Orbital AI Computing: Carbon Tradeoffs Across Satellite Scale
| Source: arXiv AI
Tags: orbital computing, green AI, carbon footprint, LEO, satellite AI, sustainability, DGX H100
Accelerator-aware carbon analysis of Low Earth Orbit AI computing finds that launch emissions are a fixed overhead — high-performance DGX H100 systems amortize this more effectively per computation unit than lightweight Jetson AGX Orin hardware, challenging the intuition that smaller always means greener.
Details
As satellite computing capabilities grow with cheaper launch platforms and constellations, running AI inference in Low Earth Orbit is becoming feasible. The sustainability calculus is non-obvious. This paper extends the ESpaS carbon lifecycle framework with accelerator-aware modeling to compare two representative systems: Jetson AGX Orin (small satellites) and DGX H100 (large-payload platforms).\n\nThe key finding: launch emissions are a fixed carbon overhead that scales with mass, not compute performance. A Jetson-class system emits fewer absolute carbon emissions but delivers less compute per kg launched. A DGX H100, while heavier, amortizes launch carbon more effectively per unit of computation — resulting in lower carbon intensity per inference.\n\nThe space-ground tradeoff — whether to run inference in orbit or beam data down for ground processing — is therefore highly sensitive to hardware choice. Prior frameworks using generic datacenter configurations missed this dependency; the accelerator-aware extension reveals it.\n\nFor AI infrastructure teams and sustainability researchers: this challenges simple green-computing intuitions and adds rigor to the orbital AI computing space as launch costs continue falling.