A conference talk by Richard Campbell (of .NET Rocks) exploring the feasibility of building AI data centers in space. The talk walks through the engineering math: using an Nvidia NVL72 rack as the compute unit (135 kW, 1.3 metric tons, ~720 petaflops), calculating the solar panel area (~450 sq meters), cooling requirements (ammonia radiators comparable to the ISS), and total satellite mass (~12–15 metric tons). It covers orbital mechanics (LEO vs. geostationary vs. sun-synchronous), communication constraints (100 Gbps laser links vs. InfiniBand speeds), and why only inference workloads are feasible in space. The talk also addresses growing orbital congestion, Kessler syndrome risk, debris management, and the limits of LEO capacity (~100,000 satellites per MITRE research). Campbell concludes that space-based AI data centers are technically possible but practically marginal — likely contributing less than 1% of total compute — and suggests the hype is largely driven by SpaceX's upcoming IPO and investor narrative rather than genuine infrastructure need.