As far as I can tell, Data centres in space only seem viable because their advocates insist on comparing them to standard terrestrial data centres.

And nobody ever calls them out on it.

Data centres which are optimised for reliability, redundancy, density, repairability, connectivity and latency. Most of the savings come not from placing the data centre in space, but the fact that advocates have argued away the need for absolutely everything that modern data centres are designed to supply, except for the compute.

If they can really build a space data centre satellite for as cheap as they claim, why launch it? Just drive it out into the middle of the desert and dump it there. It can access the internet via starlink, and already has solar panels for power and radiators for cooling. IMO, If it can cool itself in direct sunlight in space, it can cool itself in the desert.

The main thing that space gains you over setting up the same satellite in the desert is ~23 hours of power, vs the ~12 hours of power on the ground. And you suddenly gain the ability to repair the satellite. The cost of the launch would have to be extremely cheap before the extra 11ish hours of runtime per day outweighed the cost of a launch; Just build twice as many "ground satellites".

And that's with a space optimised design. We can gain even more cost savings by designing proper distributed datacenter elements. You don't need lightweight materials, just use steel. You can get rid of the large radiators and become more reliant on air cooling. You can built each element bigger, because you don't have to fit the rocket dimensions. You could even add a wind turbine, so your daily runtime isn't dependant on daylight hours. Might even be worth getting rid of solar and optimising for wind power instead.

An actual ground optimised design should be able to deliver the same functionality as the space data centre, for much cheaper costs. And it's this ground optimised distributed design that space data centres should be compared to, not today's datacenter which are hyper-optimised for pre-AI use cases.