Communication latency and bandwidth would present big issues for AI satellites used for training. But if they are inference-only, then the latency and bandwidth demands are much, much lower. Starlink already demonstrates the required parameters.

Indeed ... yet the servers for that already exist and are largely ground-based - "edge" compute of the large CDNs through which much of the internet's traffic flows anyway.

Will starlink build its own edge ? Maybe. We'll see. Whether it makes sense to have this infrastructure "colocated in space" if you're not also the owner of the "connection infrastructure in space" though? It'll come down to cost. If I were anyone else but SpaceX ... that's where my eyes would be.