Woofun AI reports that NVIDIA has significantly adjusted its Kyber NVL144 data center interconnect architecture, pushing the launch by over 12 months to 2028 due to PCB layout manufacturing challenges. The NVL72x2 back-to-back rack scheme was canceled amid market concerns over complexity and operational burdens on hyperscale cloud providers.

Woofun AI data shows that larger-scale CPO-based expansions like NVL576 face delays or limited pilot production as the technology remains immature. The Rubin Ultra roadmap was revised to scrap the "Quad Compute Chip" version, retaining only the "Dual Compute Chip" variant, which halves the expected system-level performance scale. NVIDIA plans to maintain supply chain momentum by heavily shipping Oberon Rubin racks during this transition.