SpaceX Explores Custom Server Rack Design to Bypass NVIDIA Standard Boards
SpaceX is developing a custom AI server rack named Kessel, which purchases raw NVIDIA GPU dies directly while sourcing supporting hardware independently to bypass standard L6 computation boards. This strategy highlights a major trend among hyperscalers to optimize AI hardware supply chains and thermal/spatial architectures, allowing greater autonomy in cost and deployment for both ground and orbital data centers. The Kessel rack features 72 GPU dies and uses a passive central circuit board midplane instead of internal copper cables, requiring an 84-layer design built with high-grade M10 materials.
## BACKGROUND
A GPU die is the raw silicon chip cut from a semiconductor wafer where actual computing takes place, which companies can procure directly to build proprietary hardware rather than buying pre-packaged boards. Midplane architectures use a central circuit board to connect cards from both sides, offering optimized signal routing and system flexibility compared to traditional cable setups.