SpaceX developing custom Nvidia Vera Rubin rack dubbed Kessel for its orbital data centers – report - Data Center Dynamics

SpaceX's "Kessel" rack—built on Nvidia's Vera Rubin platform for orbital data centers—marks a structural decoupling of compute infrastructure from terrestrial physics. Technical cascade: vacuum cooling (no convection), radiation hardening, and microgravity thermal management force HBM stacks and CoWoS advanced packaging (TSMC, Taiwan, China) to add radiation-redundancy layers, inflating per-chip BOM by 30-50%. The deeper signal: Nvidia is migrating from "selling GPUs" to "selling orbital compute systems"—the same system-level logic as DGX H100, but the end customer is now an orbital operator. Compliance exposure: Vera Rubin's process node is tightly coupled to Taiwan, China's advanced packaging capacity. Spectrum licensing, space-debris liability, and EAR export-control compliance for cross-border orbital services will pull Nvidia's data-center business into aerospace regulatory frameworks for the first time, structurally raising compliance costs. Competitive dynamics: AWS's Project Kuiper and Google's Starboard target satellite edge compute, but neither has placed GPU-class silicon in orbit. AMD's window to ship a space-grade MI400 is roughly 18 months. 12-24-month outlook: First orbital inference nodes enter LEO validation by Q2 2026. The true long-tail effect: once compute escapes the ceiling of terrestrial power and cooling, the boundary between AI training and inference will be redrawn.