Why I have eight servers at home and six outside

Eight servers at home and six rented outside. Put like that, it sounds like an exaggeration, or like someone who doesn't know when to stop. There's a bit of both, but every machine has its reason, and here I'll tell you what it is.

When I mention it, the reaction is usually the same. A friend once asked me whether my electricity bill came in volumes, and another whether I ran the air conditioning in January to cool the machines. The picture people have is a room full of blinking cabinets. The reality is far more modest: several towers and small boxes spread between a study and a cupboard, with more cables than I'd like to admit.

There are two promises doing the rounds. One says "put everything in the cloud and forget about it". The other, the exact opposite, says "run everything at home and depend on nobody". I tried something close to both, and neither worked for me on its own.

This article explains what each thing is, how I got here, what runs where and how I decide. At the end you'll take away four questions for deciding where to put your own stuff.

For beginners, two definitions.

A VPS, or virtual private server, is a slice of a big server that you rent from a provider. To you it behaves like a whole computer, on day and night and connected to the internet. Several VPSs from different customers share the same physical machine, and each is allocated the processor, memory and disk set by the plan you've paid for. How truly reserved they are depends on the provider: some hand out to their customers more than the machine actually has, counting on not everyone using it at once. You're in charge: you install whatever you like. But you also maintain it.

An internal server is a machine of your own, at home or in your office, inside your own network. Nobody rents it to you and nobody switches it off, except the electricity company.

In my case, "server" is a generous word. Of the eight, only one has a graphics card (the server that runs the models), another is my workstation, two are low-power mini PCs that are always on, three are recycled old machines and one is a storage server with several disks.

A comparison to pin the idea down: the VPS is a rented flat in the city centre, well connected and always accessible, but where you can't knock down walls. The internal server is your own house on the outskirts: you do what you like, but you fix the leaks yourself.

None of this was a plan.

The first machine was the Linux server with the RTX 4070, the first of my life, which I set up in December 2023 with ChatGPT's help without being very sure what I wanted it for. The first VPS arrived that same year, to host a website for a client who didn't want it on any shared hosting.

From there, the growth was partly out of need and partly out of curiosity. In 2024 local AI arrived: the 4070 server finally found its job. That year three more VPSs came along too, for client services. By 2025 there were six, and between 2025 and 2026 agents and services that had to be available at all times filled out the internal network to eight machines.

To be honest, half of it was planned and the other half was putting out fires. Something was needed, it went wherever there was room, and it would all get sorted out later.

The moment that forced me to sort it out was a power cut that lasted a whole afternoon at home. That day I discovered that a service a client used depended on a machine in my study. My client was without service for four hours and I was without excuses. Since then, what has to be always available lives outside, and what has to be close lives inside.

The internal network is home, above all, to local AI. That's where the programs I use to run models on my own machines live, Ollama, LM Studio, llama.cpp and Unsloth, and it's where I've tested more than fifty models. Also the data I don't want leaving the house, the tests for everything that later ends up outside, and the VPS backups, which come down every night to the storage machine.

Why inside? For three reasons. The graphics card: models need graphics memory, and renting it outside is expensive. Privacy: some client data must not leave a controlled network. Mind you, "no cloud" isn't the same as "data under control": it also matters which programs send statistics out, who has access to each machine, what the contract with the client says and which connections can leave the house. That's why the machines holding client data can only connect outwards to update themselves, and only my user can log into them. And not paying per use: if a task is repeated thousands of times, at home it only costs electricity.

An example of real work that only makes sense at home: summarising and classifying a client's internal documents every week with a local model. By contract, those documents can't go through external services, so the alternative wasn't doing it in the cloud, it was not doing it at all.

I won't go machine by machine, because I cover that in What runs on each of my machines.

The six VPSs are home to whatever has to face the internet: the websites, mine and some clients', the public services that serve people outside, the agents that have to be available at all times and a couple of internal tools I access from my phone when I'm away from home.

This always raises the question of whether a VPS is any good for AI. Yes, with caveats. It's worth separating two jobs: the one that thinks, which is the model, and the one that organises, which is the program that receives requests, calls the model and does things with the answer. The organiser lives very happily on a modest VPS. The thinker, if it's a big model, doesn't: ordinary VPSs don't come with a graphics card, and a model without one answers at a snail's pace.

An example of something that lives outside and couldn't live at home: an agent that handles enquiries on a client's website at any hour. It calls a cloud model to do the thinking and runs from a VPS that doesn't depend on my connection or my electricity. If it lived in my study, every power cut would mean an angry client.

Over time I've boiled the decision down to four questions. It's the most practical way I know to settle the own server or VPS question.

1. Data. Can it leave the house? If not, inside.
2. Cost. What works out cheaper, counting the hours too? Your own server looks free once bought, but it has electricity, breakdowns and maintenance.
3. Availability. Does it have to be on all the time? If a power cut at home is a problem, outside. And it's not just the power: the UPS holds out for a few minutes, enough for a flicker but not for a long cut; the home fibre also goes down now and then, and everything outside that needs something inside goes through the VPS acting as the gateway to the internal network. If that gateway fails, the outside keeps working, but without access to home until I bring it back up.
4. Exposure. Does it need to be visible from the internet? Opening a home machine to the world is possible, but it's one more door to watch.

Two examples of mine. The client's weekly document summary: the data can't leave, the cost at home is just electricity, it doesn't have to be always available and it isn't visible from the internet. Four answers, one destination: inside. The website enquiries agent: the data is public, the cost of a small VPS is low, it has to work around the clock and it has to be visible from the internet. Outside, no argument.

It almost never comes out as "all inside" or "all outside". The normal thing is to combine: the outside agent, for example, stores its logs on an inside machine.

The figures, as promised. The six VPSs cost me about €92 a month in total, including tax. The electricity for the internal network, which I measured with a plug-in meter for a month, comes to around €50 a month; a bit over a quarter of it is down to the server with the graphics card. And you have to add depreciation of the hardware, about €97 a month spreading its price over five years, UPS included; breakdowns and spare parts, about €11; and the fibre upgrade with a fixed IP, €15. In total, about €173 a month. Backups go to the data server at home, so their cost is already in those figures; what isn't there is the time it takes to recover something when it fails. And maintenance, between eight and ten hours a month, between updates, backups, checking logs and the occasional scare.

That's not nothing. Whether it's worth it I go into calmly in Is it worth having your own servers?.

The mistake I made at the start: I left an internal server's admin panel exposed to the internet, with a short password, "just while I was setting it up". I forgot about it for weeks. When I checked the logs, there were thousands of login attempts from half the world. They didn't get in, as far as I know, but I spent a weekend checking everything and changing passwords. First I cut off access from outside. Then I changed the panel password and, just in case, every password stored on that machine or similar to it, plus the access keys. I went through the logs for any attempts that didn't end in "access denied", the users and the running processes. I found nothing odd, but since I couldn't be completely sure, I reinstalled the service from scratch and restored its data from a backup taken before the exposure.

The lesson is the usual one: temporary things stay.

If you want more detail, it's in How I set up a new VPS, Linux or Windows on my servers, What runs on each of my machines and Is it worth having your own servers?.