SpaceX develops its own ‘Kessel’ rack-scale system featuring 72 Nvidia Rubin GPUs for data centers in space, report claims

SpaceX has developed its own rack-scale system, called Kessel, built around 72 Nvidia Rubin GPUs, according to an analysis by SemiAnalysis reported by Tom's Hardware. The system keeps the basic layout of Nvidia's Vera Rubin NVL72 platform, which spreads 72 GPU packages across 18 compute trays and nine NVLink switch trays. SemiAnalysis says the design is intended for both terrestrial and orbital deployments, giving the US space and data-center sectors a stake in how it performs.

Rather than using Nvidia's usual internal NVLink cable cartridges, SpaceX plans to place compute and switch trays back-to-back around a passive PCB midplane. The arrangement resembles a smaller version of Nvidia's Kyber design, which carries 576 AI GPUs in 144 packages and also uses a PCB-based interconnect instead of conventional cable assemblies. SemiAnalysis notes the shared reliance on a PCB interconnect rather than cabling.

The most distinctive element of Kessel is its 84-layer PCB midplane, which uses ultra-low-loss M10 copper-clad laminates. Nvidia's Kyber architecture instead relied on a backplane built from a far more complex 156-layer PCB using M9 laminates. Cutting the layer count by nearly half could improve manufacturability and production yields, but the more advanced laminate will almost certainly raise material costs. An 84-layer PCB is, in any case, exceptionally difficult to build.

Moving to a PCB midplane could simplify assembly by removing many high-speed cable connections, though it would not necessarily lower manufacturing costs. An 84-layer board using ultra-low-loss M10 laminate is an exceptionally expensive component that demands sophisticated manufacturing and precise quality control. For context, modern high-end server motherboards use 15 to 20 layers, while server development boards use more than 20. Yields could also become a major obstacle, since a defect in an internal layer may ruin an entire board. Whether Kessel is economically viable will depend on whether savings from eliminating cables and buying components independently offset the midplane's cost.

Kessel also departs from Nvidia's conventional system-level procurement model. Instead of buying Nvidia's assembled L6 compute boards or the L10 trays Nvidia plans to sell with its Vera Rubin generation, SpaceX reportedly intends to buy AI accelerators, CPUs, switches and networking components directly from Nvidia and obtain other parts from third-party suppliers. The report raises the question of how easily SpaceX or xAI could procure memory amid current shortages. The approach would let SpaceX choose its own PCB manufacturers, component vendors and manufacturing partners, tune hardware for its infrastructure, which matters for orbital deployments, and possibly cut intermediary costs. SpaceX would take on more responsibility for system design, validation, manufacturing quality and supply-chain management, as most hyperscalers have done in recent years.

For Nvidia, the implications are less straightforward. Selling CPU and AI GPU packages directly could reduce its revenue from complete compute-board assemblies, but the arrangement could also encourage other large customers to build their own Nvidia-based systems and ultimately deploy more Nvidia-based machines, which may matter as custom silicon from hyperscalers gains traction. For that to happen, Kessel must offer clear advantages over systems Nvidia designs.

More hardware news from TechManNews.