and a plan that still has to survive the budget.

It's past 8 PM on a Thursday. Good enough.