Build Your Rig
Drag to orbit the backpack. The sun's height follows your sun hours slider, panels glow when they're generating, and the battery bar on the pack fills or drains in real time.
drag to rotate
—
Wh harvested
—
Wh consumed
—
est. tokens/day
—
daily balance
—
The honest math
Rules of thumb behind the calculator:
- Backpack panels are small. Flexible panels that fit a pack are typically 10–60 W each. Real-world output is ~70–80% of the rating (heat, angle, clouds) — the calculator applies a 75% derate.
- "Sun hours" ≠ daylight hours. A sunny mid-latitude day gives ~4–5 equivalent full-sun hours; overcast winter days can give under 1.
- A 27B model needs real hardware. A quantized ~27B model on a 64 GB unified-memory mini-PC draws roughly 40–90 W during generation and produces on the order of 10–25 tokens/second — the calculator assumes ~15 tok/s at your chosen wattage.
- Idle isn't free. Even "sleeping," the PC sips 5–10 W. Budget a battery (~500 Wh) to buffer night and clouds.
| Scenario | Harvest | Consumption | Result |
|---|---|---|---|
| 4 × 30 W, sunny day (4.5 h), 3 h of AI at 55 W | 4 × 30 × 4.5 × 0.75 = 405 Wh | 3 × 55 + idle ≈ 270 Wh | Surplus — the claim can work ☀️ |
| Same rig, overcast (1.5 h sun), 6 h of AI | 4 × 30 × 1.5 × 0.75 = 135 Wh | 6 × 55 + idle ≈ 435 Wh | Deficit — battery drains in ~2 days |
Verdict: "$0/month" is plausible for a few hours of daily inference in sunny climates — but it does not replace a frontier-model subscription on capability. A local 27B model is great for private drafting, summarizing, and coding help; cloud models still win on hard reasoning.
Local vs. cloud: what $200/month actually buys
- Local pros: privacy (nothing leaves the pack), zero marginal cost, works offline, no rate limits.
- Local cons: ~$800–2,500 upfront hardware, model quality ceiling, you are the sysadmin.
- Break-even: a $1,500 rig vs. a $200/mo subscription pays for itself in ~7.5 months — if the smaller model genuinely covers your workload.
- Hybrid is common: local model for 80% of routine tasks, cheap API calls for the hard 20%.