The Zhitong Finance App learned that SemiAnalysis, an independent research institute focusing on artificial intelligence and semiconductors, revealed on X that SpaceX (SPCX.US) is doing what only hyperscale cloud vendors have tried before: keep the NVDA.US (NVDA.US) chip, but redo the frame surrounding the chip itself. The internal code name for this self-developed rack is Kessel.
One of the most substantial changes was at the procurement level. According to SemiAnalysis, SpaceX plans to purchase the chip package directly from Nvidia and purchase the rest of the components itself, thereby skipping Nvidia's L6 calculation board. The agency wrote, “It also changed who buys what parts: SpaceX plans to buy chip packages from Nvidia and directly procure the rest, bypassing Nvidia's L6 calculation board. This will give SpaceX more control over supplier selection and parts procurement.”
In Nvidia's hierarchical system, L6 is a board-level component, followed by an L10 cabinet. Skipping L6 means SpaceX has obtained design and procurement rights for the board layer — PCB, connectors, power supply, and heat dissipation, and changed from “designated by Nvidia” to “SpaceX's own choice.”
Replace the copper cable with an entire circuit board
The hardware layout has also changed. The Nvidia standard Vera Rubin NVL72 rack uses internal copper “wire boxes” to connect computing and exchange hardware; Kessel replaced these wire boxes with a central printed circuit board, which SemiAnalysis called a passive PCB midplane.
It retains the Oberon architecture's 72 GPU packages, 18 compute trays, and 9 exchange trays, but arranges the trays back-to-back around this midplane. To sum it up in one sentence: Kessel is a smaller version of Nvidia's Kyber design—Kyber also uses a circuit board as the midplane, but it's more dense.
Small has small benefits: SemiAnalysis says Kessel's design may be easier to manufacture at the cost of requiring an 84-layer circuit board with M10 material; while the earlier Kyber design plans to use 156 layers of M9 material.
The reason for Kyber's repeated extensions was this midboard — in July of this year, SemiAnalysis revealed that the Kyber NVL144 was delayed by more than 12 months to 2028 due to difficulties in manufacturing midboard PCBs. At the time, it was said that the midboard was made of 3 26-layer boards pressing 78 layers. The number of layers and material grade are directly linked to yield: the more layers, the more advanced the material (M9 to M10 is an increase in the grade of copper-clad plates), the higher the scrap rate of a single board, and the more difficult it is to increase production capacity. SpaceX chose “smaller frame+fewer layers”, essentially trading density for manufacturability.
Why is this more important than a rack?
The cabinet floor is where Nvidia has increased its value over the past three years. From selling GPUs, to selling HGX boards, to selling the NVL72 cabinet, every time Nvidia moves up, it incorporates more components into its own specifications and supply chain, and the customer's replacement space narrows.
Kessel is the first client-initiated reverse movement in this direction. It has at least two consequences:
First, the transfer of options in the supply chain. The supplier list for PCBs, connectors, power supplies, and cooling changed from Nvidia's AVL (Qualified Supplier List) to SpaceX's own list. For upstream manufacturers, this means an additional channel that does not go through the Nvidia certification system.
Second, there is an interface for “changing the chip.” SpaceX put it bluntly on its Starmind page: “We are agnostic about AI chip vendors. Our system architecture supports computing modules from any vendor.” Since the rack was designed by oneself, silicon wafers are not 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 part of the Starbase rocket park. However, electricity is still a bottleneck: the agency's previously estimated electricity supply is lower than the demand for these data centers.
Nvidia and SpaceX set their sights beyond Earth
The same set of logic is being moved to orbit. Nvidia and SpaceX are already collaborating to develop AI computers for space; in August of this year, they plan to adapt the Vera Rubin NVL72 to the first Starmind AI satellite. Nvidia executive Ian Buck (Ian Buck) stated: “SpaceXAI is taking this architecture from a huge AI factory to the new frontier of orbital computing.” He added that this work will reprocess power supply, heat dissipation, reliability and communications for the space environment, while retaining Nvidia's architecture and software.
According to the caliber announced by SpaceX, AI1 has a peak calculated load of 250 kW and an average of 175 kW; the calculation results are transmitted back through the laser link and through the Starlink network.
The capacity-side move is the Gigasat plant, which is located in Bastrop (Bastrop), Texas. SpaceX plans to begin production and deployment of thousands of AI satellites by the end of 2027 as early as the end of 2027. Additionally, SpaceX plans to expand the computing power infrastructure behind Grok with Vera Rubin.
On Stocktwits, SpaceX's retail sentiment rose from “neutral” a week ago to “bullish,” and the discussion was at a “normal” level. The stock fell 11% during the year.
Kessel's technical details (84-layer, M10, passive midplane) will eventually be covered by the next generation, but what it demonstrates won't be: when the cabinet becomes a bottleneck, customers start defining the cabinet themselves. Kyber extended it to 2028, giving this demonstration a window of time — Nvidia's next generation cabinet hasn't arrived yet, and SpaceX has taken away the L6 layer using a board with fewer layers and updated materials.
The real restraint isn't on the board; it's on the electricity. SemiAnalysis points to the same conclusion on both sides of the ground and track: the next wall of computing power is power supply, and SpaceX has yet to come up with a solution for this wall.