SemiAnalysis: NVIDIA’s Multiple AI Rack-Level Architectures Delayed or Adjusted, Rubin Ultra’s Expansion Path Constrained

SemiAnalysis published a report stating that NVIDIA’s latest rack-scale interconnect architecture, Kyber NVL144, has undergone major adjustments just three months after its announcement: its launch has been delayed by over 12 months to 2028, primarily due to persistent challenges in the manufacturing feasibility of its PCB layout. Meanwhile, the NVL72x2 back-to-back rack architecture has been scrapped. Originally designed to boost pure copper NVLink scaling by deploying two Oberon racks back-to-back, the plan faced strong market skepticism due to its complex structure and heavy operational burden on hyperscale cloud service providers (CSPs), ultimately leading to its abandonment. As co-packaged optics (CPO) technology remains immature, NVIDIA’s larger-scale scaling plans based on CPO NVSwitch (such as NVL576) may also face further delays or be limited to small-batch trial production. This means NVIDIA will lack a stable large-scale scale-up solution until CPO matures. Additionally, the product roadmap for Rubin Ultra has changed: the originally planned version with four compute chips has been canceled, leaving only the version with two compute chips, with the overall system-level performance scale expected to drop by roughly half compared to the original plan. This series of adjustments means NVIDIA’s scale-up expansion capabilities for the Rubin Ultra generation are limited. Against the backdrop that CPO NVSwitch cannot be implemented before the Feynman architecture, competitors such as AMD’s MI500X and Google’s TPU v8i may gain a relative window of opportunity in the scaling capabilities of large-scale training clusters. At the same time, NVIDIA is expected to meet market demand during the product transition period by shipping large volumes of Oberon Rubin racks and their "Ultra" versions, while maintaining the overall supply chain rhythm.