SpaceX is reportedly developing Kessel, a custom rack-scale system built around NVIDIA's Vera Rubin platform. The interesting part is not the 72 GPUs it retains. It is what SpaceX plans to remove.
Kessel reportedly replaces NVIDIA's conventional NVLink cable cartridges with an 84-layer passive PCB midplane using M10 ultra-low-loss copper-clad laminate. The compute and switch trays are arranged back-to-back around this central board.
The reported configuration retains 72 Rubin GPUs, 36 Vera CPUs, 18 compute trays and nine NVLink switch trays. NVIDIA's standard NVL72 uses the same basic compute configuration, connected through a passive copper-cable backplane.
This is a significant architectural change.
1. The midplane is the real engineering challenge
A high-speed interconnect is not simply a collection of wires connecting GPUs and switches. It is an electrical channel whose performance depends on insertion loss, return loss, crosstalk, impedance discontinuities and timing skew.
Replacing cable assemblies with a multilayer PCB changes the entire signal path.
The advantage is structural integration: fewer cable assemblies, fewer individual connections and potentially simpler rack-level assembly.
The cost is that the midplane itself becomes a highly demanding electrical and manufacturing component. At 84 layers, material consistency, lamination accuracy, via design and layer registration become critical. A defect in an internal layer can compromise the entire board.
M10 laminate is intended to reduce high-frequency signal loss, but better material does not automatically guarantee a better system. The complete channel must still meet the electrical requirements of the NVLink topology.
2. Why 84 layers instead of 156?
SemiAnalysis reports that Kessel uses an 84-layer M10 midplane, compared with a 156-layer M9 design planned for NVIDIA's larger Kyber architecture.
This is not simply a reduction in manufacturing complexity.
Fewer layers can reduce lamination steps and improve the potential manufacturing path. But the higher-grade M10 material increases material costs, while the larger board and demanding routing requirements continue to create yield challenges.
The real engineering objective is to achieve the required electrical performance with fewer layers, without making material costs and production yield uneconomic.
This is a trade-off between signal integrity, manufacturability and system cost.
3. SpaceX is changing the procurement boundary
The more consequential change may be commercial rather than electrical.
Under the reported arrangement, SpaceX would purchase NVIDIA's chip packages directly while sourcing more of the surrounding components independently, bypassing NVIDIA's L6 compute boards.
This shifts responsibility for board design, component qualification, system validation and manufacturing yield toward SpaceX and its manufacturing partners.
NVIDIA retains the value of its GPUs, CPUs and interconnect technology, but potentially gives up part of the system integration and hardware supply-chain value.
For large customers with sufficient engineering resources, this creates an alternative to buying a fully integrated rack.
4. Why this matters for AI infrastructure
The industry has spent years optimizing GPU performance. But at rack scale, the limiting factors increasingly include power distribution, cooling, signal integrity, packaging and manufacturing yield.
A more compact architecture is not automatically a more economical one. Eliminating cable assemblies only creates value if the savings exceed the cost of the midplane, additional validation and manufacturing losses.
This is particularly important for orbital deployment, where thermal rejection, radiation exposure, launch mass and limited repairability impose constraints that terrestrial data centers do not face.
A ground-qualified rack is not automatically space-qualified hardware.
The investment implications
Kessel could create opportunities for suppliers capable of producing ultra-low-loss laminates, advanced multilayer PCBs, high-speed connectors and reliable power and cooling systems.
But technical relevance is not evidence of a purchase order. The decisive indicators will be supplier qualification, production yields, actual rack deployment and repeat orders.
The larger strategic implication is more important:
NVIDIA may continue to own the computing platform without owning every layer of the physical system.
As hyperscalers and large infrastructure operators gain the ability to design their own racks, value could gradually migrate from integrated system assembly toward the specialized components and manufacturing capabilities that make those systems possible.
Kessel is therefore not necessarily an attempt to replace NVIDIA. It is a test of how much of NVIDIA's system-level value customers can take back without sacrificing the performance and reliability of its computing platform.
