Technical writing from the Maargin research and engineering team — on orbital compute, onboard AI, and the infrastructure of distributed inference.
How we built an autonomous AI workload scheduler running aboard a LEO satellite node using a fine-tuned Qwen 2.5-1.5B model, trained on a synthetic dataset generated with PASEOS — making real-time decisions on task execution, resource allocation, and escalation entirely onboard at 1000 km altitude.
Launch costs have collapsed twenty-fold — but like the shipping container before it, cheap orbit only moves the binding constraint somewhere else. The real prize is orchestration: the control layer that turns a swarm of intermittently lit, thermally constrained, laser-linked satellites into something a customer can simply rent.
Abundant sunlight, no land constraints, no grid bottlenecks — the orbital data centre narrative is seductive. But treat "space DC" as a real design brief rather than a slide-deck meme, and the constraints compound: thermodynamics, power delivery at the accelerator, reliability in a hostile environment, data gravity, and governance.