The reason why is because this system has essentially no moving parts, no particles flying around and takes up 0 space on earth. If you can ever get it to work (and there's nothing preventing it from being physically possible) you'll have nothing other than the upfront costs, and maybe a gpu upgrade every 10 or so years if you want. Not to mention that energy consumption on earth has hard upper limits you can never surpass without cooking the plannet. I doubt they'll use heat pumps as fluid boil off would be the opposite of low maintenance. . They'll likely just use very conductive material to get it to the radiator. But maybe the math works out that sending up fluid is cheaper than the extra mass for all these conductive elements, hard to predict. >Also he doesnt explicitly mention temp, but it's in his estimate of 3m2/kW. Yes, but then his temp is wrong. At 50c it's closer to 1.8m2, why would you run it at less than that? At 1.8m2, the area is significantly less than that of the solar panels, so it won't increase footprint.

---

I’m not parroting fanboys, I’m an engineer who’s semi interested in this. I’ve worked on deep learning stuff since before GPT existed. Being at least somewhat accurate when doing napkin math checks on feasibility is important, because you’re trying to predict the future, and your result changes what you think is likely to happen. I think this seems feasible, and might actually be the best way forward due to terrestrial power constraints. 3 m^2/kw is the ISS figure, if you run them at 85C, which is fine for hardware, no heat pump, it’s 2m^2/kw, if you go hotter, it’s even less, so mentioning temp scaling does matter. When you’re trying to show something is infeasible, you have to steelman it/give the best case and show it’s still impossible. Re cosmic radiation, bitflips matter a lot with mission critical computers running deterministic code, they don’t matter as much in non-deterministic computations. if one weight out of a trillion changes value a bit, that’s not going to change the end result much. So it’s probably not as big a focus as with eg the space shuttle and its voting system, where if the code errors out, computations halt and everyone could die. If it hits an important bit and crashes, they can always restart the execution. I’m guessing the actual control hardware will be hardened as usual. The thing he seems to be trying to solve for is lack of available power on the ground. People have already seen power bills rising, and a popular backlash stirring already, and this is early innings in computation deployment. He’s talking about putting up terawatts, more than current US electrical generation capacity, so you can see how that might be problematic. There’s a huge backlog on generation construction in the US, and red tape has been holding back massive amounts of solar deployment (call your congressperson to support EPRA/Energy Permitting Reform Act). If you think AI is stupid, I feel for you, you’re going to have to reckon with that viewpoint, because things are going to get a lot more absurd from here.

---