The solar power argument for orbital data centers is correct. The heat rejection argument is correct. The lack of permitting friction is a genuine advantage. But the most rigorous case for space-based compute infrastructure does not rest on those advantages — it rests on whether launch costs fall far enough, fast enough, to make the economics competitive with terrestrial facilities that keep getting cheaper to build and operate.
CEPA researcher James Lewis mapped the actual obstacles in February: the IT equipment for a small-to-medium data center weighs approximately 2,000 tons. At current launch economics, putting that in orbit costs at least $200 million and requires multiple Starship missions. Distributed satellite architectures reduce per-launch risk but do not yet change the aggregate cost equation enough to be competitive. The cost variable is not solved. It is moving in the right direction. Those are different statements.
Small-to-medium data center IT hardware weight: approximately 2,000 tons. Current estimated launch cost per ton to LEO: $100+ per kilogram at current SpaceX rates. Total launch cost for a single mid-scale facility: $200 million minimum, not including solar arrays, radiative cooling systems, communications hardware, or structural components. Hardware obsolescence cycle in orbit: 2–3 years, same as on Earth — but without the option to swap equipment without a launch.
Terrestrial data centers fail constantly. Drives fail. Network cards fail. Cooling pumps fail. The operations model relies on technicians who can physically access the hardware within hours of a failure event. The Northern Virginia interconnect outage that disrupted AWS for a day in 2025 was resolved by manual intervention on-site. In orbit, that intervention requires either a crewed mission or robotic repair capability that does not yet exist at commercial scale.
Musk's stated solution is to deploy cheap satellites in sufficient quantity that individual failures are statistically irrelevant — replace the satellite rather than repair it. This is how Starlink handles constellation maintenance. It works for communications satellites where each unit handles a small slice of traffic. It is a harder architecture to apply to stateful computing workloads where data residency, consistency, and hardware-software co-design matter.
Placing a data center in LEO adds latency. Signals traveling between Earth and a satellite 550 km overhead and back incur approximately 5 to 10 milliseconds of round-trip overhead. For latency-sensitive workloads — financial trading, gaming, real-time inference, interactive applications — that additional delay is disqualifying. The addressable market for orbital compute is workloads that are either latency-tolerant (batch AI training, archival storage, remote sensing processing) or that can be co-located within the satellite network itself (inter-satellite compute).
This is not a fatal constraint. Latency-tolerant workloads are large and growing. Bezos specifically called out AI training as the workload that makes sense for space. Training jobs are multi-day batch operations with no real-time requirements. The model trains in orbit; the inference serves on Earth. That split architecture is coherent.
Orbital data centers will eventually make commercial sense. The physics, the energy availability, and the cooling simplicity all favor them at scale. The obstacles — launch economics, maintenance, latency for specific workloads, radiation hardening requirements — are engineering and economic problems, not fundamental physics barriers. They are solvable with enough capital and time.
The cooling industry's relevant question is not whether orbital data centers will happen, but when the launch cost curve crosses the terrestrial construction cost curve for the relevant workload classes. China's 2030 target is aggressive. Bezos's 10-to-20-year window is probably more realistic. The terrestrial cooling market has a long runway ahead of it. It also has a ceiling that did not exist five years ago.