An off-grid AI data center has a problem ordinary data centers do not: tonight’s energy is what is in the battery now, plus the sun that will arrive tomorrow. If the GPUs draw too much in the evening, the system shuts down at four in the morning. If they draw too little, midday sun is wasted.
Barrel OS solves this with a single rule, recalculated every ten seconds: what is the maximum power I can give compute right now so that battery and expected sun last until PV covers the minimum service again, with a safety margin?
From the answer come four modes:
- Solar surplus: energy is ahead; deferrable jobs start (model adaptation, archive indexing, bulk transcriptions).
- Full service: GPUs run at full power for users.
- Saving: GPU power is limited just enough to reach the morning.
- Reserve: only critical service stays on; below the minimum threshold compute shuts down in order, without losing data.
Barrel OS does not need the internet: it estimates today’s sun by comparing measured production with the expected curve, and for tomorrow it prudently uses the worst-month yield. It will be published as open source.