100 TW as a comp target is absurd enough to be useful: it forces the question of whether Starship can collapse launch cost faster than Earth can permit power, substations and cooling. The hard bottleneck is bandwidth and radiator mass, so early orbital AI looks less like AWS in space and more like Celestia/EigenDA-style compression for workloads where data is born in orbit or can be reduced before downlink. Crypto should recognize the pattern: cheaper blockspace created new markets, but the rents moved to sequencers, DA, routing and whoever controls the scarce coordination layer.

Top comment by @Benthic