Built from 22 evidence rows collected from 22 independent sources — one row per host — covering real projects (Axiom, Starcloud, Lonestar), advertised 2026 rideshare launch prices, radiator-sizing engineering rules and cost benchmarks. Where no public all-in price exists, cost and risk are shown qualitatively and marked as such.
Qualitative ranking (1 = lowest, 4 = highest) from the collected evidence; public all-in mission prices do not exist for any path, so bars show relative position, not dollars. Hover a bar for what each path includes.
Host the payload on an existing LEO spacecraft, station or optical-relay platform and reuse its power, thermal control, attitude control and communications, as Axiom's ODC nodes do on the ISS and on Kepler's optical relay constellation. A dedicated CubeSat or free-flyer is the fallback, not the first choice.
A rugged commercial off-the-shelf edge server, FPGA or GPU with SSD storage — the Jetson-class 10–75 W and 100–500 W multi-module payloads already flown or specified — rather than megawatt ambitions.
Checkpointing, ECC, watchdogs and redundant storage instead of radiation-hardening every component; schedule intensive work in sunlight if the host allows it.
Process space-generated data — satellite imagery, Earth observation, secure delayed storage, experiments — locally, so bandwidth stays modest. Not general-purpose cloud for users on Earth.
Radiator area needed per payload power, m² (bars show published low–high engineering range; spacecraft radiators reject 100–350 W/m²). Vacuum gives no convection: cooling converts an electricity bill into a mass-and-area budget at 5–10 kg/m², which is why radiator area — not server purchase price — decides what is buildable.
Orbital Data Center nodes hosted on the ISS and on Kepler Communications' optical relay constellation: AWS Snowcone, AxDCU-1 with Red Hat Device Edge, commercial off-the-shelf hardware, Skyloom optical comms — the hosted-payload pattern this report recommends.
Free-flyer satellite that carried an Nvidia H100 into orbit in November 2025 and trained the first AI model in space — proof COTS GPUs run in orbit, and the fallback path when no host is available.
Hosted payloads on Intuitive Machines IM-1 and IM-2 lunar landers with a Skycorp multi-core RISC-V space server, after an edge data-centre test on the ISS — the long-lived archive niche.
| Evidence / option | Key number or note | Source |
|---|
Method: 22 evidence rows collected in early 2026 from 22 independent web sources (one row per host), each describing an orbital data-centre option, project, launch price, thermal rule or cost benchmark. Cost and risk in the lead chart are qualitative ranks because no public all-in mission prices exist; radiator areas are published engineering ranges in m². Long architecture descriptions were shortened for space; every row links its source.