Nvidia's Kyber AI rack has been postponed until 2028.

      Nvidia's upcoming flagship AI machine has encountered a setback, with a single circuit board being the main issue. The Kyber rack, designed to accommodate its 2027 Rubin Ultra chips, has been pushed back to 2028.

      Research firm SemiAnalysis highlighted this delay, which was reported by CNBC. The Kyber is not a chip; it is a server cabinet that houses 144 of Nvidia's most powerful GPUs in one unit, enabling them to function as a single massive computer. This high density is crucial for training and running the largest AI models.

      One board, one significant issue

      The delay is rooted in a less glamorous aspect. The Kyber installs its chips in vertical trays to maximize capacity and reduce latency. At the core is a multi-layer circuit board known as the PCB midplane. According to SemiAnalysis, this board "remains challenging from a manufacturability standpoint." Simply put, Nvidia is not yet able to produce it at scale.

      The repercussions extend beyond this issue. A larger system, the NVL576, connects eight racks via optical cables and is likely facing delays or limited production. Nvidia has not provided any comments.

      No alternative plan

      Nvidia did have a contingency plan, intending to combine two current-generation racks to achieve similar performance. However, cloud customers disliked this approach, criticizing the "odd design and heavy operational burden," which led Nvidia to abandon the idea.

      This situation creates a gap. SemiAnalysis indicates that Nvidia currently has “no proven solution” for scaling up its most advanced Rubin Ultra systems. For a company that typically releases a new architecture each year, failing to progress is unusual.

      An opportunity for competitors

      The timing is advantageous for competitors. Google and AMD have already secured projects with leading AI labs using their own chips. A stumble at the top end creates a unique technical opportunity for them, according to SemiAnalysis. This situation also contributes to the current unease in the AI chip market, as Asian technology and circuit-board stocks fell following the news, as noted by Bloomberg.

      Nonetheless, this setback does not impact Nvidia's short-term dominance. Its existing Rubin systems are in full production, with shipments set to begin this autumn to eight cloud partners, including AWS, Microsoft Azure, and Google Cloud. SemiAnalysis even anticipates Nvidia's data center compute revenue to exceed Wall Street projections by 20 percent in the latter half of its 2027 financial year. Nvidia's stock remained relatively stable.

      The real challenge

      The underlying issue is manufacturing. Nvidia's annual release schedule is clashing with the capacity limitations of its suppliers. This includes everything from advanced circuit boards to the memory that powers each GPU.

      The factories producing these racks are already operating at full capacity. Competitors creating their own custom silicon believe that the production pace, rather than the design itself, is where Nvidia is most vulnerable.

      For once, the toughest aspect of the AI boom isn't the chip itself, but the enclosure surrounding it.