Orbital Data Centers: Constraints, Resilience, and Break-Even Economics

AI-driven deployments are pushing U.S. data-center electricity demand into a system-planning regime, with projections rising from 176 TWh yr^−1 (2023) to 325–580 TWh yr^−1 by 2028, corresponding to 37–66 GW average load. Motivated by grid interconnection and time-to-power constraints, orbital data centers propose self-powered compute using photovoltaic (PV) generation and radiative heat rejection. We develop a physics-based sizing framework that maps delivered IT power PIT (i.e., compute payload electrical power at the regulated interface) into PV area/mass, eclipse storage, radiator area/mass, duty-cycle-limited space–ground throughput, and lifetime penalties from radiation and correlated risk. This yields a compact economic discriminator: Corbit ≃ (L$/kg + B$/kg )mkW + CGS + Cops, where mkW is deployed mass per delivered IT kW, also L$/kg be marginal launch cost to the operational orbit (including mission-unique delivery requirements), and B$/kg be the fully burdened spacecraft build, integration, and qualification cost per delivered kg. For an illustrative PIT = 1 MW node with optimistic but credible parameters (αOH = 1.25, f⊙ = 0.95, SPPV = 100 W kg^−1, Trad = 350 K, σrad = 5 kg m^−2), we obtain APV ≈ 4.4 × 10^3 m^2, Arad ≈ 1.9 × 10^3 m^2, and mkW ∼ 30–50 kg/kW once fixed spacecraft mass is included. The resulting break-even condition implies that Earth-serving general workloads (C3) are not cost-robust unless (L$/kg + B$/kg ) falls to the low-10^2 –low-10^3 $/kg range and communications and replenishment penalties remain modest. By contrast, the most defensible near-term business cases are space-native preprocessing/autonomy (C1) and communications-bundled services (C2), where location value and reduced downlink burden can justify higher orbital infrastructure cost.