A custom take on Nvidia's platform
According to SemiAnalysis, SpaceX/xAI is working on a rack-scale system codenamed Kessel that builds on Nvidia's Vera Rubin NVL72 platform but changes the interconnect architecture. It keeps 36 Vera CPUs, 72 Rubin GPUs and nine NVLink switch trays, while dropping conventional NVLink cable assemblies.
An 84-layer PCB midplane
The standout feature is an 84-layer printed circuit board midplane that places compute and switch trays back-to-back. It uses ultra-low-loss M10 copper-clad laminates. Nvidia's Kyber design, by contrast, relied on a more complex 156-layer backplane with M9 laminates. Halving the layer count could improve manufacturability, but the more advanced laminate likely raises material costs.
A shift in procurement
Rather than buying Nvidia's assembled compute boards or trays, SpaceX reportedly plans to buy AI accelerators, CPUs, switches and networking components directly from Nvidia and source other parts from third-party suppliers. That gives the company more control over PCB manufacturers and manufacturing partners, but also more responsibility for system design, validation and quality.
Orbital data center ambitions
The design is aimed at both terrestrial and orbital data centers. Whether Kessel is economically viable will depend on whether savings from eliminating cables and independent sourcing offset the cost of the complex midplane. Modern high-end server motherboards typically use 15-20 layers, underscoring how unusual an 84-layer board is.
