SpaceX Kessel rack reportedly replaces Nvidia’s cables with an 84-layer board
SpaceX is reportedly developing the SpaceX Kessel rack, a custom AI system that connects 72 Nvidia Rubin GPUs through an 84-layer circuit board. GPUs, the processors that handle parallel AI calculations, normally communicate across Nvidia’s rack through cable assemblies. Kessel would change that physical arrangement while retaining Nvidia’s core computing hardware.
According to SemiAnalysis, the design targets both terrestrial and orbital deployment. Tom’s Hardware covered the claim on November 11, 2026. The immediate tension concerns engineering ownership: SpaceX would take responsibility for more of the system surrounding Nvidia’s chips.
That puts a custom rack architecture against Nvidia’s integrated reference design, the documented configuration manufacturers use to build compatible systems. It does not establish a working orbital computer. The available reporting describes a development effort, leaving manufacturing readiness, operating performance, and space qualification unresolved.
What the SpaceX Kessel rack would actually change
The reported change concerns how the rack connects its processors, rather than a replacement for Nvidia’s computing platform.
Tom’s Hardware describes a custom Rubin rack retaining 36 Vera CPUs, the general-purpose processors coordinating workloads, alongside Rubin accelerators. Its account attributes the underlying design information to SemiAnalysis. Those details should therefore remain reported specifications, rather than independently tested hardware characteristics.
The central component is a printed circuit board, or PCB, which carries electrical connections through patterned conductive layers. A midplane is a board positioned between assemblies so they can connect through it. In the reported arrangement, computing and switching trays face each other across that board.
NVLink, Nvidia’s high-speed communication technology for connected processors, remains central to the system. Replacing its cable cartridges does not imply replacing the communication protocol. Nor does it imply that Kessel eliminates every cable, connector, or external network connection.
This distinction matters because “cable-free” can describe different parts of a machine. A processor tray can eliminate internal wiring while the rack still uses cables between trays. Headlines that merge those levels make the reported modification sound broader than the evidence supports.
The public description concerns the rack’s internal communication path. It does not establish a new GPU architecture, a new memory technology, or a substitute for Nvidia’s software. Readers should resist treating a different physical connection system as evidence of higher application performance.
The phrase “SpaceX Kessel explained” therefore starts with a boundary: the processor platform stays recognizable, while its physical integration changes. The engineering question is whether that rearrangement improves the complete machine. A diagram alone cannot establish the answer.
Consider a training job distributed across connected accelerators. Each processor performs calculations, but the job also depends on data arriving when needed. A revised connection path matters only if it supports those exchanges reliably under sustained operation.
The same reasoning applies to an AI service answering requests. Installing the same processors does not automatically produce identical service behavior across different rack implementations. Cooling, communication reliability, maintenance procedures, and system software all influence what operators can deliver.
Nothing in the public Kessel description settles those operational questions. It establishes a specific architecture to investigate. The useful starting point is the proposed hardware change, followed by evidence about whether the resulting system works as intended.
That also keeps terrestrial and orbital interpretations separate. A custom interconnect can deserve attention before any spacecraft carries it. Its relevance on Earth does not depend on accepting every proposed space application.
Direct procurement would move responsibility toward SpaceX
Greater control over component sourcing also means accepting more responsibility for integration and support.
SemiAnalysis’s November 11 public thread describes direct component procurement, with SpaceX buying Nvidia chip packages and sourcing other components independently. It says this would bypass Nvidia’s L6 compute boards, the board assemblies surrounding the purchased computing packages.
That is a proposed purchasing arrangement, rather than a disclosed supply contract. The accessible thread does not establish order quantities, named manufacturing partners, or delivery commitments. Its linked subscription research was not available for this analysis.
The business implication follows from that boundary. Buying a more complete assembly concentrates integration work with its provider. Buying components separately gives the customer more choices, while requiring someone to make those choices function together.
For SpaceX, the reported approach would expand the decisions under its control. Those decisions include supplier selection and the physical architecture linking the purchased hardware. Control becomes useful when an operator has requirements that standard assemblies do not adequately serve.
