A rack rebuilt around one board
The design was outlined by SemiAnalysis in a post about SpaceX’s Kessel rack, which describes the system as a custom version of Nvidia’s Vera Rubin NVL72 platform. Kessel would retain 72 GPU packages, 18 compute trays and nine switch trays, the same broad component count associated with the reference system.
Read Nvidia’s Vera Rubin Could Redefine the Economics of AI Inference
The significant change would be inside the rack. Instead of connecting the compute and switching trays through internal NVLink cable cartridges, SpaceX would arrange them back-to-back around an 84-layer passive printed-circuit-board midplane. In practical terms, the midplane would act like a dense electrical backbone, with the accelerator hardware facing the same central board from opposite sides.
SemiAnalysis compared the arrangement to a smaller version of Nvidia’s Kyber architecture. That comparison is useful because it shifts attention away from the individual GPU and toward the physical environment around it. At these densities, the rack is not merely a box containing processors. Its board layout, connectors, cooling paths and power system all influence how much computing can be delivered in a given space.
The post also says SpaceX would purchase Nvidia chip packages while sourcing other components directly. That would give the company more control over the rack’s mechanical and electrical design than it would have with a standard, fully assembled Nvidia platform.
Why SpaceX may want its own architecture
A custom rack could matter if SpaceX intends to put large amounts of computing in places where conventional data-center equipment is difficult to use. The company’s official Starmind page describes an orbital AI-computing effort and says its system architecture can support compute modules from any provider.
That flexibility points to a broader strategy. Rather than designing an orbital system around one fixed accelerator configuration, SpaceX could build an infrastructure layer that accepts different types of computing hardware. Kessel would fit that approach by treating Nvidia’s GPU packages as one part of a larger system that SpaceX controls.
In a future orbital installation, the visible result might not be a familiar data-center rack. It could be a tightly integrated computing module connected to power systems, communications equipment and large external radiators. Workloads could be divided between Earth and space, with machines in orbit handling selected AI tasks while people on the ground interact with the results through ordinary software.
That scenario remains speculative. Neither the SemiAnalysis post nor the supplied official Starmind material confirms that Kessel has entered production, that it will fly, or that SpaceX has selected a deployment site.
The engineering tradeoff
Replacing cable cartridges with a passive midplane could simplify the internal arrangement and reduce the number of cable assemblies that must be routed through the rack. A board running through the system may also make the design more compact and give engineers a more controlled path between GPUs and switches.
But a midplane is not automatically a simpler solution. It concentrates a large number of high-speed connections into a single manufactured component. Any flaw in its layers, traces or connectors could affect a substantial portion of the rack. Repairing or replacing that board could also be more difficult than changing an individual cable assembly.
The design would have to manage signal integrity, power delivery and heat while maintaining tight manufacturing tolerances. Those challenges become more severe when the rack is expected to operate in a constrained environment, where cooling and physical access may be limited. A passive board does not remove complexity. It moves complexity into the board’s design, fabrication and validation.
The biggest unanswered question is scale
The most ambitious interpretation comes from an estimate rather than a SpaceX announcement. Wccftech reported that GF Securities analyst Jeff Pu estimates SpaceX could deploy roughly 10,000 compact Kessel racks for orbital AI computing by 2028. The report also describes a space-oriented NVL72 design using high-power GPUs and external-radiator cooling.
That forecast would imply an enormous manufacturing and launch effort. It would also make rack efficiency more than an engineering preference. Small improvements in size, serviceability and power management could multiply across thousands of systems.
However, the estimate should not be confused with a confirmed company plan. FourWeekMBA has emphasized that SpaceX has not publicly confirmed Kessel, its component-sourcing strategy or deployment volumes. It also notes that the more detailed supplier analysis from SemiAnalysis is behind a paywall, limiting what outsiders can independently verify.
That criticism is central to the story. The Kessel post provides a specific architecture, but not a production schedule, contract, launch plan or order volume. The public evidence supports the existence of a reported design effort, not the conclusion that thousands of orbital racks are already on their way.
For now, Kessel is best understood as a sign of where large AI operators may be heading. As accelerator systems become denser and more specialized, companies may increasingly buy the core chips while designing the surrounding machines themselves. SpaceX’s reported midplane rack shows how the next contest in AI infrastructure could be fought not only over processors, but also over the physical architecture that allows them to work together.
Maya Lindqvist is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Maya Lindqvist's usual length and in Maya Lindqvist's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.