Orbital AI Datacenter: An Architectural Analysis

Compute demand for frontier AI training is growing at ∼5×∼5×/year, straining not only compute hardware but also terrestrial power grids and cooling infrastructure. Motivated by abundant solar energy and effectively free radiative cooling, several recent proposals have suggested moving AI training into orbit, ranging from disaggregated satellite platforms to distributed constellations and monolithic orbital mega-structures. However, the feasibility and economics of orbital AI training remain largely unexplored. This paper presents an architectural and techno-economic analysis that unfolds the benefits and limitations of orbital AI datacenters. We characterize the key constraints that distinguish in-space AI training from its terrestrial counterpart, including thermal management, radiation, and deployment logistics, and show that orbital datacenters cannot scale to meaningfully offset terrestrial demand. We develop a bottom-up, compute-centric cost model to quantify the overhead of sustaining compute in orbit, and find that orbital AI datacenters are far from achieving cost parity even under optimistic and forward-looking assumptions. Our performance analysis further shows that communication bottlenecks and radiation-induced errors lead to severely underutilized compute, compounding the cost disadvantage. We further show that scaling up individual satellites to improve co-location and amortize structural overhead is necessary to mitigate both cost and performance penalties, but that debris avoidance, construction complexity, and single-point-of-failure risk render this approach untenable beyond megawatt scales. Even at the cost-optimal design point, the gap to terrestrial infrastructure remains large. We conclude that space computing is better suited not to frontier training but to edge inference on space-generated data, where the interconnect, synchronization, and reliability demands that defeat training are absent.