Run Local AI on Solar: The Small-Business Starter Stack
You need less than you think: a Mac mini (30-50W, silent, runs excellent local models) or a used GPU workstation (300-600W, runs bigger ones), Ollama as the software on-ramp, ONE repeated workload wired first, and your heavy jobs batched to the midday solar window. Cost: roughly $600-$2,500 once, near-zero marginal energy on owned solar, versus cloud AI's $20-60 per seat per month forever. Honest limit: local models trail the frontier on the hardest tasks — the design decision (daytime-only vs battery-aware vs cloud-fallback) is the part worth getting right.
Surplus solar should do work. This is the page where the thesis becomes a parts list.
The Two Hardware Paths
| Path | What it is | Honest fit |
|---|---|---|
| The quiet path: Mac mini | M-series, 16-32GB unified memory, 30-50W, silent, ~$600-1,400 (used to new) | Drafting, research, summarization, extraction — most small-business AI work. Sips from any solar system without scheduling gymnastics |
| The muscle path: GPU workstation | Used tower + one consumer GPU, 300-600W under load, ~$1,200-2,500 | Bigger models, faster generation, batch-heavy workloads. This is the box that earns sun-scheduling |
| The wrong path: overbuying | A rack of GPUs because a YouTube video said so | A small business's repeated workloads rarely justify it. Start small; the workload will tell you if it needs more |
The Starter Stack, In Order
| Step | The move |
|---|---|
| 1 · Pick ONE workload | The repeated task that eats the most hours: proposal drafting, document summarizing, intake triage, first-pass research. One. The lane's pilot rule applies to you too |
| 2 · Install Ollama, pull a capable model | One command each. You now have a private assistant with no meter running |
| 3 · Wire the workload | A prompt template and a folder convention beat a fancy app. Make the task repeatable before making it pretty |
| 4 · Schedule the heavy jobs to the sun | Batch runs at 10am-3pm eat the surplus your export credit undervalues. Mac mini users can skip this step entirely |
| 5 · Decide the fallback | Daytime-only, battery-aware, or cloud-fallback for the hard 20%. This design call is the difference between a toy and a tool |
The Honest Verdict
This setup will not replace the frontier models on the hardest reasoning work, and anyone who tells you otherwise is selling hardware. What it does: run 70-80% of a small business's repeated AI workload privately, at near-zero marginal cost on owned solar, with your data never leaving the building — in the county with the best solar economics and the highest retail rates in America. San Diego operators with panels on the roof are already paid up for the compute era; most just haven't plugged it in yet. If you want the version scoped to your actual system and workload, that's the Solar AI Hour: $150, one operator, workload picked, power math done, stack recommended, first workflow mapped.
