Three RTX 2080 Ti blowers, 64GB of DDR3, and a case that was already in the shop. 91.5 tokens per second on a 35B model, running 24/7 on solar, for $771 out of pocket.
V1 build page & photos| Component | Part | Notes |
|---|---|---|
| Motherboard | Gigabyte GA-X79-UD5 | LGA2011, the consumer-side cousin of V2's server board |
| CPU | Intel Core i7-3930K | 6 cores / 12 threads, the overclocker's chip from 2011 |
| GPU | 3x NVIDIA GeForce RTX 2080 Ti Blower Edition | 33GB VRAM total; blowers because three cards in one case need exhaust, not recirculation |
| RAM | 64GB DDR3 (8x 8GB G.Skill Ripjaws X) | Same generation V2 runs on — the family trait |
| PSU | Corsair RM1000e, refurbished | 1kW of refurbished supply for a 550W load |
| Storage | 240GB SATA SSD (OS) + 3x 1TB SATA SSD (AI array) | V2 moved model storage to HDDs; V1 kept it on SSDs |
| Case | Advantech Industrial Computer 510 | Was already around. It's packed in there. |
| Layer | Detail |
|---|---|
| OS | Ubuntu Server 26.04 LTS, headless |
| Networking | Netplan + NetBird VPN for remote access |
| AI stack | Unsloth Studio, Hermes Agent, Docker |
| Model | Qwen3.6-35B-A3B-UD-Q4_K_M |
| Throughput | ~91.5 tokens/sec — impressive for a DDR3 platform and a 35B model |
| Power | ~100W idle, 550W+ across the three GPUs during inference — all of it solar |
V1 proved the model fits the business model. V2 scales it: 33GB VRAM becomes 64GB, a 35B model becomes 120B, 3 consumer GPUs become 4 datacenter accelerators, and the serving stack hardened into the machine that runs the house today. The constants across both builds: DDR3, the used market, solar power, Hermes Agent, and a refusal to pay per token.
V2 is slower per token and 3.4x the model. That trade is the whole point of a 120B machine: answers you can trust, not answers that arrive fast.
The original V1 page with build photos: lydie.cc/solar-ai-server.html — V1's own self-built site lives at lydie.cc/purpleai.