A traditional Earth-observation satellite must capture a full-resolution image and wait for it to reach the ground before anyone can analyze it — bounded by the next ground-station pass, which can be hours away, and burning heavy downlink bandwidth on scenes that are often unremarkable. Running a lightweight neural network directly onboard flips that: the moment a captured frame contains a recognizable hotspot signature, a tiny high-priority alert with coordinates and a small image chip can go out over any available low-bandwidth link immediately, while the bulk image waits for a later high-bandwidth pass. This simulator races both strategies against the same randomly-igniting wildfire so the gap in time-to-alert and bandwidth use is visible directly.