Engineering perspective from the Sophia Space team — on-orbit inference, satellite data pipeline constraints, downlink bandwidth economics, and the hardware tradeoffs of building compute for the radiation environment of LEO.
Every satellite operator knows their ground pass windows. Few have calculated what those windows actually cost in time-critical decision latency.
Modern imaging satellites produce more data per orbit than a downlink window can carry. The limiting factor isn't sensor resolution — it's the pipeline between sensor and ground.
Radiation tolerance, power budgets, and deterministic processing — the three constraints that dominate every hardware decision when building compute for low Earth orbit.
Downlink bandwidth isn't free. As constellation sizes grow, the cost structure of ground station networks becomes a non-trivial line item.
The radiation-hardened processor market has not kept pace with commercial silicon. Understanding the tradeoffs changes how you design for orbit.
Running machine learning inference on orbit isn't about reproducing a cloud platform in space. It's about a constrained problem: fast classification on a tight power budget.
Not every workload belongs on orbit. But the ones that do — time-critical classification, sensor fusion, alert triggers — are where edge compute earns its place.
Change detection has always been an EO superpower. Running it on orbit, at sensor rate, with no downlink dependency — that's a different class of capability.