
GPU average power consumption if it's doing inference over the course of 24 hours is ~2/3 of its peak power, even if you're super efficient.
SpaceX AI Sat V1 peak power spec has been raised to ~250kW (battery-assisted), with average power of ~160kW. Will be able to handle an NVL72 Ruben rack.
The satellite is designed to handle one full NVL72 Ruben (Rubin/Vera Rubin-era) rack, whose power draw is in the ~190-230+ kW range depending on the exact configuration and power profile (Max Q vs. Max P).
100 kW extra power, at 220 Wh/kg average pack density, ~40% usable DoD for safety/longevity).
10-minute boost (600 s, common short inference spike or thermal buffer):
Energy = 100 kW × (10/60) h ≈ 16.7 kWh
Battery mass ≈ 16.7 kWh / 0.22 kWh/kg ≈ 75-85 kg (depending on exact DoD and overhead).
30-minute boost (more conservative for sustained high-load periods)
Energy ≈ 50 kWh
Battery mass ≈ 225-280 kg.
5-minute boost (very short peaks):
Energy ≈ 8.3 kWh → ~35-45 kg.