SpaceX Develops Custom ‘Kessel’ AI Rack With 84-Layer Midplane for Earth and Orbit

image 538
🔊 Listen to the summary

SpaceX and xAI are building a bespoke rack-scale AI computing system codenamed Kessel, replacing standard NVLink cabling with an 84-layer printed circuit board designed for deployment both on the ground and in orbit.

According to analysis from research firm SemiAnalysis, the Kessel system adapts the core blueprint of Nvidia’s Vera Rubin NVL72 platform. The setup retains 72 Rubin GPUs, 36 Vera CPUs, and nine NVLink switch trays arranged across 18 compute units. However, instead of using Nvidia’s standard internal cable cartridges to bridge the hardware, SpaceX positions the compute and switch trays back-to-back against a central, passive midplane.

The core of the design is its 84-layer midplane constructed from ultra-low-loss M10 copper-clad laminates. While this significantly trims the 156-layer design seen in Nvidia’s experimental Kyber interconnect concept, fabricating an 84-layer board remains an extreme engineering challenge. Standard high-end server motherboards generally require only 15 to 20 layers. Fabricating boards with such dense internal structures creates substantial yield risks, as a single internal flaw can render the entire component unusable.

The custom architecture also signals a departure from Nvidia’s turnkey hardware sales. Rather than purchasing fully assembled compute boards or pre-integrated server trays, SpaceX plans to buy bare processors, GPUs, and networking silicon directly from Nvidia. Elon Musk’s company will independently source circuit boards and peripheral hardware from third-party manufacturers, granting it greater authority over component validation and physical optimizations tailored for orbital conditions.

Source: Tom’s Hardware