Only radiation rejects heat in orbit: a closed liquid loop feeding huge deployable radiators is the credible way to cool a space data centre
Asked:
“What is the most likely way to do cooling for data centre effectively in space?”
Six independent sources — NASA spacecraft thermal guidance, an ESA-linked European concept, the Thales Alenia Space ASCEND study, Starcloud’s developer white paper, an independent IEEE Spectrum engineering analysis, and a JPL-style physics model — converge on one architecture: direct-to-chip liquid cooling (two-phase where its complexity pays) carrying heat to large, segmented, sun-shielded deployable radiators that emit infrared to deep space. Starcloud and ASCEND are proponents’ feasibility concepts, not operational proof.
The thermal chain every source shares
Hover or tap a stage for what it does and which sources back it. Inside the craft heat only moves; only the last stage rejects it.
The physics tax: radiator area grows with power
Radiator area, m² (log scale). Radiated power scales with T⁴, so hotter coolant shrinks area — but chip temperature limits caps the tactic.
One 700 W H100 held at 60 °C needs about 1.4 m² of radiator, rising toward 2.0 m² as UV and atomic oxygen age the coating — a 40 kW rack needs roughly 80 m². spectrum.ieee.org
A physics-based model puts a 1 MW node at roughly 1,500–2,500 m² of radiator depending on temperature and absorbed environmental heat, and calls for segmentation against punctures at MW scale. arxiv.org
NASA guidance is explicit that heaters, cryocoolers and thermoelectric coolers only move heat; rejection always ends at high-emissivity, low-solar-absorptivity radiator surfaces facing deep space. nasa.gov
Starcloud’s 5 GW concept banks on two-phase loops feeding radiators that net-radiate 633 W/m² at 20 °C, with optional heat pumps to raise radiator temperature — a feasibility claim, not flight heritage. starcloud.com
Evidence, source by source
Source
Evidence type
Thermal architecture
Scale or performance
Main constraint
Six source rows (one per host), compiled from public pages and papers on orbital data-centre thermal control; each row gives evidence type, architecture, stated scale or performance, and constraint. Areas are the sources’ own estimates in m² of radiator per stated heat load; long source text is trimmed for space, full context at the links. No geographic data applies.