Finally, a good and accurate article about space-based data centers. I’ve said from the beginning that heat is a major problem, a mostly unsolvable problem with current tech, for space-based AI systems. So much of the media just repeats the lies and exaggerations Musk puts out on X, but the NPR article about Google’s effort gets it right. TPUs generate a lot of heat, and you need a lot of tightly-linked TPUs for an LLM, and there’s no easy way to get rid of heat in airless space. Plus the sun-facing side of a satellite gets massive heat added. The frontier AI companies see space-based AI as an easier win than building massive, expensive, and unpopular terrestrial data canters, but orbital launch mass and heat are problems just as large.

Give it a decade and the TPUs will be higher performing (Moore’s Law, anyone?) and terrestrial data centers about 100 times smaller and less objectionable. Problem solved. Or…just build them in the high desert with micro nukes and the NIMBY problems will go away. Building data centers on fertile farmland is/was/will always be a dumb idea.

Speaking of heat, we had a power outage in our Socal neighborhood during the heat wave. 101 degree temps for days on end, and the power was out for about 12 hours. Thankfully, we have a fully automated generator system that kicked in and kept things normal-ish even with us absent. Our refrigerated food thanks us. SDGE is being weird and secretive about the source of the outage.