However, direct purchasing does not automatically remove constraints. A buyer can choose suppliers and still depend on their production capacity, qualification processes, and replacement inventory. Moving a purchasing decision does not eliminate the underlying manufacturing requirement.
The pressure therefore falls first on the team choosing customization. It must justify the added engineering work against the benefits of its particular deployment. That is a demanding comparison even before orbital requirements enter the discussion.
Nvidia faces a different question: how much of the surrounding system a major customer wants to obtain through its established design. Kessel would still depend on Nvidia processors. The report does not support describing it as SpaceX abandoning Nvidia.
Nor does the available evidence identify a commercial loss for a particular board maker or cable supplier. Component arrangements can change during development. Naming winners and losers would require sourcing disclosures beyond the public architecture description.
A useful purchasing evaluation would follow responsibility through the entire machine. Who approves a board revision? Who investigates an intermittent communication fault? Who provides a replacement assembly when a system stops meeting its operating requirements?
Those questions become more consequential when a custom configuration differs from a widely deployed reference system. The operator needs a clear support boundary between processor behavior and its own integration choices. Otherwise, troubleshooting can cross organizational boundaries without a clear owner.
For enterprise buyers, that is the practical lesson. A recognizable accelerator model tells only part of the infrastructure story. Buyers also need to understand who validates the assembled system and who remains accountable after deployment.
Kessel’s reported procurement strategy makes that division unusually visible. Its potential advantage is architectural freedom. Its corresponding obligation is to turn that freedom into a machine that can be built, operated, repaired, and supported consistently.
Kessel vs Nvidia NVL72 is an integration choice
Nvidia’s documented rack provides an operational reference that a custom implementation would need to match or improve.
Nvidia’s NVLink rack architecture uses passive copper cable cartridges between computing and switching assemblies. Its hardware guide specifies 18 compute trays and nine switch trays. That documentation provides an official baseline for interpreting the reported Kessel changes.
Nvidia also describes a PCB inside its redesigned compute tray. Its Rubin tray design connects computing assemblies with front-mounted components. That existing board should not be confused with the reported Kessel midplane replacing rack-level NVLink cable cartridges.
The distinction resolves an apparent contradiction. Nvidia can describe simplified wiring inside a tray while documenting copper cable cartridges elsewhere in the rack. Kessel’s reported modification concerns a different physical boundary.
The comparison is therefore narrower than “boards versus cables” across the entire computer. Both approaches contain boards and connectors. The relevant question concerns where the high-speed paths run and how the surrounding assemblies connect.
A cable cartridge packages many connections into a replaceable assembly. A midplane consolidates paths within a board. Those choices create different manufacturing, inspection, installation, and replacement tasks, even when they serve the same communication purpose.
Neither format deserves an automatic performance advantage. Signal integrity, meaning the preservation of usable electrical signals, depends on the complete path. That path includes the transmitting hardware, connectors, conductive routes, receiving hardware, and operating conditions.
Application performance adds another layer of evaluation. A successful electrical connection is necessary, but operators ultimately care about completed workloads. A rack needs to maintain useful operation while its processors exchange data under realistic demand.
Serviceability also belongs in the comparison. Engineers need access to faulty parts, repeatable diagnostic procedures, and a workable replacement sequence. A layout that simplifies initial assembly can still create difficult maintenance tasks later.
This makes Kessel vs Nvidia NVL72 a question about the entire operating lifecycle. The custom approach must earn its place through measurable advantages relevant to SpaceX. A different arrangement is a design decision, not a result.
The reference design has a different appeal. It gives manufacturers and operators a shared configuration around which documentation, procedures, and compatibility work can develop. Customization changes how much of that shared work an operator can reuse.
That does not make a reference configuration universally preferable. A specialized deployment can justify changes that would be unnecessary elsewhere. The important requirement is to state the deployment constraint and show that the modification addresses it.
For developers, the physical layout matters indirectly. They need stable computing capacity, predictable job behavior, and reliable recovery when hardware fails. The architecture becomes valuable when those outcomes improve, rather than when its board specification sounds unusual.
