SpaceX Develops Kessel: 72-GPU NVIDIA Rubin Rack with 84-Layer PCB for Orbital AI Data Centers
SpaceX develops Kessel, a custom 72-GPU NVIDIA Vera Rubin NVL72 rack replacing internal copper cabling with an 84-layer PCB midplane for terrestrial and space-based AI data centers.
SpaceX and xAI are developing a bespoke rack-scale AI supercomputing architecture codenamed Kessel, engineered around NVIDIA next-generation Vera Rubin NVL72 platform. Designed for dual deployment across terrestrial data centers and orbital AI satellite constellations, Kessel replaces traditional copper NVLink cable harnesses with an unprecedented 84-layer printed circuit board midplane, enabling extreme compute density and direct component sourcing outside standard OEM supply channels.
The Kessel Architecture: Replacing Copper Cables with an 84-Layer PCB Midplane
According to technical analyses from SemiAnalysis and corroborated by industry reporting, Kessel preserves the core computational blueprint of NVIDIA flagship NVL72—incorporating 36 Vera CPUs, 72 Rubin GPUs, and nine NVLink switch trays. However, SpaceX has eliminated the complex, bulky copper cable cartridges used in standard racks. Instead, compute and switch trays mount back-to-back into an 84-layer high-frequency printed circuit board midplane manufactured from M10 low-loss laminate, creating a solid-state interconnect fabric capable of withstanding the extreme mechanical vibrations of rocket launches.
Supply Chain Independence: Sourcing Direct Silicon and Bypassing NVIDIA L6 Assemblies
By developing its own chassis, power delivery, and interconnect backplane, SpaceX is shifting its procurement model: the company plans to purchase packaged Vera and Rubin silicon directly from NVIDIA while designing and assembling its own motherboards and compute trays. Bypassing NVIDIA standard L6 integrated server assemblies gives Elon Musk enterprise direct control over component vendors, manufacturing yields, and cooling designs, significantly compressing infrastructure costs across multibillion-dollar GPU cluster buildouts.
Orbital Deployment: Starmind AI Satellites and the Race for Terrestrial Power Alternatives
Beyond terrestrial data centers supporting xAI Grok models, Kessel is engineered as the primary compute payload for SpaceX upcoming Starmind AI1 satellite fleet, scheduled to launch on Starship in late 2027. Operating in low Earth orbit with 175 kW average and up to 400 kW peak power dissipation managed by closed-loop liquid cooling and massive 160-square-meter deployable space radiators, Kessel sidesteps terrestrial electrical grid constraints and local permitting moratoria, paving the way for gigawatt-scale orbital AI superclusters.