Reportedly, SpaceX (SPCX.US) is attempting "de-NVIDIA Corporation-ization" at the rack level: developing the Kessel rack to bypass the L6 compute board and buy only chips.

SemiAnalysis, an independent research firm focused on AI and semiconductors, disclosed on X that SpaceX (SPCX.US) is doing something previously attempted only by hyperscale cloud providers: keeping NVIDIA Corporation's (NVDA.US) chips but redoing the rack around them itself. The internal codename for this self-developed rack is Kessel. The most substantive change is at the procurement level. SemiAnalysis says SpaceX plans to buy chip packages directly from NVIDIA Corporation and procure the remaining components itself, thereby bypassing NVIDIA Corporation's L6 compute boards. The firm wrote: "This also changes who buys which parts: SpaceX plans to buy chip packages from NVIDIA Corporation and source the rest directly, bypassing NVIDIA Corporation's L6 compute boards. This would give SpaceX more control over supplier selection and component procurement." In NVIDIA Corporation's tiered system, L6 is the board-level component, and above that is the L10 full rack. Skipping L6 means SpaceX gains design and procurement control at the board levelPCB, connectors, power delivery, and cooling shift from "specified by NVIDIA Corporation" to "chosen by SpaceX." Replacing copper cables with a single circuit board The hardware layout has also changed. NVIDIA Corporation's standard Vera Rubin NVL72 rack uses internal copper cable "cartridges" to connect compute and switch hardware; Kessel replaces these cartridges with a centrally located printed circuit board, which SemiAnalysis calls a passive PCB midplane. It retains the Oberon architecture's 72 GPU packages, 18 compute trays, and 9 switch trays, but arranges the trays back-to-back around this midplane. In one sentence: Kessel is a smaller version of NVIDIA Corporation's Kyber designKyber also uses a circuit board as a midplane, but with higher density. Smaller has its benefits: SemiAnalysis says Kessel's design may be easier to manufacture, at the cost of requiring an 84-layer board using M10 material; the earlier Kyber design was planned to use 156-layer M9 material. Kyber's repeated delays have been bottlenecked precisely by this midplanein July this year, SemiAnalysis disclosed that Kyber NVL144 was delayed by more than 12 months to 2028 due to difficulties manufacturing the midplane PCB, with the midplane at the time described as three 26-layer boards pressed into 78 layers. Layer count and material grade are directly tied to yield: the more layers and the higher the material grade (M9 to M10 represents an upgrade in copper-clad laminate grade), the higher the scrap rate per board and the harder it is to ramp capacity. SpaceX's choice of "a smaller rack plus fewer layers" is essentially trading density for manufacturability. Why does this matter more than a single rack? The rack level is where NVIDIA Corporation has been capturing value upward over the past three years. From selling GPUs to selling HGX boards to selling NVL72 full racks, each step up has brought more components into its own specifications and supply chain, narrowing customers' room to substitute. Kessel is the first customer-initiated counter-move in this direction. It brings at least two consequences: First, the transfer of supply chain choice. The supplier lists for PCB, connectors, power, and cooling shift from NVIDIA Corporation's AVL (approved vendor list) to SpaceX's own list. For upstream vendors, this means an additional channel that does not go through NVIDIA Corporation's certification system. Second, it leaves an interface for "swapping chips." SpaceX states plainly on its Starmind page: "We are silicon-agnostic on AI chips. Our system architecture supports compute modules from any vendor." Since the rack is self-designed, the silicon is not irreplaceablethis is both a negotiating chip and a long-term option. SemiAnalysis also mentions that SpaceX is expanding around Colossus and plans to build a data center on part of the Starbase rocket campus. But power remains the bottleneck: the firm's earlier estimate of power supply is below the demand of these data centers. NVIDIA Corporation and SpaceX set their sights beyond Earth The same logic is being moved into orbit. NVIDIA Corporation and SpaceX have been collaborating on an AI computer for space; in August this year, the two plan to adapt Vera Rubin NVL72 to the first Starmind AI satellite. NVIDIA Corporation executive Ian Buck said: "SpaceXAI is taking this architecture from massive AI factories to the new frontier of orbital computing." He added that the work requires re-engineering power, cooling, reliability, and communications for the space environment while retaining NVIDIA Corporation's architecture and software. According to SpaceX's published figures, AI1's compute payload peaks at 250 kilowatts and averages 175 kilowatts; compute results are transmitted back via laser links through the Starlink network. On the production side, the action is the Gigasat factory, located in Bastrop, Texas. SpaceX plans to begin production and deploy thousands of AI satellites as early as the end of 2027. In addition, SpaceX plans to use Vera Rubin to expand the compute infrastructure behind Grok. On Stocktwits, SpaceX's retail sentiment rose from "neutral" a week ago to "bullish," with discussion volume at a "normal" level. The stock is down 11% year to date. Kessel's technical details (84 layers, M10, passive midplane) will eventually be superseded by the next generation of products, but the example it sets will not: when the rack becomes the bottleneck, customers start defining the rack themselves. Kyber's delay to 2028 gave this demonstration a time windowbefore NVIDIA Corporation's next-generation full rack arrives, SpaceX has already taken away the L6 layer with a board that has fewer layers and newer material. The real constraint is not on the board, but on power. SemiAnalysis points to the same conclusion on both the ground and orbital sides: the next wall for compute is power supply, and SpaceX has not yet offered a solution to this wall.