In all our simulations the weak point has been the night. The sun is ahead, but the battery is expensive and runs out. Storing a kWh in a battery is costly; storing an answer already computed costs almost nothing: a few kilobytes on a disk.
AI compute is the only load that can be moved in time in both directions. Forward, by deferring work to tomorrow’s sun: Barrel OS already does that. And backward, by preparing today with sun the work that will be needed tonight or tomorrow morning.
Institutions are predictable. A hospital does ward rounds at 7, discharges late morning, changes shift at 19:30. A town hall protocols on Monday. If Barrel OS learns what will be needed and when, it can prepare with sun the chart summaries for rounds, draft discharge letters, translated forms. Non-urgent requests get one more question: when do you need it?
The test
One Barrel node (one Planet, 250 kWp, 261 kWh) serves the compute of a healthcare facility: 480 kWh per day. The sun is the real hourly PVGIS 2023 series for Lampedusa. Requests are our assumptions: 45% interactive, which do not move; 20% deferrable within 24 hours, such as coding and statistics; 35% predictable, such as summaries for 7, forms for 9, letters for 11 and handovers at 19:30.
Anticipation is imperfect on purpose: it predicts 80% of products, a quarter of prepared work is unused, and every product is refreshed at the hour of use with 20% extra compute. Anticipation only runs when the sun is ahead.
| Management | Energy from the grid or genset in one year | Night compute | Compute served directly from sun |
|---|---|---|---|
| Without orchestrator | 48.7 MWh | 92 MWh | 44% |
| Barrel OS, deferral only | 13.2 MWh | 57 MWh | 64% |
| Barrel OS, deferral and anticipation | 10.8 MWh | 41 MWh | 76% |
What we learned
- The big jump comes from deferral: from 48.7 to 13.2 MWh.
- Anticipation adds about 18%. With a single cabinet it is worth about 40% of a second battery cabinet: part of storage becomes software.
- With two cabinets the energy advantage disappears: anticipation matters where the battery is small and compute is heavy.
- It costs energy: total compute rises 13%, between unused products and refreshes. But it is sun that would otherwise be lost.
- Predicting better does not cut diesel: with 80% or 95% correct forecasts the result is the same, because what remains concentrates in a few cloudy winter days. Night compute and the sun share improve instead.
- The strongest value is not energy, it is service: at 7 a.m. the doctor finds the summary already ready, with no wait, computed with sun.