For infrastructure teams, the comparison should begin with the same workload and operating assumptions. Without a common baseline, claims about assembly convenience or communication efficiency can obscure differences in cooling, software, maintenance, or test conditions.
The reported project raises a worthwhile engineering question. It does not yet provide the operating evidence needed to answer that question in Kessel’s favor.
An 84-layer board shifts the manufacturing challenge
Reducing cable assemblies transfers engineering work into the board and its connections.
SemiAnalysis contrasts Kessel’s reported 84-layer M10 midplane with an earlier Kyber proposal using 156 layers and M9 material. Those material designations describe different laminate grades, the layered materials used to build the board. The comparison is an analyst’s manufacturing assessment, not disclosed production validation.
Fewer layers can make a proposed design appear more approachable relative to that earlier plan. But layer count alone does not reveal whether a board can be produced consistently. It also does not establish whether its overall manufacturing cost improves.
This is where the distinction between design feasibility and production readiness becomes essential. An engineering team can specify a board before its suppliers demonstrate repeatable output. A prototype can work without establishing a reliable path to sustained manufacturing.
For a board-based interconnect, evaluation must reach beyond whether the finished assembly powers on. Engineers need to determine whether its connections remain within requirements across operating conditions. They also need procedures for detecting defects before the board enters service.
Manufacturing yield, the share of production meeting specifications, would be particularly useful evidence. No public Kessel yield figures appear in the sources reviewed here. Inferring them from the difference between the proposed layer counts would be unjustified.
Replacement policy presents another unresolved issue. Consolidating connections can simplify parts of assembly while making the central board more consequential. The operator needs to know what a board fault means for diagnosis, repair access, and system downtime.
That is a general integration concern, not a documented Kessel failure. The public reporting does not describe a failed prototype. It also does not establish that the proposed layout has passed a defined reliability program.
An informative demonstration would identify what was tested and under which conditions. A board photograph would establish physical hardware. A functioning rack would establish more, while sustained workload results would answer a different set of questions.
Even successful workload results need context. The duration of operation, cooling arrangement, software configuration, and treatment of faults all affect interpretation. A brief demonstration cannot support every claim an operator might make about continuous service.
The manufacturing argument should therefore stay conditional. If the design reduces assembly work while preserving reliability and repair access, customization gains a practical justification. If production or maintenance becomes harder, the apparent simplification may move effort elsewhere.
For readers seeking SpaceX Kessel explained, this is the central mechanism. The reported design relocates communication paths into a demanding central component. Its value depends on the consequences of that relocation across manufacturing and operation.
There is no need to declare that approach impractical. There is also no basis for describing its manufacturability as established. The public evidence supports a concrete proposal with identifiable validation questions.
Those questions remain relevant even if Kessel serves only terrestrial installations. Orbital use would add another set of requirements, rather than resolve the manufacturing ones.
Orbital deployment adds requirements the rack report does not settle
A rack intended for space still needs evidence that its complete operating environment has been addressed.
The space connection gives this story much of its attention. However, intended deployment describes a design objective. It does not establish a launch manifest, a flight-qualified assembly, or an operational orbital data center.
Thermal control illustrates the gap. NASA’s spacecraft thermal control guidance explains that vacuum removes convection with the surrounding environment. Heat must move through the spacecraft and ultimately leave through radiation.
That principle does not rule out computing in orbit. It establishes a requirement that a computing architecture must accommodate. Changing an internal electrical connection does not, by itself, explain how the spacecraft rejects the resulting heat.
The same separation applies to liquid cooling. Circulating fluid can move heat away from computing components. The broader spacecraft still needs a way to release that heat into its external environment.
An orbital design also needs a defined mechanical arrangement. The system must accommodate its launch and operating conditions. The accessible Kessel description does not provide the test evidence required to assess those conditions.
Reliability takes on a different operational meaning when technicians cannot approach the equipment as they would in a data center. Fault detection and recovery need to fit the actual service model. A terrestrial replacement procedure cannot simply be assumed available.
