SpaceX's Reported Kessel Design Trades Nvidia Racks for an 84-Layer Orbital Compute Midplane
Market
Orbital and terrestrial AI infrastructure hardware/software co-design / hyperscalers choosing between integrated racks and custom supply chains.
Trend
This is a reported design, not a benchmark or deployed orbital system. Kessel retains 36 Vera CPUs, 72 Rubin GPUs and nine NVLink switch trays, but SpaceX/xAI would buy chips and switches directly while sourcing the rest itself—trading Nvidia's system integration for control over orbital packaging, manufacturing partners and intermediary cost.
Tech Highlight
The design replaces conventional NVLink cable cartridges with an 84-layer passive PCB midplane using ultra-low-loss M10 laminate and back-to-back compute and switch trays. Fewer cable assemblies may simplify integration, but the board is roughly four times the layer count of a high-end server motherboard; internal-layer defects can scrap the whole unit, making yield, thermal validation, launch loads and radiation behavior the real hardware/software partition risks.
6-Month Outlook
Watch for a physical prototype, signal-integrity and thermal results, manufacturing yield, radiation testing and a terrestrial-versus-orbital bill of materials. Measured reliability or cost advantages over Nvidia's integrated rack would validate vertical integration; continued reporting without hardware evidence would keep the architecture at proposal maturity.