* They assume 1 satalite = 1 GPU. This is quite funny, actually. A single H200 floating in space with a solar panel and an antenna. In reality, a satellite would pack as many chips as the heat/power allows. A Starlink-sized satellite should be able to hold 40 or so chips. There's no reason why a larger satellitte couldn't hold, say, 1024. * They mention training, but sampling is what makes sense here. Training is a different beast, and requires high reliability, high bandwidth, low latency, and a lot of IO. Space would not be ideal for this. I'd expect training to remain terrestrial and just do sampling in space. (FWIW, sampling will be most of the compute allocation). * Also, no one upgrades GPUs in datacenters, they just add new nodes and leave the old ones there. Google still has their P100 nodes running. Not being able to fix them, though, is a significant concern.