A report says SpaceX is developing a custom Nvidia Vera Rubin rack, dubbed Kessel, for its orbital data centers, keeping the underlying configuration while changing the PCB.

SpaceX is developing a custom Nvidia Vera Rubin NL72 rack, dubbed Kessel, according to SemiAnalysis.

The research firm reports that the bespoke rack will keep the NVL72's underlying configuration, which comprises 72 GPU packages distributed across 18 compute trays and nine switch trays. However, it will replace the system's NVLink cable cartridges with an 84-layer printed circuit board (PCB).

SpaceX plans to use the custom Kessel racks "for both terrestrial and orbital deployment," the report said.

SemiAnalysis described the setup as a "mini Kyber, with a smaller scale-up domain." Nvidia's Kyber rack is set to contain 576 AI GPUs in 144 packages. The key difference between the two is that the Kyber architecture has a backplane featuring a 156-layer PCB with M9 copper-clad laminates (CCLs), substantially more than the 84-layer M10 midplane specified for Kessel.

The research firm said fewer layers will give Kessel a "more feasible manufacturing path." While SpaceX will buy the chip packages from Nvidia, the company plans to bypass Nvidia's L6 compute boards and procure the remaining parts directly. This gives SpaceX greater control over supplier selection and component sourcing.

In July 2026, SemiAnalysis separately reported that Nvidia had delayed production of its Kyber racks by more than 12 months due to manufacturing issues with the PCB midplane that connects the modules. At the time, SemiAnalysis said the manufacturing issues would also affect the planned NVL576 system. As a result of the delays, Nvidia was pushing the hardware's release date into 2028.