Microgrids & Private AI in San Diego (2026): The Operator Math
Yes, a microgrid or even a normal San Diego rooftop solar-plus-battery system can power private AI, and it's a better fit than almost any other load: an inference machine draws roughly 300-800W (less than a hot tub), it can run when the sun produces surplus, pause when the battery has better things to do, and keep working through outages if your system islands. With SD retail rates among the highest in the country, self-generated electrons running local AI is one of the county's quietest structural edges.
Panels are a commodity. Batteries are getting there. The new question is what your surplus electrons should DO, and the best answer we've found is: think.
Why AI Inference Is the Perfect Dispatchable Load
| Property | Why it matters to a solar/microgrid owner |
|---|---|
| Interruptible | AI work pauses and resumes cleanly. Battery needed elsewhere? Compute waits. No other 'productive' load is this polite |
| Schedulable | Batch jobs (research, page generation, report assembly) run when the sun is up or rates are lowest, not when a human happens to click |
| Surplus-hungry | Midday overproduction that would export at weak credit or curtail becomes finished work product instead |
| Small but valuable | 300-800W under load, a fraction of an EV charger, yet the output is the most valuable thing a kilowatt-hour can currently buy |
| Outage-resilient | On an islanding system, local AI keeps working when the grid and the cloud connection don't |
The San Diego Math (Honest Ranges, 2026)
Ranges, not quotes; check your own rate plan. The pattern is what matters: SD grid power is expensive, SD sun is excellent, and the spread is the edge.
| Item | Typical figure | What it means |
|---|---|---|
| Local AI machine draw | ~300-800W under load, near-idle when quiet | A GPU workstation or mini server, not a data center |
| Daily energy at 8h/day working load | ~2.5-6 kWh | Roughly one dishwasher-to-EV-commute of energy |
| SD grid cost for that | ~$1.20-$3.00/day at 40-50¢/kWh plans | ~$450-$1,100/yr on grid power |
| On self-generated solar | Marginal cost approaching zero | The machine eats surplus you'd otherwise export at weak credit |
| Typical rooftop system | 5-10 kW solar · 10-15 kWh battery | Runs an inference box as a scheduled load without noticing |
The Operator Verdict
Start with one box on the system you already own. One inference machine, scheduled to the sun, running the workloads that don't need to be instant: research, drafting, page generation, report assembly, overnight batch jobs. That configuration works today, costs almost nothing to run, and keeps working through outages. The bigger version, microgrids and virtual power plants treating aggregated AI compute as a coordinated dispatchable load, is where this is heading, and San Diego's combination of rates, sun, and operators makes it the natural proving ground. The playbook lives at the flagship page below; the microgrid-scale conversation is an open door.
