 is developing a custom version of [Nvidia](https://baxtel.com/data-centers/nvidia)'s Vera Rubin NVL72 rack for orbital and terrestrial data centers,...](https://cdn.baxtel.com/data-center/spacex-designs-kessel-rack-with-72-nvidia-rubin-gpus-for-orbital-data-centers/logo/large_spacex-designs-kessel-rack-with-72-nvidia-rubin-gpus-for-orbital-data-centers-featured.jpg){kind=logo}
SpaceX is developing a custom version of Nvidia's Vera Rubin NVL72 rack for orbital and terrestrial data centers, research firm SemiAnalysis reported on November 11. The system, codenamed Kessel, keeps the NVL72 layout of 72 GPU packages, 36 Vera CPUs and nine NVLink switch trays, and swaps the rack's internal NVLink cable cartridges for a passive printed circ
In-article: Close-up view of an assembled NVL72 rack-scale server system displaying individual compute and switch trays.uit board midplane.
Nvidia supplies the chip packages, and SpaceX designs and builds the rest of the rack itself, bypassing the L6 compute boards Nvidia normally sells, according to SemiAnalysis. [Data Center Dynamics reported](https://www.datacenterdynamics.com/en/news/spacex-developing-custom-nvidia-vera-rubin-rack-dubbed-kessel-for-its-orbital-data-centers-report) that the direct-sourcing approach gives SpaceX more control over supplier selection and component sourcing. [Tom's Hardware reported](https://www.tomshardware.com/tech-industry/artificial-intelligence/spacex-develops-its-own-kessel-rack-scale-system-featuring-72-nvidia-rubin-gpus-for-data-centers-in-space-report-claims-replaces-conventional-nvlink-cable-assemblies-with-a-massive-84-layer-printed-circuit-board) that SpaceX also takes on design validation, manufacturing quality and supply-chain risk.
The midplane is the key difference from Nvidia's own design. Kessel specifies 84 layers of M10 laminate, a higher-grade ultra-low-loss material, against the 156 layers of M9 laminate planned for Nvidia's larger Kyber rack. SemiAnalysis described the layout as "a mini Kyber, with a smaller scale-up domain," and said the lower layer count gives Kessel a more feasible manufacturing path. Tom's Hardware noted that high-end server motherboards use 15 to 20 layers, and that a defect in any internal layer can scrap an entire board.
SemiAnalysis reported separately in July 2026 that Nvidia had delayed production of its Kyber racks by more than 12 months because of manufacturing issues with the midplane that connects the modules. Data Center Dynamics reported that the problem also affects the planned NVL576 system and pushed its release into 2028.
Kessel fits SpaceX's stated orbital plans. In August, chief executive Elon Musk said SpaceX and Nvidia had designed a space-optimized Vera Rubin NVL72 system for the first Starmind AI1 satellite, targeted for launch in the fourth quarter of 2027. R&D World reported that Musk rated that system at 175 kW of average compute and up to 250 kW at peak, and that Musk called the design
In-article: Close-up view of an assembled NVL72 rack-scale server system displaying individual compute and switch trays.
"significantly simpler, lower cost, denser and lighter than a traditional rack." On SpaceX's August earnings call, Musk said the company was reworking the NVL72 rack for orbital and terrestrial use and called it "a radical simplification of the normal NVL72 rack," as Baxtel reported.
Neither SpaceX nor Nvidia has commented publicly on Kessel, and MLQ.ai noted that SemiAnalysis did not say whether Kessel is the system Musk described. Baxtel lists SpaceX as an AI end user with one tracked data center site, the Starbase campus in South Texas, where 500 MW is recorded under construction for 2027 and a further 500 MW as planned.
Related on Baxtel
