Space Data Centers (SDCs): Rethinking Computing Infrastructure Beyond Earth

Abstract

Falling launch costs, together with advances in radiation-tolerant electronics, optical inter-satellite links, and high-efficiency power and thermal systems, have made Space Data Centers (SDCs) an increasingly credible proposition, whether for deep-space autonomy, on-board processing of Earth-observation data, or as orbital capacity for large-scale artificial intelligence workloads. Power and thermal budgets remain the binding constraint, and conventional GPU, TPU, and NPU architectures are limited by the energy cost of data movement rather than by arithmetic. Compute-in-Memory (CiM) addresses that bottleneck directly and is therefore an attractive substrate for space-based inference, but its accuracy is a function of time since programming rather than of architecture alone. This paper examines SDCs from both architectural and computing perspectives, and presents a CiM case study that quantifies that cost: a remote-sensing foundation model is mapped onto simulated non-volatile memory crossbars and evaluated over a twelvemonth mission under conductance drift, Single-Event Upsets (SEU), Total Ionizing Dose (TID), and operating temperature. A naive mapping onto phase-change memory loses 84 percentage points of top-1 accuracy and is unusable within hours, while a three-step mitigation stack recovers to 94.0% of the floating-point reference. Two findings generalize beyond the workload: mitigation effectiveness is axis-dependent, with only periodic recalibration improving all four stress axes, and the cell mapping outranks the choice of device technology at mission timescales.