[SemiAnalysis: NVIDIA's Multiple AI Rack-Level Architectures Delayed or Adjusted, Rubin Ultra Expansion Path Limited] BlockBeats News, July 6 — SemiAnalysis published an article stating that NVIDIA's latest rack-level interconnect architecture, Kyber NVL144, underwent significant adjustments just three months after its release. The original plan has been postponed by over 12 months to 2028, primarily due to challenges in the manufacturability of PCB planar designs. Additionally, the NVL72x2 back-to-back rack architecture has been canceled. This design was originally intended to enhance pure copper NVLink expansion capabilities by deploying two Oberon racks back-to-back, but due to its structural complexity and the high operational burden it imposed on hyperscale cloud service providers (CSPs), it faced strong market skepticism and was ultimately abandoned. Since CPO (Co-Packaged Optics) technology remains immature, NVIDIA's larger-scale expansion plans based on CPO NVSwitch (such as NVL576) may also face continued delays or be limited to small-scale trial production. This indicates that until CPO technology matures, NVIDIA lacks a stable large-scale scale-up solution. Furthermore, the Rubin Ultra product roadmap has changed: the originally planned "four compute chip" version has been canceled, leaving only the "dual compute chip" version, with overall system-level performance expected to drop to about half of the original plan. These adjustments mean that NVIDIA's scale-up expansion capabilities in the Rubin Ultra generation are constrained. In the absence of CPO NVSwitch deployment before the Feynman architecture, competitors such as AMD MI500X or Google TPU v8i may gain a relative window of opportunity in scaling large training clusters. Meanwhile, NVIDIA is expected to meet market demand during the product transition period and maintain overall supply chain rhythm by shipping large volumes of Oberon Rubin racks and their "Ultra" versions. [Original Link]