These are engineering questions, rather than evidence that SpaceX has overlooked them. The distinction matters. Missing public details establish a verification gap; they do not reveal everything happening inside a development program.
Likewise, an Earth-based installation would not confirm orbital readiness. It could establish useful evidence about assembly, software compatibility, communication behavior, and sustained workloads. Space deployment would still require evidence tied to the spacecraft configuration.
The reverse distinction is equally important. If the orbital application takes longer, the terrestrial design does not automatically lose its purpose. A custom rack can be evaluated against ground-based requirements independently of a future flight program.
For AI product users, orbital hardware claims also sit far from immediate service outcomes. A model answering requests depends on available computing capacity, networking, software, and operations. A proposed space-compatible rack does not establish changes in those user-facing results.
The strongest interpretation of the SpaceX Kessel rack report therefore remains bounded. It describes an effort to customize Nvidia-based infrastructure for particular deployment ambitions. It leaves the complete orbital system and its validation outside the available evidence.
That boundary helps avoid another common mistake: treating every SpaceX computing development as part of the same deployment. A terrestrial campus, a rack prototype, and an orbital payload answer different questions. Progress in one should not be silently substituted for proof of another.
The space ambition deserves scrutiny precisely because it adds requirements. Those requirements make published test conditions and operating results more useful than broad declarations about where the hardware is intended to go.
Watch confirmation, sustained operation, and space qualification
The next useful signals are disclosures that move the story from a reported design toward verifiable hardware.
Over the coming months, the first signal to watch is direct confirmation from SpaceX or Nvidia. Useful confirmation would define Kessel’s scope and the division of responsibility. It should distinguish development intentions from purchase commitments and deployment plans.
Such a disclosure would strengthen confidence that the public description reflects the actual program. A changed specification would instead narrow or revise the original account. Silence would leave the claim unresolved, rather than establish either success or failure.
The second signal is evidence of sustained operation in a complete rack. That should include a clear description of the tested configuration and workload. Readers need enough context to understand what the demonstration establishes.
A meaningful operating disclosure would address communication reliability, cooling conditions, and recovery from faults. Manufacturing evidence would also help connect a working example with a repeatable production process. These are related milestones, but they should remain distinguishable.
This is where Kessel vs Nvidia NVL72 becomes testable rather than descriptive. A comparison using consistent workloads and conditions could identify practical differences. Without that consistency, isolated performance claims would be difficult to interpret.
Evidence of stable operation would strengthen the case for the custom integration strategy. Persistent manufacturing or maintenance difficulties would weaken it. Either outcome would be more informative than speculation based only on the central board’s layer count.
The third signal is space-specific qualification tied to an identified hardware configuration. Useful disclosures would describe relevant environmental testing and the intended flight arrangement. A connection to an actual payload or mission would make the deployment claim more concrete.
That evidence would strengthen the orbital interpretation of the project. A terrestrial deployment announcement would support a narrower conclusion. It would show ground-based progress without resolving the additional requirements associated with flight.
Enterprise buyers can apply the same sequence to their own evaluations. Establish who stands behind the architecture, examine sustained operating evidence, and verify suitability for the intended environment. Those questions remain useful regardless of the vendor or deployment location.
Teams following the story should also preserve the distinction between a supplier statement, an analyst inference, and a measured result. A shared searchable knowledge base can keep technical documents and later revisions connected. That makes it easier to notice when a proposed specification becomes confirmed hardware, or quietly changes.
Follow the SpaceX Kessel rack through those evidence milestones before drawing conclusions about its eventual role. Start with the reported architecture, then look for accountable confirmation and sustained workload results. Treat orbital readiness as a separate claim requiring its own disclosures. When a new announcement arrives, ask what it actually establishes: a design decision, a manufactured board, an operating rack, or flight-qualified equipment. Keep those categories distinct when discussing capacity plans or evaluating infrastructure suppliers. The most useful next question is specific: what newly published evidence shows that this custom integration performs reliably in the environment SpaceX intends to use?



