04.07.2026
So I have started this year to be much more active with having my own personal "sovereignty" when it comes to my systems, as an example, I have ordered last December (around Christmas time) a NAS to have my own media, I basically do not want to actually spend the obscene money for iCloud or Google Drive to host my own media and backups, therefore it was something that I wanted to have. I did manage to build it up, and that's when I entered the rabbit hole.
The specs were okay but nothing out of this world, yet one thing was important for me when selecting the NAS, I wanted it expandable (being a NAS of course storage is expandable) so I looked for a NAS that would let me upgrade at least the amount of RAM. Some of you might be thinking why specially for pre-built NAS hardware normally they come with just a sort of barebones multimedia server type of thing, but I was lucky enough to also find something that lets me run some nice Docker containers, any of my choosing on it therefore RAM upgradability was important. I ended up going with a multi-core CPU (nothing fancy) and then ordering some disks and scavanged classifieds to find a nice 32GB RAM stick to add. The ultimate goal for the NAS which is yet to be reached is to host some scheduled services on it, things that fetch periodic data that I might find useful such as weather, sports outcomes, appointments, events and whatnot. Then whenever I am home I can sync these, or I can also host on the NAS some form of Push Notification to send it to my devices. So this is something that I have been very focused on and even though the full potential is yet to be achieved it is something that I really wanted to invest in.
Fast forward a few months and, I start to realise that I also would like to have some more alternatives to my hobby coding, so I start to do some research and think, what can I do next to also get rid of my dependency on 3rd party LLM services? So I did what a non-sensible coder would do and decided that the real solution is to also find a way to self-host it. Not really sure why but the first thought was to not use any Macbook or Apple hardware and maybe go for Linux, my university past-self shed a teer for sure, and so I started again to scavenge around classifieds, also using Gemini to "brainstorm" what I needed and the perfect combinations, and I found a nice and cheap Thinkpad L14, "brick" shaped and platicky but, as a common theme, it was expandable. Of course, I cannot change the CPU but I saw that a battery sawp is possible, in case it goes bad, and both the storage and RAM were also upgradable, so again I embarked on another classifieds hunt so that I could "spec out" this laptop. Long story short it has 64GB of RAM and 1TB of SSD, and it runs Linux, more concretely PopOS! Some of you might be wondering how did reliance on third party LLMs got me to a new PopOS Thinkpad, and here is the reason (probably stupid), I want to now have a sort of external GPU setup to run an LLM on an external graphics card, or cards, so that I can have my own.
This is a very good question, and probably it will involve quite a bit of "copium" on my side, but the truth is that I have been playing around with pi.dev and in order to have an LLM to connect to my pi.dev harness I have embarked on a journey through DeepSeek v4. Journey of which I am enjoying very much, but you see the timing was just perfect to create the current predicament, v4 had just come out and DeepSeek labs was running a very nice promotion where the Pro version of the model was very cheap (I recall it was the same price as the less-powerful but still very capable Flash version) therefore my 10$ of initial investment was very slowly draining while I prompted everything and anything. Fast-forward a couple of weeks and the price for Pro increased (stopped being on sale) and I naturally shift towards Flash, specially since most of my prompting is quite specific, and things kept going well for another couple of weeks, and now we reach the culmination of why I want to self-host: DeepSeek is following suite of the American companies and introducing Peak Hour Windows, where they will double yet again the pricing for both Pro and Flash models. Since I am in Europe, and the peak hours seem to be mainly coinciding with workdays in China I am half safe. What this means is that in the morning it would get doubled but after 12:00 the pricing comes back to normal.
Of course pricing is the real reason to switch to self-hosting, and of course it is still not horrible but there is a chance that the pricing doubles again, it is mainly a question of time, specially in these bellic times that we are currently going through. The world is becoming more uncerstain about the future, therefore the best way to handle it is by controlling as much of your life as possible.
So what am I going to try to do, you might be thinking? Well I will try to find some nice GPU and eGPU setup on classifieds to build it up. This should not be very hard to be quite frank but I am hoping that the open-source models improve but also become more efficient. I will continue to use DeepSeek, I honestly do not blame them or have anything against what they did. To me it's perfectly understandable and it continues to be the cheapest OSS model that is provided via an API and it does work wonderfully for me and my setup, therefore I will continue to use it, specially outside of the "peak hours" they announced.
The main reason why I am writing this is to share a bit of my journey, specially for people that don't know and are a bit more afraid to thinker around and mess up with hardware. Honestly there is nothing to fear and overall you can always consult with ChatGPT, Gemini or DeepSeek (any LLM) about how to self-host it or thinker with hardware.
Later, I will go into a nice deepdive about pi.dev, I think that is an article in and of itself and what I have been modifying on it to maximise my productivity. Have a great rest of the day!