Probably the world's first quad-GPU, DDR3-based rig using HDDs for LLM storage at this model size. It cost about a grand. Here's the reasoning.
| Line | What | Approx. cost |
|---|---|---|
| Board + CPUs + RAM | Z9PE-D16, 2x E5-2620 v2, 128GB DDR3 — one retired-server ecosystem, bought as a unit | $150–250 |
| GPUs | 4x Tesla P100 16GB from the decommission wave | $600–700 |
| Storage | 120GB SSD + 4x Velociraptor 500GB | $50–80 |
| Chassis, PSU, misc | The parts that make it not a fire | $100–150 |
| Power | 100% solar — 28.94kW array + 85kWh battery, paid off Oct 2023 | $0.00 / token |
Total: about a grand. The same 64GB of inference VRAM rented from a cloud provider is a subscription you never stop paying. This is a one-time purchase with a power bill.
V1 proved the concept on less: an i7-3930K with three RTX 2080 Ti blowers and 64GB of DDR3, 91.5 tps on a 35B model, $771 out of pocket, running 24/7 on solar. V2 is the version that runs the house full-time: four GPUs instead of three, the model size more than tripled, the serving stack hardened, the HDD array dedicated to models so the OS SSD stays boring and reliable. The full V1 build page lives at lydie.cc/solar-ai-server.html; the solar system that powers both machines is documented at lydie.cc/solar.html.
The parts are all from the shelf the industry walks away from — DDR3, LGA2011, Pascal. That shelf is the point. Every component was surplus, which means every component was cheap, which means the machine exists at a price a household can pay. New-generation hardware can't make this argument: the same build in current parts is five to ten times the money for tokens that arrive faster than anyone at this scale needs them.
A grand, a kilowatt of sunshine, and a house that thinks for itself.