NVIDIA Delays Multiple AI Rack-Level Architectures; Rubin Ultra Expansion Path Restricted

Huo Xing Finance reports that on July 6, SemiAnalysis stated that NVIDIA’s latest rack-scale interconnect architecture, Kyber NVL144, has undergone significant adjustments just three months after its announcement, with its launch now delayed by over 12 months until 2028, primarily due to ongoing challenges in the manufacturability of its PCB planar design. Additionally, the NVL72x2 back-to-back rack architecture has been canceled. This design originally aimed to enhance pure copper NVLink scalability by deploying two Oberon racks back-to-back, but faced strong market skepticism due to its structural complexity and high operational burden on hyperscale cloud service providers (CSPs), leading to its eventual abandonment. Since CPO (Co-Packaged Optics) technology remains immature, NVIDIA’s larger-scale expansion solutions based on CPO NVSwitch—such as NVL576—may also face further delays or be limited to small-scale pilot production. This implies that, prior to CPO maturity, NVIDIA lacks a stable large-scale scale-up solution. Furthermore, the Rubin Ultra product roadmap has changed: the originally planned “quad-chip” version has been canceled, leaving only the “dual-chip” version, with overall system-level performance expected to drop to approximately half of the original plan. These adjustments indicate that NVIDIA’s scale-up capabilities in the Rubin Ultra generation are constrained. With CPO NVSwitch unlikely to be implemented before the Feynman architecture, competitors such as AMD’s MI500X or Google’s TPU v8i may gain a relative window of opportunity in scaling large-scale training clusters. Meanwhile, NVIDIA is expected to meet market demand during this product transition period and maintain overall supply chain momentum by heavily shipping Oberon Rubin racks and their “Ultra” variants.