Trying to preserve my hobbies
Computer hardware and the AI memory crunch
Around November or December of 2025, I went on a bit of a spending spree with computer hardware - I purchased two different GMKtec mini PCs with respectable internals, an external GPU dock, a new laptop (my first gaming laptop with a dedicated GPU... it's about time), and a new ATX chassis + CPU cooler to refurbish an older Ryzen 5800X3D machine that I had. On top of that, last Christmas, I was given a little pink-colored Intel N100 low-power mini PC, which now sits in the network rack as a Proxmox node.
In most other contexts, this string of purchases would seem excessive. In the current moment, though, I now feel somewhat justified. The AI crunch is destroying the hobbies of video game and PC-building enthusiasts everywhere right now. It sucks, and the long-term implications of more and more average people getting priced out of meaningfully-powerful local compute makes me very sad. ("You will own nothing, and you will like it," etc.)
So for now, I feel like I have all the devices I need (if not a few extra).
With the prices of generative AI subscriptions going up, however, I've been wondering if maybe I should invest in some kind of hardware that would be better at running local LLMs. I take a somewhat moderate stance on generative AI: while I think it ultimately sucks at doing most things, there are a few things it can do well sometimes, like coding and certain types of data analysis. Instead of having to rely on an expensive 3rd-party to charge me a monthly fee and make available their own LLM which I have little control over (and is possibly sucking up all of my data), it's seeming more and more preferable these days to have a self-hosted solution (just like it is for so many other types of services).
If it's not clear already that I have a weird fascination with miniature PCs, then I will say at this point that I am now looking at more mini PCs, this time on AMD's Strix Halo platform, specifically the AI Max+ 395 CPU. It's not the most powerful platform for running local LLMs, but with a specially-tuned efficient Mixture of Experts (MoE) model taking advantage of the large amount of shared memory (128GB), it can certainly punch above it's weight.
There's the problem: It needs 128GB to really shine running local models. And it's not November 2025 anymore; it is now June 2026. Prices have gone absolutely insane over the past 6 months.
Popular laptop and Chinese mini PC models that utilize the Ryzen AI Max+ 395 CPU have, quite literally, gone up in price by over a thousand dollars since initial release. Most models, like the GMKtec EVO-X2 mini PC (kitted out with 128GB of memory) now demand over **$3,000 USD***.
So that's a tough pill to swallow. Meanwhile, I have this refurbished 5800X3D sitting in the corner of my room, basically good as new, not being used. The AM4 platform is still popular and has a lot of life left in it. Also, that tower in the corner has a Radeon RX 7800 XT in it, along with 32GB of DDR4 RAM. Should I sell that thing and use it toward a new AI computer? Is it too late, and everything is too expensive anymore? Am I out of my mind for wanting to be able to self-host LLMs in my own house?
As the middle class gets eviscerated, the "K-shaped" economy grows, and wealthy elites attempt to claw everything for themselves, I don't think I'm actually that crazy. Still noodling on what my local setup would look like. If it's small and efficient enough, I could potentially hook it up to a solar generator and have it entirely solar-powered. Think about that: A carbon-negative generative AI server for your home, serving up your own models.
I wish our politicians were smart enough to be this forward-thinking. Instead we are building gargantuan datacenters across the state, harming just about everyone downstream of the powers who build them and run them, and calling them "AI factories" to make them sound more palatable (Does this really work? Do people actually fall for that term, given the current mainstream sentiment around generative AI? Honest question, because I don't actually know).