What about, never? The fuel cost of launching these are immense. I'm not seeing how this is 'cheaper' this doesn't even take into consideration cooling.

BUT it doesn't need to be cheaper for there to be value. Its much harder to shoot down a space GPU farm than a terrestrial one. And its closer to space assets which might want it for targeting during times of war.

> Its much harder to shoot down a space GPU farm than a terrestrial one.

Not by as much as you may hope. Getting a rocket up to altitude is much, much easier than getting one to orbital speed:

https://what-if.xkcd.com/58/

Hitting satellites in orbit is still hard, but ASAT was demonstrated (albeit the target was designed for the test) by the USSR no later than 1968, and on a more arbitrary (end of life) satellite by the USA no later than 1985: https://en.wikipedia.org/wiki/Anti-satellite_weapon#List_of_...

The severity of one single missile depends on the orbital configuration; I'm expecting everyone to declare "we have adjusted our strategic priorities elsewhere" (i.e. "this was a bad idea") well before there's a relevant quantity of them.

Given how cheap laser welding has become, I also foresee ground-to-orbit lasers being credible. They still need quite a large aperture and corrective optics, but that's had known solutions for a while now thanks to astronomy.

> And its closer to space assets which might want it for targeting during times of war.

Not so far as I can see. If you want really low-latency you have edge compute on the weapon systems themselves; otherwise you use the existing satellite comms (e.g. Starlink) to just talk to something hardened outside weapons range.