The heat  is on

Space has always been hostile to electronics, but the industry is now confronting a thermal challenge that threatens to define what is possible in orbit. The vacuum of space eliminates two of the three mechanisms by which heat leaves a system – conduction and convection – leaving only radiation, the least efficient of the three. For decades, engineers have managed this constraint. Today, it is becoming a design driver, writes Peter Donaldson.

Two emerging applications are throwing the problem into sharp relief: nuclear propulsion for outer-system missions and Earth-orbiting AI data centres. In nuclear thermal propulsion, liquid hydrogen passes through the reactor core, absorbing heat and expanding through a nozzle. The propellant itself acts as a coolant – an elegant co-location of problem and solution. But residual radiation continues to generate heat after shutdown, and that heat must be radiated to prevent core meltdown. Nuclear electric propulsion faces a comparable challenge in normal operation as does nuclear thermal propulsion post-shutdown, so all spacecraft with nuclear reactors will need large radiator arrays.

For orbital data centres, the numbers are sobering. A 1 MW facility would require radiators with an area of approximately 1200 to 1600 m2,

weighing around 100 tonnes, while its compute hardware would weigh only 10 tonnes. Even mounting radiators on the backs of solar panels – which naturally face away from the Sun – provides only partial relief. The radiators still require two to four times the area and five to 20 times the weight of the solar arrays, depending on technology choices. Moreover, the panels themselves generate waste heat, Earth reflects solar radiation back at the radiators, and the thermal gradient between Sun-facing panels and deep-space-facing radiators is extreme.

These constraints are forcing a fundamental rethinking of compute hardware. The dominant AI platform – GPUs with high-bandwidth memory – concentrates power on a single die, creating hotspots that necessitate throttling under limited radiator capacity. Compute-in-memory (CIM) architectures offer a compelling alternative. By distributing processing capabilities among memory arrays rather than concentrating them in dedicated processor chips, CIM achieves a more uniform thermal profile. Simulations show that CIM accelerators deliver over 10 to 40 times higher effective TOPS than GPU-HBM systems under tight radiator budgets, while maintaining thermally safe operation.

Thermally aware software adds another layer of optimisation, mapping high-power layers away from each other and dynamically assigning complex computations to cooler regions of the chip.

CIM is not without competition. Neuromorphic computing promises graceful degradation under radiation but faces material-level thermal stability challenges. Photonic computing offers dissipationless computation but requires active thermal tuning that adds power and complexity. Established radiation-tolerant GPUs remain the pragmatic choice for shorter LEO missions.

What is clear is that thermal management is no longer a secondary concern. It is the constraint that will determine which architectures succeed in orbit. For engineers designing the next generation of space hardware, the question is not simply how much compute can be delivered, but how much heat can be rejected. CIM’s thermal advantage makes it a strong contender – but the broader lesson is that in space, thermodynamics always has the final word.