Why We Should Move AI Data Centers Into Space

[badskywalker]

This is taken from the source listed in the article's cooling section. I will say that trying to run these chips at anything above ~70-75 C continously is pretty uneconomical for total lifespan and will start to see chip failure within the 5 year operating window. Second, a 5 year lifespan is far too short to be economical (personal opinion explained below), and far too long for the advancement of such chips (most are outdated in the space of 2-3 years).

The ISS (largest thing in space) has a surface area of ~2,500 square meters. That would get you about 1.75 MW of dissipation. Now the cost of putting something into space has decreased quite a lot thanks to SpaceX (~1/10 of the historical). If we were to apply this we would get ~17.5 MW, and if we are being super gernous and say a further 90% of the tonnage of the ISS is purely for human habitation and expirementation, you could get 175 MW. This would cost ~150 Billion dollars applying the ISS budget.

Per congress there's about 176TWh used annually by data centers in the US at current levels. 175 MW would equate to about 1.53 TWh (Straight past GWh, whihc I was surprised by). The discrepancy is more than a hundred times, meaning you'd need an utterly ludicrous investment. Even if you were to entertain that you could cut off 99% of the tonnage of the ISS for just the cooling array, it is simply an insane amount of money to get current levels.

Note: I just realized that I accidently did not actually adjust square meters to 70C and used the assumption of 1 square meter per 700 Watts which is 85 C fresh launch, not even the end of life condition! I just don't feel like redoing the math lol. This would make the number ~1.5 times higher.

As someone who got a B.S. in Physics and works in Electrical Engineering, I see way too many challenges in the project for it to be completed in any reasonable amount of time (before 2035) with any reasonable budget (under $100 billion for 1 TWh).