SpaceX reportedly builds custom 72-GPU Nvidia Rubin rack for space data centers
According to SemiAnalysis, SpaceX and xAI are developing a modified version of Nvidia's Vera Rubin NVL72 rack system, codenamed Kessel, that swaps standard NVLink cable assemblies for an 84-layer printed circuit board midplane. The system keeps the core configuration of 36 Vera CPUs and 72 Rubin GPUs across 18 compute trays and nine NVLink switch trays, but uses a different solid-state interconnect design intended for both ground and orbital data centers.
GoKawiil's interpretation of the reporting above, not reported fact.
By sourcing non-Nvidia components independently while still buying Nvidia chips directly, SpaceX could gain more control over its AI hardware supply chain, SemiAnalysis suggests. The simplified 84-layer PCB compared to Nvidia's 156-layer Kyber design may ease manufacturing, though pricier M10 laminates could offset some of those savings, according to the report's analysis.
- SpaceX/xAI reportedly customizing Nvidia's Vera Rubin NVL72 rack system under the codename Kessel
- Design replaces NVLink cable assemblies with an 84-layer PCB midplane for compute and switch trays
- Intended for both terrestrial and orbital data center deployments, per SemiAnalysis
Source: tomshardware.com — Anton Shilov, 2026-11-11
Published there as: “SpaceX develops its own ‘Kessel’ rack-scale system featuring 72 Nvidia Rubin GPUs for data centers in space, report claims”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.