Hibiscus Media · Infrastructure Notes

The little box on my shelf runs AI that people buy $5,000 rigs for

How I got a compact mini PC in rural Japan running 32-billion-parameter AI models at full GPU speed — in under 30 minutes, using nothing but Ubuntu’s own packages.

There’s a machine the size of a lunchbox sitting on a shelf in my office in Uken Village — a small farming village on Amami Oshima, a subtropical island down the bottom of Japan. It’s a GEEKOM A9 Mega: compact, quiet, sips a few dollars a month of electricity, and it runs my whole AI setup locally. No cloud, no per-use fees, no client data ever leaving the building.

It’s got serious hardware inside — an AMD Ryzen AI Max+ 395, 128GB of unified memory, and a Radeon 8060S GPU. On paper it should chew through big AI models. But when I checked, every model was running on the CPU alone. The GPU — the fast bit, the whole reason you’d buy this chip — was sitting there doing nothing.

Here’s how I fixed it, why it was broken in the first place, and what it means if you’d like this kind of setup for your own business.

The short version

The GPU was invisible because the operating system’s kernel was older than the silicon. The chip is so new that the version of Linux I was running didn’t recognise it yet. Installing Ubuntu’s newer “HWE” kernel let the driver connect to the chip — and then, to my surprise, everything else sorted itself out automatically. Three commands and a reboot. That was the whole job.

Working out why the GPU was hiding

First rule of this kind of work: measure before you touch anything, so success has a number. On CPU only, the box generated text at 7.55 tokens a second and read through prompts at 43.26 tokens a second. Fine, but slow — especially when you’re feeding it long documents.

A quick check showed the driver was installed but never attached to the chip. The graphics tools could only see “llvmpipe” — a software stand-in that fakes a GPU using the CPU. No real GPU in sight. The running kernel had shipped before this hardware existed, so it simply didn’t know what it was looking at.

The fix: three commands and a reboot

Install Ubuntu’s Hardware Enablement kernel — Canonical’s official channel for exactly this situation, new hardware on a stable release. The old kernel stays put as a fallback, so it’s reversible by design:

sudo apt install --install-recommends linux-generic-hwe-24.04
sudo reboot

After the reboot, the one line I was hoping for appeared: “Kernel driver in use: amdgpu.” That’s the whole game — the GPU was finally connected. A quick permissions tweak to let the AI software talk to it, and I was ready to test.

The results

Same model, same machine, same test — the only thing that changed was where the work happened:

The 3.5× jump in reading is the one that counts in real life. Anything genuinely useful — analysing a document, working through a pile of records, running an AI agent — is read-heavy. That’s the difference between the AI feeling snappy and feeling like it’s wading through treacle.

So this little box now serves 32-billion-parameter models at 100% GPU, around the clock, for a few dollars a month in electricity — the kind of workload people buy dedicated GPU rigs for. And every tool on the machine inherited the speedup the moment I restarted it.

What this means if you run a business

Here’s why I bother writing this down. This same box — quiet, cheap to run, owned outright — can run your whole business privately: your files, your accounting, your office documents, and genuinely private AI, all on hardware you own, reachable securely from anywhere, with nothing flowing through anyone else’s cloud.

That’s the Privacy Stack: I spec the machine, build the whole thing remotely, hand you a plain-English manual, and you own it from day one. No monthly rent on your own data.

Private infrastructure, built and run from Amami Oshima, Japan.