CS Colloquium: Nathaniel Bleier

Please join us in ECCR 265 or on Zoom for this talk! Abstract: Tech companies are racing to put AI datacenters into orbit, pitching them as a way to grow compute without straining terrestrial power grids. This talk asks what orbital computing can and cannot deliver. Drawing on architectural models of launch capacity, inter-satellite laser links, radiation, and servicing, I show that orbital clusters cannot keep pace with projected AI compute demand: launch capacity alone falls short, and the latency of re-aiming laser links leaves accelerators idle. Radiation-induced soft errors and the difficulty of servicing hardware in orbit further erode the economics. Computing in orbit does pay off where the data is generated in orbit. I will close with one such case --- memory-centric accelerators for high-order wavefront sensing and control on the coronagraph of NASA's Habitable Worlds Observatory, an exoplanet-imaging mission --- and the open architecture problems that space-native workloads p...