SpaceX Replaces NVLink Cables with a Single 84-Layer Board in Its Kessel AI Rack

Normally an AI rack is assembled like a construction set: compute nodes, switches, and bundles of copper NVLink cables in between. SpaceX and xAI decided that such a set does not suit their needs, and they are drawing up the Kessel rack, which has no cables at all. In their place is one gigantic 84-layer printed circuit board.

What is inside Kessel

According to SemiAnalysis, the compute portion stays the same as in the Vera Rubin NVL72: 36 Vera processors and 72 Rubin accelerators across 18 compute nodes, plus 9 NVLink switch modules. The only thing that changes is how it is all connected.

In a standard NVIDIA rack, nodes and switches are linked by cable cartridges. In Kessel, the trays are placed back to back around a passive midplane. Picture a bookshelf where, instead of wires between the books, there is one shared board with contacts: the nodes simply plug into it from both sides. NVIDIA uses a similar idea in its own Kyber rack, where modules also connect without cables through a shared panel.

Why 84 layers: both bold and risky

For comparison: ordinary server motherboards have 14–20 layers, and test boards have only slightly more than 20. In Kyber, the backplane has 156 layers on M9-class material. Kessel has 84 layers, but on the more expensive M10 dielectric with ultra-low signal loss. Fewer layers mean simpler production and a higher yield of good boards, but the material itself costs more.

The risk is that such a board cannot be repaired piece by piece. An etching or alignment defect on an inner layer means the entire panel is scrapped, and the more layers there are, the costlier each mistake. Whether Kessel wins economically depends on whether the savings on cables and assembly cover the cost of the complex board. There are no public figures on this yet.

The second shift: chips without ready-made integration

The second part of the story concerns procurement. With the release of Vera Rubin, NVIDIA plans to supply partners mainly with ready-made L6-level compute boards or fully assembled L10-level server nodes. SpaceX is taking a different path: it takes packaged processors, accelerators, switch chips, and network controllers from NVIDIA, while designing the boards and buying the remaining components itself.

What this gives it: freedom in choosing board manufacturers, the ability to adapt the design to orbit (heat dissipation in a vacuum, vibration at launch), and a smaller integration markup. The price of this: validation, quality control, and supply chains, including scarce memory, fall on SpaceX's shoulders.

What this means in practice

For most developers and product managers, Kessel is not something you can buy or rent tomorrow. But the signal is important:

• Large customers are starting to move away from ready-made turnkey racks and to build platforms from individual chips. If the approach works, other hyperscalers may follow SpaceX. NVIDIA would lose margin on every rack but keep its position as the main chip supplier.
• Physical constraints (cables, heat dissipation, vibration) are becoming part of AI infrastructure architecture. Those planning their own clusters should count not only the price of the GPU but also the cost of integration.
• Orbital data centers remain a claim for now. The project is designed for both terrestrial and orbital conditions, but there is no data on real-world operation.

The main question for the coming year: will the 84-layer board scale in volume production? If so, cable cartridges in top-tier racks will start to look like a relic. If not, Kessel will remain an expensive experiment.

Limitations: all of this is based on SemiAnalysis material, not on official SpaceX or NVIDIA specifications. The cost and yield of good boards have not been disclosed, so conclusions about the project's economics remain a hypothesis.