SpaceX develops Kessel AI rack, buying chip packages from NVIDIA only

On the morning of November 11, 2026, semiconductor research firm SemiAnalysis posted on X that SpaceX is developing a custom AI server rack codenamed Kessel. According to the disclosure, SpaceX plans to keep using NVIDIA chips but buy chip packages directly from NVIDIA and source all other components on its own, bypassing NVIDIA's L6 compute board. SemiAnalysis described the move as something previously attempted only by hyperscale cloud providers, and said it hands SpaceX design and procurement control at the board level. PCBs, connectors, power delivery and cooling would move from NVIDIA-specified to SpaceX-selected, giving the company more control over supplier choice and component purchasing.

The same disclosure adds that Kessel is intended for both ground-based data centers and orbital computing. The rack's internal layout would also be redesigned. NVIDIA's standard Vera Rubin NVL72 rack relies on internal copper cable modules to connect compute devices and switches, whereas Kessel would drop those modules and use a passive circuit board at the center of the rack as a midplane. Kessel would still use the 72 GPU packages, 18 compute trays and 9 switch trays of the Oberon design, but mount the trays back-to-back on either side of the central board. The overall design is said to resemble NVIDIA's Kyber rack, which also uses a circuit-board midplane; the material stops at the scale comparison and no further details have been disclosed since.