SpaceX (SPCX.US) reportedly attempts “de-Nvidiaization” at the rack level: developing Kessel cabinets to bypass L6 compute boards, purchasing only chips

English Planet Daily APP has learned that SemiAnalysis, an independent research institution focused on artificial intelligence and semiconductors, disclosed on X that SpaceX (SPCX.US) is attempting something previously only tried by hyperscale cloud providers: retaining Nvidia (NVDA.US) chips but completely redesigning the rack around the chips. This in-house rack is internally codenamed Kessel.

The most fundamental change is at the procurement level. SemiAnalysis stated that SpaceX plans to purchase chip packages directly from Nvidia, source the rest of the components on their own, and thereby bypass Nvidia’s L6 compute board. The institution wrote: “This also changes who buys which parts: SpaceX plans to buy chip packages from Nvidia and procure other components directly, skipping Nvidia's L6 compute board. This gives SpaceX more control over supplier selection and component procurement.”

Within Nvidia’s hierarchical system, L6 is a board-level component, above which is the L10 full rack. Bypassing L6 means SpaceX gains both design and procurement rights at the board level—PCB, connectors, power supply, and cooling—changing these links from “Nvidia specified” to “SpaceX chosen.”

Replacing Copper Cables with a Whole Circuit Board

The hardware layout has also been modified. Nvidia’s standard Vera Rubin NVL72 rack uses internal copper cable “wire boxes” to connect compute and switching hardware; Kessel replaces these wire boxes with a centrally positioned printed circuit board, which SemiAnalysis refers to as a passive PCB mid-board.

It retains the Oberon architecture’s 72 GPU packages, 18 compute trays, and 9 switch trays but arranges them back-to-back around the mid-board. To summarize: Kessel is a smaller version of Nvidia’s Kyber design—Kyber also uses a circuit board as its mid-board but at a much higher density.

Smaller has its advantages: SemiAnalysis says Kessel’s design is likely easier to manufacture, with the tradeoff being a need for an 84-layer circuit board made from M10 material; earlier Kyber designs planned to use a 156-layer board made from M9 material.

The repeated delays to Kyber are precisely because of this mid-board—SemiAnalysis reported in July that Kyber NVL144 was delayed more than 12 months to 2028 due to difficulties in manufacturing its PCB mid-board, which at the time was described as three 26-layer plates laminated to form 78 layers. Layer count, material grade, and yield are directly linked: The more layers, the higher the material grade (M9 to M10 is an upgrade in copper-clad laminate), the higher the scrap rate of a single board and the harder it is to scale production. SpaceX’s choice of “smaller rack + fewer layers” essentially trades density for manufacturability.

Why Is This More Than Just a Rack?

The rack level has been where Nvidia has recaptured value in the past three years. From selling GPUs, to selling HGX boards, to selling NVL72 full racks, every step Nvidia takes upwards brings more parts into its own specifications and supply chain, narrowing customers’ ability to substitute components.

Kessel is the first counter-move initiated by a customer. It brings at least two outcomes:

First, the transfer of supply chain choice. The supplier list for PCB, connectors, power supplies, and cooling shifts from Nvidia’s AVL (Approved Vendor List) to SpaceX’s own list. For upstream manufacturers, this means there’s now a channel that bypasses Nvidia’s certification system.

Second, it leaves an interface for “chip replacement”. SpaceX states very directly on its Starmind page: “We are agnostic about AI chip suppliers. Our system architecture supports compute modules from any supplier.” Since the rack is self-designed, the silicon chips aren’t irreplaceable—this is both a bargaining chip and a long-term option.

SemiAnalysis also mentioned that SpaceX is expanding around Colossus and plans to build a data center in a section of the Starbase rocket complex. But electricity remains a bottleneck: The previously estimated power supply by the institution is below what these data centers require.

Nvidia and SpaceX Look Beyond Earth

The same logic is being moved to orbit. Nvidia and SpaceX are already collaborating to develop AI computers for space; in August, they plan to adapt Vera Rubin NVL72 for the first Starmind AI satellite. Nvidia executive Ian Buck stated: “SpaceXAI is bringing this architecture from massive AI factories to the new frontier of orbital computing.” He added that this effort requires reworking power, cooling, reliability, and communication for the space environment, while still preserving Nvidia’s architecture and software.

According to SpaceX’s official statements, AI1’s compute payload peaks at 250 kW, averaging 175 kW; computation results are transmitted via laser links and relayed back through the Starlink network.

The production move is the Gigasat factory in Bastrop, Texas. SpaceX plans to begin manufacturing and deploying thousands of AI satellites by as early as the end of 2027. Additionally, SpaceX plans to use Vera Rubin racks to expand the computing infrastructure behind Grok.

On Stocktwits, retail sentiment toward SpaceX has risen from “neutral” a week ago to “bullish,” with the discussion heat at “normal” levels. The stock has fallen 11% year-to-date.

Kessel’s technical specifics (84 layers, M10, passive mid-board) will ultimately be superseded by next-generation products, but its demonstration will not: When racks become a bottleneck, customers start to define racks themselves. Kyber's delay to 2028 has given a window—before Nvidia’s next full rack arrives, SpaceX has already taken the L6 layer away using a board with fewer layers and newer materials.

The real constraint is not the board, but electricity. SemiAnalysis points to the same conclusion both on the ground and in orbit: the next wall for computing power is power supply, and so far SpaceX has not provided a solution for it.