Edge inference means running an AI model where data is produced — on a satellite, camera, or phone — rather than shipping raw data to a distant server.

The math explains why this matters in orbit: a single radar image can be hundreds of megabytes, and downlink bandwidth is scarce. Beaming everything down would saturate the link instantly. Run AI onboard, and only the result is transmitted — a flag, a bounding box, an anomaly alert — compressing the payload by 99% or more, while also eliminating the round-trip latency to a ground station.

The catch is that space hardware must survive radiation, extreme temperatures, and strict power budgets that standard AI chips aren't designed for. Satlyt raised funding this week specifically to solve that problem, and chip designers are now building radiation-tolerant AI accelerators purpose-built for orbit. Space is the sharpest version of a question the whole AI industry is asking: how do you move the model closer to the data?