Nvidia's next-gen AI rack system delayed to 2028 on manufacturing snags, SemiAnalysis says

Nvidia's highly anticipated Kyber rack-scale architecture, intended to integrate its 2027 Rubin Ultra chips, has encountered a significant delay, now expected to launch in 2028. This news, reported by the research firm SemiAnalysis, raises concerns about the company's future product strategies amidst a series of setbacks. The Kyber system is designed to house 144 of Nvidia's most advanced chips within a single server cabinet, enabling immense computational power for AI companies to train and operate their complex models. Its innovative design features vertically mounted graphics processing units in compute trays, enhancing density and minimizing latency. Initially, the Kyber was slated to debut alongside the Rubin Ultra system in 2027. According to SemiAnalysis, the delay is attributed to challenges in manufacturing a critical circuit board essential to the system. "The Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint," the firm noted in a recent update. Additionally, the NVL576 system, which connects eight racks through optical links, may also face delays or be limited in production volume. Nvidia has not yet commented on these developments. The delays highlight increasing pressures on Nvidia’s product lines, suggesting that the company’s rapid annual release cycle may be hitting manufacturing constraints. A proposed alternative solution that involved combining two existing racks to achieve similar performance has also been abandoned, following significant criticism from cloud service providers regarding its impractical design and high operational costs. Consequently, Nvidia is left without a reliable method to enhance the scalability of the Rubin Ultra systems. This situation could potentially open doors for competitors like Advanced Micro Devices and Google, who are already making strides with their in-house chips in securing contracts with leading AI laboratories. Currently, Nvidia's generation of Rubin systems is in full production, expected to ship this fall to several cloud partners, including Amazon Web Services, Microsoft Azure, and Google Cloud. Moreover, SemiAnalysis anticipates that Nvidia’s data-center compute revenue will surpass Wall Street expectations by 20% in the latter half of fiscal 2027. In premarket trading, Nvidia's shares experienced minor fluctuations, recently down by less than 0.1% to $194.79.