SpaceX is reportedly developing Kessel, a custom rack-scale system built around 72 Nvidia Rubin GPUs, with a design intended for both terrestrial and orbital AI data centres. The striking move is not a new chip, but a different way to connect and assemble the hardware around it.
xAI Watch analysis
What happened
Tom’s Hardware reports, citing SemiAnalysis, that Kessel keeps the broad configuration of Nvidia’s Vera Rubin NVL72 system: 36 Vera CPUs, 72 Rubin GPUs and nine NVLink switch trays. Instead of conventional internal NVLink cable assemblies, the reported design places compute and switch trays around a passive 84-layer printed circuit board.
The report says the system is intended for ground-based and orbital deployments. SpaceX would reportedly buy major components such as GPUs, CPUs and switches from Nvidia, while sourcing other parts and managing more of the system design and supply chain itself. These are reported plans, not a confirmed product launch.
Why it matters
A custom rack could let SpaceX tailor AI infrastructure to its own needs, including a proposed orbital setting, while relying on Nvidia’s core processors. Replacing many high-speed cable connections with a board-based interconnect may simplify assembly, but the unusually complex board could also be costly and difficult to manufacture reliably. That is a substantial engineering trade, not a free shortcut.
Our read
Kessel is an intriguing example of AI infrastructure moving beyond buying ready-made racks. The orbital angle gives the story its cinematic flourish; the hard question is whether the engineering and economics work on Earth first. SpaceX is reportedly taking on more responsibility for validation, manufacturing quality and supply-chain management, so the useful proof will be a working system, not a memorable codename.
What to watch
- Whether SpaceX or xAI confirms the reported Kessel design and its intended use.
- Whether a working system is demonstrated, and what workloads it can run.
- Whether the custom interconnect proves manufacturable at scale and delivers practical benefits.
Discussion spark: Would you trust a custom AI rack built around Nvidia parts to deliver meaningful advantages, or is the added design and manufacturing burden a needless complication?
Sources and evidence
not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor