Orbital data centers and edge compute

Running substantial computation in orbit is a recent proposition with a specific rationale and a specific hard limit. The rationale is that downlink is scarce. The limit is that heat has nowhere to go.

Why compute in orbit at all

The argument is not that space is a good place for computers. It is that some data is generated in space and moving it is expensive.

From Earth observation, a high-resolution optical satellite generates roughly 10 Gbit/s while imaging and can downlink around 30 GB per pass, four to six times a day. The gap is more than an order of magnitude, and the standard responses are to duty-cycle the sensor and compress aggressively, both of which discard information.

Processing onboard inverts this. If the useful output is a ship position, a change flag or a cloud mask, then downlinking kilobytes instead of gigabytes removes the bottleneck entirely.

The latency column matters more than the volume for several applications. Maritime surveillance, disaster response and defence tasking all lose value rapidly with age. A detection delivered in minutes over a low-rate link can be worth more than an image delivered in hours.

There are two further arguments, both weaker.

Data sovereignty: computation performed in orbit is arguably outside any national jurisdiction, which is a legally untested claim rather than an established one. See Policy and regulation.

Solar power and radiative cooling as free resources: this is the argument most often made and it is the one that does not survive examination, for the reason below.

The thermal limit

The binding constraint on orbital computation is not power generation. It is heat rejection.

From Thermal control, radiation is the only path for heat to leave a spacecraft, and the area required is

Roughly 2.6 m² of radiator per kilowatt at room temperature. A terrestrial data centre rack draws 10 to 40 kW, so one rack needs 26 to 100 m² of radiator, and that is before accounting for the fact that a radiator facing the Sun or a warm Earth rejects far less than one facing deep space.

For comparison, the solar array to generate that kilowatt is about 2.9 m² from Power systems. Generation and rejection require comparable area, so the spacecraft is roughly twice as large as the naive power calculation suggests, and the radiator is the part that cannot be made more efficient with better technology, because it is set by the Stefan-Boltzmann law.

The often-repeated claim that space offers free cooling is therefore wrong in the way that matters. Space is cold in the sense of having a low background temperature and is an excellent insulator, which is the opposite of what a cooling problem needs. There is no convection and no conduction to a medium; there is only radiation, and radiation is slow.

Radiation

Commercial processors are not radiation-hardened, and a data centre built from rad-hard parts would be three orders of magnitude behind terrestrial performance, which defeats the purpose. See Avionics and onboard computing.

The workable approach is commercial silicon with architectural fault tolerance:

The last row is the architectural insight, and it is borrowed directly from terrestrial cloud practice. Hyperscale data centres already assume nodes fail constantly and build reliability at the fleet level rather than the node level. That assumption transfers well to orbit, where the failure rate is higher but the response is the same.

Below the belts, at 500 to 600 km, annual dose behind modest shielding is a fraction of a krad, which screened commercial parts tolerate for the few-year lifetimes that constellation satellites already accept.

What scales and what does not

The distinction between edge processing and a data centre is one of degree that becomes one of kind.

Edge processing at tens to hundreds of watts is established and sensible. An inference accelerator running a detection model on a satellite already flies, and its thermal load is a fraction of the bus power. See Edge computing in space.

Kilowatt-scale processing is a design problem but a tractable one. The radiator is large but comparable to the solar array already needed.

Megawatt-scale orbital data centres run into the thermal wall. A megawatt needs on the order of 2600 m² of radiator, which is a structure larger than the ISS’s entire array area devoted purely to cooling, requiring in-orbit assembly and continuous attitude management to keep the radiators edge-on to the Sun and away from Earth’s infrared.

That is not a reason to dismiss the concept, but it is the number that any serious proposal has to address first, and it is frequently absent from the public versions.

The honest case

The defensible proposition is processing data that is already in orbit, where the alternative is an expensive downlink. The indefensible one is moving terrestrial workloads to orbit, where the alternative is a building with a chiller.

References

• Denby, B. and Lucia, B. “Orbital Edge Computing: Nanosatellite Constellations as a New Class of Computer System”, ASPLOS 2020.
• Furano, G. et al. “Towards the Use of Artificial Intelligence on the Edge in Space Systems: Challenges and Opportunities”, IEEE Aerospace and Electronic Systems Magazine, 35:44, 2020.
• Giuffrida, G. et al. “The Phi-Sat-1 Mission: The First On-Board Deep Neural Network Demonstrator for Satellite Earth Observation”, IEEE Transactions on Geoscience and Remote Sensing, 60:1, 2022.
• Gilmore, D. G. (ed.) Spacecraft Thermal Control Handbook, Aerospace Press, 2002, for the radiator sizing relations.