SpaceX and xAI are building their own version of Nvidia's Vera Rubin NVL72 AI rack, with a huge circuit board where the internal cables used to be, according to SemiAnalysis. The system is codenamed Kessel. The account comes from that single analyst firm, and no SpaceX confirmation accompanies it.
Kessel keeps the core layout of the Nvidia design: 36 Vera CPUs, 72 Rubin GPUs, and nine NVLink switch trays. NVLink is Nvidia's high-speed link that lets GPUs share data. In Nvidia's version, cable cartridges carry that traffic. In Kessel, SemiAnalysis wrote, compute and switch trays sit back-to-back around a passive printed circuit board (PCB) midplane. SemiAnalysis compares it to a smaller version of Nvidia's Kyber design, which also uses a PCB and packs 576 GPUs into 144 packages. The design is aimed at both ground-based and orbital data centers.
The board is the striking part. It has 84 layers and uses M10 laminate, a very low-loss grade of copper-clad material. Kyber's backplane used 156 layers of M9. Fewer layers could improve manufacturability and yields, but the newer laminate will almost certainly cost more. For scale, high-end server motherboards typically use 15 to 20 layers, and a defect in an inner layer can scrap the whole board.
The bigger shift is procurement. SpaceX would reportedly buy GPUs, CPUs, switches, and networking parts from Nvidia, then source everything else itself, rather than buying Nvidia's assembled compute boards or trays. That means choosing its own PCB makers and taking on design, validation, and quality control, a job most hyperscalers already do. Memory shortages could make that harder.
Nvidia could lose board revenue on each Kessel-style deal. It could also gain if other large buyers copy the approach and deploy more Nvidia silicon. That only happens if Kessel clearly beats the racks Nvidia designs itself, which makes the cost of that 84-layer board the number to watch.