There Are Now Computers Doing Real Work in Orbit, and a Long Way to Go

Starcloud raised $170 million at a $1.1 billion valuation, which is the clearest signal that private capital regards this as an investable sector rather than a research programme.

The problem is the ride up

Everything about this depends on launch cost, and independent analysts are consistent that orbital data centres do not become cost-effective unless launch costs fall substantially from where they are.

That is not a small caveat. It is the entire business case. Every kilogram of server, structure, solar array and radiator has to be lifted, and the hardware cannot be walked out to by a technician when a drive fails. Maintenance is not a thing that happens; replacement is.

The counter-argument is that launch costs have fallen by a large factor over the past fifteen years and the people building the data centres are, in one case, also the people building the rockets. Vertical integration of that kind changes the arithmetic in a way that an external customer buying launch services cannot replicate.

Is this actually better for the environment?

Not obviously, and this is where the marketing gets ahead of the analysis. The solar power is genuinely clean and continuous. But independent researchers have warned that orbital facilities could end up with larger environmental and climate effects than terrestrial ones once the launches are counted, because rocket emissions are deposited directly into the upper atmosphere where their effects differ from ground-level emissions. There is also the question of orbital debris and what happens to a decommissioned compute satellite. Describing this as green compute requires ignoring the launch, which is the largest single input.

What would you actually run in orbit?

Not anything latency-sensitive. The round trip to orbit and back adds delay that makes interactive workloads unattractive, and a user in Dubai querying a model does not want the request going to space and back. The plausible fit is batch work: model training, large-scale scientific computation, and above all processing data that is already collected in orbit. Earth observation satellites currently downlink enormous volumes of raw imagery to be processed on the ground; doing that processing where the data already is, and sending down only the results, is the most immediately sensible use of the idea.

When will orbital data centres be commercially viable?

Nobody credible is giving a firm date, and claims that 2027 is the year should be read as ambition rather than forecast. The prototypes flying now are demonstrations, the Google prototypes are due in early 2027, and viability at scale requires launch economics that do not yet exist. A reasonable reading of the current evidence is that specialised orbital processing of orbital data becomes commercially normal well before general-purpose compute does, and that the second may take a decade or may not arrive at all.

Why it is worth watching anyway

Because the constraint driving it is not going away. If terrestrial grid connection queues keep lengthening and communities keep objecting to water use, the relative appeal of somewhere with free power and no neighbours increases even if the absolute costs stay high.

That is the honest case for orbital compute. Not that it is cheap, but that the alternative is getting more expensive and more contested every year.