SpaceX Develops Kessel AI System for Orbital Data Centers
SpaceX customizes Nvidia Rubin hardware using an 84-layer PCB midplane for space deployment.
1Custom AI Hardware Development
SpaceX and xAI are developing a custom AI platform codenamed Kessel, based on Nvidia’s Vera Rubin NVL72 architecture. This system is specifically designed for both terrestrial and orbital data center environments. By modifying the standard design, the companies aim to optimize hardware performance for unique operational requirements in space and on Earth.
2Innovative PCB Midplane Design
The Kessel system replaces traditional NVLink cable assemblies with a massive 84-layer printed circuit board midplane. This architecture utilizes ultra-low-loss M10 copper-clad laminates to manage high-speed data connections. This design shift simplifies system assembly by eliminating numerous cable connections, potentially improving structural integrity for hardware intended for deployment in orbital conditions.
3Complexity and Manufacturing Challenges
Manufacturing an 84-layer PCB is exceptionally difficult compared to standard high-end server motherboards, which typically utilize only 15 to 20 layers. Any defect in the internal layers of this complex board could render the entire unit unusable. Consequently, production yields represent a significant technical and economic obstacle for the project.
4Strategic Supply Chain Shift
SpaceX plans to acquire AI accelerators, CPUs, and switches directly from Nvidia while sourcing other components independently. This strategy allows the company to select its own PCB manufacturers and vendors. By moving away from Nvidia’s pre-assembled L10 trays, SpaceX gains greater control over its hardware supply chain and system validation processes.
5Economic Viability and Trade-offs
The financial success of the Kessel system depends on whether the cost savings from eliminating cable assemblies can offset the high price of the complex 84-layer PCB. While the advanced laminate material increases costs, independent procurement might reduce intermediary expenses. SpaceX must balance these factors to ensure the system remains economically competitive.
6Comparison to Kyber Architecture
The Kessel design shares similarities with Nvidia’s Kyber architecture, which also utilizes a PCB-based interconnect instead of cables. However, Kyber relies on a significantly more complex 156-layer PCB using M9 laminates. By reducing the layer count to 84, SpaceX aims to improve manufacturability while maintaining the necessary performance for high-end AI workloads.
7Implications for Nvidia Partnerships
For Nvidia, this arrangement represents a shift away from selling complete compute-board assemblies. While this might reduce revenue from integrated systems, it could encourage other large hyperscalers to develop custom Nvidia-based platforms. This trend is critical as custom silicon solutions from major tech companies continue to gain traction in the global market.
8Future of Orbital Computing
Kessel represents a departure from conventional procurement models, placing more responsibility on SpaceX for system design and quality management. If successful, this platform could set a new standard for high-performance computing in space. The project’s ultimate success hinges on proving that custom-engineered hardware can outperform standard, off-the-shelf enterprise server solutions.
Original article: Tom's Hardware — SpaceX Develops Kessel AI System for Orbital Data Centers