SpaceX is designing Kessel, a 72-GPU Nvidia Rubin rack for data centers in space, report says - MLQ.ai
SpaceX is developing its own version of Nvidia's Vera Rubin NVL72 rack, codenamed Kessel, designed for both terrestrial and orbital data centers, according to research firm SemiAnalysis.
The Kessel design maintains the Vera Rubin NVL72 layout of 72 GPU packages, 18 compute trays and nine NVLink switch trays, but mounts them on a passive printed-circuit-board midplane instead of using internal NVLink cable cartridges. SemiAnalysis described it as a "mini Kyber", referring to Nvidia's larger midplane-based rack design.
Advanced Manufacturing
The midplane features 84 layers of M10 laminate, compared to the 156 layers of M9 planned for Nvidia's Kyber rack. SemiAnalysis noted that fewer layers provide a more feasible manufacturing path, though high-end server motherboards typically use only 15 to 20 layers, and defects in inner layers can scrap an entire board.
Supply Chain Strategy
SpaceX plans to purchase chip packages directly from Nvidia while sourcing the remaining components itself, bypassing Nvidia's L6 compute boards. This approach gives SpaceX more control over suppliers but shifts design validation, manufacturing quality and supply-chain risk onto the company.
Kessel keeps the 72 Rubin GPU packages, 18 compute trays and nine switch trays of Nvidia's standard Oberon rack but arranges the trays back-to-back around a passive printed circuit board midplane in place of internal NVLink cable cartridges.
Orbital Applications
The rack aligns with SpaceX's orbital plans. This system is rated at 175 kW average and 250 kW peak power and described as "significantly simpler, lower cost, denser and lighter than a traditional rack."
SpaceX recently acquired a solar cell line from Singulus for its Starmind satellites, and financial institutions are selling portions of its $40 billion Nvidia chip lease agreement. SemiAnalysis did not confirm whether Kessel represents the same system Musk described.
Related
